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 #include <trace/events/sock.h>
  50
  51 #include <linux/uaccess.h>
  52
  53 #include "heartbeat.h"
  54 #include "tcp.h"
  55 #include "nodemanager.h"
  56 #define MLOG_MASK_PREFIX ML_TCP
  57 #include "masklog.h"
  58 #include "quorum.h"
  59
  60 #include "tcp_internal.h"
  61
  62 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
  63 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
  64                           &sc->sc_node->nd_ipv4_address,                \
  65                           ntohs(sc->sc_node->nd_ipv4_port)
  66
  67 /*
  68  * In the following two log macros, the whitespace after the ',' just
  69  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  70  * previous token if args expands to nothing.
  71  */
  72 #define msglog(hdr, fmt, args...) do {                                  \
  73         typeof(hdr) __hdr = (hdr);                                      \
  74         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
  75              "key %08x num %u] " fmt,                                   \
  76              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
  77              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
  78              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
  79              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
  80 } while (0)
  81
  82 #define sclog(sc, fmt, args...) do {                                    \
  83         typeof(sc) __sc = (sc);                                         \
  84         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
  85              "pg_off %zu] " fmt, __sc,                                  \
  86              kref_read(&__sc->sc_kref), __sc->sc_sock,  \
  87             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
  88             ##args);                                                    \
  89 } while (0)
  90
  91 static DEFINE_RWLOCK(o2net_handler_lock);
  92 static struct rb_root o2net_handler_tree = RB_ROOT;
  93
  94 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
  95
  96 /* XXX someday we'll need better accounting */
  97 static struct socket *o2net_listen_sock;
  98
  99 /*
 100  * listen work is only queued by the listening socket callbacks on the
 101  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 102  * quorum work is queued as sock containers are shutdown.. stop_listening
 103  * tears down all the node's sock containers, preventing future shutdowns
 104  * and queued quroum work, before canceling delayed quorum work and
 105  * destroying the work queue.
 106  */
 107 static struct workqueue_struct *o2net_wq;
 108 static struct work_struct o2net_listen_work;
 109
 110 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 111 #define O2NET_HB_PRI 0x1
 112
 113 static struct o2net_handshake *o2net_hand;
 114 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 115
 116 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 117                 {[O2NET_ERR_NONE]       = 0,
 118                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
 119                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
 120                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
 121
 122 /* can't quite avoid *all* internal declarations :/ */
 123 static void o2net_sc_connect_completed(struct work_struct *work);
 124 static void o2net_rx_until_empty(struct work_struct *work);
 125 static void o2net_shutdown_sc(struct work_struct *work);
 126 static void o2net_listen_data_ready(struct sock *sk);
 127 static void o2net_sc_send_keep_req(struct work_struct *work);
 128 static void o2net_idle_timer(struct timer_list *t);
 129 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 130 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 131
 132 #ifdef CONFIG_DEBUG_FS
 133 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 134                            u32 msgkey, struct task_struct *task, u8 node)
 135 {
 136         INIT_LIST_HEAD(&nst->st_net_debug_item);
 137         nst->st_task = task;
 138         nst->st_msg_type = msgtype;
 139         nst->st_msg_key = msgkey;
 140         nst->st_node = node;
 141 }
 142
 143 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 144 {
 145         nst->st_sock_time = ktime_get();
 146 }
 147
 148 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 149 {
 150         nst->st_send_time = ktime_get();
 151 }
 152
 153 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 154 {
 155         nst->st_status_time = ktime_get();
 156 }
 157
 158 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 159                                                 struct o2net_sock_container *sc)
 160 {
 161         nst->st_sc = sc;
 162 }
 163
 164 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 165                                         u32 msg_id)
 166 {
 167         nst->st_id = msg_id;
 168 }
 169
 170 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
 171 {
 172         sc->sc_tv_timer = ktime_get();
 173 }
 174
 175 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
 176 {
 177         sc->sc_tv_data_ready = ktime_get();
 178 }
 179
 180 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
 181 {
 182         sc->sc_tv_advance_start = ktime_get();
 183 }
 184
 185 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
 186 {
 187         sc->sc_tv_advance_stop = ktime_get();
 188 }
 189
 190 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
 191 {
 192         sc->sc_tv_func_start = ktime_get();
 193 }
 194
 195 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
 196 {
 197         sc->sc_tv_func_stop = ktime_get();
 198 }
 199
 200 #else  /* CONFIG_DEBUG_FS */
 201 # define o2net_init_nst(a, b, c, d, e)
 202 # define o2net_set_nst_sock_time(a)
 203 # define o2net_set_nst_send_time(a)
 204 # define o2net_set_nst_status_time(a)
 205 # define o2net_set_nst_sock_container(a, b)
 206 # define o2net_set_nst_msg_id(a, b)
 207 # define o2net_set_sock_timer(a)
 208 # define o2net_set_data_ready_time(a)
 209 # define o2net_set_advance_start_time(a)
 210 # define o2net_set_advance_stop_time(a)
 211 # define o2net_set_func_start_time(a)
 212 # define o2net_set_func_stop_time(a)
 213 #endif /* CONFIG_DEBUG_FS */
 214
 215 #ifdef CONFIG_OCFS2_FS_STATS
 216 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
 217 {
 218         return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
 219 }
 220
 221 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
 222                                     struct o2net_sock_container *sc)
 223 {
 224         sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
 225                                            ktime_sub(ktime_get(),
 226                                                      nst->st_status_time));
 227         sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
 228                                          ktime_sub(nst->st_status_time,
 229                                                    nst->st_send_time));
 230         sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
 231                                             ktime_sub(nst->st_send_time,
 232                                                       nst->st_sock_time));
 233         sc->sc_send_count++;
 234 }
 235
 236 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
 237 {
 238         sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
 239                                             o2net_get_func_run_time(sc));
 240         sc->sc_recv_count++;
 241 }
 242
 243 #else
 244
 245 # define o2net_update_send_stats(a, b)
 246
 247 # define o2net_update_recv_stats(sc)
 248
 249 #endif /* CONFIG_OCFS2_FS_STATS */
 250
 251 static inline unsigned int o2net_reconnect_delay(void)
 252 {
 253         return o2nm_single_cluster->cl_reconnect_delay_ms;
 254 }
 255
 256 static inline unsigned int o2net_keepalive_delay(void)
 257 {
 258         return o2nm_single_cluster->cl_keepalive_delay_ms;
 259 }
 260
 261 static inline unsigned int o2net_idle_timeout(void)
 262 {
 263         return o2nm_single_cluster->cl_idle_timeout_ms;
 264 }
 265
 266 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 267 {
 268         int trans;
 269         BUG_ON(err >= O2NET_ERR_MAX);
 270         trans = o2net_sys_err_translations[err];
 271
 272         /* Just in case we mess up the translation table above */
 273         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 274         return trans;
 275 }
 276
 277 static struct o2net_node * o2net_nn_from_num(u8 node_num)
 278 {
 279         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 280         return &o2net_nodes[node_num];
 281 }
 282
 283 static u8 o2net_num_from_nn(struct o2net_node *nn)
 284 {
 285         BUG_ON(nn == NULL);
 286         return nn - o2net_nodes;
 287 }
 288
 289 /* ------------------------------------------------------------ */
 290
 291 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 292 {
 293         int ret;
 294
 295         spin_lock(&nn->nn_lock);
 296         ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
 297         if (ret >= 0) {
 298                 nsw->ns_id = ret;
 299                 list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
 300         }
 301         spin_unlock(&nn->nn_lock);
 302         if (ret < 0)
 303                 return ret;
 304
 305         init_waitqueue_head(&nsw->ns_wq);
 306         nsw->ns_sys_status = O2NET_ERR_NONE;
 307         nsw->ns_status = 0;
 308         return 0;
 309 }
 310
 311 static void o2net_complete_nsw_locked(struct o2net_node *nn,
 312                                       struct o2net_status_wait *nsw,
 313                                       enum o2net_system_error sys_status,
 314                                       s32 status)
 315 {
 316         assert_spin_locked(&nn->nn_lock);
 317
 318         if (!list_empty(&nsw->ns_node_item)) {
 319                 list_del_init(&nsw->ns_node_item);
 320                 nsw->ns_sys_status = sys_status;
 321                 nsw->ns_status = status;
 322                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
 323                 wake_up(&nsw->ns_wq);
 324         }
 325 }
 326
 327 static void o2net_complete_nsw(struct o2net_node *nn,
 328                                struct o2net_status_wait *nsw,
 329                                u64 id, enum o2net_system_error sys_status,
 330                                s32 status)
 331 {
 332         spin_lock(&nn->nn_lock);
 333         if (nsw == NULL) {
 334                 if (id > INT_MAX)
 335                         goto out;
 336
 337                 nsw = idr_find(&nn->nn_status_idr, id);
 338                 if (nsw == NULL)
 339                         goto out;
 340         }
 341
 342         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 343
 344 out:
 345         spin_unlock(&nn->nn_lock);
 346         return;
 347 }
 348
 349 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 350 {
 351         struct o2net_status_wait *nsw, *tmp;
 352         unsigned int num_kills = 0;
 353
 354         assert_spin_locked(&nn->nn_lock);
 355
 356         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 357                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 358                 num_kills++;
 359         }
 360
 361         mlog(0, "completed %d messages for node %u\n", num_kills,
 362              o2net_num_from_nn(nn));
 363 }
 364
 365 static int o2net_nsw_completed(struct o2net_node *nn,
 366                                struct o2net_status_wait *nsw)
 367 {
 368         int completed;
 369         spin_lock(&nn->nn_lock);
 370         completed = list_empty(&nsw->ns_node_item);
 371         spin_unlock(&nn->nn_lock);
 372         return completed;
 373 }
 374
 375 /* ------------------------------------------------------------ */
 376
 377 static void sc_kref_release(struct kref *kref)
 378 {
 379         struct o2net_sock_container *sc = container_of(kref,
 380                                         struct o2net_sock_container, sc_kref);
 381         BUG_ON(timer_pending(&sc->sc_idle_timeout));
 382
 383         sclog(sc, "releasing\n");
 384
 385         if (sc->sc_sock) {
 386                 sock_release(sc->sc_sock);
 387                 sc->sc_sock = NULL;
 388         }
 389
 390         o2nm_undepend_item(&sc->sc_node->nd_item);
 391         o2nm_node_put(sc->sc_node);
 392         sc->sc_node = NULL;
 393
 394         o2net_debug_del_sc(sc);
 395
 396         if (sc->sc_page)
 397                 __free_page(sc->sc_page);
 398         kfree(sc);
 399 }
 400
 401 static void sc_put(struct o2net_sock_container *sc)
 402 {
 403         sclog(sc, "put\n");
 404         kref_put(&sc->sc_kref, sc_kref_release);
 405 }
 406 static void sc_get(struct o2net_sock_container *sc)
 407 {
 408         sclog(sc, "get\n");
 409         kref_get(&sc->sc_kref);
 410 }
 411 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 412 {
 413         struct o2net_sock_container *sc, *ret = NULL;
 414         struct page *page = NULL;
 415         int status = 0;
 416
 417         page = alloc_page(GFP_NOFS);
 418         sc = kzalloc(sizeof(*sc), GFP_NOFS);
 419         if (sc == NULL || page == NULL)
 420                 goto out;
 421
 422         kref_init(&sc->sc_kref);
 423         o2nm_node_get(node);
 424         sc->sc_node = node;
 425
 426         /* pin the node item of the remote node */
 427         status = o2nm_depend_item(&node->nd_item);
 428         if (status) {
 429                 mlog_errno(status);
 430                 o2nm_node_put(node);
 431                 goto out;
 432         }
 433         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 434         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 435         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 436         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 437
 438         timer_setup(&sc->sc_idle_timeout, o2net_idle_timer, 0);
 439
 440         sclog(sc, "alloced\n");
 441
 442         ret = sc;
 443         sc->sc_page = page;
 444         o2net_debug_add_sc(sc);
 445         sc = NULL;
 446         page = NULL;
 447
 448 out:
 449         if (page)
 450                 __free_page(page);
 451         kfree(sc);
 452
 453         return ret;
 454 }
 455
 456 /* ------------------------------------------------------------ */
 457
 458 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 459                                 struct work_struct *work)
 460 {
 461         sc_get(sc);
 462         if (!queue_work(o2net_wq, work))
 463                 sc_put(sc);
 464 }
 465 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 466                                         struct delayed_work *work,
 467                                         int delay)
 468 {
 469         sc_get(sc);
 470         if (!queue_delayed_work(o2net_wq, work, delay))
 471                 sc_put(sc);
 472 }
 473 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 474                                          struct delayed_work *work)
 475 {
 476         if (cancel_delayed_work(work))
 477                 sc_put(sc);
 478 }
 479
 480 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 481
 482 int o2net_num_connected_peers(void)
 483 {
 484         return atomic_read(&o2net_connected_peers);
 485 }
 486
 487 static void o2net_set_nn_state(struct o2net_node *nn,
 488                                struct o2net_sock_container *sc,
 489                                unsigned valid, int err)
 490 {
 491         int was_valid = nn->nn_sc_valid;
 492         int was_err = nn->nn_persistent_error;
 493         struct o2net_sock_container *old_sc = nn->nn_sc;
 494
 495         assert_spin_locked(&nn->nn_lock);
 496
 497         if (old_sc && !sc)
 498                 atomic_dec(&o2net_connected_peers);
 499         else if (!old_sc && sc)
 500                 atomic_inc(&o2net_connected_peers);
 501
 502         /* the node num comparison and single connect/accept path should stop
 503          * an non-null sc from being overwritten with another */
 504         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 505         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 506         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 507
 508         if (was_valid && !valid && err == 0)
 509                 err = -ENOTCONN;
 510
 511         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 512              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 513              nn->nn_persistent_error, err);
 514
 515         nn->nn_sc = sc;
 516         nn->nn_sc_valid = valid ? 1 : 0;
 517         nn->nn_persistent_error = err;
 518
 519         /* mirrors o2net_tx_can_proceed() */
 520         if (nn->nn_persistent_error || nn->nn_sc_valid)
 521                 wake_up(&nn->nn_sc_wq);
 522
 523         if (was_valid && !was_err && nn->nn_persistent_error) {
 524                 o2quo_conn_err(o2net_num_from_nn(nn));
 525                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
 526                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 527         }
 528
 529         if (was_valid && !valid) {
 530                 if (old_sc)
 531                         printk(KERN_NOTICE "o2net: No longer connected to "
 532                                 SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 533                 o2net_complete_nodes_nsw(nn);
 534         }
 535
 536         if (!was_valid && valid) {
 537                 o2quo_conn_up(o2net_num_from_nn(nn));
 538                 cancel_delayed_work(&nn->nn_connect_expired);
 539                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 540                        o2nm_this_node() > sc->sc_node->nd_num ?
 541                        "Connected to" : "Accepted connection from",
 542                        SC_NODEF_ARGS(sc));
 543         }
 544
 545         /* trigger the connecting worker func as long as we're not valid,
 546          * it will back off if it shouldn't connect.  This can be called
 547          * from node config teardown and so needs to be careful about
 548          * the work queue actually being up. */
 549         if (!valid && o2net_wq) {
 550                 unsigned long delay;
 551                 /* delay if we're within a RECONNECT_DELAY of the
 552                  * last attempt */
 553                 delay = (nn->nn_last_connect_attempt +
 554                          msecs_to_jiffies(o2net_reconnect_delay()))
 555                         - jiffies;
 556                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 557                         delay = 0;
 558                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 559                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 560
 561                 /*
 562                  * Delay the expired work after idle timeout.
 563                  *
 564                  * We might have lots of failed connection attempts that run
 565                  * through here but we only cancel the connect_expired work when
 566                  * a connection attempt succeeds.  So only the first enqueue of
 567                  * the connect_expired work will do anything.  The rest will see
 568                  * that it's already queued and do nothing.
 569                  */
 570                 delay += msecs_to_jiffies(o2net_idle_timeout());
 571                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 572         }
 573
 574         /* keep track of the nn's sc ref for the caller */
 575         if ((old_sc == NULL) && sc)
 576                 sc_get(sc);
 577         if (old_sc && (old_sc != sc)) {
 578                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 579                 sc_put(old_sc);
 580         }
 581 }
 582
 583 /* see o2net_register_callbacks() */
 584 static void o2net_data_ready(struct sock *sk)
 585 {
 586         void (*ready)(struct sock *sk);
 587         struct o2net_sock_container *sc;
 588
 589         trace_sk_data_ready(sk);
 590
 591         read_lock_bh(&sk->sk_callback_lock);
 592         sc = sk->sk_user_data;
 593         if (sc) {
 594                 sclog(sc, "data_ready hit\n");
 595                 o2net_set_data_ready_time(sc);
 596                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
 597                 ready = sc->sc_data_ready;
 598         } else {
 599                 ready = sk->sk_data_ready;
 600         }
 601         read_unlock_bh(&sk->sk_callback_lock);
 602
 603         ready(sk);
 604 }
 605
 606 /* see o2net_register_callbacks() */
 607 static void o2net_state_change(struct sock *sk)
 608 {
 609         void (*state_change)(struct sock *sk);
 610         struct o2net_sock_container *sc;
 611
 612         read_lock_bh(&sk->sk_callback_lock);
 613         sc = sk->sk_user_data;
 614         if (sc == NULL) {
 615                 state_change = sk->sk_state_change;
 616                 goto out;
 617         }
 618
 619         sclog(sc, "state_change to %d\n", sk->sk_state);
 620
 621         state_change = sc->sc_state_change;
 622
 623         switch(sk->sk_state) {
 624         /* ignore connecting sockets as they make progress */
 625         case TCP_SYN_SENT:
 626         case TCP_SYN_RECV:
 627                 break;
 628         case TCP_ESTABLISHED:
 629                 o2net_sc_queue_work(sc, &sc->sc_connect_work);
 630                 break;
 631         default:
 632                 printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
 633                         " shutdown, state %d\n",
 634                         SC_NODEF_ARGS(sc), sk->sk_state);
 635                 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 636                 break;
 637         }
 638 out:
 639         read_unlock_bh(&sk->sk_callback_lock);
 640         state_change(sk);
 641 }
 642
 643 /*
 644  * we register callbacks so we can queue work on events before calling
 645  * the original callbacks.  our callbacks our careful to test user_data
 646  * to discover when they've reaced with o2net_unregister_callbacks().
 647  */
 648 static void o2net_register_callbacks(struct sock *sk,
 649                                      struct o2net_sock_container *sc)
 650 {
 651         write_lock_bh(&sk->sk_callback_lock);
 652
 653         /* accepted sockets inherit the old listen socket data ready */
 654         if (sk->sk_data_ready == o2net_listen_data_ready) {
 655                 sk->sk_data_ready = sk->sk_user_data;
 656                 sk->sk_user_data = NULL;
 657         }
 658
 659         BUG_ON(sk->sk_user_data != NULL);
 660         sk->sk_user_data = sc;
 661         sc_get(sc);
 662
 663         sc->sc_data_ready = sk->sk_data_ready;
 664         sc->sc_state_change = sk->sk_state_change;
 665         sk->sk_data_ready = o2net_data_ready;
 666         sk->sk_state_change = o2net_state_change;
 667
 668         mutex_init(&sc->sc_send_lock);
 669
 670         write_unlock_bh(&sk->sk_callback_lock);
 671 }
 672
 673 static int o2net_unregister_callbacks(struct sock *sk,
 674                                    struct o2net_sock_container *sc)
 675 {
 676         int ret = 0;
 677
 678         write_lock_bh(&sk->sk_callback_lock);
 679         if (sk->sk_user_data == sc) {
 680                 ret = 1;
 681                 sk->sk_user_data = NULL;
 682                 sk->sk_data_ready = sc->sc_data_ready;
 683                 sk->sk_state_change = sc->sc_state_change;
 684         }
 685         write_unlock_bh(&sk->sk_callback_lock);
 686
 687         return ret;
 688 }
 689
 690 /*
 691  * this is a little helper that is called by callers who have seen a problem
 692  * with an sc and want to detach it from the nn if someone already hasn't beat
 693  * them to it.  if an error is given then the shutdown will be persistent
 694  * and pending transmits will be canceled.
 695  */
 696 static void o2net_ensure_shutdown(struct o2net_node *nn,
 697                                    struct o2net_sock_container *sc,
 698                                    int err)
 699 {
 700         spin_lock(&nn->nn_lock);
 701         if (nn->nn_sc == sc)
 702                 o2net_set_nn_state(nn, NULL, 0, err);
 703         spin_unlock(&nn->nn_lock);
 704 }
 705
 706 /*
 707  * This work queue function performs the blocking parts of socket shutdown.  A
 708  * few paths lead here.  set_nn_state will trigger this callback if it sees an
 709  * sc detached from the nn.  state_change will also trigger this callback
 710  * directly when it sees errors.  In that case we need to call set_nn_state
 711  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 712  * itself.
 713  */
 714 static void o2net_shutdown_sc(struct work_struct *work)
 715 {
 716         struct o2net_sock_container *sc =
 717                 container_of(work, struct o2net_sock_container,
 718                              sc_shutdown_work);
 719         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 720
 721         sclog(sc, "shutting down\n");
 722
 723         /* drop the callbacks ref and call shutdown only once */
 724         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 725                 /* we shouldn't flush as we're in the thread, the
 726                  * races with pending sc work structs are harmless */
 727                 del_timer_sync(&sc->sc_idle_timeout);
 728                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 729                 sc_put(sc);
 730                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 731         }
 732
 733         /* not fatal so failed connects before the other guy has our
 734          * heartbeat can be retried */
 735         o2net_ensure_shutdown(nn, sc, 0);
 736         sc_put(sc);
 737 }
 738
 739 /* ------------------------------------------------------------ */
 740
 741 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 742                              u32 key)
 743 {
 744         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 745
 746         if (ret == 0)
 747                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 748
 749         return ret;
 750 }
 751
 752 static struct o2net_msg_handler *
 753 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 754                           struct rb_node **ret_parent)
 755 {
 756         struct rb_node **p = &o2net_handler_tree.rb_node;
 757         struct rb_node *parent = NULL;
 758         struct o2net_msg_handler *nmh, *ret = NULL;
 759         int cmp;
 760
 761         while (*p) {
 762                 parent = *p;
 763                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 764                 cmp = o2net_handler_cmp(nmh, msg_type, key);
 765
 766                 if (cmp < 0)
 767                         p = &(*p)->rb_left;
 768                 else if (cmp > 0)
 769                         p = &(*p)->rb_right;
 770                 else {
 771                         ret = nmh;
 772                         break;
 773                 }
 774         }
 775
 776         if (ret_p != NULL)
 777                 *ret_p = p;
 778         if (ret_parent != NULL)
 779                 *ret_parent = parent;
 780
 781         return ret;
 782 }
 783
 784 static void o2net_handler_kref_release(struct kref *kref)
 785 {
 786         struct o2net_msg_handler *nmh;
 787         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 788
 789         kfree(nmh);
 790 }
 791
 792 static void o2net_handler_put(struct o2net_msg_handler *nmh)
 793 {
 794         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 795 }
 796
 797 /* max_len is protection for the handler func.  incoming messages won't
 798  * be given to the handler if their payload is longer than the max. */
 799 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 800                            o2net_msg_handler_func *func, void *data,
 801                            o2net_post_msg_handler_func *post_func,
 802                            struct list_head *unreg_list)
 803 {
 804         struct o2net_msg_handler *nmh = NULL;
 805         struct rb_node **p, *parent;
 806         int ret = 0;
 807
 808         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 809                 mlog(0, "max_len for message handler out of range: %u\n",
 810                         max_len);
 811                 ret = -EINVAL;
 812                 goto out;
 813         }
 814
 815         if (!msg_type) {
 816                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 817                 ret = -EINVAL;
 818                 goto out;
 819
 820         }
 821         if (!func) {
 822                 mlog(0, "no message handler provided: %u, %p\n",
 823                        msg_type, func);
 824                 ret = -EINVAL;
 825                 goto out;
 826         }
 827
 828         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 829         if (nmh == NULL) {
 830                 ret = -ENOMEM;
 831                 goto out;
 832         }
 833
 834         nmh->nh_func = func;
 835         nmh->nh_func_data = data;
 836         nmh->nh_post_func = post_func;
 837         nmh->nh_msg_type = msg_type;
 838         nmh->nh_max_len = max_len;
 839         nmh->nh_key = key;
 840         /* the tree and list get this ref.. they're both removed in
 841          * unregister when this ref is dropped */
 842         kref_init(&nmh->nh_kref);
 843         INIT_LIST_HEAD(&nmh->nh_unregister_item);
 844
 845         write_lock(&o2net_handler_lock);
 846         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 847                 ret = -EEXIST;
 848         else {
 849                 rb_link_node(&nmh->nh_node, parent, p);
 850                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 851                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
 852
 853                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 854                      func, msg_type, key);
 855                 /* we've had some trouble with handlers seemingly vanishing. */
 856                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 857                                                           &parent) == NULL,
 858                                 "couldn't find handler we *just* registered "
 859                                 "for type %u key %08x\n", msg_type, key);
 860         }
 861         write_unlock(&o2net_handler_lock);
 862
 863 out:
 864         if (ret)
 865                 kfree(nmh);
 866
 867         return ret;
 868 }
 869 EXPORT_SYMBOL_GPL(o2net_register_handler);
 870
 871 void o2net_unregister_handler_list(struct list_head *list)
 872 {
 873         struct o2net_msg_handler *nmh, *n;
 874
 875         write_lock(&o2net_handler_lock);
 876         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 877                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 878                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 879                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
 880                 list_del_init(&nmh->nh_unregister_item);
 881                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 882         }
 883         write_unlock(&o2net_handler_lock);
 884 }
 885 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 886
 887 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 888 {
 889         struct o2net_msg_handler *nmh;
 890
 891         read_lock(&o2net_handler_lock);
 892         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 893         if (nmh)
 894                 kref_get(&nmh->nh_kref);
 895         read_unlock(&o2net_handler_lock);
 896
 897         return nmh;
 898 }
 899
 900 /* ------------------------------------------------------------ */
 901
 902 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 903 {
 904         struct kvec vec = { .iov_len = len, .iov_base = data, };
 905         struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
 906         iov_iter_kvec(&msg.msg_iter, ITER_DEST, &vec, 1, len);
 907         return sock_recvmsg(sock, &msg, MSG_DONTWAIT);
 908 }
 909
 910 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 911                               size_t veclen, size_t total)
 912 {
 913         int ret;
 914         struct msghdr msg = {.msg_flags = 0,};
 915
 916         if (sock == NULL) {
 917                 ret = -EINVAL;
 918                 goto out;
 919         }
 920
 921         ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
 922         if (likely(ret == total))
 923                 return 0;
 924         mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
 925         if (ret >= 0)
 926                 ret = -EPIPE; /* should be smarter, I bet */
 927 out:
 928         mlog(0, "returning error: %d\n", ret);
 929         return ret;
 930 }
 931
 932 static void o2net_sendpage(struct o2net_sock_container *sc,
 933                            void *kmalloced_virt,
 934                            size_t size)
 935 {
 936         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 937         ssize_t ret;
 938
 939         while (1) {
 940                 mutex_lock(&sc->sc_send_lock);
 941                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 942                                                  virt_to_page(kmalloced_virt),
 943                                                  offset_in_page(kmalloced_virt),
 944                                                  size, MSG_DONTWAIT);
 945                 mutex_unlock(&sc->sc_send_lock);
 946                 if (ret == size)
 947                         break;
 948                 if (ret == (ssize_t)-EAGAIN) {
 949                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 950                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 951                         cond_resched();
 952                         continue;
 953                 }
 954                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 955                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 956                 o2net_ensure_shutdown(nn, sc, 0);
 957                 break;
 958         }
 959 }
 960
 961 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 962 {
 963         memset(msg, 0, sizeof(struct o2net_msg));
 964         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 965         msg->data_len = cpu_to_be16(data_len);
 966         msg->msg_type = cpu_to_be16(msg_type);
 967         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 968         msg->status = 0;
 969         msg->key = cpu_to_be32(key);
 970 }
 971
 972 static int o2net_tx_can_proceed(struct o2net_node *nn,
 973                                 struct o2net_sock_container **sc_ret,
 974                                 int *error)
 975 {
 976         int ret = 0;
 977
 978         spin_lock(&nn->nn_lock);
 979         if (nn->nn_persistent_error) {
 980                 ret = 1;
 981                 *sc_ret = NULL;
 982                 *error = nn->nn_persistent_error;
 983         } else if (nn->nn_sc_valid) {
 984                 kref_get(&nn->nn_sc->sc_kref);
 985
 986                 ret = 1;
 987                 *sc_ret = nn->nn_sc;
 988                 *error = 0;
 989         }
 990         spin_unlock(&nn->nn_lock);
 991
 992         return ret;
 993 }
 994
 995 /* Get a map of all nodes to which this node is currently connected to */
 996 void o2net_fill_node_map(unsigned long *map, unsigned int bits)
 997 {
 998         struct o2net_sock_container *sc;
 999         int node, ret;
1000
1001         bitmap_zero(map, bits);
1002         for (node = 0; node < O2NM_MAX_NODES; ++node) {
1003                 if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1004                         continue;
1005                 if (!ret) {
1006                         set_bit(node, map);
1007                         sc_put(sc);
1008                 }
1009         }
1010 }
1011 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1012
1013 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1014                            size_t caller_veclen, u8 target_node, int *status)
1015 {
1016         int ret = 0;
1017         struct o2net_msg *msg = NULL;
1018         size_t veclen, caller_bytes = 0;
1019         struct kvec *vec = NULL;
1020         struct o2net_sock_container *sc = NULL;
1021         struct o2net_node *nn = o2net_nn_from_num(target_node);
1022         struct o2net_status_wait nsw = {
1023                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1024         };
1025         struct o2net_send_tracking nst;
1026
1027         o2net_init_nst(&nst, msg_type, key, current, target_node);
1028
1029         if (o2net_wq == NULL) {
1030                 mlog(0, "attempt to tx without o2netd running\n");
1031                 ret = -ESRCH;
1032                 goto out;
1033         }
1034
1035         if (caller_veclen == 0) {
1036                 mlog(0, "bad kvec array length\n");
1037                 ret = -EINVAL;
1038                 goto out;
1039         }
1040
1041         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1042         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1043                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1044                 ret = -EINVAL;
1045                 goto out;
1046         }
1047
1048         if (target_node == o2nm_this_node()) {
1049                 ret = -ELOOP;
1050                 goto out;
1051         }
1052
1053         o2net_debug_add_nst(&nst);
1054
1055         o2net_set_nst_sock_time(&nst);
1056
1057         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1058         if (ret)
1059                 goto out;
1060
1061         o2net_set_nst_sock_container(&nst, sc);
1062
1063         veclen = caller_veclen + 1;
1064         vec = kmalloc_array(veclen, sizeof(struct kvec), GFP_ATOMIC);
1065         if (vec == NULL) {
1066                 mlog(0, "failed to %zu element kvec!\n", veclen);
1067                 ret = -ENOMEM;
1068                 goto out;
1069         }
1070
1071         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1072         if (!msg) {
1073                 mlog(0, "failed to allocate a o2net_msg!\n");
1074                 ret = -ENOMEM;
1075                 goto out;
1076         }
1077
1078         o2net_init_msg(msg, caller_bytes, msg_type, key);
1079
1080         vec[0].iov_len = sizeof(struct o2net_msg);
1081         vec[0].iov_base = msg;
1082         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1083
1084         ret = o2net_prep_nsw(nn, &nsw);
1085         if (ret)
1086                 goto out;
1087
1088         msg->msg_num = cpu_to_be32(nsw.ns_id);
1089         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1090
1091         o2net_set_nst_send_time(&nst);
1092
1093         /* finally, convert the message header to network byte-order
1094          * and send */
1095         mutex_lock(&sc->sc_send_lock);
1096         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1097                                  sizeof(struct o2net_msg) + caller_bytes);
1098         mutex_unlock(&sc->sc_send_lock);
1099         msglog(msg, "sending returned %d\n", ret);
1100         if (ret < 0) {
1101                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1102                 goto out;
1103         }
1104
1105         /* wait on other node's handler */
1106         o2net_set_nst_status_time(&nst);
1107         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1108
1109         o2net_update_send_stats(&nst, sc);
1110
1111         /* Note that we avoid overwriting the callers status return
1112          * variable if a system error was reported on the other
1113          * side. Callers beware. */
1114         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1115         if (status && !ret)
1116                 *status = nsw.ns_status;
1117
1118         mlog(0, "woken, returning system status %d, user status %d\n",
1119              ret, nsw.ns_status);
1120 out:
1121         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1122         if (sc)
1123                 sc_put(sc);
1124         kfree(vec);
1125         kfree(msg);
1126         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1127         return ret;
1128 }
1129 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1130
1131 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1132                        u8 target_node, int *status)
1133 {
1134         struct kvec vec = {
1135                 .iov_base = data,
1136                 .iov_len = len,
1137         };
1138         return o2net_send_message_vec(msg_type, key, &vec, 1,
1139                                       target_node, status);
1140 }
1141 EXPORT_SYMBOL_GPL(o2net_send_message);
1142
1143 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1144                                    enum o2net_system_error syserr, int err)
1145 {
1146         struct kvec vec = {
1147                 .iov_base = hdr,
1148                 .iov_len = sizeof(struct o2net_msg),
1149         };
1150
1151         BUG_ON(syserr >= O2NET_ERR_MAX);
1152
1153         /* leave other fields intact from the incoming message, msg_num
1154          * in particular */
1155         hdr->sys_status = cpu_to_be32(syserr);
1156         hdr->status = cpu_to_be32(err);
1157         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1158         hdr->data_len = 0;
1159
1160         msglog(hdr, "about to send status magic %d\n", err);
1161         /* hdr has been in host byteorder this whole time */
1162         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1163 }
1164
1165 /* this returns -errno if the header was unknown or too large, etc.
1166  * after this is called the buffer us reused for the next message */
1167 static int o2net_process_message(struct o2net_sock_container *sc,
1168                                  struct o2net_msg *hdr)
1169 {
1170         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1171         int ret = 0, handler_status;
1172         enum  o2net_system_error syserr;
1173         struct o2net_msg_handler *nmh = NULL;
1174         void *ret_data = NULL;
1175
1176         msglog(hdr, "processing message\n");
1177
1178         o2net_sc_postpone_idle(sc);
1179
1180         switch(be16_to_cpu(hdr->magic)) {
1181                 case O2NET_MSG_STATUS_MAGIC:
1182                         /* special type for returning message status */
1183                         o2net_complete_nsw(nn, NULL,
1184                                            be32_to_cpu(hdr->msg_num),
1185                                            be32_to_cpu(hdr->sys_status),
1186                                            be32_to_cpu(hdr->status));
1187                         goto out;
1188                 case O2NET_MSG_KEEP_REQ_MAGIC:
1189                         o2net_sendpage(sc, o2net_keep_resp,
1190                                        sizeof(*o2net_keep_resp));
1191                         goto out;
1192                 case O2NET_MSG_KEEP_RESP_MAGIC:
1193                         goto out;
1194                 case O2NET_MSG_MAGIC:
1195                         break;
1196                 default:
1197                         msglog(hdr, "bad magic\n");
1198                         ret = -EINVAL;
1199                         goto out;
1200         }
1201
1202         /* find a handler for it */
1203         handler_status = 0;
1204         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1205                                 be32_to_cpu(hdr->key));
1206         if (!nmh) {
1207                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1208                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1209                 syserr = O2NET_ERR_NO_HNDLR;
1210                 goto out_respond;
1211         }
1212
1213         syserr = O2NET_ERR_NONE;
1214
1215         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1216                 syserr = O2NET_ERR_OVERFLOW;
1217
1218         if (syserr != O2NET_ERR_NONE)
1219                 goto out_respond;
1220
1221         o2net_set_func_start_time(sc);
1222         sc->sc_msg_key = be32_to_cpu(hdr->key);
1223         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1224         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1225                                              be16_to_cpu(hdr->data_len),
1226                                         nmh->nh_func_data, &ret_data);
1227         o2net_set_func_stop_time(sc);
1228
1229         o2net_update_recv_stats(sc);
1230
1231 out_respond:
1232         /* this destroys the hdr, so don't use it after this */
1233         mutex_lock(&sc->sc_send_lock);
1234         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1235                                       handler_status);
1236         mutex_unlock(&sc->sc_send_lock);
1237         hdr = NULL;
1238         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1239              handler_status, syserr, ret);
1240
1241         if (nmh) {
1242                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1243                 if (nmh->nh_post_func)
1244                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1245                                             ret_data);
1246         }
1247
1248 out:
1249         if (nmh)
1250                 o2net_handler_put(nmh);
1251         return ret;
1252 }
1253
1254 static int o2net_check_handshake(struct o2net_sock_container *sc)
1255 {
1256         struct o2net_handshake *hand = page_address(sc->sc_page);
1257         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1258
1259         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1260                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1261                        "protocol version %llu but %llu is required. "
1262                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1263                        (unsigned long long)be64_to_cpu(hand->protocol_version),
1264                        O2NET_PROTOCOL_VERSION);
1265
1266                 /* don't bother reconnecting if its the wrong version. */
1267                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1268                 return -1;
1269         }
1270
1271         /*
1272          * Ensure timeouts are consistent with other nodes, otherwise
1273          * we can end up with one node thinking that the other must be down,
1274          * but isn't. This can ultimately cause corruption.
1275          */
1276         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1277                                 o2net_idle_timeout()) {
1278                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1279                        "idle timeout of %u ms, but we use %u ms locally. "
1280                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1281                        be32_to_cpu(hand->o2net_idle_timeout_ms),
1282                        o2net_idle_timeout());
1283                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1284                 return -1;
1285         }
1286
1287         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1288                         o2net_keepalive_delay()) {
1289                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1290                        "delay of %u ms, but we use %u ms locally. "
1291                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1292                        be32_to_cpu(hand->o2net_keepalive_delay_ms),
1293                        o2net_keepalive_delay());
1294                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1295                 return -1;
1296         }
1297
1298         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1299                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1300                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1301                        "timeout of %u ms, but we use %u ms locally. "
1302                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1303                        be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1304                        O2HB_MAX_WRITE_TIMEOUT_MS);
1305                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1306                 return -1;
1307         }
1308
1309         sc->sc_handshake_ok = 1;
1310
1311         spin_lock(&nn->nn_lock);
1312         /* set valid and queue the idle timers only if it hasn't been
1313          * shut down already */
1314         if (nn->nn_sc == sc) {
1315                 o2net_sc_reset_idle_timer(sc);
1316                 atomic_set(&nn->nn_timeout, 0);
1317                 o2net_set_nn_state(nn, sc, 1, 0);
1318         }
1319         spin_unlock(&nn->nn_lock);
1320
1321         /* shift everything up as though it wasn't there */
1322         sc->sc_page_off -= sizeof(struct o2net_handshake);
1323         if (sc->sc_page_off)
1324                 memmove(hand, hand + 1, sc->sc_page_off);
1325
1326         return 0;
1327 }
1328
1329 /* this demuxes the queued rx bytes into header or payload bits and calls
1330  * handlers as each full message is read off the socket.  it returns -error,
1331  * == 0 eof, or > 0 for progress made.*/
1332 static int o2net_advance_rx(struct o2net_sock_container *sc)
1333 {
1334         struct o2net_msg *hdr;
1335         int ret = 0;
1336         void *data;
1337         size_t datalen;
1338
1339         sclog(sc, "receiving\n");
1340         o2net_set_advance_start_time(sc);
1341
1342         if (unlikely(sc->sc_handshake_ok == 0)) {
1343                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1344                         data = page_address(sc->sc_page) + sc->sc_page_off;
1345                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1346                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1347                         if (ret > 0)
1348                                 sc->sc_page_off += ret;
1349                 }
1350
1351                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1352                         o2net_check_handshake(sc);
1353                         if (unlikely(sc->sc_handshake_ok == 0))
1354                                 ret = -EPROTO;
1355                 }
1356                 goto out;
1357         }
1358
1359         /* do we need more header? */
1360         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1361                 data = page_address(sc->sc_page) + sc->sc_page_off;
1362                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1363                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1364                 if (ret > 0) {
1365                         sc->sc_page_off += ret;
1366                         /* only swab incoming here.. we can
1367                          * only get here once as we cross from
1368                          * being under to over */
1369                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1370                                 hdr = page_address(sc->sc_page);
1371                                 if (be16_to_cpu(hdr->data_len) >
1372                                     O2NET_MAX_PAYLOAD_BYTES)
1373                                         ret = -EOVERFLOW;
1374                         }
1375                 }
1376                 if (ret <= 0)
1377                         goto out;
1378         }
1379
1380         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1381                 /* oof, still don't have a header */
1382                 goto out;
1383         }
1384
1385         /* this was swabbed above when we first read it */
1386         hdr = page_address(sc->sc_page);
1387
1388         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1389
1390         /* do we need more payload? */
1391         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1392                 /* need more payload */
1393                 data = page_address(sc->sc_page) + sc->sc_page_off;
1394                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1395                           sc->sc_page_off;
1396                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1397                 if (ret > 0)
1398                         sc->sc_page_off += ret;
1399                 if (ret <= 0)
1400                         goto out;
1401         }
1402
1403         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1404                 /* we can only get here once, the first time we read
1405                  * the payload.. so set ret to progress if the handler
1406                  * works out. after calling this the message is toast */
1407                 ret = o2net_process_message(sc, hdr);
1408                 if (ret == 0)
1409                         ret = 1;
1410                 sc->sc_page_off = 0;
1411         }
1412
1413 out:
1414         sclog(sc, "ret = %d\n", ret);
1415         o2net_set_advance_stop_time(sc);
1416         return ret;
1417 }
1418
1419 /* this work func is triggerd by data ready.  it reads until it can read no
1420  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1421  * our work the work struct will be marked and we'll be called again. */
1422 static void o2net_rx_until_empty(struct work_struct *work)
1423 {
1424         struct o2net_sock_container *sc =
1425                 container_of(work, struct o2net_sock_container, sc_rx_work);
1426         int ret;
1427
1428         do {
1429                 ret = o2net_advance_rx(sc);
1430         } while (ret > 0);
1431
1432         if (ret <= 0 && ret != -EAGAIN) {
1433                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1434                 sclog(sc, "saw error %d, closing\n", ret);
1435                 /* not permanent so read failed handshake can retry */
1436                 o2net_ensure_shutdown(nn, sc, 0);
1437         }
1438
1439         sc_put(sc);
1440 }
1441
1442 static void o2net_initialize_handshake(void)
1443 {
1444         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1445                 O2HB_MAX_WRITE_TIMEOUT_MS);
1446         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1447         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1448                 o2net_keepalive_delay());
1449         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1450                 o2net_reconnect_delay());
1451 }
1452
1453 /* ------------------------------------------------------------ */
1454
1455 /* called when a connect completes and after a sock is accepted.  the
1456  * rx path will see the response and mark the sc valid */
1457 static void o2net_sc_connect_completed(struct work_struct *work)
1458 {
1459         struct o2net_sock_container *sc =
1460                 container_of(work, struct o2net_sock_container,
1461                              sc_connect_work);
1462
1463         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1464               (unsigned long long)O2NET_PROTOCOL_VERSION,
1465               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1466
1467         o2net_initialize_handshake();
1468         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1469         sc_put(sc);
1470 }
1471
1472 /* this is called as a work_struct func. */
1473 static void o2net_sc_send_keep_req(struct work_struct *work)
1474 {
1475         struct o2net_sock_container *sc =
1476                 container_of(work, struct o2net_sock_container,
1477                              sc_keepalive_work.work);
1478
1479         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1480         sc_put(sc);
1481 }
1482
1483 /* socket shutdown does a del_timer_sync against this as it tears down.
1484  * we can't start this timer until we've got to the point in sc buildup
1485  * where shutdown is going to be involved */
1486 static void o2net_idle_timer(struct timer_list *t)
1487 {
1488         struct o2net_sock_container *sc = from_timer(sc, t, sc_idle_timeout);
1489         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1490 #ifdef CONFIG_DEBUG_FS
1491         unsigned long msecs = ktime_to_ms(ktime_get()) -
1492                 ktime_to_ms(sc->sc_tv_timer);
1493 #else
1494         unsigned long msecs = o2net_idle_timeout();
1495 #endif
1496
1497         printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1498                "idle for %lu.%lu secs.\n",
1499                SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1500
1501         /* idle timerout happen, don't shutdown the connection, but
1502          * make fence decision. Maybe the connection can recover before
1503          * the decision is made.
1504          */
1505         atomic_set(&nn->nn_timeout, 1);
1506         o2quo_conn_err(o2net_num_from_nn(nn));
1507         queue_delayed_work(o2net_wq, &nn->nn_still_up,
1508                         msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1509
1510         o2net_sc_reset_idle_timer(sc);
1511
1512 }
1513
1514 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1515 {
1516         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1517         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1518                       msecs_to_jiffies(o2net_keepalive_delay()));
1519         o2net_set_sock_timer(sc);
1520         mod_timer(&sc->sc_idle_timeout,
1521                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1522 }
1523
1524 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1525 {
1526         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1527
1528         /* clear fence decision since the connection recover from timeout*/
1529         if (atomic_read(&nn->nn_timeout)) {
1530                 o2quo_conn_up(o2net_num_from_nn(nn));
1531                 cancel_delayed_work(&nn->nn_still_up);
1532                 atomic_set(&nn->nn_timeout, 0);
1533         }
1534
1535         /* Only push out an existing timer */
1536         if (timer_pending(&sc->sc_idle_timeout))
1537                 o2net_sc_reset_idle_timer(sc);
1538 }
1539
1540 /* this work func is kicked whenever a path sets the nn state which doesn't
1541  * have valid set.  This includes seeing hb come up, losing a connection,
1542  * having a connect attempt fail, etc. This centralizes the logic which decides
1543  * if a connect attempt should be made or if we should give up and all future
1544  * transmit attempts should fail */
1545 static void o2net_start_connect(struct work_struct *work)
1546 {
1547         struct o2net_node *nn =
1548                 container_of(work, struct o2net_node, nn_connect_work.work);
1549         struct o2net_sock_container *sc = NULL;
1550         struct o2nm_node *node = NULL, *mynode = NULL;
1551         struct socket *sock = NULL;
1552         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1553         int ret = 0, stop;
1554         unsigned int timeout;
1555         unsigned int nofs_flag;
1556
1557         /*
1558          * sock_create allocates the sock with GFP_KERNEL. We must
1559          * prevent the filesystem from being reentered by memory reclaim.
1560          */
1561         nofs_flag = memalloc_nofs_save();
1562         /* if we're greater we initiate tx, otherwise we accept */
1563         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1564                 goto out;
1565
1566         /* watch for racing with tearing a node down */
1567         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1568         if (node == NULL)
1569                 goto out;
1570
1571         mynode = o2nm_get_node_by_num(o2nm_this_node());
1572         if (mynode == NULL)
1573                 goto out;
1574
1575         spin_lock(&nn->nn_lock);
1576         /*
1577          * see if we already have one pending or have given up.
1578          * For nn_timeout, it is set when we close the connection
1579          * because of the idle time out. So it means that we have
1580          * at least connected to that node successfully once,
1581          * now try to connect to it again.
1582          */
1583         timeout = atomic_read(&nn->nn_timeout);
1584         stop = (nn->nn_sc ||
1585                 (nn->nn_persistent_error &&
1586                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1587         spin_unlock(&nn->nn_lock);
1588         if (stop)
1589                 goto out;
1590
1591         nn->nn_last_connect_attempt = jiffies;
1592
1593         sc = sc_alloc(node);
1594         if (sc == NULL) {
1595                 mlog(0, "couldn't allocate sc\n");
1596                 ret = -ENOMEM;
1597                 goto out;
1598         }
1599
1600         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1601         if (ret < 0) {
1602                 mlog(0, "can't create socket: %d\n", ret);
1603                 goto out;
1604         }
1605         sc->sc_sock = sock; /* freed by sc_kref_release */
1606
1607         sock->sk->sk_allocation = GFP_ATOMIC;
1608         sock->sk->sk_use_task_frag = false;
1609
1610         myaddr.sin_family = AF_INET;
1611         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1612         myaddr.sin_port = htons(0); /* any port */
1613
1614         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1615                               sizeof(myaddr));
1616         if (ret) {
1617                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1618                      ret, &mynode->nd_ipv4_address);
1619                 goto out;
1620         }
1621
1622         tcp_sock_set_nodelay(sc->sc_sock->sk);
1623         tcp_sock_set_user_timeout(sock->sk, O2NET_TCP_USER_TIMEOUT);
1624
1625         o2net_register_callbacks(sc->sc_sock->sk, sc);
1626
1627         spin_lock(&nn->nn_lock);
1628         /* handshake completion will set nn->nn_sc_valid */
1629         o2net_set_nn_state(nn, sc, 0, 0);
1630         spin_unlock(&nn->nn_lock);
1631
1632         remoteaddr.sin_family = AF_INET;
1633         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1634         remoteaddr.sin_port = node->nd_ipv4_port;
1635
1636         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1637                                         (struct sockaddr *)&remoteaddr,
1638                                         sizeof(remoteaddr),
1639                                         O_NONBLOCK);
1640         if (ret == -EINPROGRESS)
1641                 ret = 0;
1642
1643 out:
1644         if (ret && sc) {
1645                 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1646                        " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1647                 /* 0 err so that another will be queued and attempted
1648                  * from set_nn_state */
1649                 o2net_ensure_shutdown(nn, sc, 0);
1650         }
1651         if (sc)
1652                 sc_put(sc);
1653         if (node)
1654                 o2nm_node_put(node);
1655         if (mynode)
1656                 o2nm_node_put(mynode);
1657
1658         memalloc_nofs_restore(nofs_flag);
1659         return;
1660 }
1661
1662 static void o2net_connect_expired(struct work_struct *work)
1663 {
1664         struct o2net_node *nn =
1665                 container_of(work, struct o2net_node, nn_connect_expired.work);
1666
1667         spin_lock(&nn->nn_lock);
1668         if (!nn->nn_sc_valid) {
1669                 printk(KERN_NOTICE "o2net: No connection established with "
1670                        "node %u after %u.%u seconds, check network and"
1671                        " cluster configuration.\n",
1672                      o2net_num_from_nn(nn),
1673                      o2net_idle_timeout() / 1000,
1674                      o2net_idle_timeout() % 1000);
1675
1676                 o2net_set_nn_state(nn, NULL, 0, 0);
1677         }
1678         spin_unlock(&nn->nn_lock);
1679 }
1680
1681 static void o2net_still_up(struct work_struct *work)
1682 {
1683         struct o2net_node *nn =
1684                 container_of(work, struct o2net_node, nn_still_up.work);
1685
1686         o2quo_hb_still_up(o2net_num_from_nn(nn));
1687 }
1688
1689 /* ------------------------------------------------------------ */
1690
1691 void o2net_disconnect_node(struct o2nm_node *node)
1692 {
1693         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1694
1695         /* don't reconnect until it's heartbeating again */
1696         spin_lock(&nn->nn_lock);
1697         atomic_set(&nn->nn_timeout, 0);
1698         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1699         spin_unlock(&nn->nn_lock);
1700
1701         if (o2net_wq) {
1702                 cancel_delayed_work(&nn->nn_connect_expired);
1703                 cancel_delayed_work(&nn->nn_connect_work);
1704                 cancel_delayed_work(&nn->nn_still_up);
1705                 flush_workqueue(o2net_wq);
1706         }
1707 }
1708
1709 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1710                                   void *data)
1711 {
1712         o2quo_hb_down(node_num);
1713
1714         if (!node)
1715                 return;
1716
1717         if (node_num != o2nm_this_node())
1718                 o2net_disconnect_node(node);
1719
1720         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1721 }
1722
1723 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1724                                 void *data)
1725 {
1726         struct o2net_node *nn = o2net_nn_from_num(node_num);
1727
1728         o2quo_hb_up(node_num);
1729
1730         BUG_ON(!node);
1731
1732         /* ensure an immediate connect attempt */
1733         nn->nn_last_connect_attempt = jiffies -
1734                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1735
1736         if (node_num != o2nm_this_node()) {
1737                 /* believe it or not, accept and node heartbeating testing
1738                  * can succeed for this node before we got here.. so
1739                  * only use set_nn_state to clear the persistent error
1740                  * if that hasn't already happened */
1741                 spin_lock(&nn->nn_lock);
1742                 atomic_set(&nn->nn_timeout, 0);
1743                 if (nn->nn_persistent_error)
1744                         o2net_set_nn_state(nn, NULL, 0, 0);
1745                 spin_unlock(&nn->nn_lock);
1746         }
1747 }
1748
1749 void o2net_unregister_hb_callbacks(void)
1750 {
1751         o2hb_unregister_callback(NULL, &o2net_hb_up);
1752         o2hb_unregister_callback(NULL, &o2net_hb_down);
1753 }
1754
1755 int o2net_register_hb_callbacks(void)
1756 {
1757         int ret;
1758
1759         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1760                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1761         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1762                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1763
1764         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1765         if (ret == 0)
1766                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1767
1768         if (ret)
1769                 o2net_unregister_hb_callbacks();
1770
1771         return ret;
1772 }
1773
1774 /* ------------------------------------------------------------ */
1775
1776 static int o2net_accept_one(struct socket *sock, int *more)
1777 {
1778         int ret;
1779         struct sockaddr_in sin;
1780         struct socket *new_sock = NULL;
1781         struct o2nm_node *node = NULL;
1782         struct o2nm_node *local_node = NULL;
1783         struct o2net_sock_container *sc = NULL;
1784         struct o2net_node *nn;
1785         unsigned int nofs_flag;
1786
1787         /*
1788          * sock_create_lite allocates the sock with GFP_KERNEL. We must
1789          * prevent the filesystem from being reentered by memory reclaim.
1790          */
1791         nofs_flag = memalloc_nofs_save();
1792
1793         BUG_ON(sock == NULL);
1794         *more = 0;
1795         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1796                                sock->sk->sk_protocol, &new_sock);
1797         if (ret)
1798                 goto out;
1799
1800         new_sock->type = sock->type;
1801         new_sock->ops = sock->ops;
1802         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, false);
1803         if (ret < 0)
1804                 goto out;
1805
1806         *more = 1;
1807         new_sock->sk->sk_allocation = GFP_ATOMIC;
1808
1809         tcp_sock_set_nodelay(new_sock->sk);
1810         tcp_sock_set_user_timeout(new_sock->sk, O2NET_TCP_USER_TIMEOUT);
1811
1812         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin, 1);
1813         if (ret < 0)
1814                 goto out;
1815
1816         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1817         if (node == NULL) {
1818                 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1819                        "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1820                        ntohs(sin.sin_port));
1821                 ret = -EINVAL;
1822                 goto out;
1823         }
1824
1825         if (o2nm_this_node() >= node->nd_num) {
1826                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1827                 if (local_node)
1828                         printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1829                                         "seen at node '%s' (%u, %pI4:%d) from "
1830                                         "node '%s' (%u, %pI4:%d)\n",
1831                                         local_node->nd_name, local_node->nd_num,
1832                                         &(local_node->nd_ipv4_address),
1833                                         ntohs(local_node->nd_ipv4_port),
1834                                         node->nd_name,
1835                                         node->nd_num, &sin.sin_addr.s_addr,
1836                                         ntohs(sin.sin_port));
1837                 ret = -EINVAL;
1838                 goto out;
1839         }
1840
1841         /* this happens all the time when the other node sees our heartbeat
1842          * and tries to connect before we see their heartbeat */
1843         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1844                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1845                      "%pI4:%d but it isn't heartbeating\n",
1846                      node->nd_name, &sin.sin_addr.s_addr,
1847                      ntohs(sin.sin_port));
1848                 ret = -EINVAL;
1849                 goto out;
1850         }
1851
1852         nn = o2net_nn_from_num(node->nd_num);
1853
1854         spin_lock(&nn->nn_lock);
1855         if (nn->nn_sc)
1856                 ret = -EBUSY;
1857         else
1858                 ret = 0;
1859         spin_unlock(&nn->nn_lock);
1860         if (ret) {
1861                 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1862                        "at %pI4:%d but it already has an open connection\n",
1863                        node->nd_name, &sin.sin_addr.s_addr,
1864                        ntohs(sin.sin_port));
1865                 goto out;
1866         }
1867
1868         sc = sc_alloc(node);
1869         if (sc == NULL) {
1870                 ret = -ENOMEM;
1871                 goto out;
1872         }
1873
1874         sc->sc_sock = new_sock;
1875         new_sock = NULL;
1876
1877         spin_lock(&nn->nn_lock);
1878         atomic_set(&nn->nn_timeout, 0);
1879         o2net_set_nn_state(nn, sc, 0, 0);
1880         spin_unlock(&nn->nn_lock);
1881
1882         o2net_register_callbacks(sc->sc_sock->sk, sc);
1883         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1884
1885         o2net_initialize_handshake();
1886         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1887
1888 out:
1889         if (new_sock)
1890                 sock_release(new_sock);
1891         if (node)
1892                 o2nm_node_put(node);
1893         if (local_node)
1894                 o2nm_node_put(local_node);
1895         if (sc)
1896                 sc_put(sc);
1897
1898         memalloc_nofs_restore(nofs_flag);
1899         return ret;
1900 }
1901
1902 /*
1903  * This function is invoked in response to one or more
1904  * pending accepts at softIRQ level. We must drain the
1905  * entire que before returning.
1906  */
1907
1908 static void o2net_accept_many(struct work_struct *work)
1909 {
1910         struct socket *sock = o2net_listen_sock;
1911         int     more;
1912
1913         /*
1914          * It is critical to note that due to interrupt moderation
1915          * at the network driver level, we can't assume to get a
1916          * softIRQ for every single conn since tcp SYN packets
1917          * can arrive back-to-back, and therefore many pending
1918          * accepts may result in just 1 softIRQ. If we terminate
1919          * the o2net_accept_one() loop upon seeing an err, what happens
1920          * to the rest of the conns in the queue? If no new SYN
1921          * arrives for hours, no softIRQ  will be delivered,
1922          * and the connections will just sit in the queue.
1923          */
1924
1925         for (;;) {
1926                 o2net_accept_one(sock, &more);
1927                 if (!more)
1928                         break;
1929                 cond_resched();
1930         }
1931 }
1932
1933 static void o2net_listen_data_ready(struct sock *sk)
1934 {
1935         void (*ready)(struct sock *sk);
1936
1937         trace_sk_data_ready(sk);
1938
1939         read_lock_bh(&sk->sk_callback_lock);
1940         ready = sk->sk_user_data;
1941         if (ready == NULL) { /* check for teardown race */
1942                 ready = sk->sk_data_ready;
1943                 goto out;
1944         }
1945
1946         /* This callback may called twice when a new connection
1947          * is  being established as a child socket inherits everything
1948          * from a parent LISTEN socket, including the data_ready cb of
1949          * the parent. This leads to a hazard. In o2net_accept_one()
1950          * we are still initializing the child socket but have not
1951          * changed the inherited data_ready callback yet when
1952          * data starts arriving.
1953          * We avoid this hazard by checking the state.
1954          * For the listening socket,  the state will be TCP_LISTEN; for the new
1955          * socket, will be  TCP_ESTABLISHED. Also, in this case,
1956          * sk->sk_user_data is not a valid function pointer.
1957          */
1958
1959         if (sk->sk_state == TCP_LISTEN) {
1960                 queue_work(o2net_wq, &o2net_listen_work);
1961         } else {
1962                 ready = NULL;
1963         }
1964
1965 out:
1966         read_unlock_bh(&sk->sk_callback_lock);
1967         if (ready != NULL)
1968                 ready(sk);
1969 }
1970
1971 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1972 {
1973         struct socket *sock = NULL;
1974         int ret;
1975         struct sockaddr_in sin = {
1976                 .sin_family = PF_INET,
1977                 .sin_addr = { .s_addr = addr },
1978                 .sin_port = port,
1979         };
1980
1981         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1982         if (ret < 0) {
1983                 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
1984                 goto out;
1985         }
1986
1987         sock->sk->sk_allocation = GFP_ATOMIC;
1988
1989         write_lock_bh(&sock->sk->sk_callback_lock);
1990         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1991         sock->sk->sk_data_ready = o2net_listen_data_ready;
1992         write_unlock_bh(&sock->sk->sk_callback_lock);
1993
1994         o2net_listen_sock = sock;
1995         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1996
1997         sock->sk->sk_reuse = SK_CAN_REUSE;
1998         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1999         if (ret < 0) {
2000                 printk(KERN_ERR "o2net: Error %d while binding socket at "
2001                        "%pI4:%u\n", ret, &addr, ntohs(port));
2002                 goto out;
2003         }
2004
2005         ret = sock->ops->listen(sock, 64);
2006         if (ret < 0)
2007                 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2008                        ret, &addr, ntohs(port));
2009
2010 out:
2011         if (ret) {
2012                 o2net_listen_sock = NULL;
2013                 if (sock)
2014                         sock_release(sock);
2015         }
2016         return ret;
2017 }
2018
2019 /*
2020  * called from node manager when we should bring up our network listening
2021  * socket.  node manager handles all the serialization to only call this
2022  * once and to match it with o2net_stop_listening().  note,
2023  * o2nm_this_node() doesn't work yet as we're being called while it
2024  * is being set up.
2025  */
2026 int o2net_start_listening(struct o2nm_node *node)
2027 {
2028         int ret = 0;
2029
2030         BUG_ON(o2net_wq != NULL);
2031         BUG_ON(o2net_listen_sock != NULL);
2032
2033         mlog(ML_KTHREAD, "starting o2net thread...\n");
2034         o2net_wq = alloc_ordered_workqueue("o2net", WQ_MEM_RECLAIM);
2035         if (o2net_wq == NULL) {
2036                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2037                 return -ENOMEM; /* ? */
2038         }
2039
2040         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2041                                         node->nd_ipv4_port);
2042         if (ret) {
2043                 destroy_workqueue(o2net_wq);
2044                 o2net_wq = NULL;
2045         } else
2046                 o2quo_conn_up(node->nd_num);
2047
2048         return ret;
2049 }
2050
2051 /* again, o2nm_this_node() doesn't work here as we're involved in
2052  * tearing it down */
2053 void o2net_stop_listening(struct o2nm_node *node)
2054 {
2055         struct socket *sock = o2net_listen_sock;
2056         size_t i;
2057
2058         BUG_ON(o2net_wq == NULL);
2059         BUG_ON(o2net_listen_sock == NULL);
2060
2061         /* stop the listening socket from generating work */
2062         write_lock_bh(&sock->sk->sk_callback_lock);
2063         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2064         sock->sk->sk_user_data = NULL;
2065         write_unlock_bh(&sock->sk->sk_callback_lock);
2066
2067         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2068                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2069                 if (node) {
2070                         o2net_disconnect_node(node);
2071                         o2nm_node_put(node);
2072                 }
2073         }
2074
2075         /* finish all work and tear down the work queue */
2076         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2077         destroy_workqueue(o2net_wq);
2078         o2net_wq = NULL;
2079
2080         sock_release(o2net_listen_sock);
2081         o2net_listen_sock = NULL;
2082
2083         o2quo_conn_err(node->nd_num);
2084 }
2085
2086 /* ------------------------------------------------------------ */
2087
2088 int o2net_init(void)
2089 {
2090         unsigned long i;
2091
2092         o2quo_init();
2093
2094         o2net_debugfs_init();
2095
2096         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2097         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2098         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2099         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp)
2100                 goto out;
2101
2102         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2103         o2net_hand->connector_id = cpu_to_be64(1);
2104
2105         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2106         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2107
2108         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2109                 struct o2net_node *nn = o2net_nn_from_num(i);
2110
2111                 atomic_set(&nn->nn_timeout, 0);
2112                 spin_lock_init(&nn->nn_lock);
2113                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2114                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2115                                   o2net_connect_expired);
2116                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2117                 /* until we see hb from a node we'll return einval */
2118                 nn->nn_persistent_error = -ENOTCONN;
2119                 init_waitqueue_head(&nn->nn_sc_wq);
2120                 idr_init(&nn->nn_status_idr);
2121                 INIT_LIST_HEAD(&nn->nn_status_list);
2122         }
2123
2124         return 0;
2125
2126 out:
2127         kfree(o2net_hand);
2128         kfree(o2net_keep_req);
2129         kfree(o2net_keep_resp);
2130         o2net_debugfs_exit();
2131         o2quo_exit();
2132         return -ENOMEM;
2133 }
2134
2135 void o2net_exit(void)
2136 {
2137         o2quo_exit();
2138         kfree(o2net_hand);
2139         kfree(o2net_keep_req);
2140         kfree(o2net_keep_resp);
2141         o2net_debugfs_exit();
2142 }