fs/ocfs2/cluster/tcp.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  *
   4  * Copyright (C) 2004 Oracle.  All rights reserved.
   5  *
   6  * ----
   7  *
   8  * Callers for this were originally written against a very simple synchronus
   9  * API.  This implementation reflects those simple callers.  Some day I'm sure
  10  * we'll need to move to a more robust posting/callback mechanism.
  11  *
  12  * Transmit calls pass in kernel virtual addresses and block copying this into
  13  * the socket's tx buffers via a usual blocking sendmsg.  They'll block waiting
  14  * for a failed socket to timeout.  TX callers can also pass in a poniter to an
  15  * 'int' which gets filled with an errno off the wire in response to the
  16  * message they send.
  17  *
  18  * Handlers for unsolicited messages are registered.  Each socket has a page
  19  * that incoming data is copied into.  First the header, then the data.
  20  * Handlers are called from only one thread with a reference to this per-socket
  21  * page.  This page is destroyed after the handler call, so it can't be
  22  * referenced beyond the call.  Handlers may block but are discouraged from
  23  * doing so.
  24  *
  25  * Any framing errors (bad magic, large payload lengths) close a connection.
  26  *
  27  * Our sock_container holds the state we associate with a socket.  It's current
  28  * framing state is held there as well as the refcounting we do around when it
  29  * is safe to tear down the socket.  The socket is only finally torn down from
  30  * the container when the container loses all of its references -- so as long
  31  * as you hold a ref on the container you can trust that the socket is valid
  32  * for use with kernel socket APIs.
  33  *
  34  * Connections are initiated between a pair of nodes when the node with the
  35  * higher node number gets a heartbeat callback which indicates that the lower
  36  * numbered node has started heartbeating.  The lower numbered node is passive
  37  * and only accepts the connection if the higher numbered node is heartbeating.
  38  */
  39
  40 #include <linux/kernel.h>
  41 #include <linux/sched/mm.h>
  42 #include <linux/jiffies.h>
  43 #include <linux/slab.h>
  44 #include <linux/idr.h>
  45 #include <linux/kref.h>
  46 #include <linux/net.h>
  47 #include <linux/export.h>
  48 #include <net/tcp.h>
  49
  50 #include <linux/uaccess.h>
  51
  52 #include "heartbeat.h"
  53 #include "tcp.h"
  54 #include "nodemanager.h"
  55 #define MLOG_MASK_PREFIX ML_TCP
  56 #include "masklog.h"
  57 #include "quorum.h"
  58
  59 #include "tcp_internal.h"
  60
  61 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
  62 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
  63                           &sc->sc_node->nd_ipv4_address,                \
  64                           ntohs(sc->sc_node->nd_ipv4_port)
  65
  66 /*
  67  * In the following two log macros, the whitespace after the ',' just
  68  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  69  * previous token if args expands to nothing.
  70  */
  71 #define msglog(hdr, fmt, args...) do {                                  \
  72         typeof(hdr) __hdr = (hdr);                                      \
  73         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
  74              "key %08x num %u] " fmt,                                   \
  75              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
  76              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
  77              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
  78              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
  79 } while (0)
  80
  81 #define sclog(sc, fmt, args...) do {                                    \
  82         typeof(sc) __sc = (sc);                                         \
  83         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
  84              "pg_off %zu] " fmt, __sc,                                  \
  85              kref_read(&__sc->sc_kref), __sc->sc_sock,  \
  86             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
  87             ##args);                                                    \
  88 } while (0)
  89
  90 static DEFINE_RWLOCK(o2net_handler_lock);
  91 static struct rb_root o2net_handler_tree = RB_ROOT;
  92
  93 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
  94
  95 /* XXX someday we'll need better accounting */
  96 static struct socket *o2net_listen_sock;
  97
  98 /*
  99  * listen work is only queued by the listening socket callbacks on the
 100  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 101  * quorum work is queued as sock containers are shutdown.. stop_listening
 102  * tears down all the node's sock containers, preventing future shutdowns
 103  * and queued quroum work, before canceling delayed quorum work and
 104  * destroying the work queue.
 105  */
 106 static struct workqueue_struct *o2net_wq;
 107 static struct work_struct o2net_listen_work;
 108
 109 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 110 #define O2NET_HB_PRI 0x1
 111
 112 static struct o2net_handshake *o2net_hand;
 113 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 114
 115 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 116                 {[O2NET_ERR_NONE]       = 0,
 117                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
 118                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
 119                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
 120
 121 /* can't quite avoid *all* internal declarations :/ */
 122 static void o2net_sc_connect_completed(struct work_struct *work);
 123 static void o2net_rx_until_empty(struct work_struct *work);
 124 static void o2net_shutdown_sc(struct work_struct *work);
 125 static void o2net_listen_data_ready(struct sock *sk);
 126 static void o2net_sc_send_keep_req(struct work_struct *work);
 127 static void o2net_idle_timer(struct timer_list *t);
 128 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 129 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 130
 131 #ifdef CONFIG_DEBUG_FS
 132 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 133                            u32 msgkey, struct task_struct *task, u8 node)
 134 {
 135         INIT_LIST_HEAD(&nst->st_net_debug_item);
 136         nst->st_task = task;
 137         nst->st_msg_type = msgtype;
 138         nst->st_msg_key = msgkey;
 139         nst->st_node = node;
 140 }
 141
 142 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 143 {
 144         nst->st_sock_time = ktime_get();
 145 }
 146
 147 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 148 {
 149         nst->st_send_time = ktime_get();
 150 }
 151
 152 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 153 {
 154         nst->st_status_time = ktime_get();
 155 }
 156
 157 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 158                                                 struct o2net_sock_container *sc)
 159 {
 160         nst->st_sc = sc;
 161 }
 162
 163 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 164                                         u32 msg_id)
 165 {
 166         nst->st_id = msg_id;
 167 }
 168
 169 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
 170 {
 171         sc->sc_tv_timer = ktime_get();
 172 }
 173
 174 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
 175 {
 176         sc->sc_tv_data_ready = ktime_get();
 177 }
 178
 179 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
 180 {
 181         sc->sc_tv_advance_start = ktime_get();
 182 }
 183
 184 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
 185 {
 186         sc->sc_tv_advance_stop = ktime_get();
 187 }
 188
 189 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
 190 {
 191         sc->sc_tv_func_start = ktime_get();
 192 }
 193
 194 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
 195 {
 196         sc->sc_tv_func_stop = ktime_get();
 197 }
 198
 199 #else  /* CONFIG_DEBUG_FS */
 200 # define o2net_init_nst(a, b, c, d, e)
 201 # define o2net_set_nst_sock_time(a)
 202 # define o2net_set_nst_send_time(a)
 203 # define o2net_set_nst_status_time(a)
 204 # define o2net_set_nst_sock_container(a, b)
 205 # define o2net_set_nst_msg_id(a, b)
 206 # define o2net_set_sock_timer(a)
 207 # define o2net_set_data_ready_time(a)
 208 # define o2net_set_advance_start_time(a)
 209 # define o2net_set_advance_stop_time(a)
 210 # define o2net_set_func_start_time(a)
 211 # define o2net_set_func_stop_time(a)
 212 #endif /* CONFIG_DEBUG_FS */
 213
 214 #ifdef CONFIG_OCFS2_FS_STATS
 215 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
 216 {
 217         return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
 218 }
 219
 220 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
 221                                     struct o2net_sock_container *sc)
 222 {
 223         sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
 224                                            ktime_sub(ktime_get(),
 225                                                      nst->st_status_time));
 226         sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
 227                                          ktime_sub(nst->st_status_time,
 228                                                    nst->st_send_time));
 229         sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
 230                                             ktime_sub(nst->st_send_time,
 231                                                       nst->st_sock_time));
 232         sc->sc_send_count++;
 233 }
 234
 235 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
 236 {
 237         sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
 238                                             o2net_get_func_run_time(sc));
 239         sc->sc_recv_count++;
 240 }
 241
 242 #else
 243
 244 # define o2net_update_send_stats(a, b)
 245
 246 # define o2net_update_recv_stats(sc)
 247
 248 #endif /* CONFIG_OCFS2_FS_STATS */
 249
 250 static inline unsigned int o2net_reconnect_delay(void)
 251 {
 252         return o2nm_single_cluster->cl_reconnect_delay_ms;
 253 }
 254
 255 static inline unsigned int o2net_keepalive_delay(void)
 256 {
 257         return o2nm_single_cluster->cl_keepalive_delay_ms;
 258 }
 259
 260 static inline unsigned int o2net_idle_timeout(void)
 261 {
 262         return o2nm_single_cluster->cl_idle_timeout_ms;
 263 }
 264
 265 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 266 {
 267         int trans;
 268         BUG_ON(err >= O2NET_ERR_MAX);
 269         trans = o2net_sys_err_translations[err];
 270
 271         /* Just in case we mess up the translation table above */
 272         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 273         return trans;
 274 }
 275
 276 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 277 {
 278         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 279         return &o2net_nodes[node_num];
 280 }
 281
 282 static u8 o2net_num_from_nn(struct o2net_node *nn)
 283 {
 284         BUG_ON(nn == NULL);
 285         return nn - o2net_nodes;
 286 }
 287
 288 /* ------------------------------------------------------------ */
 289
 290 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 291 {
 292         int ret;
 293
 294         spin_lock(&nn->nn_lock);
 295         ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
 296         if (ret >= 0) {
 297                 nsw->ns_id = ret;
 298                 list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
 299         }
 300         spin_unlock(&nn->nn_lock);
 301         if (ret < 0)
 302                 return ret;
 303
 304         init_waitqueue_head(&nsw->ns_wq);
 305         nsw->ns_sys_status = O2NET_ERR_NONE;
 306         nsw->ns_status = 0;
 307         return 0;
 308 }
 309
 310 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 311                                       struct o2net_status_wait *nsw,
 312                                       enum o2net_system_error sys_status,
 313                                       s32 status)
 314 {
 315         assert_spin_locked(&nn->nn_lock);
 316
 317         if (!list_empty(&nsw->ns_node_item)) {
 318                 list_del_init(&nsw->ns_node_item);
 319                 nsw->ns_sys_status = sys_status;
 320                 nsw->ns_status = status;
 321                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 322                 wake_up(&nsw->ns_wq);
 323         }
 324 }
 325
 326 static void o2net_complete_nsw(struct o2net_node *nn,
 327                                struct o2net_status_wait *nsw,
 328                                u64 id, enum o2net_system_error sys_status,
 329                                s32 status)
 330 {
 331         spin_lock(&nn->nn_lock);
 332         if (nsw == NULL) {
 333                 if (id > INT_MAX)
 334                         goto out;
 335
 336                 nsw = idr_find(&nn->nn_status_idr, id);
 337                 if (nsw == NULL)
 338                         goto out;
 339         }
 340
 341         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 342
 343 out:
 344         spin_unlock(&nn->nn_lock);
 345         return;
 346 }
 347
 348 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 349 {
 350         struct o2net_status_wait *nsw, *tmp;
 351         unsigned int num_kills = 0;
 352
 353         assert_spin_locked(&nn->nn_lock);
 354
 355         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 356                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 357                 num_kills++;
 358         }
 359
 360         mlog(0, "completed %d messages for node %u\n", num_kills,
 361              o2net_num_from_nn(nn));
 362 }
 363
 364 static int o2net_nsw_completed(struct o2net_node *nn,
 365                                struct o2net_status_wait *nsw)
 366 {
 367         int completed;
 368         spin_lock(&nn->nn_lock);
 369         completed = list_empty(&nsw->ns_node_item);
 370         spin_unlock(&nn->nn_lock);
 371         return completed;
 372 }
 373
 374 /* ------------------------------------------------------------ */
 375
 376 static void sc_kref_release(struct kref *kref)
 377 {
 378         struct o2net_sock_container *sc = container_of(kref,
 379                                         struct o2net_sock_container, sc_kref);
 380         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 381
 382         sclog(sc, "releasing\n");
 383
 384         if (sc->sc_sock) {
 385                 sock_release(sc->sc_sock);
 386                 sc->sc_sock = NULL;
 387         }
 388
 389         o2nm_undepend_item(&sc->sc_node->nd_item);
 390         o2nm_node_put(sc->sc_node);
 391         sc->sc_node = NULL;
 392
 393         o2net_debug_del_sc(sc);
 394
 395         if (sc->sc_page)
 396                 __free_page(sc->sc_page);
 397         kfree(sc);
 398 }
 399
 400 static void sc_put(struct o2net_sock_container *sc)
 401 {
 402         sclog(sc, "put\n");
 403         kref_put(&sc->sc_kref, sc_kref_release);
 404 }
 405 static void sc_get(struct o2net_sock_container *sc)
 406 {
 407         sclog(sc, "get\n");
 408         kref_get(&sc->sc_kref);
 409 }
 410 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 411 {
 412         struct o2net_sock_container *sc, *ret = NULL;
 413         struct page *page = NULL;
 414         int status = 0;
 415
 416         page = alloc_page(GFP_NOFS);
 417         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 418         if (sc == NULL || page == NULL)
 419                 goto out;
 420
 421         kref_init(&sc->sc_kref);
 422         o2nm_node_get(node);
 423         sc->sc_node = node;
 424
 425         /* pin the node item of the remote node */
 426         status = o2nm_depend_item(&node->nd_item);
 427         if (status) {
 428                 mlog_errno(status);
 429                 o2nm_node_put(node);
 430                 goto out;
 431         }
 432         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 433         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 434         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 435         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 436
 437         timer_setup(&sc->sc_idle_timeout, o2net_idle_timer, 0);
 438
 439         sclog(sc, "alloced\n");
 440
 441         ret = sc;
 442         sc->sc_page = page;
 443         o2net_debug_add_sc(sc);
 444         sc = NULL;
 445         page = NULL;
 446
 447 out:
 448         if (page)
 449                 __free_page(page);
 450         kfree(sc);
 451
 452         return ret;
 453 }
 454
 455 /* ------------------------------------------------------------ */
 456
 457 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 458                                 struct work_struct *work)
 459 {
 460         sc_get(sc);
 461         if (!queue_work(o2net_wq, work))
 462                 sc_put(sc);
 463 }
 464 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 465                                         struct delayed_work *work,
 466                                         int delay)
 467 {
 468         sc_get(sc);
 469         if (!queue_delayed_work(o2net_wq, work, delay))
 470                 sc_put(sc);
 471 }
 472 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 473                                          struct delayed_work *work)
 474 {
 475         if (cancel_delayed_work(work))
 476                 sc_put(sc);
 477 }
 478
 479 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 480
 481 int o2net_num_connected_peers(void)
 482 {
 483         return atomic_read(&o2net_connected_peers);
 484 }
 485
 486 static void o2net_set_nn_state(struct o2net_node *nn,
 487                                struct o2net_sock_container *sc,
 488                                unsigned valid, int err)
 489 {
 490         int was_valid = nn->nn_sc_valid;
 491         int was_err = nn->nn_persistent_error;
 492         struct o2net_sock_container *old_sc = nn->nn_sc;
 493
 494         assert_spin_locked(&nn->nn_lock);
 495
 496         if (old_sc && !sc)
 497                 atomic_dec(&o2net_connected_peers);
 498         else if (!old_sc && sc)
 499                 atomic_inc(&o2net_connected_peers);
 500
 501         /* the node num comparison and single connect/accept path should stop
 502          * an non-null sc from being overwritten with another */
 503         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 504         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 505         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 506
 507         if (was_valid && !valid && err == 0)
 508                 err = -ENOTCONN;
 509
 510         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 511              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 512              nn->nn_persistent_error, err);
 513
 514         nn->nn_sc = sc;
 515         nn->nn_sc_valid = valid ? 1 : 0;
 516         nn->nn_persistent_error = err;
 517
 518         /* mirrors o2net_tx_can_proceed() */
 519         if (nn->nn_persistent_error || nn->nn_sc_valid)
 520                 wake_up(&nn->nn_sc_wq);
 521
 522         if (was_valid && !was_err && nn->nn_persistent_error) {
 523                 o2quo_conn_err(o2net_num_from_nn(nn));
 524                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 525                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 526         }
 527
 528         if (was_valid && !valid) {
 529                 if (old_sc)
 530                         printk(KERN_NOTICE "o2net: No longer connected to "
 531                                 SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 532                 o2net_complete_nodes_nsw(nn);
 533         }
 534
 535         if (!was_valid && valid) {
 536                 o2quo_conn_up(o2net_num_from_nn(nn));
 537                 cancel_delayed_work(&nn->nn_connect_expired);
 538                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 539                        o2nm_this_node() > sc->sc_node->nd_num ?
 540                        "Connected to" : "Accepted connection from",
 541                        SC_NODEF_ARGS(sc));
 542         }
 543
 544         /* trigger the connecting worker func as long as we're not valid,
 545          * it will back off if it shouldn't connect.  This can be called
 546          * from node config teardown and so needs to be careful about
 547          * the work queue actually being up. */
 548         if (!valid && o2net_wq) {
 549                 unsigned long delay;
 550                 /* delay if we're within a RECONNECT_DELAY of the
 551                  * last attempt */
 552                 delay = (nn->nn_last_connect_attempt +
 553                          msecs_to_jiffies(o2net_reconnect_delay()))
 554                         - jiffies;
 555                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 556                         delay = 0;
 557                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 558                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 559
 560                 /*
 561                  * Delay the expired work after idle timeout.
 562                  *
 563                  * We might have lots of failed connection attempts that run
 564                  * through here but we only cancel the connect_expired work when
 565                  * a connection attempt succeeds.  So only the first enqueue of
 566                  * the connect_expired work will do anything.  The rest will see
 567                  * that it's already queued and do nothing.
 568                  */
 569                 delay += msecs_to_jiffies(o2net_idle_timeout());
 570                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 571         }
 572
 573         /* keep track of the nn's sc ref for the caller */
 574         if ((old_sc == NULL) && sc)
 575                 sc_get(sc);
 576         if (old_sc && (old_sc != sc)) {
 577                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 578                 sc_put(old_sc);
 579         }
 580 }
 581
 582 /* see o2net_register_callbacks() */
 583 static void o2net_data_ready(struct sock *sk)
 584 {
 585         void (*ready)(struct sock *sk);
 586         struct o2net_sock_container *sc;
 587
 588         read_lock_bh(&sk->sk_callback_lock);
 589         sc = sk->sk_user_data;
 590         if (sc) {
 591                 sclog(sc, "data_ready hit\n");
 592                 o2net_set_data_ready_time(sc);
 593                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 594                 ready = sc->sc_data_ready;
 595         } else {
 596                 ready = sk->sk_data_ready;
 597         }
 598         read_unlock_bh(&sk->sk_callback_lock);
 599
 600         ready(sk);
 601 }
 602
 603 /* see o2net_register_callbacks() */
 604 static void o2net_state_change(struct sock *sk)
 605 {
 606         void (*state_change)(struct sock *sk);
 607         struct o2net_sock_container *sc;
 608
 609         read_lock_bh(&sk->sk_callback_lock);
 610         sc = sk->sk_user_data;
 611         if (sc == NULL) {
 612                 state_change = sk->sk_state_change;
 613                 goto out;
 614         }
 615
 616         sclog(sc, "state_change to %d\n", sk->sk_state);
 617
 618         state_change = sc->sc_state_change;
 619
 620         switch(sk->sk_state) {
 621         /* ignore connecting sockets as they make progress */
 622         case TCP_SYN_SENT:
 623         case TCP_SYN_RECV:
 624                 break;
 625         case TCP_ESTABLISHED:
 626                 o2net_sc_queue_work(sc, &sc->sc_connect_work);
 627                 break;
 628         default:
 629                 printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
 630                         " shutdown, state %d\n",
 631                         SC_NODEF_ARGS(sc), sk->sk_state);
 632                 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 633                 break;
 634         }
 635 out:
 636         read_unlock_bh(&sk->sk_callback_lock);
 637         state_change(sk);
 638 }
 639
 640 /*
 641  * we register callbacks so we can queue work on events before calling
 642  * the original callbacks.  our callbacks our careful to test user_data
 643  * to discover when they've reaced with o2net_unregister_callbacks().
 644  */
 645 static void o2net_register_callbacks(struct sock *sk,
 646                                      struct o2net_sock_container *sc)
 647 {
 648         write_lock_bh(&sk->sk_callback_lock);
 649
 650         /* accepted sockets inherit the old listen socket data ready */
 651         if (sk->sk_data_ready == o2net_listen_data_ready) {
 652                 sk->sk_data_ready = sk->sk_user_data;
 653                 sk->sk_user_data = NULL;
 654         }
 655
 656         BUG_ON(sk->sk_user_data != NULL);
 657         sk->sk_user_data = sc;
 658         sc_get(sc);
 659
 660         sc->sc_data_ready = sk->sk_data_ready;
 661         sc->sc_state_change = sk->sk_state_change;
 662         sk->sk_data_ready = o2net_data_ready;
 663         sk->sk_state_change = o2net_state_change;
 664
 665         mutex_init(&sc->sc_send_lock);
 666
 667         write_unlock_bh(&sk->sk_callback_lock);
 668 }
 669
 670 static int o2net_unregister_callbacks(struct sock *sk,
 671                                    struct o2net_sock_container *sc)
 672 {
 673         int ret = 0;
 674
 675         write_lock_bh(&sk->sk_callback_lock);
 676         if (sk->sk_user_data == sc) {
 677                 ret = 1;
 678                 sk->sk_user_data = NULL;
 679                 sk->sk_data_ready = sc->sc_data_ready;
 680                 sk->sk_state_change = sc->sc_state_change;
 681         }
 682         write_unlock_bh(&sk->sk_callback_lock);
 683
 684         return ret;
 685 }
 686
 687 /*
 688  * this is a little helper that is called by callers who have seen a problem
 689  * with an sc and want to detach it from the nn if someone already hasn't beat
 690  * them to it.  if an error is given then the shutdown will be persistent
 691  * and pending transmits will be canceled.
 692  */
 693 static void o2net_ensure_shutdown(struct o2net_node *nn,
 694                                    struct o2net_sock_container *sc,
 695                                    int err)
 696 {
 697         spin_lock(&nn->nn_lock);
 698         if (nn->nn_sc == sc)
 699                 o2net_set_nn_state(nn, NULL, 0, err);
 700         spin_unlock(&nn->nn_lock);
 701 }
 702
 703 /*
 704  * This work queue function performs the blocking parts of socket shutdown.  A
 705  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 706  * sc detached from the nn.  state_change will also trigger this callback
 707  * directly when it sees errors.  In that case we need to call set_nn_state
 708  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 709  * itself.
 710  */
 711 static void o2net_shutdown_sc(struct work_struct *work)
 712 {
 713         struct o2net_sock_container *sc =
 714                 container_of(work, struct o2net_sock_container,
 715                              sc_shutdown_work);
 716         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 717
 718         sclog(sc, "shutting down\n");
 719
 720         /* drop the callbacks ref and call shutdown only once */
 721         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 722                 /* we shouldn't flush as we're in the thread, the
 723                  * races with pending sc work structs are harmless */
 724                 del_timer_sync(&sc->sc_idle_timeout);
 725                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 726                 sc_put(sc);
 727                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 728         }
 729
 730         /* not fatal so failed connects before the other guy has our
 731          * heartbeat can be retried */
 732         o2net_ensure_shutdown(nn, sc, 0);
 733         sc_put(sc);
 734 }
 735
 736 /* ------------------------------------------------------------ */
 737
 738 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 739                              u32 key)
 740 {
 741         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 742
 743         if (ret == 0)
 744                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 745
 746         return ret;
 747 }
 748
 749 static struct o2net_msg_handler *
 750 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 751                           struct rb_node **ret_parent)
 752 {
 753         struct rb_node **p = &o2net_handler_tree.rb_node;
 754         struct rb_node *parent = NULL;
 755         struct o2net_msg_handler *nmh, *ret = NULL;
 756         int cmp;
 757
 758         while (*p) {
 759                 parent = *p;
 760                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 761                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 762
 763                 if (cmp < 0)
 764                         p = &(*p)->rb_left;
 765                 else if (cmp > 0)
 766                         p = &(*p)->rb_right;
 767                 else {
 768                         ret = nmh;
 769                         break;
 770                 }
 771         }
 772
 773         if (ret_p != NULL)
 774                 *ret_p = p;
 775         if (ret_parent != NULL)
 776                 *ret_parent = parent;
 777
 778         return ret;
 779 }
 780
 781 static void o2net_handler_kref_release(struct kref *kref)
 782 {
 783         struct o2net_msg_handler *nmh;
 784         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 785
 786         kfree(nmh);
 787 }
 788
 789 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 790 {
 791         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 792 }
 793
 794 /* max_len is protection for the handler func.  incoming messages won't
 795  * be given to the handler if their payload is longer than the max. */
 796 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 797                            o2net_msg_handler_func *func, void *data,
 798                            o2net_post_msg_handler_func *post_func,
 799                            struct list_head *unreg_list)
 800 {
 801         struct o2net_msg_handler *nmh = NULL;
 802         struct rb_node **p, *parent;
 803         int ret = 0;
 804
 805         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 806                 mlog(0, "max_len for message handler out of range: %u\n",
 807                         max_len);
 808                 ret = -EINVAL;
 809                 goto out;
 810         }
 811
 812         if (!msg_type) {
 813                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 814                 ret = -EINVAL;
 815                 goto out;
 816
 817         }
 818         if (!func) {
 819                 mlog(0, "no message handler provided: %u, %p\n",
 820                        msg_type, func);
 821                 ret = -EINVAL;
 822                 goto out;
 823         }
 824
 825         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 826         if (nmh == NULL) {
 827                 ret = -ENOMEM;
 828                 goto out;
 829         }
 830
 831         nmh->nh_func = func;
 832         nmh->nh_func_data = data;
 833         nmh->nh_post_func = post_func;
 834         nmh->nh_msg_type = msg_type;
 835         nmh->nh_max_len = max_len;
 836         nmh->nh_key = key;
 837         /* the tree and list get this ref.. they're both removed in
 838          * unregister when this ref is dropped */
 839         kref_init(&nmh->nh_kref);
 840         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 841
 842         write_lock(&o2net_handler_lock);
 843         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 844                 ret = -EEXIST;
 845         else {
 846                 rb_link_node(&nmh->nh_node, parent, p);
 847                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 848                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 849
 850                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 851                      func, msg_type, key);
 852                 /* we've had some trouble with handlers seemingly vanishing. */
 853                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 854                                                           &parent) == NULL,
 855                                 "couldn't find handler we *just* registered "
 856                                 "for type %u key %08x\n", msg_type, key);
 857         }
 858         write_unlock(&o2net_handler_lock);
 859
 860 out:
 861         if (ret)
 862                 kfree(nmh);
 863
 864         return ret;
 865 }
 866 EXPORT_SYMBOL_GPL(o2net_register_handler);
 867
 868 void o2net_unregister_handler_list(struct list_head *list)
 869 {
 870         struct o2net_msg_handler *nmh, *n;
 871
 872         write_lock(&o2net_handler_lock);
 873         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 874                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 875                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 876                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 877                 list_del_init(&nmh->nh_unregister_item);
 878                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 879         }
 880         write_unlock(&o2net_handler_lock);
 881 }
 882 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 883
 884 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 885 {
 886         struct o2net_msg_handler *nmh;
 887
 888         read_lock(&o2net_handler_lock);
 889         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 890         if (nmh)
 891                 kref_get(&nmh->nh_kref);
 892         read_unlock(&o2net_handler_lock);
 893
 894         return nmh;
 895 }
 896
 897 /* ------------------------------------------------------------ */
 898
 899 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 900 {
 901         struct kvec vec = { .iov_len = len, .iov_base = data, };
 902         struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
 903         iov_iter_kvec(&msg.msg_iter, READ, &vec, 1, len);
 904         return sock_recvmsg(sock, &msg, MSG_DONTWAIT);
 905 }
 906
 907 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 908                               size_t veclen, size_t total)
 909 {
 910         int ret;
 911         struct msghdr msg = {.msg_flags = 0,};
 912
 913         if (sock == NULL) {
 914                 ret = -EINVAL;
 915                 goto out;
 916         }
 917
 918         ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
 919         if (likely(ret == total))
 920                 return 0;
 921         mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
 922         if (ret >= 0)
 923                 ret = -EPIPE; /* should be smarter, I bet */
 924 out:
 925         mlog(0, "returning error: %d\n", ret);
 926         return ret;
 927 }
 928
 929 static void o2net_sendpage(struct o2net_sock_container *sc,
 930                            void *kmalloced_virt,
 931                            size_t size)
 932 {
 933         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 934         ssize_t ret;
 935
 936         while (1) {
 937                 mutex_lock(&sc->sc_send_lock);
 938                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 939                                                  virt_to_page(kmalloced_virt),
 940                                                  offset_in_page(kmalloced_virt),
 941                                                  size, MSG_DONTWAIT);
 942                 mutex_unlock(&sc->sc_send_lock);
 943                 if (ret == size)
 944                         break;
 945                 if (ret == (ssize_t)-EAGAIN) {
 946                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 947                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 948                         cond_resched();
 949                         continue;
 950                 }
 951                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 952                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 953                 o2net_ensure_shutdown(nn, sc, 0);
 954                 break;
 955         }
 956 }
 957
 958 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 959 {
 960         memset(msg, 0, sizeof(struct o2net_msg));
 961         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 962         msg->data_len = cpu_to_be16(data_len);
 963         msg->msg_type = cpu_to_be16(msg_type);
 964         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 965         msg->status = 0;
 966         msg->key = cpu_to_be32(key);
 967 }
 968
 969 static int o2net_tx_can_proceed(struct o2net_node *nn,
 970                                 struct o2net_sock_container **sc_ret,
 971                                 int *error)
 972 {
 973         int ret = 0;
 974
 975         spin_lock(&nn->nn_lock);
 976         if (nn->nn_persistent_error) {
 977                 ret = 1;
 978                 *sc_ret = NULL;
 979                 *error = nn->nn_persistent_error;
 980         } else if (nn->nn_sc_valid) {
 981                 kref_get(&nn->nn_sc->sc_kref);
 982
 983                 ret = 1;
 984                 *sc_ret = nn->nn_sc;
 985                 *error = 0;
 986         }
 987         spin_unlock(&nn->nn_lock);
 988
 989         return ret;
 990 }
 991
 992 /* Get a map of all nodes to which this node is currently connected to */
 993 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
 994 {
 995         struct o2net_sock_container *sc;
 996         int node, ret;
 997
 998         BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
 999
1000         memset(map, 0, bytes);
1001         for (node = 0; node < O2NM_MAX_NODES; ++node) {
1002                 if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1003                         continue;
1004                 if (!ret) {
1005                         set_bit(node, map);
1006                         sc_put(sc);
1007                 }
1008         }
1009 }
1010 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1011
1012 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1013                            size_t caller_veclen, u8 target_node, int *status)
1014 {
1015         int ret = 0;
1016         struct o2net_msg *msg = NULL;
1017         size_t veclen, caller_bytes = 0;
1018         struct kvec *vec = NULL;
1019         struct o2net_sock_container *sc = NULL;
1020         struct o2net_node *nn = o2net_nn_from_num(target_node);
1021         struct o2net_status_wait nsw = {
1022                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1023         };
1024         struct o2net_send_tracking nst;
1025
1026         o2net_init_nst(&nst, msg_type, key, current, target_node);
1027
1028         if (o2net_wq == NULL) {
1029                 mlog(0, "attempt to tx without o2netd running\n");
1030                 ret = -ESRCH;
1031                 goto out;
1032         }
1033
1034         if (caller_veclen == 0) {
1035                 mlog(0, "bad kvec array length\n");
1036                 ret = -EINVAL;
1037                 goto out;
1038         }
1039
1040         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1041         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1042                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1043                 ret = -EINVAL;
1044                 goto out;
1045         }
1046
1047         if (target_node == o2nm_this_node()) {
1048                 ret = -ELOOP;
1049                 goto out;
1050         }
1051
1052         o2net_debug_add_nst(&nst);
1053
1054         o2net_set_nst_sock_time(&nst);
1055
1056         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1057         if (ret)
1058                 goto out;
1059
1060         o2net_set_nst_sock_container(&nst, sc);
1061
1062         veclen = caller_veclen + 1;
1063         vec = kmalloc_array(veclen, sizeof(struct kvec), GFP_ATOMIC);
1064         if (vec == NULL) {
1065                 mlog(0, "failed to %zu element kvec!\n", veclen);
1066                 ret = -ENOMEM;
1067                 goto out;
1068         }
1069
1070         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1071         if (!msg) {
1072                 mlog(0, "failed to allocate a o2net_msg!\n");
1073                 ret = -ENOMEM;
1074                 goto out;
1075         }
1076
1077         o2net_init_msg(msg, caller_bytes, msg_type, key);
1078
1079         vec[0].iov_len = sizeof(struct o2net_msg);
1080         vec[0].iov_base = msg;
1081         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1082
1083         ret = o2net_prep_nsw(nn, &nsw);
1084         if (ret)
1085                 goto out;
1086
1087         msg->msg_num = cpu_to_be32(nsw.ns_id);
1088         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1089
1090         o2net_set_nst_send_time(&nst);
1091
1092         /* finally, convert the message header to network byte-order
1093          * and send */
1094         mutex_lock(&sc->sc_send_lock);
1095         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1096                                  sizeof(struct o2net_msg) + caller_bytes);
1097         mutex_unlock(&sc->sc_send_lock);
1098         msglog(msg, "sending returned %d\n", ret);
1099         if (ret < 0) {
1100                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1101                 goto out;
1102         }
1103
1104         /* wait on other node's handler */
1105         o2net_set_nst_status_time(&nst);
1106         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1107
1108         o2net_update_send_stats(&nst, sc);
1109
1110         /* Note that we avoid overwriting the callers status return
1111          * variable if a system error was reported on the other
1112          * side. Callers beware. */
1113         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1114         if (status && !ret)
1115                 *status = nsw.ns_status;
1116
1117         mlog(0, "woken, returning system status %d, user status %d\n",
1118              ret, nsw.ns_status);
1119 out:
1120         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1121         if (sc)
1122                 sc_put(sc);
1123         kfree(vec);
1124         kfree(msg);
1125         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1126         return ret;
1127 }
1128 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1129
1130 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1131                        u8 target_node, int *status)
1132 {
1133         struct kvec vec = {
1134                 .iov_base = data,
1135                 .iov_len = len,
1136         };
1137         return o2net_send_message_vec(msg_type, key, &vec, 1,
1138                                       target_node, status);
1139 }
1140 EXPORT_SYMBOL_GPL(o2net_send_message);
1141
1142 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1143                                    enum o2net_system_error syserr, int err)
1144 {
1145         struct kvec vec = {
1146                 .iov_base = hdr,
1147                 .iov_len = sizeof(struct o2net_msg),
1148         };
1149
1150         BUG_ON(syserr >= O2NET_ERR_MAX);
1151
1152         /* leave other fields intact from the incoming message, msg_num
1153          * in particular */
1154         hdr->sys_status = cpu_to_be32(syserr);
1155         hdr->status = cpu_to_be32(err);
1156         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1157         hdr->data_len = 0;
1158
1159         msglog(hdr, "about to send status magic %d\n", err);
1160         /* hdr has been in host byteorder this whole time */
1161         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1162 }
1163
1164 /* this returns -errno if the header was unknown or too large, etc.
1165  * after this is called the buffer us reused for the next message */
1166 static int o2net_process_message(struct o2net_sock_container *sc,
1167                                  struct o2net_msg *hdr)
1168 {
1169         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1170         int ret = 0, handler_status;
1171         enum  o2net_system_error syserr;
1172         struct o2net_msg_handler *nmh = NULL;
1173         void *ret_data = NULL;
1174
1175         msglog(hdr, "processing message\n");
1176
1177         o2net_sc_postpone_idle(sc);
1178
1179         switch(be16_to_cpu(hdr->magic)) {
1180                 case O2NET_MSG_STATUS_MAGIC:
1181                         /* special type for returning message status */
1182                         o2net_complete_nsw(nn, NULL,
1183                                            be32_to_cpu(hdr->msg_num),
1184                                            be32_to_cpu(hdr->sys_status),
1185                                            be32_to_cpu(hdr->status));
1186                         goto out;
1187                 case O2NET_MSG_KEEP_REQ_MAGIC:
1188                         o2net_sendpage(sc, o2net_keep_resp,
1189                                        sizeof(*o2net_keep_resp));
1190                         goto out;
1191                 case O2NET_MSG_KEEP_RESP_MAGIC:
1192                         goto out;
1193                 case O2NET_MSG_MAGIC:
1194                         break;
1195                 default:
1196                         msglog(hdr, "bad magic\n");
1197                         ret = -EINVAL;
1198                         goto out;
1199         }
1200
1201         /* find a handler for it */
1202         handler_status = 0;
1203         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1204                                 be32_to_cpu(hdr->key));
1205         if (!nmh) {
1206                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1207                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1208                 syserr = O2NET_ERR_NO_HNDLR;
1209                 goto out_respond;
1210         }
1211
1212         syserr = O2NET_ERR_NONE;
1213
1214         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1215                 syserr = O2NET_ERR_OVERFLOW;
1216
1217         if (syserr != O2NET_ERR_NONE)
1218                 goto out_respond;
1219
1220         o2net_set_func_start_time(sc);
1221         sc->sc_msg_key = be32_to_cpu(hdr->key);
1222         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1223         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1224                                              be16_to_cpu(hdr->data_len),
1225                                         nmh->nh_func_data, &ret_data);
1226         o2net_set_func_stop_time(sc);
1227
1228         o2net_update_recv_stats(sc);
1229
1230 out_respond:
1231         /* this destroys the hdr, so don't use it after this */
1232         mutex_lock(&sc->sc_send_lock);
1233         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1234                                       handler_status);
1235         mutex_unlock(&sc->sc_send_lock);
1236         hdr = NULL;
1237         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1238              handler_status, syserr, ret);
1239
1240         if (nmh) {
1241                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1242                 if (nmh->nh_post_func)
1243                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1244                                             ret_data);
1245         }
1246
1247 out:
1248         if (nmh)
1249                 o2net_handler_put(nmh);
1250         return ret;
1251 }
1252
1253 static int o2net_check_handshake(struct o2net_sock_container *sc)
1254 {
1255         struct o2net_handshake *hand = page_address(sc->sc_page);
1256         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1257
1258         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1259                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1260                        "protocol version %llu but %llu is required. "
1261                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1262                        (unsigned long long)be64_to_cpu(hand->protocol_version),
1263                        O2NET_PROTOCOL_VERSION);
1264
1265                 /* don't bother reconnecting if its the wrong version. */
1266                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1267                 return -1;
1268         }
1269
1270         /*
1271          * Ensure timeouts are consistent with other nodes, otherwise
1272          * we can end up with one node thinking that the other must be down,
1273          * but isn't. This can ultimately cause corruption.
1274          */
1275         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1276                                 o2net_idle_timeout()) {
1277                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1278                        "idle timeout of %u ms, but we use %u ms locally. "
1279                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1280                        be32_to_cpu(hand->o2net_idle_timeout_ms),
1281                        o2net_idle_timeout());
1282                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1283                 return -1;
1284         }
1285
1286         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1287                         o2net_keepalive_delay()) {
1288                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1289                        "delay of %u ms, but we use %u ms locally. "
1290                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1291                        be32_to_cpu(hand->o2net_keepalive_delay_ms),
1292                        o2net_keepalive_delay());
1293                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1294                 return -1;
1295         }
1296
1297         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1298                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1299                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1300                        "timeout of %u ms, but we use %u ms locally. "
1301                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1302                        be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1303                        O2HB_MAX_WRITE_TIMEOUT_MS);
1304                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1305                 return -1;
1306         }
1307
1308         sc->sc_handshake_ok = 1;
1309
1310         spin_lock(&nn->nn_lock);
1311         /* set valid and queue the idle timers only if it hasn't been
1312          * shut down already */
1313         if (nn->nn_sc == sc) {
1314                 o2net_sc_reset_idle_timer(sc);
1315                 atomic_set(&nn->nn_timeout, 0);
1316                 o2net_set_nn_state(nn, sc, 1, 0);
1317         }
1318         spin_unlock(&nn->nn_lock);
1319
1320         /* shift everything up as though it wasn't there */
1321         sc->sc_page_off -= sizeof(struct o2net_handshake);
1322         if (sc->sc_page_off)
1323                 memmove(hand, hand + 1, sc->sc_page_off);
1324
1325         return 0;
1326 }
1327
1328 /* this demuxes the queued rx bytes into header or payload bits and calls
1329  * handlers as each full message is read off the socket.  it returns -error,
1330  * == 0 eof, or > 0 for progress made.*/
1331 static int o2net_advance_rx(struct o2net_sock_container *sc)
1332 {
1333         struct o2net_msg *hdr;
1334         int ret = 0;
1335         void *data;
1336         size_t datalen;
1337
1338         sclog(sc, "receiving\n");
1339         o2net_set_advance_start_time(sc);
1340
1341         if (unlikely(sc->sc_handshake_ok == 0)) {
1342                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1343                         data = page_address(sc->sc_page) + sc->sc_page_off;
1344                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1345                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1346                         if (ret > 0)
1347                                 sc->sc_page_off += ret;
1348                 }
1349
1350                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1351                         o2net_check_handshake(sc);
1352                         if (unlikely(sc->sc_handshake_ok == 0))
1353                                 ret = -EPROTO;
1354                 }
1355                 goto out;
1356         }
1357
1358         /* do we need more header? */
1359         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1360                 data = page_address(sc->sc_page) + sc->sc_page_off;
1361                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1362                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1363                 if (ret > 0) {
1364                         sc->sc_page_off += ret;
1365                         /* only swab incoming here.. we can
1366                          * only get here once as we cross from
1367                          * being under to over */
1368                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1369                                 hdr = page_address(sc->sc_page);
1370                                 if (be16_to_cpu(hdr->data_len) >
1371                                     O2NET_MAX_PAYLOAD_BYTES)
1372                                         ret = -EOVERFLOW;
1373                         }
1374                 }
1375                 if (ret <= 0)
1376                         goto out;
1377         }
1378
1379         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1380                 /* oof, still don't have a header */
1381                 goto out;
1382         }
1383
1384         /* this was swabbed above when we first read it */
1385         hdr = page_address(sc->sc_page);
1386
1387         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1388
1389         /* do we need more payload? */
1390         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1391                 /* need more payload */
1392                 data = page_address(sc->sc_page) + sc->sc_page_off;
1393                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1394                           sc->sc_page_off;
1395                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1396                 if (ret > 0)
1397                         sc->sc_page_off += ret;
1398                 if (ret <= 0)
1399                         goto out;
1400         }
1401
1402         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1403                 /* we can only get here once, the first time we read
1404                  * the payload.. so set ret to progress if the handler
1405                  * works out. after calling this the message is toast */
1406                 ret = o2net_process_message(sc, hdr);
1407                 if (ret == 0)
1408                         ret = 1;
1409                 sc->sc_page_off = 0;
1410         }
1411
1412 out:
1413         sclog(sc, "ret = %d\n", ret);
1414         o2net_set_advance_stop_time(sc);
1415         return ret;
1416 }
1417
1418 /* this work func is triggerd by data ready.  it reads until it can read no
1419  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1420  * our work the work struct will be marked and we'll be called again. */
1421 static void o2net_rx_until_empty(struct work_struct *work)
1422 {
1423         struct o2net_sock_container *sc =
1424                 container_of(work, struct o2net_sock_container, sc_rx_work);
1425         int ret;
1426
1427         do {
1428                 ret = o2net_advance_rx(sc);
1429         } while (ret > 0);
1430
1431         if (ret <= 0 && ret != -EAGAIN) {
1432                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1433                 sclog(sc, "saw error %d, closing\n", ret);
1434                 /* not permanent so read failed handshake can retry */
1435                 o2net_ensure_shutdown(nn, sc, 0);
1436         }
1437
1438         sc_put(sc);
1439 }
1440
1441 static void o2net_initialize_handshake(void)
1442 {
1443         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1444                 O2HB_MAX_WRITE_TIMEOUT_MS);
1445         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1446         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1447                 o2net_keepalive_delay());
1448         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1449                 o2net_reconnect_delay());
1450 }
1451
1452 /* ------------------------------------------------------------ */
1453
1454 /* called when a connect completes and after a sock is accepted.  the
1455  * rx path will see the response and mark the sc valid */
1456 static void o2net_sc_connect_completed(struct work_struct *work)
1457 {
1458         struct o2net_sock_container *sc =
1459                 container_of(work, struct o2net_sock_container,
1460                              sc_connect_work);
1461
1462         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1463               (unsigned long long)O2NET_PROTOCOL_VERSION,
1464               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1465
1466         o2net_initialize_handshake();
1467         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1468         sc_put(sc);
1469 }
1470
1471 /* this is called as a work_struct func. */
1472 static void o2net_sc_send_keep_req(struct work_struct *work)
1473 {
1474         struct o2net_sock_container *sc =
1475                 container_of(work, struct o2net_sock_container,
1476                              sc_keepalive_work.work);
1477
1478         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1479         sc_put(sc);
1480 }
1481
1482 /* socket shutdown does a del_timer_sync against this as it tears down.
1483  * we can't start this timer until we've got to the point in sc buildup
1484  * where shutdown is going to be involved */
1485 static void o2net_idle_timer(struct timer_list *t)
1486 {
1487         struct o2net_sock_container *sc = from_timer(sc, t, sc_idle_timeout);
1488         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1489 #ifdef CONFIG_DEBUG_FS
1490         unsigned long msecs = ktime_to_ms(ktime_get()) -
1491                 ktime_to_ms(sc->sc_tv_timer);
1492 #else
1493         unsigned long msecs = o2net_idle_timeout();
1494 #endif
1495
1496         printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1497                "idle for %lu.%lu secs.\n",
1498                SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1499
1500         /* idle timerout happen, don't shutdown the connection, but
1501          * make fence decision. Maybe the connection can recover before
1502          * the decision is made.
1503          */
1504         atomic_set(&nn->nn_timeout, 1);
1505         o2quo_conn_err(o2net_num_from_nn(nn));
1506         queue_delayed_work(o2net_wq, &nn->nn_still_up,
1507                         msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1508
1509         o2net_sc_reset_idle_timer(sc);
1510
1511 }
1512
1513 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1514 {
1515         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1516         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1517                       msecs_to_jiffies(o2net_keepalive_delay()));
1518         o2net_set_sock_timer(sc);
1519         mod_timer(&sc->sc_idle_timeout,
1520                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1521 }
1522
1523 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1524 {
1525         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1526
1527         /* clear fence decision since the connection recover from timeout*/
1528         if (atomic_read(&nn->nn_timeout)) {
1529                 o2quo_conn_up(o2net_num_from_nn(nn));
1530                 cancel_delayed_work(&nn->nn_still_up);
1531                 atomic_set(&nn->nn_timeout, 0);
1532         }
1533
1534         /* Only push out an existing timer */
1535         if (timer_pending(&sc->sc_idle_timeout))
1536                 o2net_sc_reset_idle_timer(sc);
1537 }
1538
1539 /* this work func is kicked whenever a path sets the nn state which doesn't
1540  * have valid set.  This includes seeing hb come up, losing a connection,
1541  * having a connect attempt fail, etc. This centralizes the logic which decides
1542  * if a connect attempt should be made or if we should give up and all future
1543  * transmit attempts should fail */
1544 static void o2net_start_connect(struct work_struct *work)
1545 {
1546         struct o2net_node *nn =
1547                 container_of(work, struct o2net_node, nn_connect_work.work);
1548         struct o2net_sock_container *sc = NULL;
1549         struct o2nm_node *node = NULL, *mynode = NULL;
1550         struct socket *sock = NULL;
1551         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1552         int ret = 0, stop;
1553         unsigned int timeout;
1554         unsigned int nofs_flag;
1555
1556         /*
1557          * sock_create allocates the sock with GFP_KERNEL. We must
1558          * prevent the filesystem from being reentered by memory reclaim.
1559          */
1560         nofs_flag = memalloc_nofs_save();
1561         /* if we're greater we initiate tx, otherwise we accept */
1562         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1563                 goto out;
1564
1565         /* watch for racing with tearing a node down */
1566         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1567         if (node == NULL)
1568                 goto out;
1569
1570         mynode = o2nm_get_node_by_num(o2nm_this_node());
1571         if (mynode == NULL)
1572                 goto out;
1573
1574         spin_lock(&nn->nn_lock);
1575         /*
1576          * see if we already have one pending or have given up.
1577          * For nn_timeout, it is set when we close the connection
1578          * because of the idle time out. So it means that we have
1579          * at least connected to that node successfully once,
1580          * now try to connect to it again.
1581          */
1582         timeout = atomic_read(&nn->nn_timeout);
1583         stop = (nn->nn_sc ||
1584                 (nn->nn_persistent_error &&
1585                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1586         spin_unlock(&nn->nn_lock);
1587         if (stop)
1588                 goto out;
1589
1590         nn->nn_last_connect_attempt = jiffies;
1591
1592         sc = sc_alloc(node);
1593         if (sc == NULL) {
1594                 mlog(0, "couldn't allocate sc\n");
1595                 ret = -ENOMEM;
1596                 goto out;
1597         }
1598
1599         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1600         if (ret < 0) {
1601                 mlog(0, "can't create socket: %d\n", ret);
1602                 goto out;
1603         }
1604         sc->sc_sock = sock; /* freed by sc_kref_release */
1605
1606         sock->sk->sk_allocation = GFP_ATOMIC;
1607
1608         myaddr.sin_family = AF_INET;
1609         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1610         myaddr.sin_port = htons(0); /* any port */
1611
1612         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1613                               sizeof(myaddr));
1614         if (ret) {
1615                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1616                      ret, &mynode->nd_ipv4_address);
1617                 goto out;
1618         }
1619
1620         tcp_sock_set_nodelay(sc->sc_sock->sk);
1621         tcp_sock_set_user_timeout(sock->sk, O2NET_TCP_USER_TIMEOUT);
1622
1623         o2net_register_callbacks(sc->sc_sock->sk, sc);
1624
1625         spin_lock(&nn->nn_lock);
1626         /* handshake completion will set nn->nn_sc_valid */
1627         o2net_set_nn_state(nn, sc, 0, 0);
1628         spin_unlock(&nn->nn_lock);
1629
1630         remoteaddr.sin_family = AF_INET;
1631         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1632         remoteaddr.sin_port = node->nd_ipv4_port;
1633
1634         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1635                                         (struct sockaddr *)&remoteaddr,
1636                                         sizeof(remoteaddr),
1637                                         O_NONBLOCK);
1638         if (ret == -EINPROGRESS)
1639                 ret = 0;
1640
1641 out:
1642         if (ret && sc) {
1643                 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1644                        " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1645                 /* 0 err so that another will be queued and attempted
1646                  * from set_nn_state */
1647                 o2net_ensure_shutdown(nn, sc, 0);
1648         }
1649         if (sc)
1650                 sc_put(sc);
1651         if (node)
1652                 o2nm_node_put(node);
1653         if (mynode)
1654                 o2nm_node_put(mynode);
1655
1656         memalloc_nofs_restore(nofs_flag);
1657         return;
1658 }
1659
1660 static void o2net_connect_expired(struct work_struct *work)
1661 {
1662         struct o2net_node *nn =
1663                 container_of(work, struct o2net_node, nn_connect_expired.work);
1664
1665         spin_lock(&nn->nn_lock);
1666         if (!nn->nn_sc_valid) {
1667                 printk(KERN_NOTICE "o2net: No connection established with "
1668                        "node %u after %u.%u seconds, check network and"
1669                        " cluster configuration.\n",
1670                      o2net_num_from_nn(nn),
1671                      o2net_idle_timeout() / 1000,
1672                      o2net_idle_timeout() % 1000);
1673
1674                 o2net_set_nn_state(nn, NULL, 0, 0);
1675         }
1676         spin_unlock(&nn->nn_lock);
1677 }
1678
1679 static void o2net_still_up(struct work_struct *work)
1680 {
1681         struct o2net_node *nn =
1682                 container_of(work, struct o2net_node, nn_still_up.work);
1683
1684         o2quo_hb_still_up(o2net_num_from_nn(nn));
1685 }
1686
1687 /* ------------------------------------------------------------ */
1688
1689 void o2net_disconnect_node(struct o2nm_node *node)
1690 {
1691         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1692
1693         /* don't reconnect until it's heartbeating again */
1694         spin_lock(&nn->nn_lock);
1695         atomic_set(&nn->nn_timeout, 0);
1696         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1697         spin_unlock(&nn->nn_lock);
1698
1699         if (o2net_wq) {
1700                 cancel_delayed_work(&nn->nn_connect_expired);
1701                 cancel_delayed_work(&nn->nn_connect_work);
1702                 cancel_delayed_work(&nn->nn_still_up);
1703                 flush_workqueue(o2net_wq);
1704         }
1705 }
1706
1707 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1708                                   void *data)
1709 {
1710         o2quo_hb_down(node_num);
1711
1712         if (!node)
1713                 return;
1714
1715         if (node_num != o2nm_this_node())
1716                 o2net_disconnect_node(node);
1717
1718         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1719 }
1720
1721 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1722                                 void *data)
1723 {
1724         struct o2net_node *nn = o2net_nn_from_num(node_num);
1725
1726         o2quo_hb_up(node_num);
1727
1728         BUG_ON(!node);
1729
1730         /* ensure an immediate connect attempt */
1731         nn->nn_last_connect_attempt = jiffies -
1732                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1733
1734         if (node_num != o2nm_this_node()) {
1735                 /* believe it or not, accept and node heartbeating testing
1736                  * can succeed for this node before we got here.. so
1737                  * only use set_nn_state to clear the persistent error
1738                  * if that hasn't already happened */
1739                 spin_lock(&nn->nn_lock);
1740                 atomic_set(&nn->nn_timeout, 0);
1741                 if (nn->nn_persistent_error)
1742                         o2net_set_nn_state(nn, NULL, 0, 0);
1743                 spin_unlock(&nn->nn_lock);
1744         }
1745 }
1746
1747 void o2net_unregister_hb_callbacks(void)
1748 {
1749         o2hb_unregister_callback(NULL, &o2net_hb_up);
1750         o2hb_unregister_callback(NULL, &o2net_hb_down);
1751 }
1752
1753 int o2net_register_hb_callbacks(void)
1754 {
1755         int ret;
1756
1757         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1758                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1759         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1760                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1761
1762         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1763         if (ret == 0)
1764                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1765
1766         if (ret)
1767                 o2net_unregister_hb_callbacks();
1768
1769         return ret;
1770 }
1771
1772 /* ------------------------------------------------------------ */
1773
1774 static int o2net_accept_one(struct socket *sock, int *more)
1775 {
1776         int ret;
1777         struct sockaddr_in sin;
1778         struct socket *new_sock = NULL;
1779         struct o2nm_node *node = NULL;
1780         struct o2nm_node *local_node = NULL;
1781         struct o2net_sock_container *sc = NULL;
1782         struct o2net_node *nn;
1783         unsigned int nofs_flag;
1784
1785         /*
1786          * sock_create_lite allocates the sock with GFP_KERNEL. We must
1787          * prevent the filesystem from being reentered by memory reclaim.
1788          */
1789         nofs_flag = memalloc_nofs_save();
1790
1791         BUG_ON(sock == NULL);
1792         *more = 0;
1793         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1794                                sock->sk->sk_protocol, &new_sock);
1795         if (ret)
1796                 goto out;
1797
1798         new_sock->type = sock->type;
1799         new_sock->ops = sock->ops;
1800         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, false);
1801         if (ret < 0)
1802                 goto out;
1803
1804         *more = 1;
1805         new_sock->sk->sk_allocation = GFP_ATOMIC;
1806
1807         tcp_sock_set_nodelay(new_sock->sk);
1808         tcp_sock_set_user_timeout(new_sock->sk, O2NET_TCP_USER_TIMEOUT);
1809
1810         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin, 1);
1811         if (ret < 0)
1812                 goto out;
1813
1814         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1815         if (node == NULL) {
1816                 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1817                        "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1818                        ntohs(sin.sin_port));
1819                 ret = -EINVAL;
1820                 goto out;
1821         }
1822
1823         if (o2nm_this_node() >= node->nd_num) {
1824                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1825                 if (local_node)
1826                         printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1827                                         "seen at node '%s' (%u, %pI4:%d) from "
1828                                         "node '%s' (%u, %pI4:%d)\n",
1829                                         local_node->nd_name, local_node->nd_num,
1830                                         &(local_node->nd_ipv4_address),
1831                                         ntohs(local_node->nd_ipv4_port),
1832                                         node->nd_name,
1833                                         node->nd_num, &sin.sin_addr.s_addr,
1834                                         ntohs(sin.sin_port));
1835                 ret = -EINVAL;
1836                 goto out;
1837         }
1838
1839         /* this happens all the time when the other node sees our heartbeat
1840          * and tries to connect before we see their heartbeat */
1841         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1842                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1843                      "%pI4:%d but it isn't heartbeating\n",
1844                      node->nd_name, &sin.sin_addr.s_addr,
1845                      ntohs(sin.sin_port));
1846                 ret = -EINVAL;
1847                 goto out;
1848         }
1849
1850         nn = o2net_nn_from_num(node->nd_num);
1851
1852         spin_lock(&nn->nn_lock);
1853         if (nn->nn_sc)
1854                 ret = -EBUSY;
1855         else
1856                 ret = 0;
1857         spin_unlock(&nn->nn_lock);
1858         if (ret) {
1859                 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1860                        "at %pI4:%d but it already has an open connection\n",
1861                        node->nd_name, &sin.sin_addr.s_addr,
1862                        ntohs(sin.sin_port));
1863                 goto out;
1864         }
1865
1866         sc = sc_alloc(node);
1867         if (sc == NULL) {
1868                 ret = -ENOMEM;
1869                 goto out;
1870         }
1871
1872         sc->sc_sock = new_sock;
1873         new_sock = NULL;
1874
1875         spin_lock(&nn->nn_lock);
1876         atomic_set(&nn->nn_timeout, 0);
1877         o2net_set_nn_state(nn, sc, 0, 0);
1878         spin_unlock(&nn->nn_lock);
1879
1880         o2net_register_callbacks(sc->sc_sock->sk, sc);
1881         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1882
1883         o2net_initialize_handshake();
1884         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1885
1886 out:
1887         if (new_sock)
1888                 sock_release(new_sock);
1889         if (node)
1890                 o2nm_node_put(node);
1891         if (local_node)
1892                 o2nm_node_put(local_node);
1893         if (sc)
1894                 sc_put(sc);
1895
1896         memalloc_nofs_restore(nofs_flag);
1897         return ret;
1898 }
1899
1900 /*
1901  * This function is invoked in response to one or more
1902  * pending accepts at softIRQ level. We must drain the
1903  * entire que before returning.
1904  */
1905
1906 static void o2net_accept_many(struct work_struct *work)
1907 {
1908         struct socket *sock = o2net_listen_sock;
1909         int     more;
1910
1911         /*
1912          * It is critical to note that due to interrupt moderation
1913          * at the network driver level, we can't assume to get a
1914          * softIRQ for every single conn since tcp SYN packets
1915          * can arrive back-to-back, and therefore many pending
1916          * accepts may result in just 1 softIRQ. If we terminate
1917          * the o2net_accept_one() loop upon seeing an err, what happens
1918          * to the rest of the conns in the queue? If no new SYN
1919          * arrives for hours, no softIRQ  will be delivered,
1920          * and the connections will just sit in the queue.
1921          */
1922
1923         for (;;) {
1924                 o2net_accept_one(sock, &more);
1925                 if (!more)
1926                         break;
1927                 cond_resched();
1928         }
1929 }
1930
1931 static void o2net_listen_data_ready(struct sock *sk)
1932 {
1933         void (*ready)(struct sock *sk);
1934
1935         read_lock_bh(&sk->sk_callback_lock);
1936         ready = sk->sk_user_data;
1937         if (ready == NULL) { /* check for teardown race */
1938                 ready = sk->sk_data_ready;
1939                 goto out;
1940         }
1941
1942         /* This callback may called twice when a new connection
1943          * is  being established as a child socket inherits everything
1944          * from a parent LISTEN socket, including the data_ready cb of
1945          * the parent. This leads to a hazard. In o2net_accept_one()
1946          * we are still initializing the child socket but have not
1947          * changed the inherited data_ready callback yet when
1948          * data starts arriving.
1949          * We avoid this hazard by checking the state.
1950          * For the listening socket,  the state will be TCP_LISTEN; for the new
1951          * socket, will be  TCP_ESTABLISHED. Also, in this case,
1952          * sk->sk_user_data is not a valid function pointer.
1953          */
1954
1955         if (sk->sk_state == TCP_LISTEN) {
1956                 queue_work(o2net_wq, &o2net_listen_work);
1957         } else {
1958                 ready = NULL;
1959         }
1960
1961 out:
1962         read_unlock_bh(&sk->sk_callback_lock);
1963         if (ready != NULL)
1964                 ready(sk);
1965 }
1966
1967 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1968 {
1969         struct socket *sock = NULL;
1970         int ret;
1971         struct sockaddr_in sin = {
1972                 .sin_family = PF_INET,
1973                 .sin_addr = { .s_addr = addr },
1974                 .sin_port = port,
1975         };
1976
1977         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1978         if (ret < 0) {
1979                 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
1980                 goto out;
1981         }
1982
1983         sock->sk->sk_allocation = GFP_ATOMIC;
1984
1985         write_lock_bh(&sock->sk->sk_callback_lock);
1986         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1987         sock->sk->sk_data_ready = o2net_listen_data_ready;
1988         write_unlock_bh(&sock->sk->sk_callback_lock);
1989
1990         o2net_listen_sock = sock;
1991         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1992
1993         sock->sk->sk_reuse = SK_CAN_REUSE;
1994         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1995         if (ret < 0) {
1996                 printk(KERN_ERR "o2net: Error %d while binding socket at "
1997                        "%pI4:%u\n", ret, &addr, ntohs(port));
1998                 goto out;
1999         }
2000
2001         ret = sock->ops->listen(sock, 64);
2002         if (ret < 0)
2003                 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2004                        ret, &addr, ntohs(port));
2005
2006 out:
2007         if (ret) {
2008                 o2net_listen_sock = NULL;
2009                 if (sock)
2010                         sock_release(sock);
2011         }
2012         return ret;
2013 }
2014
2015 /*
2016  * called from node manager when we should bring up our network listening
2017  * socket.  node manager handles all the serialization to only call this
2018  * once and to match it with o2net_stop_listening().  note,
2019  * o2nm_this_node() doesn't work yet as we're being called while it
2020  * is being set up.
2021  */
2022 int o2net_start_listening(struct o2nm_node *node)
2023 {
2024         int ret = 0;
2025
2026         BUG_ON(o2net_wq != NULL);
2027         BUG_ON(o2net_listen_sock != NULL);
2028
2029         mlog(ML_KTHREAD, "starting o2net thread...\n");
2030         o2net_wq = alloc_ordered_workqueue("o2net", WQ_MEM_RECLAIM);
2031         if (o2net_wq == NULL) {
2032                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2033                 return -ENOMEM; /* ? */
2034         }
2035
2036         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2037                                         node->nd_ipv4_port);
2038         if (ret) {
2039                 destroy_workqueue(o2net_wq);
2040                 o2net_wq = NULL;
2041         } else
2042                 o2quo_conn_up(node->nd_num);
2043
2044         return ret;
2045 }
2046
2047 /* again, o2nm_this_node() doesn't work here as we're involved in
2048  * tearing it down */
2049 void o2net_stop_listening(struct o2nm_node *node)
2050 {
2051         struct socket *sock = o2net_listen_sock;
2052         size_t i;
2053
2054         BUG_ON(o2net_wq == NULL);
2055         BUG_ON(o2net_listen_sock == NULL);
2056
2057         /* stop the listening socket from generating work */
2058         write_lock_bh(&sock->sk->sk_callback_lock);
2059         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2060         sock->sk->sk_user_data = NULL;
2061         write_unlock_bh(&sock->sk->sk_callback_lock);
2062
2063         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2064                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2065                 if (node) {
2066                         o2net_disconnect_node(node);
2067                         o2nm_node_put(node);
2068                 }
2069         }
2070
2071         /* finish all work and tear down the work queue */
2072         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2073         destroy_workqueue(o2net_wq);
2074         o2net_wq = NULL;
2075
2076         sock_release(o2net_listen_sock);
2077         o2net_listen_sock = NULL;
2078
2079         o2quo_conn_err(node->nd_num);
2080 }
2081
2082 /* ------------------------------------------------------------ */
2083
2084 int o2net_init(void)
2085 {
2086         unsigned long i;
2087
2088         o2quo_init();
2089
2090         o2net_debugfs_init();
2091
2092         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2093         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2094         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2095         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp)
2096                 goto out;
2097
2098         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2099         o2net_hand->connector_id = cpu_to_be64(1);
2100
2101         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2102         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2103
2104         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2105                 struct o2net_node *nn = o2net_nn_from_num(i);
2106
2107                 atomic_set(&nn->nn_timeout, 0);
2108                 spin_lock_init(&nn->nn_lock);
2109                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2110                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2111                                   o2net_connect_expired);
2112                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2113                 /* until we see hb from a node we'll return einval */
2114                 nn->nn_persistent_error = -ENOTCONN;
2115                 init_waitqueue_head(&nn->nn_sc_wq);
2116                 idr_init(&nn->nn_status_idr);
2117                 INIT_LIST_HEAD(&nn->nn_status_list);
2118         }
2119
2120         return 0;
2121
2122 out:
2123         kfree(o2net_hand);
2124         kfree(o2net_keep_req);
2125         kfree(o2net_keep_resp);
2126         o2net_debugfs_exit();
2127         o2quo_exit();
2128         return -ENOMEM;
2129 }
2130
2131 void o2net_exit(void)
2132 {
2133         o2quo_exit();
2134         kfree(o2net_hand);
2135         kfree(o2net_keep_req);
2136         kfree(o2net_keep_resp);
2137         o2net_debugfs_exit();
2138 }