1// SPDX-License-Identifier: GPL-2.0-or-later
   2/* -*- mode: c; c-basic-offset: 8; -*-
   3 *
   4 * vim: noexpandtab sw=8 ts=8 sts=0:
   5 *
   6 * Copyright (C) 2004 Oracle.  All rights reserved.
   7 *
   8 * ----
   9 *
  10 * Callers for this were originally written against a very simple synchronus
  11 * API.  This implementation reflects those simple callers.  Some day I'm sure
  12 * we'll need to move to a more robust posting/callback mechanism.
  13 *
  14 * Transmit calls pass in kernel virtual addresses and block copying this into
  15 * the socket's tx buffers via a usual blocking sendmsg.  They'll block waiting
  16 * for a failed socket to timeout.  TX callers can also pass in a poniter to an
  17 * 'int' which gets filled with an errno off the wire in response to the
  18 * message they send.
  19 *
  20 * Handlers for unsolicited messages are registered.  Each socket has a page
  21 * that incoming data is copied into.  First the header, then the data.
  22 * Handlers are called from only one thread with a reference to this per-socket
  23 * page.  This page is destroyed after the handler call, so it can't be
  24 * referenced beyond the call.  Handlers may block but are discouraged from
  25 * doing so.
  26 *
  27 * Any framing errors (bad magic, large payload lengths) close a connection.
  28 *
  29 * Our sock_container holds the state we associate with a socket.  It's current
  30 * framing state is held there as well as the refcounting we do around when it
  31 * is safe to tear down the socket.  The socket is only finally torn down from
  32 * the container when the container loses all of its references -- so as long
  33 * as you hold a ref on the container you can trust that the socket is valid
  34 * for use with kernel socket APIs.
  35 *
  36 * Connections are initiated between a pair of nodes when the node with the
  37 * higher node number gets a heartbeat callback which indicates that the lower
  38 * numbered node has started heartbeating.  The lower numbered node is passive
  39 * and only accepts the connection if the higher numbered node is heartbeating.
  40 */
  41
  42#include <linux/kernel.h>
  43#include <linux/sched/mm.h>
  44#include <linux/jiffies.h>
  45#include <linux/slab.h>
  46#include <linux/idr.h>
  47#include <linux/kref.h>
  48#include <linux/net.h>
  49#include <linux/export.h>
  50#include <net/tcp.h>
  51
  52#include <linux/uaccess.h>
  53
  54#include "heartbeat.h"
  55#include "tcp.h"
  56#include "nodemanager.h"
  57#define MLOG_MASK_PREFIX ML_TCP
  58#include "masklog.h"
  59#include "quorum.h"
  60
  61#include "tcp_internal.h"
  62
  63#define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
  64#define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,	\
  65			  &sc->sc_node->nd_ipv4_address,		\
  66			  ntohs(sc->sc_node->nd_ipv4_port)
  67
  68/*
  69 * In the following two log macros, the whitespace after the ',' just
  70 * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  71 * previous token if args expands to nothing.
  72 */
  73#define msglog(hdr, fmt, args...) do {					\
  74	typeof(hdr) __hdr = (hdr);					\
  75	mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "	\
  76	     "key %08x num %u] " fmt,					\
  77	     be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len), 	\
  78	     be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),	\
  79	     be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),	\
  80	     be32_to_cpu(__hdr->msg_num) ,  ##args);			\
  81} while (0)
  82
  83#define sclog(sc, fmt, args...) do {					\
  84	typeof(sc) __sc = (sc);						\
  85	mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "	\
  86	     "pg_off %zu] " fmt, __sc,					\
  87	     kref_read(&__sc->sc_kref), __sc->sc_sock,	\
  88	    __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,	\
  89	    ##args);							\
  90} while (0)
  91
  92static DEFINE_RWLOCK(o2net_handler_lock);
  93static struct rb_root o2net_handler_tree = RB_ROOT;
  94
  95static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
  96
  97/* XXX someday we'll need better accounting */
  98static struct socket *o2net_listen_sock;
  99
 100/*
 101 * listen work is only queued by the listening socket callbacks on the
 102 * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 103 * quorum work is queued as sock containers are shutdown.. stop_listening
 104 * tears down all the node's sock containers, preventing future shutdowns
 105 * and queued quroum work, before canceling delayed quorum work and
 106 * destroying the work queue.
 107 */
 108static struct workqueue_struct *o2net_wq;
 109static struct work_struct o2net_listen_work;
 110
 111static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 112#define O2NET_HB_PRI 0x1
 113
 114static struct o2net_handshake *o2net_hand;
 115static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 116
 117static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 118		{[O2NET_ERR_NONE]	= 0,
 119		 [O2NET_ERR_NO_HNDLR]	= -ENOPROTOOPT,
 120		 [O2NET_ERR_OVERFLOW]	= -EOVERFLOW,
 121		 [O2NET_ERR_DIED]	= -EHOSTDOWN,};
 122
 123/* can't quite avoid *all* internal declarations :/ */
 124static void o2net_sc_connect_completed(struct work_struct *work);
 125static void o2net_rx_until_empty(struct work_struct *work);
 126static void o2net_shutdown_sc(struct work_struct *work);
 127static void o2net_listen_data_ready(struct sock *sk);
 128static void o2net_sc_send_keep_req(struct work_struct *work);
 129static void o2net_idle_timer(struct timer_list *t);
 130static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 131static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 132
 133#ifdef CONFIG_DEBUG_FS
 134static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 135			   u32 msgkey, struct task_struct *task, u8 node)
 136{
 137	INIT_LIST_HEAD(&nst->st_net_debug_item);
 138	nst->st_task = task;
 139	nst->st_msg_type = msgtype;
 140	nst->st_msg_key = msgkey;
 141	nst->st_node = node;
 142}
 143
 144static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 145{
 146	nst->st_sock_time = ktime_get();
 147}
 148
 149static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 150{
 151	nst->st_send_time = ktime_get();
 152}
 153
 154static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 155{
 156	nst->st_status_time = ktime_get();
 157}
 158
 159static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 160						struct o2net_sock_container *sc)
 161{
 162	nst->st_sc = sc;
 163}
 164
 165static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 166					u32 msg_id)
 167{
 168	nst->st_id = msg_id;
 169}
 170
 171static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
 172{
 173	sc->sc_tv_timer = ktime_get();
 174}
 175
 176static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
 177{
 178	sc->sc_tv_data_ready = ktime_get();
 179}
 180
 181static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
 182{
 183	sc->sc_tv_advance_start = ktime_get();
 184}
 185
 186static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
 187{
 188	sc->sc_tv_advance_stop = ktime_get();
 189}
 190
 191static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
 192{
 193	sc->sc_tv_func_start = ktime_get();
 194}
 195
 196static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
 197{
 198	sc->sc_tv_func_stop = ktime_get();
 199}
 200
 201#else  /* CONFIG_DEBUG_FS */
 202# define o2net_init_nst(a, b, c, d, e)
 203# define o2net_set_nst_sock_time(a)
 204# define o2net_set_nst_send_time(a)
 205# define o2net_set_nst_status_time(a)
 206# define o2net_set_nst_sock_container(a, b)
 207# define o2net_set_nst_msg_id(a, b)
 208# define o2net_set_sock_timer(a)
 209# define o2net_set_data_ready_time(a)
 210# define o2net_set_advance_start_time(a)
 211# define o2net_set_advance_stop_time(a)
 212# define o2net_set_func_start_time(a)
 213# define o2net_set_func_stop_time(a)
 214#endif /* CONFIG_DEBUG_FS */
 215
 216#ifdef CONFIG_OCFS2_FS_STATS
 217static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
 218{
 219	return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
 220}
 221
 222static void o2net_update_send_stats(struct o2net_send_tracking *nst,
 223				    struct o2net_sock_container *sc)
 224{
 225	sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
 226					   ktime_sub(ktime_get(),
 227						     nst->st_status_time));
 228	sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
 229					 ktime_sub(nst->st_status_time,
 230						   nst->st_send_time));
 231	sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
 232					    ktime_sub(nst->st_send_time,
 233						      nst->st_sock_time));
 234	sc->sc_send_count++;
 235}
 236
 237static void o2net_update_recv_stats(struct o2net_sock_container *sc)
 238{
 239	sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
 240					    o2net_get_func_run_time(sc));
 241	sc->sc_recv_count++;
 242}
 243
 244#else
 245
 246# define o2net_update_send_stats(a, b)
 247
 248# define o2net_update_recv_stats(sc)
 249
 250#endif /* CONFIG_OCFS2_FS_STATS */
 251
 252static inline unsigned int o2net_reconnect_delay(void)
 253{
 254	return o2nm_single_cluster->cl_reconnect_delay_ms;
 255}
 256
 257static inline unsigned int o2net_keepalive_delay(void)
 258{
 259	return o2nm_single_cluster->cl_keepalive_delay_ms;
 260}
 261
 262static inline unsigned int o2net_idle_timeout(void)
 263{
 264	return o2nm_single_cluster->cl_idle_timeout_ms;
 265}
 266
 267static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 268{
 269	int trans;
 270	BUG_ON(err >= O2NET_ERR_MAX);
 271	trans = o2net_sys_err_translations[err];
 272
 273	/* Just in case we mess up the translation table above */
 274	BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 275	return trans;
 276}
 277
 278static struct o2net_node * o2net_nn_from_num(u8 node_num)
 279{
 280	BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 281	return &o2net_nodes[node_num];
 282}
 283
 284static u8 o2net_num_from_nn(struct o2net_node *nn)
 285{
 286	BUG_ON(nn == NULL);
 287	return nn - o2net_nodes;
 288}
 289
 290/* ------------------------------------------------------------ */
 291
 292static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 293{
 294	int ret;
 295
 296	spin_lock(&nn->nn_lock);
 297	ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
 298	if (ret >= 0) {
 299		nsw->ns_id = ret;
 300		list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
 301	}
 302	spin_unlock(&nn->nn_lock);
 303	if (ret < 0)
 304		return ret;
 305
 306	init_waitqueue_head(&nsw->ns_wq);
 307	nsw->ns_sys_status = O2NET_ERR_NONE;
 308	nsw->ns_status = 0;
 309	return 0;
 310}
 311
 312static void o2net_complete_nsw_locked(struct o2net_node *nn,
 313				      struct o2net_status_wait *nsw,
 314				      enum o2net_system_error sys_status,
 315				      s32 status)
 316{
 317	assert_spin_locked(&nn->nn_lock);
 318
 319	if (!list_empty(&nsw->ns_node_item)) {
 320		list_del_init(&nsw->ns_node_item);
 321		nsw->ns_sys_status = sys_status;
 322		nsw->ns_status = status;
 323		idr_remove(&nn->nn_status_idr, nsw->ns_id);
 324		wake_up(&nsw->ns_wq);
 325	}
 326}
 327
 328static void o2net_complete_nsw(struct o2net_node *nn,
 329			       struct o2net_status_wait *nsw,
 330			       u64 id, enum o2net_system_error sys_status,
 331			       s32 status)
 332{
 333	spin_lock(&nn->nn_lock);
 334	if (nsw == NULL) {
 335		if (id > INT_MAX)
 336			goto out;
 337
 338		nsw = idr_find(&nn->nn_status_idr, id);
 339		if (nsw == NULL)
 340			goto out;
 341	}
 342
 343	o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 344
 345out:
 346	spin_unlock(&nn->nn_lock);
 347	return;
 348}
 349
 350static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 351{
 352	struct o2net_status_wait *nsw, *tmp;
 353	unsigned int num_kills = 0;
 354
 355	assert_spin_locked(&nn->nn_lock);
 356
 357	list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 358		o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 359		num_kills++;
 360	}
 361
 362	mlog(0, "completed %d messages for node %u\n", num_kills,
 363	     o2net_num_from_nn(nn));
 364}
 365
 366static int o2net_nsw_completed(struct o2net_node *nn,
 367			       struct o2net_status_wait *nsw)
 368{
 369	int completed;
 370	spin_lock(&nn->nn_lock);
 371	completed = list_empty(&nsw->ns_node_item);
 372	spin_unlock(&nn->nn_lock);
 373	return completed;
 374}
 375
 376/* ------------------------------------------------------------ */
 377
 378static void sc_kref_release(struct kref *kref)
 379{
 380	struct o2net_sock_container *sc = container_of(kref,
 381					struct o2net_sock_container, sc_kref);
 382	BUG_ON(timer_pending(&sc->sc_idle_timeout));
 383
 384	sclog(sc, "releasing\n");
 385
 386	if (sc->sc_sock) {
 387		sock_release(sc->sc_sock);
 388		sc->sc_sock = NULL;
 389	}
 390
 391	o2nm_undepend_item(&sc->sc_node->nd_item);
 392	o2nm_node_put(sc->sc_node);
 393	sc->sc_node = NULL;
 394
 395	o2net_debug_del_sc(sc);
 396
 397	if (sc->sc_page)
 398		__free_page(sc->sc_page);
 399	kfree(sc);
 400}
 401
 402static void sc_put(struct o2net_sock_container *sc)
 403{
 404	sclog(sc, "put\n");
 405	kref_put(&sc->sc_kref, sc_kref_release);
 406}
 407static void sc_get(struct o2net_sock_container *sc)
 408{
 409	sclog(sc, "get\n");
 410	kref_get(&sc->sc_kref);
 411}
 412static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 413{
 414	struct o2net_sock_container *sc, *ret = NULL;
 415	struct page *page = NULL;
 416	int status = 0;
 417
 418	page = alloc_page(GFP_NOFS);
 419	sc = kzalloc(sizeof(*sc), GFP_NOFS);
 420	if (sc == NULL || page == NULL)
 421		goto out;
 422
 423	kref_init(&sc->sc_kref);
 424	o2nm_node_get(node);
 425	sc->sc_node = node;
 426
 427	/* pin the node item of the remote node */
 428	status = o2nm_depend_item(&node->nd_item);
 429	if (status) {
 430		mlog_errno(status);
 431		o2nm_node_put(node);
 432		goto out;
 433	}
 434	INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 435	INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 436	INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 437	INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 438
 439	timer_setup(&sc->sc_idle_timeout, o2net_idle_timer, 0);
 440
 441	sclog(sc, "alloced\n");
 442
 443	ret = sc;
 444	sc->sc_page = page;
 445	o2net_debug_add_sc(sc);
 446	sc = NULL;
 447	page = NULL;
 448
 449out:
 450	if (page)
 451		__free_page(page);
 452	kfree(sc);
 453
 454	return ret;
 455}
 456
 457/* ------------------------------------------------------------ */
 458
 459static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 460				struct work_struct *work)
 461{
 462	sc_get(sc);
 463	if (!queue_work(o2net_wq, work))
 464		sc_put(sc);
 465}
 466static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 467					struct delayed_work *work,
 468					int delay)
 469{
 470	sc_get(sc);
 471	if (!queue_delayed_work(o2net_wq, work, delay))
 472		sc_put(sc);
 473}
 474static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 475					 struct delayed_work *work)
 476{
 477	if (cancel_delayed_work(work))
 478		sc_put(sc);
 479}
 480
 481static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 482
 483int o2net_num_connected_peers(void)
 484{
 485	return atomic_read(&o2net_connected_peers);
 486}
 487
 488static void o2net_set_nn_state(struct o2net_node *nn,
 489			       struct o2net_sock_container *sc,
 490			       unsigned valid, int err)
 491{
 492	int was_valid = nn->nn_sc_valid;
 493	int was_err = nn->nn_persistent_error;
 494	struct o2net_sock_container *old_sc = nn->nn_sc;
 495
 496	assert_spin_locked(&nn->nn_lock);
 497
 498	if (old_sc && !sc)
 499		atomic_dec(&o2net_connected_peers);
 500	else if (!old_sc && sc)
 501		atomic_inc(&o2net_connected_peers);
 502
 503	/* the node num comparison and single connect/accept path should stop
 504	 * an non-null sc from being overwritten with another */
 505	BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 506	mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 507	mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 508
 509	if (was_valid && !valid && err == 0)
 510		err = -ENOTCONN;
 511
 512	mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 513	     o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 514	     nn->nn_persistent_error, err);
 515
 516	nn->nn_sc = sc;
 517	nn->nn_sc_valid = valid ? 1 : 0;
 518	nn->nn_persistent_error = err;
 519
 520	/* mirrors o2net_tx_can_proceed() */
 521	if (nn->nn_persistent_error || nn->nn_sc_valid)
 522		wake_up(&nn->nn_sc_wq);
 523
 524	if (was_valid && !was_err && nn->nn_persistent_error) {
 525		o2quo_conn_err(o2net_num_from_nn(nn));
 526		queue_delayed_work(o2net_wq, &nn->nn_still_up,
 527				   msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 528	}
 529
 530	if (was_valid && !valid) {
 531		if (old_sc)
 532			printk(KERN_NOTICE "o2net: No longer connected to "
 533				SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 534		o2net_complete_nodes_nsw(nn);
 535	}
 536
 537	if (!was_valid && valid) {
 538		o2quo_conn_up(o2net_num_from_nn(nn));
 539		cancel_delayed_work(&nn->nn_connect_expired);
 540		printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 541		       o2nm_this_node() > sc->sc_node->nd_num ?
 542		       "Connected to" : "Accepted connection from",
 543		       SC_NODEF_ARGS(sc));
 544	}
 545
 546	/* trigger the connecting worker func as long as we're not valid,
 547	 * it will back off if it shouldn't connect.  This can be called
 548	 * from node config teardown and so needs to be careful about
 549	 * the work queue actually being up. */
 550	if (!valid && o2net_wq) {
 551		unsigned long delay;
 552		/* delay if we're within a RECONNECT_DELAY of the
 553		 * last attempt */
 554		delay = (nn->nn_last_connect_attempt +
 555			 msecs_to_jiffies(o2net_reconnect_delay()))
 556			- jiffies;
 557		if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 558			delay = 0;
 559		mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 560		queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 561
 562		/*
 563		 * Delay the expired work after idle timeout.
 564		 *
 565		 * We might have lots of failed connection attempts that run
 566		 * through here but we only cancel the connect_expired work when
 567		 * a connection attempt succeeds.  So only the first enqueue of
 568		 * the connect_expired work will do anything.  The rest will see
 569		 * that it's already queued and do nothing.
 570		 */
 571		delay += msecs_to_jiffies(o2net_idle_timeout());
 572		queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 573	}
 574
 575	/* keep track of the nn's sc ref for the caller */
 576	if ((old_sc == NULL) && sc)
 577		sc_get(sc);
 578	if (old_sc && (old_sc != sc)) {
 579		o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 580		sc_put(old_sc);
 581	}
 582}
 583
 584/* see o2net_register_callbacks() */
 585static void o2net_data_ready(struct sock *sk)
 586{
 587	void (*ready)(struct sock *sk);
 588	struct o2net_sock_container *sc;
 589
 590	read_lock_bh(&sk->sk_callback_lock);
 591	sc = sk->sk_user_data;
 592	if (sc) {
 593		sclog(sc, "data_ready hit\n");
 594		o2net_set_data_ready_time(sc);
 595		o2net_sc_queue_work(sc, &sc->sc_rx_work);
 596		ready = sc->sc_data_ready;
 597	} else {
 598		ready = sk->sk_data_ready;
 599	}
 600	read_unlock_bh(&sk->sk_callback_lock);
 601
 602	ready(sk);
 603}
 604
 605/* see o2net_register_callbacks() */
 606static void o2net_state_change(struct sock *sk)
 607{
 608	void (*state_change)(struct sock *sk);
 609	struct o2net_sock_container *sc;
 610
 611	read_lock_bh(&sk->sk_callback_lock);
 612	sc = sk->sk_user_data;
 613	if (sc == NULL) {
 614		state_change = sk->sk_state_change;
 615		goto out;
 616	}
 617
 618	sclog(sc, "state_change to %d\n", sk->sk_state);
 619
 620	state_change = sc->sc_state_change;
 621
 622	switch(sk->sk_state) {
 623	/* ignore connecting sockets as they make progress */
 624	case TCP_SYN_SENT:
 625	case TCP_SYN_RECV:
 626		break;
 627	case TCP_ESTABLISHED:
 628		o2net_sc_queue_work(sc, &sc->sc_connect_work);
 629		break;
 630	default:
 631		printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
 632			" shutdown, state %d\n",
 633			SC_NODEF_ARGS(sc), sk->sk_state);
 634		o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 635		break;
 636	}
 637out:
 638	read_unlock_bh(&sk->sk_callback_lock);
 639	state_change(sk);
 640}
 641
 642/*
 643 * we register callbacks so we can queue work on events before calling
 644 * the original callbacks.  our callbacks our careful to test user_data
 645 * to discover when they've reaced with o2net_unregister_callbacks().
 646 */
 647static void o2net_register_callbacks(struct sock *sk,
 648				     struct o2net_sock_container *sc)
 649{
 650	write_lock_bh(&sk->sk_callback_lock);
 651
 652	/* accepted sockets inherit the old listen socket data ready */
 653	if (sk->sk_data_ready == o2net_listen_data_ready) {
 654		sk->sk_data_ready = sk->sk_user_data;
 655		sk->sk_user_data = NULL;
 656	}
 657
 658	BUG_ON(sk->sk_user_data != NULL);
 659	sk->sk_user_data = sc;
 660	sc_get(sc);
 661
 662	sc->sc_data_ready = sk->sk_data_ready;
 663	sc->sc_state_change = sk->sk_state_change;
 664	sk->sk_data_ready = o2net_data_ready;
 665	sk->sk_state_change = o2net_state_change;
 666
 667	mutex_init(&sc->sc_send_lock);
 668
 669	write_unlock_bh(&sk->sk_callback_lock);
 670}
 671
 672static int o2net_unregister_callbacks(struct sock *sk,
 673			           struct o2net_sock_container *sc)
 674{
 675	int ret = 0;
 676
 677	write_lock_bh(&sk->sk_callback_lock);
 678	if (sk->sk_user_data == sc) {
 679		ret = 1;
 680		sk->sk_user_data = NULL;
 681		sk->sk_data_ready = sc->sc_data_ready;
 682		sk->sk_state_change = sc->sc_state_change;
 683	}
 684	write_unlock_bh(&sk->sk_callback_lock);
 685
 686	return ret;
 687}
 688
 689/*
 690 * this is a little helper that is called by callers who have seen a problem
 691 * with an sc and want to detach it from the nn if someone already hasn't beat
 692 * them to it.  if an error is given then the shutdown will be persistent
 693 * and pending transmits will be canceled.
 694 */
 695static void o2net_ensure_shutdown(struct o2net_node *nn,
 696			           struct o2net_sock_container *sc,
 697				   int err)
 698{
 699	spin_lock(&nn->nn_lock);
 700	if (nn->nn_sc == sc)
 701		o2net_set_nn_state(nn, NULL, 0, err);
 702	spin_unlock(&nn->nn_lock);
 703}
 704
 705/*
 706 * This work queue function performs the blocking parts of socket shutdown.  A
 707 * few paths lead here.  set_nn_state will trigger this callback if it sees an
 708 * sc detached from the nn.  state_change will also trigger this callback
 709 * directly when it sees errors.  In that case we need to call set_nn_state
 710 * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 711 * itself.
 712 */
 713static void o2net_shutdown_sc(struct work_struct *work)
 714{
 715	struct o2net_sock_container *sc =
 716		container_of(work, struct o2net_sock_container,
 717			     sc_shutdown_work);
 718	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 719
 720	sclog(sc, "shutting down\n");
 721
 722	/* drop the callbacks ref and call shutdown only once */
 723	if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 724		/* we shouldn't flush as we're in the thread, the
 725		 * races with pending sc work structs are harmless */
 726		del_timer_sync(&sc->sc_idle_timeout);
 727		o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 728		sc_put(sc);
 729		kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 730	}
 731
 732	/* not fatal so failed connects before the other guy has our
 733	 * heartbeat can be retried */
 734	o2net_ensure_shutdown(nn, sc, 0);
 735	sc_put(sc);
 736}
 737
 738/* ------------------------------------------------------------ */
 739
 740static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 741			     u32 key)
 742{
 743	int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 744
 745	if (ret == 0)
 746		ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 747
 748	return ret;
 749}
 750
 751static struct o2net_msg_handler *
 752o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 753			  struct rb_node **ret_parent)
 754{
 755	struct rb_node **p = &o2net_handler_tree.rb_node;
 756	struct rb_node *parent = NULL;
 757	struct o2net_msg_handler *nmh, *ret = NULL;
 758	int cmp;
 759
 760	while (*p) {
 761		parent = *p;
 762		nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 763		cmp = o2net_handler_cmp(nmh, msg_type, key);
 764
 765		if (cmp < 0)
 766			p = &(*p)->rb_left;
 767		else if (cmp > 0)
 768			p = &(*p)->rb_right;
 769		else {
 770			ret = nmh;
 771			break;
 772		}
 773	}
 774
 775	if (ret_p != NULL)
 776		*ret_p = p;
 777	if (ret_parent != NULL)
 778		*ret_parent = parent;
 779
 780	return ret;
 781}
 782
 783static void o2net_handler_kref_release(struct kref *kref)
 784{
 785	struct o2net_msg_handler *nmh;
 786	nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 787
 788	kfree(nmh);
 789}
 790
 791static void o2net_handler_put(struct o2net_msg_handler *nmh)
 792{
 793	kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 794}
 795
 796/* max_len is protection for the handler func.  incoming messages won't
 797 * be given to the handler if their payload is longer than the max. */
 798int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 799			   o2net_msg_handler_func *func, void *data,
 800			   o2net_post_msg_handler_func *post_func,
 801			   struct list_head *unreg_list)
 802{
 803	struct o2net_msg_handler *nmh = NULL;
 804	struct rb_node **p, *parent;
 805	int ret = 0;
 806
 807	if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 808		mlog(0, "max_len for message handler out of range: %u\n",
 809			max_len);
 810		ret = -EINVAL;
 811		goto out;
 812	}
 813
 814	if (!msg_type) {
 815		mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 816		ret = -EINVAL;
 817		goto out;
 818
 819	}
 820	if (!func) {
 821		mlog(0, "no message handler provided: %u, %p\n",
 822		       msg_type, func);
 823		ret = -EINVAL;
 824		goto out;
 825	}
 826
 827       	nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 828	if (nmh == NULL) {
 829		ret = -ENOMEM;
 830		goto out;
 831	}
 832
 833	nmh->nh_func = func;
 834	nmh->nh_func_data = data;
 835	nmh->nh_post_func = post_func;
 836	nmh->nh_msg_type = msg_type;
 837	nmh->nh_max_len = max_len;
 838	nmh->nh_key = key;
 839	/* the tree and list get this ref.. they're both removed in
 840	 * unregister when this ref is dropped */
 841	kref_init(&nmh->nh_kref);
 842	INIT_LIST_HEAD(&nmh->nh_unregister_item);
 843
 844	write_lock(&o2net_handler_lock);
 845	if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 846		ret = -EEXIST;
 847	else {
 848	        rb_link_node(&nmh->nh_node, parent, p);
 849		rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 850		list_add_tail(&nmh->nh_unregister_item, unreg_list);
 851
 852		mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 853		     func, msg_type, key);
 854		/* we've had some trouble with handlers seemingly vanishing. */
 855		mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 856							  &parent) == NULL,
 857			        "couldn't find handler we *just* registered "
 858				"for type %u key %08x\n", msg_type, key);
 859	}
 860	write_unlock(&o2net_handler_lock);
 861
 862out:
 863	if (ret)
 864		kfree(nmh);
 865
 866	return ret;
 867}
 868EXPORT_SYMBOL_GPL(o2net_register_handler);
 869
 870void o2net_unregister_handler_list(struct list_head *list)
 871{
 872	struct o2net_msg_handler *nmh, *n;
 873
 874	write_lock(&o2net_handler_lock);
 875	list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 876		mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 877		     nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 878		rb_erase(&nmh->nh_node, &o2net_handler_tree);
 879		list_del_init(&nmh->nh_unregister_item);
 880		kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 881	}
 882	write_unlock(&o2net_handler_lock);
 883}
 884EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 885
 886static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 887{
 888	struct o2net_msg_handler *nmh;
 889
 890	read_lock(&o2net_handler_lock);
 891	nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 892	if (nmh)
 893		kref_get(&nmh->nh_kref);
 894	read_unlock(&o2net_handler_lock);
 895
 896	return nmh;
 897}
 898
 899/* ------------------------------------------------------------ */
 900
 901static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 902{
 903	struct kvec vec = { .iov_len = len, .iov_base = data, };
 904	struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
 905	iov_iter_kvec(&msg.msg_iter, READ, &vec, 1, len);
 906	return sock_recvmsg(sock, &msg, MSG_DONTWAIT);
 907}
 908
 909static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 910			      size_t veclen, size_t total)
 911{
 912	int ret;
 913	struct msghdr msg = {.msg_flags = 0,};
 914
 915	if (sock == NULL) {
 916		ret = -EINVAL;
 917		goto out;
 918	}
 919
 920	ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
 921	if (likely(ret == total))
 922		return 0;
 923	mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
 924	if (ret >= 0)
 925		ret = -EPIPE; /* should be smarter, I bet */
 926out:
 927	mlog(0, "returning error: %d\n", ret);
 928	return ret;
 929}
 930
 931static void o2net_sendpage(struct o2net_sock_container *sc,
 932			   void *kmalloced_virt,
 933			   size_t size)
 934{
 935	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 936	ssize_t ret;
 937
 938	while (1) {
 939		mutex_lock(&sc->sc_send_lock);
 940		ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 941						 virt_to_page(kmalloced_virt),
 942						 offset_in_page(kmalloced_virt),
 943						 size, MSG_DONTWAIT);
 944		mutex_unlock(&sc->sc_send_lock);
 945		if (ret == size)
 946			break;
 947		if (ret == (ssize_t)-EAGAIN) {
 948			mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 949			     " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 950			cond_resched();
 951			continue;
 952		}
 953		mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 954		     " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 955		o2net_ensure_shutdown(nn, sc, 0);
 956		break;
 957	}
 958}
 959
 960static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 961{
 962	memset(msg, 0, sizeof(struct o2net_msg));
 963	msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 964	msg->data_len = cpu_to_be16(data_len);
 965	msg->msg_type = cpu_to_be16(msg_type);
 966	msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 967	msg->status = 0;
 968	msg->key = cpu_to_be32(key);
 969}
 970
 971static int o2net_tx_can_proceed(struct o2net_node *nn,
 972			        struct o2net_sock_container **sc_ret,
 973				int *error)
 974{
 975	int ret = 0;
 976
 977	spin_lock(&nn->nn_lock);
 978	if (nn->nn_persistent_error) {
 979		ret = 1;
 980		*sc_ret = NULL;
 981		*error = nn->nn_persistent_error;
 982	} else if (nn->nn_sc_valid) {
 983		kref_get(&nn->nn_sc->sc_kref);
 984
 985		ret = 1;
 986		*sc_ret = nn->nn_sc;
 987		*error = 0;
 988	}
 989	spin_unlock(&nn->nn_lock);
 990
 991	return ret;
 992}
 993
 994/* Get a map of all nodes to which this node is currently connected to */
 995void o2net_fill_node_map(unsigned long *map, unsigned bytes)
 996{
 997	struct o2net_sock_container *sc;
 998	int node, ret;
 999
1000	BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1001
1002	memset(map, 0, bytes);
1003	for (node = 0; node < O2NM_MAX_NODES; ++node) {
1004		if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1005			continue;
1006		if (!ret) {
1007			set_bit(node, map);
1008			sc_put(sc);
1009		}
1010	}
1011}
1012EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1013
1014int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1015			   size_t caller_veclen, u8 target_node, int *status)
1016{
1017	int ret = 0;
1018	struct o2net_msg *msg = NULL;
1019	size_t veclen, caller_bytes = 0;
1020	struct kvec *vec = NULL;
1021	struct o2net_sock_container *sc = NULL;
1022	struct o2net_node *nn = o2net_nn_from_num(target_node);
1023	struct o2net_status_wait nsw = {
1024		.ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1025	};
1026	struct o2net_send_tracking nst;
1027
1028	o2net_init_nst(&nst, msg_type, key, current, target_node);
1029
1030	if (o2net_wq == NULL) {
1031		mlog(0, "attempt to tx without o2netd running\n");
1032		ret = -ESRCH;
1033		goto out;
1034	}
1035
1036	if (caller_veclen == 0) {
1037		mlog(0, "bad kvec array length\n");
1038		ret = -EINVAL;
1039		goto out;
1040	}
1041
1042	caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1043	if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1044		mlog(0, "total payload len %zu too large\n", caller_bytes);
1045		ret = -EINVAL;
1046		goto out;
1047	}
1048
1049	if (target_node == o2nm_this_node()) {
1050		ret = -ELOOP;
1051		goto out;
1052	}
1053
1054	o2net_debug_add_nst(&nst);
1055
1056	o2net_set_nst_sock_time(&nst);
1057
1058	wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1059	if (ret)
1060		goto out;
1061
1062	o2net_set_nst_sock_container(&nst, sc);
1063
1064	veclen = caller_veclen + 1;
1065	vec = kmalloc_array(veclen, sizeof(struct kvec), GFP_ATOMIC);
1066	if (vec == NULL) {
1067		mlog(0, "failed to %zu element kvec!\n", veclen);
1068		ret = -ENOMEM;
1069		goto out;
1070	}
1071
1072	msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1073	if (!msg) {
1074		mlog(0, "failed to allocate a o2net_msg!\n");
1075		ret = -ENOMEM;
1076		goto out;
1077	}
1078
1079	o2net_init_msg(msg, caller_bytes, msg_type, key);
1080
1081	vec[0].iov_len = sizeof(struct o2net_msg);
1082	vec[0].iov_base = msg;
1083	memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1084
1085	ret = o2net_prep_nsw(nn, &nsw);
1086	if (ret)
1087		goto out;
1088
1089	msg->msg_num = cpu_to_be32(nsw.ns_id);
1090	o2net_set_nst_msg_id(&nst, nsw.ns_id);
1091
1092	o2net_set_nst_send_time(&nst);
1093
1094	/* finally, convert the message header to network byte-order
1095	 * and send */
1096	mutex_lock(&sc->sc_send_lock);
1097	ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1098				 sizeof(struct o2net_msg) + caller_bytes);
1099	mutex_unlock(&sc->sc_send_lock);
1100	msglog(msg, "sending returned %d\n", ret);
1101	if (ret < 0) {
1102		mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1103		goto out;
1104	}
1105
1106	/* wait on other node's handler */
1107	o2net_set_nst_status_time(&nst);
1108	wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1109
1110	o2net_update_send_stats(&nst, sc);
1111
1112	/* Note that we avoid overwriting the callers status return
1113	 * variable if a system error was reported on the other
1114	 * side. Callers beware. */
1115	ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1116	if (status && !ret)
1117		*status = nsw.ns_status;
1118
1119	mlog(0, "woken, returning system status %d, user status %d\n",
1120	     ret, nsw.ns_status);
1121out:
1122	o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1123	if (sc)
1124		sc_put(sc);
1125	kfree(vec);
1126	kfree(msg);
1127	o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1128	return ret;
1129}
1130EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1131
1132int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1133		       u8 target_node, int *status)
1134{
1135	struct kvec vec = {
1136		.iov_base = data,
1137		.iov_len = len,
1138	};
1139	return o2net_send_message_vec(msg_type, key, &vec, 1,
1140				      target_node, status);
1141}
1142EXPORT_SYMBOL_GPL(o2net_send_message);
1143
1144static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1145				   enum o2net_system_error syserr, int err)
1146{
1147	struct kvec vec = {
1148		.iov_base = hdr,
1149		.iov_len = sizeof(struct o2net_msg),
1150	};
1151
1152	BUG_ON(syserr >= O2NET_ERR_MAX);
1153
1154	/* leave other fields intact from the incoming message, msg_num
1155	 * in particular */
1156	hdr->sys_status = cpu_to_be32(syserr);
1157	hdr->status = cpu_to_be32(err);
1158	hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1159	hdr->data_len = 0;
1160
1161	msglog(hdr, "about to send status magic %d\n", err);
1162	/* hdr has been in host byteorder this whole time */
1163	return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1164}
1165
1166/* this returns -errno if the header was unknown or too large, etc.
1167 * after this is called the buffer us reused for the next message */
1168static int o2net_process_message(struct o2net_sock_container *sc,
1169				 struct o2net_msg *hdr)
1170{
1171	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1172	int ret = 0, handler_status;
1173	enum  o2net_system_error syserr;
1174	struct o2net_msg_handler *nmh = NULL;
1175	void *ret_data = NULL;
1176
1177	msglog(hdr, "processing message\n");
1178
1179	o2net_sc_postpone_idle(sc);
1180
1181	switch(be16_to_cpu(hdr->magic)) {
1182		case O2NET_MSG_STATUS_MAGIC:
1183			/* special type for returning message status */
1184			o2net_complete_nsw(nn, NULL,
1185					   be32_to_cpu(hdr->msg_num),
1186					   be32_to_cpu(hdr->sys_status),
1187					   be32_to_cpu(hdr->status));
1188			goto out;
1189		case O2NET_MSG_KEEP_REQ_MAGIC:
1190			o2net_sendpage(sc, o2net_keep_resp,
1191				       sizeof(*o2net_keep_resp));
1192			goto out;
1193		case O2NET_MSG_KEEP_RESP_MAGIC:
1194			goto out;
1195		case O2NET_MSG_MAGIC:
1196			break;
1197		default:
1198			msglog(hdr, "bad magic\n");
1199			ret = -EINVAL;
1200			goto out;
1201			break;
1202	}
1203
1204	/* find a handler for it */
1205	handler_status = 0;
1206	nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1207				be32_to_cpu(hdr->key));
1208	if (!nmh) {
1209		mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1210		     be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1211		syserr = O2NET_ERR_NO_HNDLR;
1212		goto out_respond;
1213	}
1214
1215	syserr = O2NET_ERR_NONE;
1216
1217	if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1218		syserr = O2NET_ERR_OVERFLOW;
1219
1220	if (syserr != O2NET_ERR_NONE)
1221		goto out_respond;
1222
1223	o2net_set_func_start_time(sc);
1224	sc->sc_msg_key = be32_to_cpu(hdr->key);
1225	sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1226	handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1227					     be16_to_cpu(hdr->data_len),
1228					nmh->nh_func_data, &ret_data);
1229	o2net_set_func_stop_time(sc);
1230
1231	o2net_update_recv_stats(sc);
1232
1233out_respond:
1234	/* this destroys the hdr, so don't use it after this */
1235	mutex_lock(&sc->sc_send_lock);
1236	ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1237				      handler_status);
1238	mutex_unlock(&sc->sc_send_lock);
1239	hdr = NULL;
1240	mlog(0, "sending handler status %d, syserr %d returned %d\n",
1241	     handler_status, syserr, ret);
1242
1243	if (nmh) {
1244		BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1245		if (nmh->nh_post_func)
1246			(nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1247					    ret_data);
1248	}
1249
1250out:
1251	if (nmh)
1252		o2net_handler_put(nmh);
1253	return ret;
1254}
1255
1256static int o2net_check_handshake(struct o2net_sock_container *sc)
1257{
1258	struct o2net_handshake *hand = page_address(sc->sc_page);
1259	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1260
1261	if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1262		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1263		       "protocol version %llu but %llu is required. "
1264		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1265		       (unsigned long long)be64_to_cpu(hand->protocol_version),
1266		       O2NET_PROTOCOL_VERSION);
1267
1268		/* don't bother reconnecting if its the wrong version. */
1269		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1270		return -1;
1271	}
1272
1273	/*
1274	 * Ensure timeouts are consistent with other nodes, otherwise
1275	 * we can end up with one node thinking that the other must be down,
1276	 * but isn't. This can ultimately cause corruption.
1277	 */
1278	if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1279				o2net_idle_timeout()) {
1280		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1281		       "idle timeout of %u ms, but we use %u ms locally. "
1282		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1283		       be32_to_cpu(hand->o2net_idle_timeout_ms),
1284		       o2net_idle_timeout());
1285		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1286		return -1;
1287	}
1288
1289	if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1290			o2net_keepalive_delay()) {
1291		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1292		       "delay of %u ms, but we use %u ms locally. "
1293		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1294		       be32_to_cpu(hand->o2net_keepalive_delay_ms),
1295		       o2net_keepalive_delay());
1296		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1297		return -1;
1298	}
1299
1300	if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1301			O2HB_MAX_WRITE_TIMEOUT_MS) {
1302		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1303		       "timeout of %u ms, but we use %u ms locally. "
1304		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1305		       be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1306		       O2HB_MAX_WRITE_TIMEOUT_MS);
1307		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1308		return -1;
1309	}
1310
1311	sc->sc_handshake_ok = 1;
1312
1313	spin_lock(&nn->nn_lock);
1314	/* set valid and queue the idle timers only if it hasn't been
1315	 * shut down already */
1316	if (nn->nn_sc == sc) {
1317		o2net_sc_reset_idle_timer(sc);
1318		atomic_set(&nn->nn_timeout, 0);
1319		o2net_set_nn_state(nn, sc, 1, 0);
1320	}
1321	spin_unlock(&nn->nn_lock);
1322
1323	/* shift everything up as though it wasn't there */
1324	sc->sc_page_off -= sizeof(struct o2net_handshake);
1325	if (sc->sc_page_off)
1326		memmove(hand, hand + 1, sc->sc_page_off);
1327
1328	return 0;
1329}
1330
1331/* this demuxes the queued rx bytes into header or payload bits and calls
1332 * handlers as each full message is read off the socket.  it returns -error,
1333 * == 0 eof, or > 0 for progress made.*/
1334static int o2net_advance_rx(struct o2net_sock_container *sc)
1335{
1336	struct o2net_msg *hdr;
1337	int ret = 0;
1338	void *data;
1339	size_t datalen;
1340
1341	sclog(sc, "receiving\n");
1342	o2net_set_advance_start_time(sc);
1343
1344	if (unlikely(sc->sc_handshake_ok == 0)) {
1345		if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1346			data = page_address(sc->sc_page) + sc->sc_page_off;
1347			datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1348			ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1349			if (ret > 0)
1350				sc->sc_page_off += ret;
1351		}
1352
1353		if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1354			o2net_check_handshake(sc);
1355			if (unlikely(sc->sc_handshake_ok == 0))
1356				ret = -EPROTO;
1357		}
1358		goto out;
1359	}
1360
1361	/* do we need more header? */
1362	if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1363		data = page_address(sc->sc_page) + sc->sc_page_off;
1364		datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1365		ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1366		if (ret > 0) {
1367			sc->sc_page_off += ret;
1368			/* only swab incoming here.. we can
1369			 * only get here once as we cross from
1370			 * being under to over */
1371			if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1372				hdr = page_address(sc->sc_page);
1373				if (be16_to_cpu(hdr->data_len) >
1374				    O2NET_MAX_PAYLOAD_BYTES)
1375					ret = -EOVERFLOW;
1376			}
1377		}
1378		if (ret <= 0)
1379			goto out;
1380	}
1381
1382	if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1383		/* oof, still don't have a header */
1384		goto out;
1385	}
1386
1387	/* this was swabbed above when we first read it */
1388	hdr = page_address(sc->sc_page);
1389
1390	msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1391
1392	/* do we need more payload? */
1393	if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1394		/* need more payload */
1395		data = page_address(sc->sc_page) + sc->sc_page_off;
1396		datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1397			  sc->sc_page_off;
1398		ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1399		if (ret > 0)
1400			sc->sc_page_off += ret;
1401		if (ret <= 0)
1402			goto out;
1403	}
1404
1405	if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1406		/* we can only get here once, the first time we read
1407		 * the payload.. so set ret to progress if the handler
1408		 * works out. after calling this the message is toast */
1409		ret = o2net_process_message(sc, hdr);
1410		if (ret == 0)
1411			ret = 1;
1412		sc->sc_page_off = 0;
1413	}
1414
1415out:
1416	sclog(sc, "ret = %d\n", ret);
1417	o2net_set_advance_stop_time(sc);
1418	return ret;
1419}
1420
1421/* this work func is triggerd by data ready.  it reads until it can read no
1422 * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1423 * our work the work struct will be marked and we'll be called again. */
1424static void o2net_rx_until_empty(struct work_struct *work)
1425{
1426	struct o2net_sock_container *sc =
1427		container_of(work, struct o2net_sock_container, sc_rx_work);
1428	int ret;
1429
1430	do {
1431		ret = o2net_advance_rx(sc);
1432	} while (ret > 0);
1433
1434	if (ret <= 0 && ret != -EAGAIN) {
1435		struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1436		sclog(sc, "saw error %d, closing\n", ret);
1437		/* not permanent so read failed handshake can retry */
1438		o2net_ensure_shutdown(nn, sc, 0);
1439	}
1440
1441	sc_put(sc);
1442}
1443
1444static int o2net_set_nodelay(struct socket *sock)
1445{
1446	int val = 1;
1447
1448	return kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
1449				    (void *)&val, sizeof(val));
1450}
1451
1452static int o2net_set_usertimeout(struct socket *sock)
1453{
1454	int user_timeout = O2NET_TCP_USER_TIMEOUT;
1455
1456	return kernel_setsockopt(sock, SOL_TCP, TCP_USER_TIMEOUT,
1457				(void *)&user_timeout, sizeof(user_timeout));
1458}
1459
1460static void o2net_initialize_handshake(void)
1461{
1462	o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1463		O2HB_MAX_WRITE_TIMEOUT_MS);
1464	o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1465	o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1466		o2net_keepalive_delay());
1467	o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1468		o2net_reconnect_delay());
1469}
1470
1471/* ------------------------------------------------------------ */
1472
1473/* called when a connect completes and after a sock is accepted.  the
1474 * rx path will see the response and mark the sc valid */
1475static void o2net_sc_connect_completed(struct work_struct *work)
1476{
1477	struct o2net_sock_container *sc =
1478		container_of(work, struct o2net_sock_container,
1479			     sc_connect_work);
1480
1481	mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1482              (unsigned long long)O2NET_PROTOCOL_VERSION,
1483	      (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1484
1485	o2net_initialize_handshake();
1486	o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1487	sc_put(sc);
1488}
1489
1490/* this is called as a work_struct func. */
1491static void o2net_sc_send_keep_req(struct work_struct *work)
1492{
1493	struct o2net_sock_container *sc =
1494		container_of(work, struct o2net_sock_container,
1495			     sc_keepalive_work.work);
1496
1497	o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1498	sc_put(sc);
1499}
1500
1501/* socket shutdown does a del_timer_sync against this as it tears down.
1502 * we can't start this timer until we've got to the point in sc buildup
1503 * where shutdown is going to be involved */
1504static void o2net_idle_timer(struct timer_list *t)
1505{
1506	struct o2net_sock_container *sc = from_timer(sc, t, sc_idle_timeout);
1507	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1508#ifdef CONFIG_DEBUG_FS
1509	unsigned long msecs = ktime_to_ms(ktime_get()) -
1510		ktime_to_ms(sc->sc_tv_timer);
1511#else
1512	unsigned long msecs = o2net_idle_timeout();
1513#endif
1514
1515	printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1516	       "idle for %lu.%lu secs.\n",
1517	       SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1518
1519	/* idle timerout happen, don't shutdown the connection, but
1520	 * make fence decision. Maybe the connection can recover before
1521	 * the decision is made.
1522	 */
1523	atomic_set(&nn->nn_timeout, 1);
1524	o2quo_conn_err(o2net_num_from_nn(nn));
1525	queue_delayed_work(o2net_wq, &nn->nn_still_up,
1526			msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1527
1528	o2net_sc_reset_idle_timer(sc);
1529
1530}
1531
1532static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1533{
1534	o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1535	o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1536		      msecs_to_jiffies(o2net_keepalive_delay()));
1537	o2net_set_sock_timer(sc);
1538	mod_timer(&sc->sc_idle_timeout,
1539	       jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1540}
1541
1542static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1543{
1544	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1545
1546	/* clear fence decision since the connection recover from timeout*/
1547	if (atomic_read(&nn->nn_timeout)) {
1548		o2quo_conn_up(o2net_num_from_nn(nn));
1549		cancel_delayed_work(&nn->nn_still_up);
1550		atomic_set(&nn->nn_timeout, 0);
1551	}
1552
1553	/* Only push out an existing timer */
1554	if (timer_pending(&sc->sc_idle_timeout))
1555		o2net_sc_reset_idle_timer(sc);
1556}
1557
1558/* this work func is kicked whenever a path sets the nn state which doesn't
1559 * have valid set.  This includes seeing hb come up, losing a connection,
1560 * having a connect attempt fail, etc. This centralizes the logic which decides
1561 * if a connect attempt should be made or if we should give up and all future
1562 * transmit attempts should fail */
1563static void o2net_start_connect(struct work_struct *work)
1564{
1565	struct o2net_node *nn =
1566		container_of(work, struct o2net_node, nn_connect_work.work);
1567	struct o2net_sock_container *sc = NULL;
1568	struct o2nm_node *node = NULL, *mynode = NULL;
1569	struct socket *sock = NULL;
1570	struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1571	int ret = 0, stop;
1572	unsigned int timeout;
1573	unsigned int noio_flag;
1574
1575	/*
1576	 * sock_create allocates the sock with GFP_KERNEL. We must set
1577	 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1578	 * by this process are done as if GFP_NOIO was specified. So we
1579	 * are not reentering filesystem while doing memory reclaim.
1580	 */
1581	noio_flag = memalloc_noio_save();
1582	/* if we're greater we initiate tx, otherwise we accept */
1583	if (o2nm_this_node() <= o2net_num_from_nn(nn))
1584		goto out;
1585
1586	/* watch for racing with tearing a node down */
1587	node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1588	if (node == NULL)
1589		goto out;
1590
1591	mynode = o2nm_get_node_by_num(o2nm_this_node());
1592	if (mynode == NULL)
1593		goto out;
1594
1595	spin_lock(&nn->nn_lock);
1596	/*
1597	 * see if we already have one pending or have given up.
1598	 * For nn_timeout, it is set when we close the connection
1599	 * because of the idle time out. So it means that we have
1600	 * at least connected to that node successfully once,
1601	 * now try to connect to it again.
1602	 */
1603	timeout = atomic_read(&nn->nn_timeout);
1604	stop = (nn->nn_sc ||
1605		(nn->nn_persistent_error &&
1606		(nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1607	spin_unlock(&nn->nn_lock);
1608	if (stop)
1609		goto out;
1610
1611	nn->nn_last_connect_attempt = jiffies;
1612
1613	sc = sc_alloc(node);
1614	if (sc == NULL) {
1615		mlog(0, "couldn't allocate sc\n");
1616		ret = -ENOMEM;
1617		goto out;
1618	}
1619
1620	ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1621	if (ret < 0) {
1622		mlog(0, "can't create socket: %d\n", ret);
1623		goto out;
1624	}
1625	sc->sc_sock = sock; /* freed by sc_kref_release */
1626
1627	sock->sk->sk_allocation = GFP_ATOMIC;
1628
1629	myaddr.sin_family = AF_INET;
1630	myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1631	myaddr.sin_port = htons(0); /* any port */
1632
1633	ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1634			      sizeof(myaddr));
1635	if (ret) {
1636		mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1637		     ret, &mynode->nd_ipv4_address);
1638		goto out;
1639	}
1640
1641	ret = o2net_set_nodelay(sc->sc_sock);
1642	if (ret) {
1643		mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1644		goto out;
1645	}
1646
1647	ret = o2net_set_usertimeout(sock);
1648	if (ret) {
1649		mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1650		goto out;
1651	}
1652
1653	o2net_register_callbacks(sc->sc_sock->sk, sc);
1654
1655	spin_lock(&nn->nn_lock);
1656	/* handshake completion will set nn->nn_sc_valid */
1657	o2net_set_nn_state(nn, sc, 0, 0);
1658	spin_unlock(&nn->nn_lock);
1659
1660	remoteaddr.sin_family = AF_INET;
1661	remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1662	remoteaddr.sin_port = node->nd_ipv4_port;
1663
1664	ret = sc->sc_sock->ops->connect(sc->sc_sock,
1665					(struct sockaddr *)&remoteaddr,
1666					sizeof(remoteaddr),
1667					O_NONBLOCK);
1668	if (ret == -EINPROGRESS)
1669		ret = 0;
1670
1671out:
1672	if (ret && sc) {
1673		printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1674		       " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1675		/* 0 err so that another will be queued and attempted
1676		 * from set_nn_state */
1677		o2net_ensure_shutdown(nn, sc, 0);
1678	}
1679	if (sc)
1680		sc_put(sc);
1681	if (node)
1682		o2nm_node_put(node);
1683	if (mynode)
1684		o2nm_node_put(mynode);
1685
1686	memalloc_noio_restore(noio_flag);
1687	return;
1688}
1689
1690static void o2net_connect_expired(struct work_struct *work)
1691{
1692	struct o2net_node *nn =
1693		container_of(work, struct o2net_node, nn_connect_expired.work);
1694
1695	spin_lock(&nn->nn_lock);
1696	if (!nn->nn_sc_valid) {
1697		printk(KERN_NOTICE "o2net: No connection established with "
1698		       "node %u after %u.%u seconds, check network and"
1699		       " cluster configuration.\n",
1700		     o2net_num_from_nn(nn),
1701		     o2net_idle_timeout() / 1000,
1702		     o2net_idle_timeout() % 1000);
1703
1704		o2net_set_nn_state(nn, NULL, 0, 0);
1705	}
1706	spin_unlock(&nn->nn_lock);
1707}
1708
1709static void o2net_still_up(struct work_struct *work)
1710{
1711	struct o2net_node *nn =
1712		container_of(work, struct o2net_node, nn_still_up.work);
1713
1714	o2quo_hb_still_up(o2net_num_from_nn(nn));
1715}
1716
1717/* ------------------------------------------------------------ */
1718
1719void o2net_disconnect_node(struct o2nm_node *node)
1720{
1721	struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1722
1723	/* don't reconnect until it's heartbeating again */
1724	spin_lock(&nn->nn_lock);
1725	atomic_set(&nn->nn_timeout, 0);
1726	o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1727	spin_unlock(&nn->nn_lock);
1728
1729	if (o2net_wq) {
1730		cancel_delayed_work(&nn->nn_connect_expired);
1731		cancel_delayed_work(&nn->nn_connect_work);
1732		cancel_delayed_work(&nn->nn_still_up);
1733		flush_workqueue(o2net_wq);
1734	}
1735}
1736
1737static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1738				  void *data)
1739{
1740	o2quo_hb_down(node_num);
1741
1742	if (!node)
1743		return;
1744
1745	if (node_num != o2nm_this_node())
1746		o2net_disconnect_node(node);
1747
1748	BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1749}
1750
1751static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1752				void *data)
1753{
1754	struct o2net_node *nn = o2net_nn_from_num(node_num);
1755
1756	o2quo_hb_up(node_num);
1757
1758	BUG_ON(!node);
1759
1760	/* ensure an immediate connect attempt */
1761	nn->nn_last_connect_attempt = jiffies -
1762		(msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1763
1764	if (node_num != o2nm_this_node()) {
1765		/* believe it or not, accept and node heartbeating testing
1766		 * can succeed for this node before we got here.. so
1767		 * only use set_nn_state to clear the persistent error
1768		 * if that hasn't already happened */
1769		spin_lock(&nn->nn_lock);
1770		atomic_set(&nn->nn_timeout, 0);
1771		if (nn->nn_persistent_error)
1772			o2net_set_nn_state(nn, NULL, 0, 0);
1773		spin_unlock(&nn->nn_lock);
1774	}
1775}
1776
1777void o2net_unregister_hb_callbacks(void)
1778{
1779	o2hb_unregister_callback(NULL, &o2net_hb_up);
1780	o2hb_unregister_callback(NULL, &o2net_hb_down);
1781}
1782
1783int o2net_register_hb_callbacks(void)
1784{
1785	int ret;
1786
1787	o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1788			    o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1789	o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1790			    o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1791
1792	ret = o2hb_register_callback(NULL, &o2net_hb_up);
1793	if (ret == 0)
1794		ret = o2hb_register_callback(NULL, &o2net_hb_down);
1795
1796	if (ret)
1797		o2net_unregister_hb_callbacks();
1798
1799	return ret;
1800}
1801
1802/* ------------------------------------------------------------ */
1803
1804static int o2net_accept_one(struct socket *sock, int *more)
1805{
1806	int ret;
1807	struct sockaddr_in sin;
1808	struct socket *new_sock = NULL;
1809	struct o2nm_node *node = NULL;
1810	struct o2nm_node *local_node = NULL;
1811	struct o2net_sock_container *sc = NULL;
1812	struct o2net_node *nn;
1813	unsigned int noio_flag;
1814
1815	/*
1816	 * sock_create_lite allocates the sock with GFP_KERNEL. We must set
1817	 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1818	 * by this process are done as if GFP_NOIO was specified. So we
1819	 * are not reentering filesystem while doing memory reclaim.
1820	 */
1821	noio_flag = memalloc_noio_save();
1822
1823	BUG_ON(sock == NULL);
1824	*more = 0;
1825	ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1826			       sock->sk->sk_protocol, &new_sock);
1827	if (ret)
1828		goto out;
1829
1830	new_sock->type = sock->type;
1831	new_sock->ops = sock->ops;
1832	ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, false);
1833	if (ret < 0)
1834		goto out;
1835
1836	*more = 1;
1837	new_sock->sk->sk_allocation = GFP_ATOMIC;
1838
1839	ret = o2net_set_nodelay(new_sock);
1840	if (ret) {
1841		mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1842		goto out;
1843	}
1844
1845	ret = o2net_set_usertimeout(new_sock);
1846	if (ret) {
1847		mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1848		goto out;
1849	}
1850
1851	ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin, 1);
1852	if (ret < 0)
1853		goto out;
1854
1855	node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1856	if (node == NULL) {
1857		printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1858		       "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1859		       ntohs(sin.sin_port));
1860		ret = -EINVAL;
1861		goto out;
1862	}
1863
1864	if (o2nm_this_node() >= node->nd_num) {
1865		local_node = o2nm_get_node_by_num(o2nm_this_node());
1866		if (local_node)
1867			printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1868					"seen at node '%s' (%u, %pI4:%d) from "
1869					"node '%s' (%u, %pI4:%d)\n",
1870					local_node->nd_name, local_node->nd_num,
1871					&(local_node->nd_ipv4_address),
1872					ntohs(local_node->nd_ipv4_port),
1873					node->nd_name,
1874					node->nd_num, &sin.sin_addr.s_addr,
1875					ntohs(sin.sin_port));
1876		ret = -EINVAL;
1877		goto out;
1878	}
1879
1880	/* this happens all the time when the other node sees our heartbeat
1881	 * and tries to connect before we see their heartbeat */
1882	if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1883		mlog(ML_CONN, "attempt to connect from node '%s' at "
1884		     "%pI4:%d but it isn't heartbeating\n",
1885		     node->nd_name, &sin.sin_addr.s_addr,
1886		     ntohs(sin.sin_port));
1887		ret = -EINVAL;
1888		goto out;
1889	}
1890
1891	nn = o2net_nn_from_num(node->nd_num);
1892
1893	spin_lock(&nn->nn_lock);
1894	if (nn->nn_sc)
1895		ret = -EBUSY;
1896	else
1897		ret = 0;
1898	spin_unlock(&nn->nn_lock);
1899	if (ret) {
1900		printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1901		       "at %pI4:%d but it already has an open connection\n",
1902		       node->nd_name, &sin.sin_addr.s_addr,
1903		       ntohs(sin.sin_port));
1904		goto out;
1905	}
1906
1907	sc = sc_alloc(node);
1908	if (sc == NULL) {
1909		ret = -ENOMEM;
1910		goto out;
1911	}
1912
1913	sc->sc_sock = new_sock;
1914	new_sock = NULL;
1915
1916	spin_lock(&nn->nn_lock);
1917	atomic_set(&nn->nn_timeout, 0);
1918	o2net_set_nn_state(nn, sc, 0, 0);
1919	spin_unlock(&nn->nn_lock);
1920
1921	o2net_register_callbacks(sc->sc_sock->sk, sc);
1922	o2net_sc_queue_work(sc, &sc->sc_rx_work);
1923
1924	o2net_initialize_handshake();
1925	o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1926
1927out:
1928	if (new_sock)
1929		sock_release(new_sock);
1930	if (node)
1931		o2nm_node_put(node);
1932	if (local_node)
1933		o2nm_node_put(local_node);
1934	if (sc)
1935		sc_put(sc);
1936
1937	memalloc_noio_restore(noio_flag);
1938	return ret;
1939}
1940
1941/*
1942 * This function is invoked in response to one or more
1943 * pending accepts at softIRQ level. We must drain the
1944 * entire que before returning.
1945 */
1946
1947static void o2net_accept_many(struct work_struct *work)
1948{
1949	struct socket *sock = o2net_listen_sock;
1950	int	more;
1951	int	err;
1952
1953	/*
1954	 * It is critical to note that due to interrupt moderation
1955	 * at the network driver level, we can't assume to get a
1956	 * softIRQ for every single conn since tcp SYN packets
1957	 * can arrive back-to-back, and therefore many pending
1958	 * accepts may result in just 1 softIRQ. If we terminate
1959	 * the o2net_accept_one() loop upon seeing an err, what happens
1960	 * to the rest of the conns in the queue? If no new SYN
1961	 * arrives for hours, no softIRQ  will be delivered,
1962	 * and the connections will just sit in the queue.
1963	 */
1964
1965	for (;;) {
1966		err = o2net_accept_one(sock, &more);
1967		if (!more)
1968			break;
1969		cond_resched();
1970	}
1971}
1972
1973static void o2net_listen_data_ready(struct sock *sk)
1974{
1975	void (*ready)(struct sock *sk);
1976
1977	read_lock_bh(&sk->sk_callback_lock);
1978	ready = sk->sk_user_data;
1979	if (ready == NULL) { /* check for teardown race */
1980		ready = sk->sk_data_ready;
1981		goto out;
1982	}
1983
1984	/* This callback may called twice when a new connection
1985	 * is  being established as a child socket inherits everything
1986	 * from a parent LISTEN socket, including the data_ready cb of
1987	 * the parent. This leads to a hazard. In o2net_accept_one()
1988	 * we are still initializing the child socket but have not
1989	 * changed the inherited data_ready callback yet when
1990	 * data starts arriving.
1991	 * We avoid this hazard by checking the state.
1992	 * For the listening socket,  the state will be TCP_LISTEN; for the new
1993	 * socket, will be  TCP_ESTABLISHED. Also, in this case,
1994	 * sk->sk_user_data is not a valid function pointer.
1995	 */
1996
1997	if (sk->sk_state == TCP_LISTEN) {
1998		queue_work(o2net_wq, &o2net_listen_work);
1999	} else {
2000		ready = NULL;
2001	}
2002
2003out:
2004	read_unlock_bh(&sk->sk_callback_lock);
2005	if (ready != NULL)
2006		ready(sk);
2007}
2008
2009static int o2net_open_listening_sock(__be32 addr, __be16 port)
2010{
2011	struct socket *sock = NULL;
2012	int ret;
2013	struct sockaddr_in sin = {
2014		.sin_family = PF_INET,
2015		.sin_addr = { .s_addr = addr },
2016		.sin_port = port,
2017	};
2018
2019	ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2020	if (ret < 0) {
2021		printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
2022		goto out;
2023	}
2024
2025	sock->sk->sk_allocation = GFP_ATOMIC;
2026
2027	write_lock_bh(&sock->sk->sk_callback_lock);
2028	sock->sk->sk_user_data = sock->sk->sk_data_ready;
2029	sock->sk->sk_data_ready = o2net_listen_data_ready;
2030	write_unlock_bh(&sock->sk->sk_callback_lock);
2031
2032	o2net_listen_sock = sock;
2033	INIT_WORK(&o2net_listen_work, o2net_accept_many);
2034
2035	sock->sk->sk_reuse = SK_CAN_REUSE;
2036	ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2037	if (ret < 0) {
2038		printk(KERN_ERR "o2net: Error %d while binding socket at "
2039		       "%pI4:%u\n", ret, &addr, ntohs(port)); 
2040		goto out;
2041	}
2042
2043	ret = sock->ops->listen(sock, 64);
2044	if (ret < 0)
2045		printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2046		       ret, &addr, ntohs(port));
2047
2048out:
2049	if (ret) {
2050		o2net_listen_sock = NULL;
2051		if (sock)
2052			sock_release(sock);
2053	}
2054	return ret;
2055}
2056
2057/*
2058 * called from node manager when we should bring up our network listening
2059 * socket.  node manager handles all the serialization to only call this
2060 * once and to match it with o2net_stop_listening().  note,
2061 * o2nm_this_node() doesn't work yet as we're being called while it
2062 * is being set up.
2063 */
2064int o2net_start_listening(struct o2nm_node *node)
2065{
2066	int ret = 0;
2067
2068	BUG_ON(o2net_wq != NULL);
2069	BUG_ON(o2net_listen_sock != NULL);
2070
2071	mlog(ML_KTHREAD, "starting o2net thread...\n");
2072	o2net_wq = alloc_ordered_workqueue("o2net", WQ_MEM_RECLAIM);
2073	if (o2net_wq == NULL) {
2074		mlog(ML_ERROR, "unable to launch o2net thread\n");
2075		return -ENOMEM; /* ? */
2076	}
2077
2078	ret = o2net_open_listening_sock(node->nd_ipv4_address,
2079					node->nd_ipv4_port);
2080	if (ret) {
2081		destroy_workqueue(o2net_wq);
2082		o2net_wq = NULL;
2083	} else
2084		o2quo_conn_up(node->nd_num);
2085
2086	return ret;
2087}
2088
2089/* again, o2nm_this_node() doesn't work here as we're involved in
2090 * tearing it down */
2091void o2net_stop_listening(struct o2nm_node *node)
2092{
2093	struct socket *sock = o2net_listen_sock;
2094	size_t i;
2095
2096	BUG_ON(o2net_wq == NULL);
2097	BUG_ON(o2net_listen_sock == NULL);
2098
2099	/* stop the listening socket from generating work */
2100	write_lock_bh(&sock->sk->sk_callback_lock);
2101	sock->sk->sk_data_ready = sock->sk->sk_user_data;
2102	sock->sk->sk_user_data = NULL;
2103	write_unlock_bh(&sock->sk->sk_callback_lock);
2104
2105	for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2106		struct o2nm_node *node = o2nm_get_node_by_num(i);
2107		if (node) {
2108			o2net_disconnect_node(node);
2109			o2nm_node_put(node);
2110		}
2111	}
2112
2113	/* finish all work and tear down the work queue */
2114	mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2115	destroy_workqueue(o2net_wq);
2116	o2net_wq = NULL;
2117
2118	sock_release(o2net_listen_sock);
2119	o2net_listen_sock = NULL;
2120
2121	o2quo_conn_err(node->nd_num);
2122}
2123
2124/* ------------------------------------------------------------ */
2125
2126int o2net_init(void)
2127{
2128	unsigned long i;
2129
2130	o2quo_init();
2131
2132	o2net_debugfs_init();
2133
2134	o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2135	o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2136	o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2137	if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp)
2138		goto out;
2139
2140	o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2141	o2net_hand->connector_id = cpu_to_be64(1);
2142
2143	o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2144	o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2145
2146	for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2147		struct o2net_node *nn = o2net_nn_from_num(i);
2148
2149		atomic_set(&nn->nn_timeout, 0);
2150		spin_lock_init(&nn->nn_lock);
2151		INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2152		INIT_DELAYED_WORK(&nn->nn_connect_expired,
2153				  o2net_connect_expired);
2154		INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2155		/* until we see hb from a node we'll return einval */
2156		nn->nn_persistent_error = -ENOTCONN;
2157		init_waitqueue_head(&nn->nn_sc_wq);
2158		idr_init(&nn->nn_status_idr);
2159		INIT_LIST_HEAD(&nn->nn_status_list);
2160	}
2161
2162	return 0;
2163
2164out:
2165	kfree(o2net_hand);
2166	kfree(o2net_keep_req);
2167	kfree(o2net_keep_resp);
2168	o2net_debugfs_exit();
2169	o2quo_exit();
2170	return -ENOMEM;
2171}
2172
2173void o2net_exit(void)
2174{
2175	o2quo_exit();
2176	kfree(o2net_hand);
2177	kfree(o2net_keep_req);
2178	kfree(o2net_keep_resp);
2179	o2net_debugfs_exit();
2180}