1/* -*- mode: c; c-basic-offset: 8; -*-
   2 *
   3 * vim: noexpandtab sw=8 ts=8 sts=0:
   4 *
   5 * Copyright (C) 2004 Oracle.  All rights reserved.
   6 *
   7 * This program is free software; you can redistribute it and/or
   8 * modify it under the terms of the GNU General Public
   9 * License as published by the Free Software Foundation; either
  10 * version 2 of the License, or (at your option) any later version.
  11 *
  12 * This program is distributed in the hope that it will be useful,
  13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15 * General Public License for more details.
  16 *
  17 * You should have received a copy of the GNU General Public
  18 * License along with this program; if not, write to the
  19 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  20 * Boston, MA 021110-1307, USA.
  21 *
  22 * ----
  23 *
  24 * Callers for this were originally written against a very simple synchronus
  25 * API.  This implementation reflects those simple callers.  Some day I'm sure
  26 * we'll need to move to a more robust posting/callback mechanism.
  27 *
  28 * Transmit calls pass in kernel virtual addresses and block copying this into
  29 * the socket's tx buffers via a usual blocking sendmsg.  They'll block waiting
  30 * for a failed socket to timeout.  TX callers can also pass in a poniter to an
  31 * 'int' which gets filled with an errno off the wire in response to the
  32 * message they send.
  33 *
  34 * Handlers for unsolicited messages are registered.  Each socket has a page
  35 * that incoming data is copied into.  First the header, then the data.
  36 * Handlers are called from only one thread with a reference to this per-socket
  37 * page.  This page is destroyed after the handler call, so it can't be
  38 * referenced beyond the call.  Handlers may block but are discouraged from
  39 * doing so.
  40 *
  41 * Any framing errors (bad magic, large payload lengths) close a connection.
  42 *
  43 * Our sock_container holds the state we associate with a socket.  It's current
  44 * framing state is held there as well as the refcounting we do around when it
  45 * is safe to tear down the socket.  The socket is only finally torn down from
  46 * the container when the container loses all of its references -- so as long
  47 * as you hold a ref on the container you can trust that the socket is valid
  48 * for use with kernel socket APIs.
  49 *
  50 * Connections are initiated between a pair of nodes when the node with the
  51 * higher node number gets a heartbeat callback which indicates that the lower
  52 * numbered node has started heartbeating.  The lower numbered node is passive
  53 * and only accepts the connection if the higher numbered node is heartbeating.
  54 */
  55
  56#include <linux/kernel.h>
  57#include <linux/jiffies.h>
  58#include <linux/slab.h>
  59#include <linux/idr.h>
  60#include <linux/kref.h>
  61#include <linux/net.h>
  62#include <linux/export.h>
  63#include <net/tcp.h>
  64
  65#include <linux/uaccess.h>
  66
  67#include "heartbeat.h"
  68#include "tcp.h"
  69#include "nodemanager.h"
  70#define MLOG_MASK_PREFIX ML_TCP
  71#include "masklog.h"
  72#include "quorum.h"
  73
  74#include "tcp_internal.h"
  75
  76#define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
  77#define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,	\
  78			  &sc->sc_node->nd_ipv4_address,		\
  79			  ntohs(sc->sc_node->nd_ipv4_port)
  80
  81/*
  82 * In the following two log macros, the whitespace after the ',' just
  83 * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  84 * previous token if args expands to nothing.
  85 */
  86#define msglog(hdr, fmt, args...) do {					\
  87	typeof(hdr) __hdr = (hdr);					\
  88	mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "	\
  89	     "key %08x num %u] " fmt,					\
  90	     be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len), 	\
  91	     be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),	\
  92	     be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),	\
  93	     be32_to_cpu(__hdr->msg_num) ,  ##args);			\
  94} while (0)
  95
  96#define sclog(sc, fmt, args...) do {					\
  97	typeof(sc) __sc = (sc);						\
  98	mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "	\
  99	     "pg_off %zu] " fmt, __sc,					\
 100	     atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,	\
 101	    __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,	\
 102	    ##args);							\
 103} while (0)
 104
 105static DEFINE_RWLOCK(o2net_handler_lock);
 106static struct rb_root o2net_handler_tree = RB_ROOT;
 107
 108static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
 109
 110/* XXX someday we'll need better accounting */
 111static struct socket *o2net_listen_sock;
 112
 113/*
 114 * listen work is only queued by the listening socket callbacks on the
 115 * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
 116 * quorum work is queued as sock containers are shutdown.. stop_listening
 117 * tears down all the node's sock containers, preventing future shutdowns
 118 * and queued quroum work, before canceling delayed quorum work and
 119 * destroying the work queue.
 120 */
 121static struct workqueue_struct *o2net_wq;
 122static struct work_struct o2net_listen_work;
 123
 124static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
 125#define O2NET_HB_PRI 0x1
 126
 127static struct o2net_handshake *o2net_hand;
 128static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
 129
 130static int o2net_sys_err_translations[O2NET_ERR_MAX] =
 131		{[O2NET_ERR_NONE]	= 0,
 132		 [O2NET_ERR_NO_HNDLR]	= -ENOPROTOOPT,
 133		 [O2NET_ERR_OVERFLOW]	= -EOVERFLOW,
 134		 [O2NET_ERR_DIED]	= -EHOSTDOWN,};
 135
 136/* can't quite avoid *all* internal declarations :/ */
 137static void o2net_sc_connect_completed(struct work_struct *work);
 138static void o2net_rx_until_empty(struct work_struct *work);
 139static void o2net_shutdown_sc(struct work_struct *work);
 140static void o2net_listen_data_ready(struct sock *sk);
 141static void o2net_sc_send_keep_req(struct work_struct *work);
 142static void o2net_idle_timer(unsigned long data);
 143static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
 144static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
 145
 146#ifdef CONFIG_DEBUG_FS
 147static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
 148			   u32 msgkey, struct task_struct *task, u8 node)
 149{
 150	INIT_LIST_HEAD(&nst->st_net_debug_item);
 151	nst->st_task = task;
 152	nst->st_msg_type = msgtype;
 153	nst->st_msg_key = msgkey;
 154	nst->st_node = node;
 155}
 156
 157static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
 158{
 159	nst->st_sock_time = ktime_get();
 160}
 161
 162static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
 163{
 164	nst->st_send_time = ktime_get();
 165}
 166
 167static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
 168{
 169	nst->st_status_time = ktime_get();
 170}
 171
 172static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
 173						struct o2net_sock_container *sc)
 174{
 175	nst->st_sc = sc;
 176}
 177
 178static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
 179					u32 msg_id)
 180{
 181	nst->st_id = msg_id;
 182}
 183
 184static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
 185{
 186	sc->sc_tv_timer = ktime_get();
 187}
 188
 189static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
 190{
 191	sc->sc_tv_data_ready = ktime_get();
 192}
 193
 194static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
 195{
 196	sc->sc_tv_advance_start = ktime_get();
 197}
 198
 199static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
 200{
 201	sc->sc_tv_advance_stop = ktime_get();
 202}
 203
 204static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
 205{
 206	sc->sc_tv_func_start = ktime_get();
 207}
 208
 209static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
 210{
 211	sc->sc_tv_func_stop = ktime_get();
 212}
 213
 214#else  /* CONFIG_DEBUG_FS */
 215# define o2net_init_nst(a, b, c, d, e)
 216# define o2net_set_nst_sock_time(a)
 217# define o2net_set_nst_send_time(a)
 218# define o2net_set_nst_status_time(a)
 219# define o2net_set_nst_sock_container(a, b)
 220# define o2net_set_nst_msg_id(a, b)
 221# define o2net_set_sock_timer(a)
 222# define o2net_set_data_ready_time(a)
 223# define o2net_set_advance_start_time(a)
 224# define o2net_set_advance_stop_time(a)
 225# define o2net_set_func_start_time(a)
 226# define o2net_set_func_stop_time(a)
 227#endif /* CONFIG_DEBUG_FS */
 228
 229#ifdef CONFIG_OCFS2_FS_STATS
 230static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
 231{
 232	return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
 233}
 234
 235static void o2net_update_send_stats(struct o2net_send_tracking *nst,
 236				    struct o2net_sock_container *sc)
 237{
 238	sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
 239					   ktime_sub(ktime_get(),
 240						     nst->st_status_time));
 241	sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
 242					 ktime_sub(nst->st_status_time,
 243						   nst->st_send_time));
 244	sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
 245					    ktime_sub(nst->st_send_time,
 246						      nst->st_sock_time));
 247	sc->sc_send_count++;
 248}
 249
 250static void o2net_update_recv_stats(struct o2net_sock_container *sc)
 251{
 252	sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
 253					    o2net_get_func_run_time(sc));
 254	sc->sc_recv_count++;
 255}
 256
 257#else
 258
 259# define o2net_update_send_stats(a, b)
 260
 261# define o2net_update_recv_stats(sc)
 262
 263#endif /* CONFIG_OCFS2_FS_STATS */
 264
 265static inline unsigned int o2net_reconnect_delay(void)
 266{
 267	return o2nm_single_cluster->cl_reconnect_delay_ms;
 268}
 269
 270static inline unsigned int o2net_keepalive_delay(void)
 271{
 272	return o2nm_single_cluster->cl_keepalive_delay_ms;
 273}
 274
 275static inline unsigned int o2net_idle_timeout(void)
 276{
 277	return o2nm_single_cluster->cl_idle_timeout_ms;
 278}
 279
 280static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
 281{
 282	int trans;
 283	BUG_ON(err >= O2NET_ERR_MAX);
 284	trans = o2net_sys_err_translations[err];
 285
 286	/* Just in case we mess up the translation table above */
 287	BUG_ON(err != O2NET_ERR_NONE && trans == 0);
 288	return trans;
 289}
 290
 291static struct o2net_node * o2net_nn_from_num(u8 node_num)
 292{
 293	BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
 294	return &o2net_nodes[node_num];
 295}
 296
 297static u8 o2net_num_from_nn(struct o2net_node *nn)
 298{
 299	BUG_ON(nn == NULL);
 300	return nn - o2net_nodes;
 301}
 302
 303/* ------------------------------------------------------------ */
 304
 305static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
 306{
 307	int ret;
 308
 309	spin_lock(&nn->nn_lock);
 310	ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
 311	if (ret >= 0) {
 312		nsw->ns_id = ret;
 313		list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
 314	}
 315	spin_unlock(&nn->nn_lock);
 316	if (ret < 0)
 317		return ret;
 318
 319	init_waitqueue_head(&nsw->ns_wq);
 320	nsw->ns_sys_status = O2NET_ERR_NONE;
 321	nsw->ns_status = 0;
 322	return 0;
 323}
 324
 325static void o2net_complete_nsw_locked(struct o2net_node *nn,
 326				      struct o2net_status_wait *nsw,
 327				      enum o2net_system_error sys_status,
 328				      s32 status)
 329{
 330	assert_spin_locked(&nn->nn_lock);
 331
 332	if (!list_empty(&nsw->ns_node_item)) {
 333		list_del_init(&nsw->ns_node_item);
 334		nsw->ns_sys_status = sys_status;
 335		nsw->ns_status = status;
 336		idr_remove(&nn->nn_status_idr, nsw->ns_id);
 337		wake_up(&nsw->ns_wq);
 338	}
 339}
 340
 341static void o2net_complete_nsw(struct o2net_node *nn,
 342			       struct o2net_status_wait *nsw,
 343			       u64 id, enum o2net_system_error sys_status,
 344			       s32 status)
 345{
 346	spin_lock(&nn->nn_lock);
 347	if (nsw == NULL) {
 348		if (id > INT_MAX)
 349			goto out;
 350
 351		nsw = idr_find(&nn->nn_status_idr, id);
 352		if (nsw == NULL)
 353			goto out;
 354	}
 355
 356	o2net_complete_nsw_locked(nn, nsw, sys_status, status);
 357
 358out:
 359	spin_unlock(&nn->nn_lock);
 360	return;
 361}
 362
 363static void o2net_complete_nodes_nsw(struct o2net_node *nn)
 364{
 365	struct o2net_status_wait *nsw, *tmp;
 366	unsigned int num_kills = 0;
 367
 368	assert_spin_locked(&nn->nn_lock);
 369
 370	list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
 371		o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
 372		num_kills++;
 373	}
 374
 375	mlog(0, "completed %d messages for node %u\n", num_kills,
 376	     o2net_num_from_nn(nn));
 377}
 378
 379static int o2net_nsw_completed(struct o2net_node *nn,
 380			       struct o2net_status_wait *nsw)
 381{
 382	int completed;
 383	spin_lock(&nn->nn_lock);
 384	completed = list_empty(&nsw->ns_node_item);
 385	spin_unlock(&nn->nn_lock);
 386	return completed;
 387}
 388
 389/* ------------------------------------------------------------ */
 390
 391static void sc_kref_release(struct kref *kref)
 392{
 393	struct o2net_sock_container *sc = container_of(kref,
 394					struct o2net_sock_container, sc_kref);
 395	BUG_ON(timer_pending(&sc->sc_idle_timeout));
 396
 397	sclog(sc, "releasing\n");
 398
 399	if (sc->sc_sock) {
 400		sock_release(sc->sc_sock);
 401		sc->sc_sock = NULL;
 402	}
 403
 404	o2nm_undepend_item(&sc->sc_node->nd_item);
 405	o2nm_node_put(sc->sc_node);
 406	sc->sc_node = NULL;
 407
 408	o2net_debug_del_sc(sc);
 409
 410	if (sc->sc_page)
 411		__free_page(sc->sc_page);
 412	kfree(sc);
 413}
 414
 415static void sc_put(struct o2net_sock_container *sc)
 416{
 417	sclog(sc, "put\n");
 418	kref_put(&sc->sc_kref, sc_kref_release);
 419}
 420static void sc_get(struct o2net_sock_container *sc)
 421{
 422	sclog(sc, "get\n");
 423	kref_get(&sc->sc_kref);
 424}
 425static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
 426{
 427	struct o2net_sock_container *sc, *ret = NULL;
 428	struct page *page = NULL;
 429	int status = 0;
 430
 431	page = alloc_page(GFP_NOFS);
 432	sc = kzalloc(sizeof(*sc), GFP_NOFS);
 433	if (sc == NULL || page == NULL)
 434		goto out;
 435
 436	kref_init(&sc->sc_kref);
 437	o2nm_node_get(node);
 438	sc->sc_node = node;
 439
 440	/* pin the node item of the remote node */
 441	status = o2nm_depend_item(&node->nd_item);
 442	if (status) {
 443		mlog_errno(status);
 444		o2nm_node_put(node);
 445		goto out;
 446	}
 447	INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
 448	INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
 449	INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
 450	INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
 451
 452	init_timer(&sc->sc_idle_timeout);
 453	sc->sc_idle_timeout.function = o2net_idle_timer;
 454	sc->sc_idle_timeout.data = (unsigned long)sc;
 455
 456	sclog(sc, "alloced\n");
 457
 458	ret = sc;
 459	sc->sc_page = page;
 460	o2net_debug_add_sc(sc);
 461	sc = NULL;
 462	page = NULL;
 463
 464out:
 465	if (page)
 466		__free_page(page);
 467	kfree(sc);
 468
 469	return ret;
 470}
 471
 472/* ------------------------------------------------------------ */
 473
 474static void o2net_sc_queue_work(struct o2net_sock_container *sc,
 475				struct work_struct *work)
 476{
 477	sc_get(sc);
 478	if (!queue_work(o2net_wq, work))
 479		sc_put(sc);
 480}
 481static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
 482					struct delayed_work *work,
 483					int delay)
 484{
 485	sc_get(sc);
 486	if (!queue_delayed_work(o2net_wq, work, delay))
 487		sc_put(sc);
 488}
 489static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
 490					 struct delayed_work *work)
 491{
 492	if (cancel_delayed_work(work))
 493		sc_put(sc);
 494}
 495
 496static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
 497
 498int o2net_num_connected_peers(void)
 499{
 500	return atomic_read(&o2net_connected_peers);
 501}
 502
 503static void o2net_set_nn_state(struct o2net_node *nn,
 504			       struct o2net_sock_container *sc,
 505			       unsigned valid, int err)
 506{
 507	int was_valid = nn->nn_sc_valid;
 508	int was_err = nn->nn_persistent_error;
 509	struct o2net_sock_container *old_sc = nn->nn_sc;
 510
 511	assert_spin_locked(&nn->nn_lock);
 512
 513	if (old_sc && !sc)
 514		atomic_dec(&o2net_connected_peers);
 515	else if (!old_sc && sc)
 516		atomic_inc(&o2net_connected_peers);
 517
 518	/* the node num comparison and single connect/accept path should stop
 519	 * an non-null sc from being overwritten with another */
 520	BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
 521	mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
 522	mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
 523
 524	if (was_valid && !valid && err == 0)
 525		err = -ENOTCONN;
 526
 527	mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
 528	     o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
 529	     nn->nn_persistent_error, err);
 530
 531	nn->nn_sc = sc;
 532	nn->nn_sc_valid = valid ? 1 : 0;
 533	nn->nn_persistent_error = err;
 534
 535	/* mirrors o2net_tx_can_proceed() */
 536	if (nn->nn_persistent_error || nn->nn_sc_valid)
 537		wake_up(&nn->nn_sc_wq);
 538
 539	if (was_valid && !was_err && nn->nn_persistent_error) {
 540		o2quo_conn_err(o2net_num_from_nn(nn));
 541		queue_delayed_work(o2net_wq, &nn->nn_still_up,
 542				   msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
 543	}
 544
 545	if (was_valid && !valid) {
 546		if (old_sc)
 547			printk(KERN_NOTICE "o2net: No longer connected to "
 548				SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
 549		o2net_complete_nodes_nsw(nn);
 550	}
 551
 552	if (!was_valid && valid) {
 553		o2quo_conn_up(o2net_num_from_nn(nn));
 554		cancel_delayed_work(&nn->nn_connect_expired);
 555		printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
 556		       o2nm_this_node() > sc->sc_node->nd_num ?
 557		       "Connected to" : "Accepted connection from",
 558		       SC_NODEF_ARGS(sc));
 559	}
 560
 561	/* trigger the connecting worker func as long as we're not valid,
 562	 * it will back off if it shouldn't connect.  This can be called
 563	 * from node config teardown and so needs to be careful about
 564	 * the work queue actually being up. */
 565	if (!valid && o2net_wq) {
 566		unsigned long delay;
 567		/* delay if we're within a RECONNECT_DELAY of the
 568		 * last attempt */
 569		delay = (nn->nn_last_connect_attempt +
 570			 msecs_to_jiffies(o2net_reconnect_delay()))
 571			- jiffies;
 572		if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
 573			delay = 0;
 574		mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
 575		queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
 576
 577		/*
 578		 * Delay the expired work after idle timeout.
 579		 *
 580		 * We might have lots of failed connection attempts that run
 581		 * through here but we only cancel the connect_expired work when
 582		 * a connection attempt succeeds.  So only the first enqueue of
 583		 * the connect_expired work will do anything.  The rest will see
 584		 * that it's already queued and do nothing.
 585		 */
 586		delay += msecs_to_jiffies(o2net_idle_timeout());
 587		queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
 588	}
 589
 590	/* keep track of the nn's sc ref for the caller */
 591	if ((old_sc == NULL) && sc)
 592		sc_get(sc);
 593	if (old_sc && (old_sc != sc)) {
 594		o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
 595		sc_put(old_sc);
 596	}
 597}
 598
 599/* see o2net_register_callbacks() */
 600static void o2net_data_ready(struct sock *sk)
 601{
 602	void (*ready)(struct sock *sk);
 603	struct o2net_sock_container *sc;
 604
 605	read_lock_bh(&sk->sk_callback_lock);
 606	sc = sk->sk_user_data;
 607	if (sc) {
 608		sclog(sc, "data_ready hit\n");
 609		o2net_set_data_ready_time(sc);
 610		o2net_sc_queue_work(sc, &sc->sc_rx_work);
 611		ready = sc->sc_data_ready;
 612	} else {
 613		ready = sk->sk_data_ready;
 614	}
 615	read_unlock_bh(&sk->sk_callback_lock);
 616
 617	ready(sk);
 618}
 619
 620/* see o2net_register_callbacks() */
 621static void o2net_state_change(struct sock *sk)
 622{
 623	void (*state_change)(struct sock *sk);
 624	struct o2net_sock_container *sc;
 625
 626	read_lock_bh(&sk->sk_callback_lock);
 627	sc = sk->sk_user_data;
 628	if (sc == NULL) {
 629		state_change = sk->sk_state_change;
 630		goto out;
 631	}
 632
 633	sclog(sc, "state_change to %d\n", sk->sk_state);
 634
 635	state_change = sc->sc_state_change;
 636
 637	switch(sk->sk_state) {
 638	/* ignore connecting sockets as they make progress */
 639	case TCP_SYN_SENT:
 640	case TCP_SYN_RECV:
 641		break;
 642	case TCP_ESTABLISHED:
 643		o2net_sc_queue_work(sc, &sc->sc_connect_work);
 644		break;
 645	default:
 646		printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
 647			" shutdown, state %d\n",
 648			SC_NODEF_ARGS(sc), sk->sk_state);
 649		o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 650		break;
 651	}
 652out:
 653	read_unlock_bh(&sk->sk_callback_lock);
 654	state_change(sk);
 655}
 656
 657/*
 658 * we register callbacks so we can queue work on events before calling
 659 * the original callbacks.  our callbacks our careful to test user_data
 660 * to discover when they've reaced with o2net_unregister_callbacks().
 661 */
 662static void o2net_register_callbacks(struct sock *sk,
 663				     struct o2net_sock_container *sc)
 664{
 665	write_lock_bh(&sk->sk_callback_lock);
 666
 667	/* accepted sockets inherit the old listen socket data ready */
 668	if (sk->sk_data_ready == o2net_listen_data_ready) {
 669		sk->sk_data_ready = sk->sk_user_data;
 670		sk->sk_user_data = NULL;
 671	}
 672
 673	BUG_ON(sk->sk_user_data != NULL);
 674	sk->sk_user_data = sc;
 675	sc_get(sc);
 676
 677	sc->sc_data_ready = sk->sk_data_ready;
 678	sc->sc_state_change = sk->sk_state_change;
 679	sk->sk_data_ready = o2net_data_ready;
 680	sk->sk_state_change = o2net_state_change;
 681
 682	mutex_init(&sc->sc_send_lock);
 683
 684	write_unlock_bh(&sk->sk_callback_lock);
 685}
 686
 687static int o2net_unregister_callbacks(struct sock *sk,
 688			           struct o2net_sock_container *sc)
 689{
 690	int ret = 0;
 691
 692	write_lock_bh(&sk->sk_callback_lock);
 693	if (sk->sk_user_data == sc) {
 694		ret = 1;
 695		sk->sk_user_data = NULL;
 696		sk->sk_data_ready = sc->sc_data_ready;
 697		sk->sk_state_change = sc->sc_state_change;
 698	}
 699	write_unlock_bh(&sk->sk_callback_lock);
 700
 701	return ret;
 702}
 703
 704/*
 705 * this is a little helper that is called by callers who have seen a problem
 706 * with an sc and want to detach it from the nn if someone already hasn't beat
 707 * them to it.  if an error is given then the shutdown will be persistent
 708 * and pending transmits will be canceled.
 709 */
 710static void o2net_ensure_shutdown(struct o2net_node *nn,
 711			           struct o2net_sock_container *sc,
 712				   int err)
 713{
 714	spin_lock(&nn->nn_lock);
 715	if (nn->nn_sc == sc)
 716		o2net_set_nn_state(nn, NULL, 0, err);
 717	spin_unlock(&nn->nn_lock);
 718}
 719
 720/*
 721 * This work queue function performs the blocking parts of socket shutdown.  A
 722 * few paths lead here.  set_nn_state will trigger this callback if it sees an
 723 * sc detached from the nn.  state_change will also trigger this callback
 724 * directly when it sees errors.  In that case we need to call set_nn_state
 725 * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 726 * itself.
 727 */
 728static void o2net_shutdown_sc(struct work_struct *work)
 729{
 730	struct o2net_sock_container *sc =
 731		container_of(work, struct o2net_sock_container,
 732			     sc_shutdown_work);
 733	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 734
 735	sclog(sc, "shutting down\n");
 736
 737	/* drop the callbacks ref and call shutdown only once */
 738	if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 739		/* we shouldn't flush as we're in the thread, the
 740		 * races with pending sc work structs are harmless */
 741		del_timer_sync(&sc->sc_idle_timeout);
 742		o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 743		sc_put(sc);
 744		kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 745	}
 746
 747	/* not fatal so failed connects before the other guy has our
 748	 * heartbeat can be retried */
 749	o2net_ensure_shutdown(nn, sc, 0);
 750	sc_put(sc);
 751}
 752
 753/* ------------------------------------------------------------ */
 754
 755static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 756			     u32 key)
 757{
 758	int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 759
 760	if (ret == 0)
 761		ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 762
 763	return ret;
 764}
 765
 766static struct o2net_msg_handler *
 767o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 768			  struct rb_node **ret_parent)
 769{
 770	struct rb_node **p = &o2net_handler_tree.rb_node;
 771	struct rb_node *parent = NULL;
 772	struct o2net_msg_handler *nmh, *ret = NULL;
 773	int cmp;
 774
 775	while (*p) {
 776		parent = *p;
 777		nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 778		cmp = o2net_handler_cmp(nmh, msg_type, key);
 779
 780		if (cmp < 0)
 781			p = &(*p)->rb_left;
 782		else if (cmp > 0)
 783			p = &(*p)->rb_right;
 784		else {
 785			ret = nmh;
 786			break;
 787		}
 788	}
 789
 790	if (ret_p != NULL)
 791		*ret_p = p;
 792	if (ret_parent != NULL)
 793		*ret_parent = parent;
 794
 795	return ret;
 796}
 797
 798static void o2net_handler_kref_release(struct kref *kref)
 799{
 800	struct o2net_msg_handler *nmh;
 801	nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 802
 803	kfree(nmh);
 804}
 805
 806static void o2net_handler_put(struct o2net_msg_handler *nmh)
 807{
 808	kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 809}
 810
 811/* max_len is protection for the handler func.  incoming messages won't
 812 * be given to the handler if their payload is longer than the max. */
 813int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 814			   o2net_msg_handler_func *func, void *data,
 815			   o2net_post_msg_handler_func *post_func,
 816			   struct list_head *unreg_list)
 817{
 818	struct o2net_msg_handler *nmh = NULL;
 819	struct rb_node **p, *parent;
 820	int ret = 0;
 821
 822	if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 823		mlog(0, "max_len for message handler out of range: %u\n",
 824			max_len);
 825		ret = -EINVAL;
 826		goto out;
 827	}
 828
 829	if (!msg_type) {
 830		mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 831		ret = -EINVAL;
 832		goto out;
 833
 834	}
 835	if (!func) {
 836		mlog(0, "no message handler provided: %u, %p\n",
 837		       msg_type, func);
 838		ret = -EINVAL;
 839		goto out;
 840	}
 841
 842       	nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 843	if (nmh == NULL) {
 844		ret = -ENOMEM;
 845		goto out;
 846	}
 847
 848	nmh->nh_func = func;
 849	nmh->nh_func_data = data;
 850	nmh->nh_post_func = post_func;
 851	nmh->nh_msg_type = msg_type;
 852	nmh->nh_max_len = max_len;
 853	nmh->nh_key = key;
 854	/* the tree and list get this ref.. they're both removed in
 855	 * unregister when this ref is dropped */
 856	kref_init(&nmh->nh_kref);
 857	INIT_LIST_HEAD(&nmh->nh_unregister_item);
 858
 859	write_lock(&o2net_handler_lock);
 860	if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 861		ret = -EEXIST;
 862	else {
 863	        rb_link_node(&nmh->nh_node, parent, p);
 864		rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 865		list_add_tail(&nmh->nh_unregister_item, unreg_list);
 866
 867		mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 868		     func, msg_type, key);
 869		/* we've had some trouble with handlers seemingly vanishing. */
 870		mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 871							  &parent) == NULL,
 872			        "couldn't find handler we *just* registered "
 873				"for type %u key %08x\n", msg_type, key);
 874	}
 875	write_unlock(&o2net_handler_lock);
 876	if (ret)
 877		goto out;
 878
 879out:
 880	if (ret)
 881		kfree(nmh);
 882
 883	return ret;
 884}
 885EXPORT_SYMBOL_GPL(o2net_register_handler);
 886
 887void o2net_unregister_handler_list(struct list_head *list)
 888{
 889	struct o2net_msg_handler *nmh, *n;
 890
 891	write_lock(&o2net_handler_lock);
 892	list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 893		mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 894		     nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 895		rb_erase(&nmh->nh_node, &o2net_handler_tree);
 896		list_del_init(&nmh->nh_unregister_item);
 897		kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 898	}
 899	write_unlock(&o2net_handler_lock);
 900}
 901EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 902
 903static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 904{
 905	struct o2net_msg_handler *nmh;
 906
 907	read_lock(&o2net_handler_lock);
 908	nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 909	if (nmh)
 910		kref_get(&nmh->nh_kref);
 911	read_unlock(&o2net_handler_lock);
 912
 913	return nmh;
 914}
 915
 916/* ------------------------------------------------------------ */
 917
 918static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 919{
 920	struct kvec vec = { .iov_len = len, .iov_base = data, };
 921	struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
 922	return kernel_recvmsg(sock, &msg, &vec, 1, len, msg.msg_flags);
 923}
 924
 925static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 926			      size_t veclen, size_t total)
 927{
 928	int ret;
 929	struct msghdr msg = {.msg_flags = 0,};
 930
 931	if (sock == NULL) {
 932		ret = -EINVAL;
 933		goto out;
 934	}
 935
 936	ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
 937	if (likely(ret == total))
 938		return 0;
 939	mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
 940	if (ret >= 0)
 941		ret = -EPIPE; /* should be smarter, I bet */
 942out:
 943	mlog(0, "returning error: %d\n", ret);
 944	return ret;
 945}
 946
 947static void o2net_sendpage(struct o2net_sock_container *sc,
 948			   void *kmalloced_virt,
 949			   size_t size)
 950{
 951	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 952	ssize_t ret;
 953
 954	while (1) {
 955		mutex_lock(&sc->sc_send_lock);
 956		ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 957						 virt_to_page(kmalloced_virt),
 958						 (long)kmalloced_virt & ~PAGE_MASK,
 959						 size, MSG_DONTWAIT);
 960		mutex_unlock(&sc->sc_send_lock);
 961		if (ret == size)
 962			break;
 963		if (ret == (ssize_t)-EAGAIN) {
 964			mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 965			     " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 966			cond_resched();
 967			continue;
 968		}
 969		mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 970		     " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 971		o2net_ensure_shutdown(nn, sc, 0);
 972		break;
 973	}
 974}
 975
 976static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 977{
 978	memset(msg, 0, sizeof(struct o2net_msg));
 979	msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 980	msg->data_len = cpu_to_be16(data_len);
 981	msg->msg_type = cpu_to_be16(msg_type);
 982	msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 983	msg->status = 0;
 984	msg->key = cpu_to_be32(key);
 985}
 986
 987static int o2net_tx_can_proceed(struct o2net_node *nn,
 988			        struct o2net_sock_container **sc_ret,
 989				int *error)
 990{
 991	int ret = 0;
 992
 993	spin_lock(&nn->nn_lock);
 994	if (nn->nn_persistent_error) {
 995		ret = 1;
 996		*sc_ret = NULL;
 997		*error = nn->nn_persistent_error;
 998	} else if (nn->nn_sc_valid) {
 999		kref_get(&nn->nn_sc->sc_kref);
1000
1001		ret = 1;
1002		*sc_ret = nn->nn_sc;
1003		*error = 0;
1004	}
1005	spin_unlock(&nn->nn_lock);
1006
1007	return ret;
1008}
1009
1010/* Get a map of all nodes to which this node is currently connected to */
1011void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1012{
1013	struct o2net_sock_container *sc;
1014	int node, ret;
1015
1016	BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1017
1018	memset(map, 0, bytes);
1019	for (node = 0; node < O2NM_MAX_NODES; ++node) {
1020		if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1021			continue;
1022		if (!ret) {
1023			set_bit(node, map);
1024			sc_put(sc);
1025		}
1026	}
1027}
1028EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1029
1030int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1031			   size_t caller_veclen, u8 target_node, int *status)
1032{
1033	int ret = 0;
1034	struct o2net_msg *msg = NULL;
1035	size_t veclen, caller_bytes = 0;
1036	struct kvec *vec = NULL;
1037	struct o2net_sock_container *sc = NULL;
1038	struct o2net_node *nn = o2net_nn_from_num(target_node);
1039	struct o2net_status_wait nsw = {
1040		.ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1041	};
1042	struct o2net_send_tracking nst;
1043
1044	o2net_init_nst(&nst, msg_type, key, current, target_node);
1045
1046	if (o2net_wq == NULL) {
1047		mlog(0, "attempt to tx without o2netd running\n");
1048		ret = -ESRCH;
1049		goto out;
1050	}
1051
1052	if (caller_veclen == 0) {
1053		mlog(0, "bad kvec array length\n");
1054		ret = -EINVAL;
1055		goto out;
1056	}
1057
1058	caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1059	if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1060		mlog(0, "total payload len %zu too large\n", caller_bytes);
1061		ret = -EINVAL;
1062		goto out;
1063	}
1064
1065	if (target_node == o2nm_this_node()) {
1066		ret = -ELOOP;
1067		goto out;
1068	}
1069
1070	o2net_debug_add_nst(&nst);
1071
1072	o2net_set_nst_sock_time(&nst);
1073
1074	wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1075	if (ret)
1076		goto out;
1077
1078	o2net_set_nst_sock_container(&nst, sc);
1079
1080	veclen = caller_veclen + 1;
1081	vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1082	if (vec == NULL) {
1083		mlog(0, "failed to %zu element kvec!\n", veclen);
1084		ret = -ENOMEM;
1085		goto out;
1086	}
1087
1088	msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1089	if (!msg) {
1090		mlog(0, "failed to allocate a o2net_msg!\n");
1091		ret = -ENOMEM;
1092		goto out;
1093	}
1094
1095	o2net_init_msg(msg, caller_bytes, msg_type, key);
1096
1097	vec[0].iov_len = sizeof(struct o2net_msg);
1098	vec[0].iov_base = msg;
1099	memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1100
1101	ret = o2net_prep_nsw(nn, &nsw);
1102	if (ret)
1103		goto out;
1104
1105	msg->msg_num = cpu_to_be32(nsw.ns_id);
1106	o2net_set_nst_msg_id(&nst, nsw.ns_id);
1107
1108	o2net_set_nst_send_time(&nst);
1109
1110	/* finally, convert the message header to network byte-order
1111	 * and send */
1112	mutex_lock(&sc->sc_send_lock);
1113	ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1114				 sizeof(struct o2net_msg) + caller_bytes);
1115	mutex_unlock(&sc->sc_send_lock);
1116	msglog(msg, "sending returned %d\n", ret);
1117	if (ret < 0) {
1118		mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1119		goto out;
1120	}
1121
1122	/* wait on other node's handler */
1123	o2net_set_nst_status_time(&nst);
1124	wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1125
1126	o2net_update_send_stats(&nst, sc);
1127
1128	/* Note that we avoid overwriting the callers status return
1129	 * variable if a system error was reported on the other
1130	 * side. Callers beware. */
1131	ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1132	if (status && !ret)
1133		*status = nsw.ns_status;
1134
1135	mlog(0, "woken, returning system status %d, user status %d\n",
1136	     ret, nsw.ns_status);
1137out:
1138	o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1139	if (sc)
1140		sc_put(sc);
1141	kfree(vec);
1142	kfree(msg);
1143	o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1144	return ret;
1145}
1146EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1147
1148int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1149		       u8 target_node, int *status)
1150{
1151	struct kvec vec = {
1152		.iov_base = data,
1153		.iov_len = len,
1154	};
1155	return o2net_send_message_vec(msg_type, key, &vec, 1,
1156				      target_node, status);
1157}
1158EXPORT_SYMBOL_GPL(o2net_send_message);
1159
1160static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1161				   enum o2net_system_error syserr, int err)
1162{
1163	struct kvec vec = {
1164		.iov_base = hdr,
1165		.iov_len = sizeof(struct o2net_msg),
1166	};
1167
1168	BUG_ON(syserr >= O2NET_ERR_MAX);
1169
1170	/* leave other fields intact from the incoming message, msg_num
1171	 * in particular */
1172	hdr->sys_status = cpu_to_be32(syserr);
1173	hdr->status = cpu_to_be32(err);
1174	hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1175	hdr->data_len = 0;
1176
1177	msglog(hdr, "about to send status magic %d\n", err);
1178	/* hdr has been in host byteorder this whole time */
1179	return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1180}
1181
1182/* this returns -errno if the header was unknown or too large, etc.
1183 * after this is called the buffer us reused for the next message */
1184static int o2net_process_message(struct o2net_sock_container *sc,
1185				 struct o2net_msg *hdr)
1186{
1187	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1188	int ret = 0, handler_status;
1189	enum  o2net_system_error syserr;
1190	struct o2net_msg_handler *nmh = NULL;
1191	void *ret_data = NULL;
1192
1193	msglog(hdr, "processing message\n");
1194
1195	o2net_sc_postpone_idle(sc);
1196
1197	switch(be16_to_cpu(hdr->magic)) {
1198		case O2NET_MSG_STATUS_MAGIC:
1199			/* special type for returning message status */
1200			o2net_complete_nsw(nn, NULL,
1201					   be32_to_cpu(hdr->msg_num),
1202					   be32_to_cpu(hdr->sys_status),
1203					   be32_to_cpu(hdr->status));
1204			goto out;
1205		case O2NET_MSG_KEEP_REQ_MAGIC:
1206			o2net_sendpage(sc, o2net_keep_resp,
1207				       sizeof(*o2net_keep_resp));
1208			goto out;
1209		case O2NET_MSG_KEEP_RESP_MAGIC:
1210			goto out;
1211		case O2NET_MSG_MAGIC:
1212			break;
1213		default:
1214			msglog(hdr, "bad magic\n");
1215			ret = -EINVAL;
1216			goto out;
1217			break;
1218	}
1219
1220	/* find a handler for it */
1221	handler_status = 0;
1222	nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1223				be32_to_cpu(hdr->key));
1224	if (!nmh) {
1225		mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1226		     be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1227		syserr = O2NET_ERR_NO_HNDLR;
1228		goto out_respond;
1229	}
1230
1231	syserr = O2NET_ERR_NONE;
1232
1233	if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1234		syserr = O2NET_ERR_OVERFLOW;
1235
1236	if (syserr != O2NET_ERR_NONE)
1237		goto out_respond;
1238
1239	o2net_set_func_start_time(sc);
1240	sc->sc_msg_key = be32_to_cpu(hdr->key);
1241	sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1242	handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1243					     be16_to_cpu(hdr->data_len),
1244					nmh->nh_func_data, &ret_data);
1245	o2net_set_func_stop_time(sc);
1246
1247	o2net_update_recv_stats(sc);
1248
1249out_respond:
1250	/* this destroys the hdr, so don't use it after this */
1251	mutex_lock(&sc->sc_send_lock);
1252	ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1253				      handler_status);
1254	mutex_unlock(&sc->sc_send_lock);
1255	hdr = NULL;
1256	mlog(0, "sending handler status %d, syserr %d returned %d\n",
1257	     handler_status, syserr, ret);
1258
1259	if (nmh) {
1260		BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1261		if (nmh->nh_post_func)
1262			(nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1263					    ret_data);
1264	}
1265
1266out:
1267	if (nmh)
1268		o2net_handler_put(nmh);
1269	return ret;
1270}
1271
1272static int o2net_check_handshake(struct o2net_sock_container *sc)
1273{
1274	struct o2net_handshake *hand = page_address(sc->sc_page);
1275	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1276
1277	if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1278		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1279		       "protocol version %llu but %llu is required. "
1280		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1281		       (unsigned long long)be64_to_cpu(hand->protocol_version),
1282		       O2NET_PROTOCOL_VERSION);
1283
1284		/* don't bother reconnecting if its the wrong version. */
1285		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1286		return -1;
1287	}
1288
1289	/*
1290	 * Ensure timeouts are consistent with other nodes, otherwise
1291	 * we can end up with one node thinking that the other must be down,
1292	 * but isn't. This can ultimately cause corruption.
1293	 */
1294	if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1295				o2net_idle_timeout()) {
1296		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1297		       "idle timeout of %u ms, but we use %u ms locally. "
1298		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1299		       be32_to_cpu(hand->o2net_idle_timeout_ms),
1300		       o2net_idle_timeout());
1301		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1302		return -1;
1303	}
1304
1305	if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1306			o2net_keepalive_delay()) {
1307		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1308		       "delay of %u ms, but we use %u ms locally. "
1309		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1310		       be32_to_cpu(hand->o2net_keepalive_delay_ms),
1311		       o2net_keepalive_delay());
1312		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1313		return -1;
1314	}
1315
1316	if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1317			O2HB_MAX_WRITE_TIMEOUT_MS) {
1318		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1319		       "timeout of %u ms, but we use %u ms locally. "
1320		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1321		       be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1322		       O2HB_MAX_WRITE_TIMEOUT_MS);
1323		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1324		return -1;
1325	}
1326
1327	sc->sc_handshake_ok = 1;
1328
1329	spin_lock(&nn->nn_lock);
1330	/* set valid and queue the idle timers only if it hasn't been
1331	 * shut down already */
1332	if (nn->nn_sc == sc) {
1333		o2net_sc_reset_idle_timer(sc);
1334		atomic_set(&nn->nn_timeout, 0);
1335		o2net_set_nn_state(nn, sc, 1, 0);
1336	}
1337	spin_unlock(&nn->nn_lock);
1338
1339	/* shift everything up as though it wasn't there */
1340	sc->sc_page_off -= sizeof(struct o2net_handshake);
1341	if (sc->sc_page_off)
1342		memmove(hand, hand + 1, sc->sc_page_off);
1343
1344	return 0;
1345}
1346
1347/* this demuxes the queued rx bytes into header or payload bits and calls
1348 * handlers as each full message is read off the socket.  it returns -error,
1349 * == 0 eof, or > 0 for progress made.*/
1350static int o2net_advance_rx(struct o2net_sock_container *sc)
1351{
1352	struct o2net_msg *hdr;
1353	int ret = 0;
1354	void *data;
1355	size_t datalen;
1356
1357	sclog(sc, "receiving\n");
1358	o2net_set_advance_start_time(sc);
1359
1360	if (unlikely(sc->sc_handshake_ok == 0)) {
1361		if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1362			data = page_address(sc->sc_page) + sc->sc_page_off;
1363			datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1364			ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1365			if (ret > 0)
1366				sc->sc_page_off += ret;
1367		}
1368
1369		if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1370			o2net_check_handshake(sc);
1371			if (unlikely(sc->sc_handshake_ok == 0))
1372				ret = -EPROTO;
1373		}
1374		goto out;
1375	}
1376
1377	/* do we need more header? */
1378	if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1379		data = page_address(sc->sc_page) + sc->sc_page_off;
1380		datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1381		ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1382		if (ret > 0) {
1383			sc->sc_page_off += ret;
1384			/* only swab incoming here.. we can
1385			 * only get here once as we cross from
1386			 * being under to over */
1387			if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1388				hdr = page_address(sc->sc_page);
1389				if (be16_to_cpu(hdr->data_len) >
1390				    O2NET_MAX_PAYLOAD_BYTES)
1391					ret = -EOVERFLOW;
1392			}
1393		}
1394		if (ret <= 0)
1395			goto out;
1396	}
1397
1398	if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1399		/* oof, still don't have a header */
1400		goto out;
1401	}
1402
1403	/* this was swabbed above when we first read it */
1404	hdr = page_address(sc->sc_page);
1405
1406	msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1407
1408	/* do we need more payload? */
1409	if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1410		/* need more payload */
1411		data = page_address(sc->sc_page) + sc->sc_page_off;
1412		datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1413			  sc->sc_page_off;
1414		ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1415		if (ret > 0)
1416			sc->sc_page_off += ret;
1417		if (ret <= 0)
1418			goto out;
1419	}
1420
1421	if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1422		/* we can only get here once, the first time we read
1423		 * the payload.. so set ret to progress if the handler
1424		 * works out. after calling this the message is toast */
1425		ret = o2net_process_message(sc, hdr);
1426		if (ret == 0)
1427			ret = 1;
1428		sc->sc_page_off = 0;
1429	}
1430
1431out:
1432	sclog(sc, "ret = %d\n", ret);
1433	o2net_set_advance_stop_time(sc);
1434	return ret;
1435}
1436
1437/* this work func is triggerd by data ready.  it reads until it can read no
1438 * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1439 * our work the work struct will be marked and we'll be called again. */
1440static void o2net_rx_until_empty(struct work_struct *work)
1441{
1442	struct o2net_sock_container *sc =
1443		container_of(work, struct o2net_sock_container, sc_rx_work);
1444	int ret;
1445
1446	do {
1447		ret = o2net_advance_rx(sc);
1448	} while (ret > 0);
1449
1450	if (ret <= 0 && ret != -EAGAIN) {
1451		struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1452		sclog(sc, "saw error %d, closing\n", ret);
1453		/* not permanent so read failed handshake can retry */
1454		o2net_ensure_shutdown(nn, sc, 0);
1455	}
1456
1457	sc_put(sc);
1458}
1459
1460static int o2net_set_nodelay(struct socket *sock)
1461{
1462	int ret, val = 1;
1463	mm_segment_t oldfs;
1464
1465	oldfs = get_fs();
1466	set_fs(KERNEL_DS);
1467
1468	/*
1469	 * Dear unsuspecting programmer,
1470	 *
1471	 * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1472	 * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1473	 * silently turn into SO_DEBUG.
1474	 *
1475	 * Yours,
1476	 * Keeper of hilariously fragile interfaces.
1477	 */
1478	ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1479				    (char __user *)&val, sizeof(val));
1480
1481	set_fs(oldfs);
1482	return ret;
1483}
1484
1485static int o2net_set_usertimeout(struct socket *sock)
1486{
1487	int user_timeout = O2NET_TCP_USER_TIMEOUT;
1488
1489	return kernel_setsockopt(sock, SOL_TCP, TCP_USER_TIMEOUT,
1490				(char *)&user_timeout, sizeof(user_timeout));
1491}
1492
1493static void o2net_initialize_handshake(void)
1494{
1495	o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1496		O2HB_MAX_WRITE_TIMEOUT_MS);
1497	o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1498	o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1499		o2net_keepalive_delay());
1500	o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1501		o2net_reconnect_delay());
1502}
1503
1504/* ------------------------------------------------------------ */
1505
1506/* called when a connect completes and after a sock is accepted.  the
1507 * rx path will see the response and mark the sc valid */
1508static void o2net_sc_connect_completed(struct work_struct *work)
1509{
1510	struct o2net_sock_container *sc =
1511		container_of(work, struct o2net_sock_container,
1512			     sc_connect_work);
1513
1514	mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1515              (unsigned long long)O2NET_PROTOCOL_VERSION,
1516	      (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1517
1518	o2net_initialize_handshake();
1519	o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1520	sc_put(sc);
1521}
1522
1523/* this is called as a work_struct func. */
1524static void o2net_sc_send_keep_req(struct work_struct *work)
1525{
1526	struct o2net_sock_container *sc =
1527		container_of(work, struct o2net_sock_container,
1528			     sc_keepalive_work.work);
1529
1530	o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1531	sc_put(sc);
1532}
1533
1534/* socket shutdown does a del_timer_sync against this as it tears down.
1535 * we can't start this timer until we've got to the point in sc buildup
1536 * where shutdown is going to be involved */
1537static void o2net_idle_timer(unsigned long data)
1538{
1539	struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1540	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1541#ifdef CONFIG_DEBUG_FS
1542	unsigned long msecs = ktime_to_ms(ktime_get()) -
1543		ktime_to_ms(sc->sc_tv_timer);
1544#else
1545	unsigned long msecs = o2net_idle_timeout();
1546#endif
1547
1548	printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1549	       "idle for %lu.%lu secs.\n",
1550	       SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1551
1552	/* idle timerout happen, don't shutdown the connection, but
1553	 * make fence decision. Maybe the connection can recover before
1554	 * the decision is made.
1555	 */
1556	atomic_set(&nn->nn_timeout, 1);
1557	o2quo_conn_err(o2net_num_from_nn(nn));
1558	queue_delayed_work(o2net_wq, &nn->nn_still_up,
1559			msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1560
1561	o2net_sc_reset_idle_timer(sc);
1562
1563}
1564
1565static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1566{
1567	o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1568	o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1569		      msecs_to_jiffies(o2net_keepalive_delay()));
1570	o2net_set_sock_timer(sc);
1571	mod_timer(&sc->sc_idle_timeout,
1572	       jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1573}
1574
1575static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1576{
1577	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1578
1579	/* clear fence decision since the connection recover from timeout*/
1580	if (atomic_read(&nn->nn_timeout)) {
1581		o2quo_conn_up(o2net_num_from_nn(nn));
1582		cancel_delayed_work(&nn->nn_still_up);
1583		atomic_set(&nn->nn_timeout, 0);
1584	}
1585
1586	/* Only push out an existing timer */
1587	if (timer_pending(&sc->sc_idle_timeout))
1588		o2net_sc_reset_idle_timer(sc);
1589}
1590
1591/* this work func is kicked whenever a path sets the nn state which doesn't
1592 * have valid set.  This includes seeing hb come up, losing a connection,
1593 * having a connect attempt fail, etc. This centralizes the logic which decides
1594 * if a connect attempt should be made or if we should give up and all future
1595 * transmit attempts should fail */
1596static void o2net_start_connect(struct work_struct *work)
1597{
1598	struct o2net_node *nn =
1599		container_of(work, struct o2net_node, nn_connect_work.work);
1600	struct o2net_sock_container *sc = NULL;
1601	struct o2nm_node *node = NULL, *mynode = NULL;
1602	struct socket *sock = NULL;
1603	struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1604	int ret = 0, stop;
1605	unsigned int timeout;
1606	unsigned int noio_flag;
1607
1608	/*
1609	 * sock_create allocates the sock with GFP_KERNEL. We must set
1610	 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1611	 * by this process are done as if GFP_NOIO was specified. So we
1612	 * are not reentering filesystem while doing memory reclaim.
1613	 */
1614	noio_flag = memalloc_noio_save();
1615	/* if we're greater we initiate tx, otherwise we accept */
1616	if (o2nm_this_node() <= o2net_num_from_nn(nn))
1617		goto out;
1618
1619	/* watch for racing with tearing a node down */
1620	node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1621	if (node == NULL)
 
1622		goto out;
 
1623
1624	mynode = o2nm_get_node_by_num(o2nm_this_node());
1625	if (mynode == NULL)
 
1626		goto out;
 
1627
1628	spin_lock(&nn->nn_lock);
1629	/*
1630	 * see if we already have one pending or have given up.
1631	 * For nn_timeout, it is set when we close the connection
1632	 * because of the idle time out. So it means that we have
1633	 * at least connected to that node successfully once,
1634	 * now try to connect to it again.
1635	 */
1636	timeout = atomic_read(&nn->nn_timeout);
1637	stop = (nn->nn_sc ||
1638		(nn->nn_persistent_error &&
1639		(nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1640	spin_unlock(&nn->nn_lock);
1641	if (stop)
1642		goto out;
1643
1644	nn->nn_last_connect_attempt = jiffies;
1645
1646	sc = sc_alloc(node);
1647	if (sc == NULL) {
1648		mlog(0, "couldn't allocate sc\n");
1649		ret = -ENOMEM;
1650		goto out;
1651	}
1652
1653	ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1654	if (ret < 0) {
1655		mlog(0, "can't create socket: %d\n", ret);
1656		goto out;
1657	}
1658	sc->sc_sock = sock; /* freed by sc_kref_release */
1659
1660	sock->sk->sk_allocation = GFP_ATOMIC;
1661
1662	myaddr.sin_family = AF_INET;
1663	myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1664	myaddr.sin_port = htons(0); /* any port */
1665
1666	ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1667			      sizeof(myaddr));
1668	if (ret) {
1669		mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1670		     ret, &mynode->nd_ipv4_address);
1671		goto out;
1672	}
1673
1674	ret = o2net_set_nodelay(sc->sc_sock);
1675	if (ret) {
1676		mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1677		goto out;
1678	}
1679
1680	ret = o2net_set_usertimeout(sock);
1681	if (ret) {
1682		mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1683		goto out;
1684	}
1685
1686	o2net_register_callbacks(sc->sc_sock->sk, sc);
1687
1688	spin_lock(&nn->nn_lock);
1689	/* handshake completion will set nn->nn_sc_valid */
1690	o2net_set_nn_state(nn, sc, 0, 0);
1691	spin_unlock(&nn->nn_lock);
1692
1693	remoteaddr.sin_family = AF_INET;
1694	remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1695	remoteaddr.sin_port = node->nd_ipv4_port;
1696
1697	ret = sc->sc_sock->ops->connect(sc->sc_sock,
1698					(struct sockaddr *)&remoteaddr,
1699					sizeof(remoteaddr),
1700					O_NONBLOCK);
1701	if (ret == -EINPROGRESS)
1702		ret = 0;
1703
1704out:
1705	if (ret && sc) {
1706		printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1707		       " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1708		/* 0 err so that another will be queued and attempted
1709		 * from set_nn_state */
1710		o2net_ensure_shutdown(nn, sc, 0);
1711	}
1712	if (sc)
1713		sc_put(sc);
1714	if (node)
1715		o2nm_node_put(node);
1716	if (mynode)
1717		o2nm_node_put(mynode);
1718
1719	memalloc_noio_restore(noio_flag);
1720	return;
1721}
1722
1723static void o2net_connect_expired(struct work_struct *work)
1724{
1725	struct o2net_node *nn =
1726		container_of(work, struct o2net_node, nn_connect_expired.work);
1727
1728	spin_lock(&nn->nn_lock);
1729	if (!nn->nn_sc_valid) {
1730		printk(KERN_NOTICE "o2net: No connection established with "
1731		       "node %u after %u.%u seconds, check network and"
1732		       " cluster configuration.\n",
1733		     o2net_num_from_nn(nn),
1734		     o2net_idle_timeout() / 1000,
1735		     o2net_idle_timeout() % 1000);
1736
1737		o2net_set_nn_state(nn, NULL, 0, 0);
1738	}
1739	spin_unlock(&nn->nn_lock);
1740}
1741
1742static void o2net_still_up(struct work_struct *work)
1743{
1744	struct o2net_node *nn =
1745		container_of(work, struct o2net_node, nn_still_up.work);
1746
1747	o2quo_hb_still_up(o2net_num_from_nn(nn));
1748}
1749
1750/* ------------------------------------------------------------ */
1751
1752void o2net_disconnect_node(struct o2nm_node *node)
1753{
1754	struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1755
1756	/* don't reconnect until it's heartbeating again */
1757	spin_lock(&nn->nn_lock);
1758	atomic_set(&nn->nn_timeout, 0);
1759	o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1760	spin_unlock(&nn->nn_lock);
1761
1762	if (o2net_wq) {
1763		cancel_delayed_work(&nn->nn_connect_expired);
1764		cancel_delayed_work(&nn->nn_connect_work);
1765		cancel_delayed_work(&nn->nn_still_up);
1766		flush_workqueue(o2net_wq);
1767	}
1768}
1769
1770static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1771				  void *data)
1772{
1773	o2quo_hb_down(node_num);
1774
1775	if (!node)
1776		return;
1777
1778	if (node_num != o2nm_this_node())
1779		o2net_disconnect_node(node);
1780
1781	BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1782}
1783
1784static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1785				void *data)
1786{
1787	struct o2net_node *nn = o2net_nn_from_num(node_num);
1788
1789	o2quo_hb_up(node_num);
1790
1791	BUG_ON(!node);
1792
1793	/* ensure an immediate connect attempt */
1794	nn->nn_last_connect_attempt = jiffies -
1795		(msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1796
1797	if (node_num != o2nm_this_node()) {
1798		/* believe it or not, accept and node hearbeating testing
1799		 * can succeed for this node before we got here.. so
1800		 * only use set_nn_state to clear the persistent error
1801		 * if that hasn't already happened */
1802		spin_lock(&nn->nn_lock);
1803		atomic_set(&nn->nn_timeout, 0);
1804		if (nn->nn_persistent_error)
1805			o2net_set_nn_state(nn, NULL, 0, 0);
1806		spin_unlock(&nn->nn_lock);
1807	}
1808}
1809
1810void o2net_unregister_hb_callbacks(void)
1811{
1812	o2hb_unregister_callback(NULL, &o2net_hb_up);
1813	o2hb_unregister_callback(NULL, &o2net_hb_down);
1814}
1815
1816int o2net_register_hb_callbacks(void)
1817{
1818	int ret;
1819
1820	o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1821			    o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1822	o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1823			    o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1824
1825	ret = o2hb_register_callback(NULL, &o2net_hb_up);
1826	if (ret == 0)
1827		ret = o2hb_register_callback(NULL, &o2net_hb_down);
1828
1829	if (ret)
1830		o2net_unregister_hb_callbacks();
1831
1832	return ret;
1833}
1834
1835/* ------------------------------------------------------------ */
1836
1837static int o2net_accept_one(struct socket *sock, int *more)
1838{
1839	int ret, slen;
1840	struct sockaddr_in sin;
1841	struct socket *new_sock = NULL;
1842	struct o2nm_node *node = NULL;
1843	struct o2nm_node *local_node = NULL;
1844	struct o2net_sock_container *sc = NULL;
1845	struct o2net_node *nn;
1846	unsigned int noio_flag;
1847
1848	/*
1849	 * sock_create_lite allocates the sock with GFP_KERNEL. We must set
1850	 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1851	 * by this process are done as if GFP_NOIO was specified. So we
1852	 * are not reentering filesystem while doing memory reclaim.
1853	 */
1854	noio_flag = memalloc_noio_save();
1855
1856	BUG_ON(sock == NULL);
1857	*more = 0;
1858	ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1859			       sock->sk->sk_protocol, &new_sock);
1860	if (ret)
1861		goto out;
1862
1863	new_sock->type = sock->type;
1864	new_sock->ops = sock->ops;
1865	ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1866	if (ret < 0)
1867		goto out;
1868
1869	*more = 1;
1870	new_sock->sk->sk_allocation = GFP_ATOMIC;
1871
1872	ret = o2net_set_nodelay(new_sock);
1873	if (ret) {
1874		mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1875		goto out;
1876	}
1877
1878	ret = o2net_set_usertimeout(new_sock);
1879	if (ret) {
1880		mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1881		goto out;
1882	}
1883
1884	slen = sizeof(sin);
1885	ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1886				       &slen, 1);
1887	if (ret < 0)
1888		goto out;
1889
1890	node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1891	if (node == NULL) {
1892		printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1893		       "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1894		       ntohs(sin.sin_port));
1895		ret = -EINVAL;
1896		goto out;
1897	}
1898
1899	if (o2nm_this_node() >= node->nd_num) {
1900		local_node = o2nm_get_node_by_num(o2nm_this_node());
1901		if (local_node)
1902			printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1903					"seen at node '%s' (%u, %pI4:%d) from "
1904					"node '%s' (%u, %pI4:%d)\n",
1905					local_node->nd_name, local_node->nd_num,
1906					&(local_node->nd_ipv4_address),
1907					ntohs(local_node->nd_ipv4_port),
1908					node->nd_name,
1909					node->nd_num, &sin.sin_addr.s_addr,
1910					ntohs(sin.sin_port));
1911		ret = -EINVAL;
1912		goto out;
1913	}
1914
1915	/* this happens all the time when the other node sees our heartbeat
1916	 * and tries to connect before we see their heartbeat */
1917	if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1918		mlog(ML_CONN, "attempt to connect from node '%s' at "
1919		     "%pI4:%d but it isn't heartbeating\n",
1920		     node->nd_name, &sin.sin_addr.s_addr,
1921		     ntohs(sin.sin_port));
1922		ret = -EINVAL;
1923		goto out;
1924	}
1925
1926	nn = o2net_nn_from_num(node->nd_num);
1927
1928	spin_lock(&nn->nn_lock);
1929	if (nn->nn_sc)
1930		ret = -EBUSY;
1931	else
1932		ret = 0;
1933	spin_unlock(&nn->nn_lock);
1934	if (ret) {
1935		printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1936		       "at %pI4:%d but it already has an open connection\n",
1937		       node->nd_name, &sin.sin_addr.s_addr,
1938		       ntohs(sin.sin_port));
1939		goto out;
1940	}
1941
1942	sc = sc_alloc(node);
1943	if (sc == NULL) {
1944		ret = -ENOMEM;
1945		goto out;
1946	}
1947
1948	sc->sc_sock = new_sock;
1949	new_sock = NULL;
1950
1951	spin_lock(&nn->nn_lock);
1952	atomic_set(&nn->nn_timeout, 0);
1953	o2net_set_nn_state(nn, sc, 0, 0);
1954	spin_unlock(&nn->nn_lock);
1955
1956	o2net_register_callbacks(sc->sc_sock->sk, sc);
1957	o2net_sc_queue_work(sc, &sc->sc_rx_work);
1958
1959	o2net_initialize_handshake();
1960	o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1961
1962out:
1963	if (new_sock)
1964		sock_release(new_sock);
1965	if (node)
1966		o2nm_node_put(node);
1967	if (local_node)
1968		o2nm_node_put(local_node);
1969	if (sc)
1970		sc_put(sc);
1971
1972	memalloc_noio_restore(noio_flag);
1973	return ret;
1974}
1975
1976/*
1977 * This function is invoked in response to one or more
1978 * pending accepts at softIRQ level. We must drain the
1979 * entire que before returning.
1980 */
1981
1982static void o2net_accept_many(struct work_struct *work)
1983{
1984	struct socket *sock = o2net_listen_sock;
1985	int	more;
1986	int	err;
1987
1988	/*
1989	 * It is critical to note that due to interrupt moderation
1990	 * at the network driver level, we can't assume to get a
1991	 * softIRQ for every single conn since tcp SYN packets
1992	 * can arrive back-to-back, and therefore many pending
1993	 * accepts may result in just 1 softIRQ. If we terminate
1994	 * the o2net_accept_one() loop upon seeing an err, what happens
1995	 * to the rest of the conns in the queue? If no new SYN
1996	 * arrives for hours, no softIRQ  will be delivered,
1997	 * and the connections will just sit in the queue.
1998	 */
1999
2000	for (;;) {
2001		err = o2net_accept_one(sock, &more);
2002		if (!more)
2003			break;
2004		cond_resched();
2005	}
2006}
2007
2008static void o2net_listen_data_ready(struct sock *sk)
2009{
2010	void (*ready)(struct sock *sk);
2011
2012	read_lock_bh(&sk->sk_callback_lock);
2013	ready = sk->sk_user_data;
2014	if (ready == NULL) { /* check for teardown race */
2015		ready = sk->sk_data_ready;
2016		goto out;
2017	}
2018
2019	/* This callback may called twice when a new connection
2020	 * is  being established as a child socket inherits everything
2021	 * from a parent LISTEN socket, including the data_ready cb of
2022	 * the parent. This leads to a hazard. In o2net_accept_one()
2023	 * we are still initializing the child socket but have not
2024	 * changed the inherited data_ready callback yet when
2025	 * data starts arriving.
2026	 * We avoid this hazard by checking the state.
2027	 * For the listening socket,  the state will be TCP_LISTEN; for the new
2028	 * socket, will be  TCP_ESTABLISHED. Also, in this case,
2029	 * sk->sk_user_data is not a valid function pointer.
2030	 */
2031
2032	if (sk->sk_state == TCP_LISTEN) {
2033		queue_work(o2net_wq, &o2net_listen_work);
2034	} else {
2035		ready = NULL;
2036	}
2037
2038out:
2039	read_unlock_bh(&sk->sk_callback_lock);
2040	if (ready != NULL)
2041		ready(sk);
2042}
2043
2044static int o2net_open_listening_sock(__be32 addr, __be16 port)
2045{
2046	struct socket *sock = NULL;
2047	int ret;
2048	struct sockaddr_in sin = {
2049		.sin_family = PF_INET,
2050		.sin_addr = { .s_addr = addr },
2051		.sin_port = port,
2052	};
2053
2054	ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2055	if (ret < 0) {
2056		printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
2057		goto out;
2058	}
2059
2060	sock->sk->sk_allocation = GFP_ATOMIC;
2061
2062	write_lock_bh(&sock->sk->sk_callback_lock);
2063	sock->sk->sk_user_data = sock->sk->sk_data_ready;
2064	sock->sk->sk_data_ready = o2net_listen_data_ready;
2065	write_unlock_bh(&sock->sk->sk_callback_lock);
2066
2067	o2net_listen_sock = sock;
2068	INIT_WORK(&o2net_listen_work, o2net_accept_many);
2069
2070	sock->sk->sk_reuse = SK_CAN_REUSE;
2071	ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2072	if (ret < 0) {
2073		printk(KERN_ERR "o2net: Error %d while binding socket at "
2074		       "%pI4:%u\n", ret, &addr, ntohs(port)); 
2075		goto out;
2076	}
2077
2078	ret = sock->ops->listen(sock, 64);
2079	if (ret < 0)
2080		printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2081		       ret, &addr, ntohs(port));
2082
2083out:
2084	if (ret) {
2085		o2net_listen_sock = NULL;
2086		if (sock)
2087			sock_release(sock);
2088	}
2089	return ret;
2090}
2091
2092/*
2093 * called from node manager when we should bring up our network listening
2094 * socket.  node manager handles all the serialization to only call this
2095 * once and to match it with o2net_stop_listening().  note,
2096 * o2nm_this_node() doesn't work yet as we're being called while it
2097 * is being set up.
2098 */
2099int o2net_start_listening(struct o2nm_node *node)
2100{
2101	int ret = 0;
2102
2103	BUG_ON(o2net_wq != NULL);
2104	BUG_ON(o2net_listen_sock != NULL);
2105
2106	mlog(ML_KTHREAD, "starting o2net thread...\n");
2107	o2net_wq = alloc_ordered_workqueue("o2net", WQ_MEM_RECLAIM);
2108	if (o2net_wq == NULL) {
2109		mlog(ML_ERROR, "unable to launch o2net thread\n");
2110		return -ENOMEM; /* ? */
2111	}
2112
2113	ret = o2net_open_listening_sock(node->nd_ipv4_address,
2114					node->nd_ipv4_port);
2115	if (ret) {
2116		destroy_workqueue(o2net_wq);
2117		o2net_wq = NULL;
2118	} else
2119		o2quo_conn_up(node->nd_num);
2120
2121	return ret;
2122}
2123
2124/* again, o2nm_this_node() doesn't work here as we're involved in
2125 * tearing it down */
2126void o2net_stop_listening(struct o2nm_node *node)
2127{
2128	struct socket *sock = o2net_listen_sock;
2129	size_t i;
2130
2131	BUG_ON(o2net_wq == NULL);
2132	BUG_ON(o2net_listen_sock == NULL);
2133
2134	/* stop the listening socket from generating work */
2135	write_lock_bh(&sock->sk->sk_callback_lock);
2136	sock->sk->sk_data_ready = sock->sk->sk_user_data;
2137	sock->sk->sk_user_data = NULL;
2138	write_unlock_bh(&sock->sk->sk_callback_lock);
2139
2140	for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2141		struct o2nm_node *node = o2nm_get_node_by_num(i);
2142		if (node) {
2143			o2net_disconnect_node(node);
2144			o2nm_node_put(node);
2145		}
2146	}
2147
2148	/* finish all work and tear down the work queue */
2149	mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2150	destroy_workqueue(o2net_wq);
2151	o2net_wq = NULL;
2152
2153	sock_release(o2net_listen_sock);
2154	o2net_listen_sock = NULL;
2155
2156	o2quo_conn_err(node->nd_num);
2157}
2158
2159/* ------------------------------------------------------------ */
2160
2161int o2net_init(void)
2162{
2163	unsigned long i;
2164
2165	o2quo_init();
2166
2167	if (o2net_debugfs_init())
2168		goto out;
2169
2170	o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2171	o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2172	o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2173	if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp)
2174		goto out;
2175
2176	o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2177	o2net_hand->connector_id = cpu_to_be64(1);
2178
2179	o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2180	o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2181
2182	for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2183		struct o2net_node *nn = o2net_nn_from_num(i);
2184
2185		atomic_set(&nn->nn_timeout, 0);
2186		spin_lock_init(&nn->nn_lock);
2187		INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2188		INIT_DELAYED_WORK(&nn->nn_connect_expired,
2189				  o2net_connect_expired);
2190		INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2191		/* until we see hb from a node we'll return einval */
2192		nn->nn_persistent_error = -ENOTCONN;
2193		init_waitqueue_head(&nn->nn_sc_wq);
2194		idr_init(&nn->nn_status_idr);
2195		INIT_LIST_HEAD(&nn->nn_status_list);
2196	}
2197
2198	return 0;
2199
2200out:
2201	kfree(o2net_hand);
2202	kfree(o2net_keep_req);
2203	kfree(o2net_keep_resp);
2204	o2net_debugfs_exit();
2205	o2quo_exit();
2206	return -ENOMEM;
2207}
2208
2209void o2net_exit(void)
2210{
2211	o2quo_exit();
2212	kfree(o2net_hand);
2213	kfree(o2net_keep_req);
2214	kfree(o2net_keep_resp);
2215	o2net_debugfs_exit();
2216}