1/* -*- mode: c; c-basic-offset: 8; -*-
   2 *
   3 * vim: noexpandtab sw=8 ts=8 sts=0:
   4 *
   5 * Copyright (C) 2004 Oracle.  All rights reserved.
   6 *
   7 * This program is free software; you can redistribute it and/or
   8 * modify it under the terms of the GNU General Public
   9 * License as published by the Free Software Foundation; either
  10 * version 2 of the License, or (at your option) any later version.
  11 *
  12 * This program is distributed in the hope that it will be useful,
  13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15 * General Public License for more details.
  16 *
  17 * You should have received a copy of the GNU General Public
  18 * License along with this program; if not, write to the
  19 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  20 * Boston, MA 021110-1307, USA.
  21 *
  22 * ----
  23 *
  24 * Callers for this were originally written against a very simple synchronus
  25 * API.  This implementation reflects those simple callers.  Some day I'm sure
  26 * we'll need to move to a more robust posting/callback mechanism.
  27 *
  28 * Transmit calls pass in kernel virtual addresses and block copying this into
  29 * the socket's tx buffers via a usual blocking sendmsg.  They'll block waiting
  30 * for a failed socket to timeout.  TX callers can also pass in a poniter to an
  31 * 'int' which gets filled with an errno off the wire in response to the
  32 * message they send.
  33 *
  34 * Handlers for unsolicited messages are registered.  Each socket has a page
  35 * that incoming data is copied into.  First the header, then the data.
  36 * Handlers are called from only one thread with a reference to this per-socket
  37 * page.  This page is destroyed after the handler call, so it can't be
  38 * referenced beyond the call.  Handlers may block but are discouraged from
  39 * doing so.
  40 *
  41 * Any framing errors (bad magic, large payload lengths) close a connection.
  42 *
  43 * Our sock_container holds the state we associate with a socket.  It's current
  44 * framing state is held there as well as the refcounting we do around when it
  45 * is safe to tear down the socket.  The socket is only finally torn down from
  46 * the container when the container loses all of its references -- so as long
  47 * as you hold a ref on the container you can trust that the socket is valid
  48 * for use with kernel socket APIs.
  49 *
  50 * Connections are initiated between a pair of nodes when the node with the
  51 * higher node number gets a heartbeat callback which indicates that the lower
  52 * numbered node has started heartbeating.  The lower numbered node is passive
  53 * and only accepts the connection if the higher numbered node is heartbeating.
  54 */
  55
  56#include <linux/kernel.h>
 
  57#include <linux/jiffies.h>
  58#include <linux/slab.h>
  59#include <linux/idr.h>
  60#include <linux/kref.h>
  61#include <linux/net.h>
  62#include <linux/export.h>
  63#include <net/tcp.h>
 
  64
  65#include <asm/uaccess.h>
  66
  67#include "heartbeat.h"
  68#include "tcp.h"
  69#include "nodemanager.h"
  70#define MLOG_MASK_PREFIX ML_TCP
  71#include "masklog.h"
  72#include "quorum.h"
  73
  74#include "tcp_internal.h"
  75
  76#define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
  77#define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,	\
  78			  &sc->sc_node->nd_ipv4_address,		\
  79			  ntohs(sc->sc_node->nd_ipv4_port)
  80
  81/*
  82 * In the following two log macros, the whitespace after the ',' just
  83 * before ##args is intentional. Otherwise, gcc 2.95 will eat the
  84 * previous token if args expands to nothing.
  85 */
  86#define msglog(hdr, fmt, args...) do {					\
  87	typeof(hdr) __hdr = (hdr);					\
  88	mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "	\
  89	     "key %08x num %u] " fmt,					\
  90	     be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len), 	\
  91	     be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),	\
  92	     be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),	\
  93	     be32_to_cpu(__hdr->msg_num) ,  ##args);			\
  94} while (0)
  95
  96#define sclog(sc, fmt, args...) do {					\
  97	typeof(sc) __sc = (sc);						\
  98	mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "	\
  99	     "pg_off %zu] " fmt, __sc,					\
 100	     atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,	\
 101	    __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,	\
 102	    ##args);							\
 103} while (0)
 104
 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
 604	read_lock(&sk->sk_callback_lock);
 605	if (sk->sk_user_data) {
 606		struct o2net_sock_container *sc = sk->sk_user_data;
 607		sclog(sc, "data_ready hit\n");
 608		o2net_set_data_ready_time(sc);
 609		o2net_sc_queue_work(sc, &sc->sc_rx_work);
 610		ready = sc->sc_data_ready;
 611	} else {
 612		ready = sk->sk_data_ready;
 613	}
 614	read_unlock(&sk->sk_callback_lock);
 615
 616	ready(sk);
 617}
 618
 619/* see o2net_register_callbacks() */
 620static void o2net_state_change(struct sock *sk)
 621{
 622	void (*state_change)(struct sock *sk);
 623	struct o2net_sock_container *sc;
 624
 625	read_lock(&sk->sk_callback_lock);
 626	sc = sk->sk_user_data;
 627	if (sc == NULL) {
 628		state_change = sk->sk_state_change;
 629		goto out;
 630	}
 631
 632	sclog(sc, "state_change to %d\n", sk->sk_state);
 633
 634	state_change = sc->sc_state_change;
 635
 636	switch(sk->sk_state) {
 637	/* ignore connecting sockets as they make progress */
 638	case TCP_SYN_SENT:
 639	case TCP_SYN_RECV:
 640		break;
 641	case TCP_ESTABLISHED:
 642		o2net_sc_queue_work(sc, &sc->sc_connect_work);
 643		break;
 644	default:
 645		printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
 646			" shutdown, state %d\n",
 647			SC_NODEF_ARGS(sc), sk->sk_state);
 648		o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
 649		break;
 650	}
 651out:
 652	read_unlock(&sk->sk_callback_lock);
 653	state_change(sk);
 654}
 655
 656/*
 657 * we register callbacks so we can queue work on events before calling
 658 * the original callbacks.  our callbacks our careful to test user_data
 659 * to discover when they've reaced with o2net_unregister_callbacks().
 660 */
 661static void o2net_register_callbacks(struct sock *sk,
 662				     struct o2net_sock_container *sc)
 663{
 664	write_lock_bh(&sk->sk_callback_lock);
 665
 666	/* accepted sockets inherit the old listen socket data ready */
 667	if (sk->sk_data_ready == o2net_listen_data_ready) {
 668		sk->sk_data_ready = sk->sk_user_data;
 669		sk->sk_user_data = NULL;
 670	}
 671
 672	BUG_ON(sk->sk_user_data != NULL);
 673	sk->sk_user_data = sc;
 674	sc_get(sc);
 675
 676	sc->sc_data_ready = sk->sk_data_ready;
 677	sc->sc_state_change = sk->sk_state_change;
 678	sk->sk_data_ready = o2net_data_ready;
 679	sk->sk_state_change = o2net_state_change;
 680
 681	mutex_init(&sc->sc_send_lock);
 682
 683	write_unlock_bh(&sk->sk_callback_lock);
 684}
 685
 686static int o2net_unregister_callbacks(struct sock *sk,
 687			           struct o2net_sock_container *sc)
 688{
 689	int ret = 0;
 690
 691	write_lock_bh(&sk->sk_callback_lock);
 692	if (sk->sk_user_data == sc) {
 693		ret = 1;
 694		sk->sk_user_data = NULL;
 695		sk->sk_data_ready = sc->sc_data_ready;
 696		sk->sk_state_change = sc->sc_state_change;
 697	}
 698	write_unlock_bh(&sk->sk_callback_lock);
 699
 700	return ret;
 701}
 702
 703/*
 704 * this is a little helper that is called by callers who have seen a problem
 705 * with an sc and want to detach it from the nn if someone already hasn't beat
 706 * them to it.  if an error is given then the shutdown will be persistent
 707 * and pending transmits will be canceled.
 708 */
 709static void o2net_ensure_shutdown(struct o2net_node *nn,
 710			           struct o2net_sock_container *sc,
 711				   int err)
 712{
 713	spin_lock(&nn->nn_lock);
 714	if (nn->nn_sc == sc)
 715		o2net_set_nn_state(nn, NULL, 0, err);
 716	spin_unlock(&nn->nn_lock);
 717}
 718
 719/*
 720 * This work queue function performs the blocking parts of socket shutdown.  A
 721 * few paths lead here.  set_nn_state will trigger this callback if it sees an
 722 * sc detached from the nn.  state_change will also trigger this callback
 723 * directly when it sees errors.  In that case we need to call set_nn_state
 724 * ourselves as state_change couldn't get the nn_lock and call set_nn_state
 725 * itself.
 726 */
 727static void o2net_shutdown_sc(struct work_struct *work)
 728{
 729	struct o2net_sock_container *sc =
 730		container_of(work, struct o2net_sock_container,
 731			     sc_shutdown_work);
 732	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 733
 734	sclog(sc, "shutting down\n");
 735
 736	/* drop the callbacks ref and call shutdown only once */
 737	if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
 738		/* we shouldn't flush as we're in the thread, the
 739		 * races with pending sc work structs are harmless */
 740		del_timer_sync(&sc->sc_idle_timeout);
 741		o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
 742		sc_put(sc);
 743		kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
 744	}
 745
 746	/* not fatal so failed connects before the other guy has our
 747	 * heartbeat can be retried */
 748	o2net_ensure_shutdown(nn, sc, 0);
 749	sc_put(sc);
 750}
 751
 752/* ------------------------------------------------------------ */
 753
 754static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
 755			     u32 key)
 756{
 757	int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
 758
 759	if (ret == 0)
 760		ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
 761
 762	return ret;
 763}
 764
 765static struct o2net_msg_handler *
 766o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
 767			  struct rb_node **ret_parent)
 768{
 769	struct rb_node **p = &o2net_handler_tree.rb_node;
 770	struct rb_node *parent = NULL;
 771	struct o2net_msg_handler *nmh, *ret = NULL;
 772	int cmp;
 773
 774	while (*p) {
 775		parent = *p;
 776		nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
 777		cmp = o2net_handler_cmp(nmh, msg_type, key);
 778
 779		if (cmp < 0)
 780			p = &(*p)->rb_left;
 781		else if (cmp > 0)
 782			p = &(*p)->rb_right;
 783		else {
 784			ret = nmh;
 785			break;
 786		}
 787	}
 788
 789	if (ret_p != NULL)
 790		*ret_p = p;
 791	if (ret_parent != NULL)
 792		*ret_parent = parent;
 793
 794	return ret;
 795}
 796
 797static void o2net_handler_kref_release(struct kref *kref)
 798{
 799	struct o2net_msg_handler *nmh;
 800	nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
 801
 802	kfree(nmh);
 803}
 804
 805static void o2net_handler_put(struct o2net_msg_handler *nmh)
 806{
 807	kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 808}
 809
 810/* max_len is protection for the handler func.  incoming messages won't
 811 * be given to the handler if their payload is longer than the max. */
 812int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
 813			   o2net_msg_handler_func *func, void *data,
 814			   o2net_post_msg_handler_func *post_func,
 815			   struct list_head *unreg_list)
 816{
 817	struct o2net_msg_handler *nmh = NULL;
 818	struct rb_node **p, *parent;
 819	int ret = 0;
 820
 821	if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
 822		mlog(0, "max_len for message handler out of range: %u\n",
 823			max_len);
 824		ret = -EINVAL;
 825		goto out;
 826	}
 827
 828	if (!msg_type) {
 829		mlog(0, "no message type provided: %u, %p\n", msg_type, func);
 830		ret = -EINVAL;
 831		goto out;
 832
 833	}
 834	if (!func) {
 835		mlog(0, "no message handler provided: %u, %p\n",
 836		       msg_type, func);
 837		ret = -EINVAL;
 838		goto out;
 839	}
 840
 841       	nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
 842	if (nmh == NULL) {
 843		ret = -ENOMEM;
 844		goto out;
 845	}
 846
 847	nmh->nh_func = func;
 848	nmh->nh_func_data = data;
 849	nmh->nh_post_func = post_func;
 850	nmh->nh_msg_type = msg_type;
 851	nmh->nh_max_len = max_len;
 852	nmh->nh_key = key;
 853	/* the tree and list get this ref.. they're both removed in
 854	 * unregister when this ref is dropped */
 855	kref_init(&nmh->nh_kref);
 856	INIT_LIST_HEAD(&nmh->nh_unregister_item);
 857
 858	write_lock(&o2net_handler_lock);
 859	if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
 860		ret = -EEXIST;
 861	else {
 862	        rb_link_node(&nmh->nh_node, parent, p);
 863		rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
 864		list_add_tail(&nmh->nh_unregister_item, unreg_list);
 865
 866		mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
 867		     func, msg_type, key);
 868		/* we've had some trouble with handlers seemingly vanishing. */
 869		mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
 870							  &parent) == NULL,
 871			        "couldn't find handler we *just* registered "
 872				"for type %u key %08x\n", msg_type, key);
 873	}
 874	write_unlock(&o2net_handler_lock);
 875	if (ret)
 876		goto out;
 877
 878out:
 879	if (ret)
 880		kfree(nmh);
 881
 882	return ret;
 883}
 884EXPORT_SYMBOL_GPL(o2net_register_handler);
 885
 886void o2net_unregister_handler_list(struct list_head *list)
 887{
 888	struct o2net_msg_handler *nmh, *n;
 889
 890	write_lock(&o2net_handler_lock);
 891	list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
 892		mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
 893		     nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
 894		rb_erase(&nmh->nh_node, &o2net_handler_tree);
 895		list_del_init(&nmh->nh_unregister_item);
 896		kref_put(&nmh->nh_kref, o2net_handler_kref_release);
 897	}
 898	write_unlock(&o2net_handler_lock);
 899}
 900EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
 901
 902static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
 903{
 904	struct o2net_msg_handler *nmh;
 905
 906	read_lock(&o2net_handler_lock);
 907	nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
 908	if (nmh)
 909		kref_get(&nmh->nh_kref);
 910	read_unlock(&o2net_handler_lock);
 911
 912	return nmh;
 913}
 914
 915/* ------------------------------------------------------------ */
 916
 917static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
 918{
 919	struct kvec vec = { .iov_len = len, .iov_base = data, };
 920	struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
 921	return kernel_recvmsg(sock, &msg, &vec, 1, len, msg.msg_flags);
 
 922}
 923
 924static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
 925			      size_t veclen, size_t total)
 926{
 927	int ret;
 928	struct msghdr msg = {.msg_flags = 0,};
 929
 930	if (sock == NULL) {
 931		ret = -EINVAL;
 932		goto out;
 933	}
 934
 935	ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
 936	if (likely(ret == total))
 937		return 0;
 938	mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
 939	if (ret >= 0)
 940		ret = -EPIPE; /* should be smarter, I bet */
 941out:
 942	mlog(0, "returning error: %d\n", ret);
 943	return ret;
 944}
 945
 946static void o2net_sendpage(struct o2net_sock_container *sc,
 947			   void *kmalloced_virt,
 948			   size_t size)
 949{
 950	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
 
 
 951	ssize_t ret;
 952
 
 
 
 953	while (1) {
 
 954		mutex_lock(&sc->sc_send_lock);
 955		ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
 956						 virt_to_page(kmalloced_virt),
 957						 (long)kmalloced_virt & ~PAGE_MASK,
 958						 size, MSG_DONTWAIT);
 959		mutex_unlock(&sc->sc_send_lock);
 
 960		if (ret == size)
 961			break;
 962		if (ret == (ssize_t)-EAGAIN) {
 963			mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
 964			     " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
 965			cond_resched();
 966			continue;
 967		}
 968		mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
 969		     " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
 970		o2net_ensure_shutdown(nn, sc, 0);
 971		break;
 972	}
 973}
 974
 975static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
 976{
 977	memset(msg, 0, sizeof(struct o2net_msg));
 978	msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
 979	msg->data_len = cpu_to_be16(data_len);
 980	msg->msg_type = cpu_to_be16(msg_type);
 981	msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
 982	msg->status = 0;
 983	msg->key = cpu_to_be32(key);
 984}
 985
 986static int o2net_tx_can_proceed(struct o2net_node *nn,
 987			        struct o2net_sock_container **sc_ret,
 988				int *error)
 989{
 990	int ret = 0;
 991
 992	spin_lock(&nn->nn_lock);
 993	if (nn->nn_persistent_error) {
 994		ret = 1;
 995		*sc_ret = NULL;
 996		*error = nn->nn_persistent_error;
 997	} else if (nn->nn_sc_valid) {
 998		kref_get(&nn->nn_sc->sc_kref);
 999
1000		ret = 1;
1001		*sc_ret = nn->nn_sc;
1002		*error = 0;
1003	}
1004	spin_unlock(&nn->nn_lock);
1005
1006	return ret;
1007}
1008
1009/* Get a map of all nodes to which this node is currently connected to */
1010void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1011{
1012	struct o2net_sock_container *sc;
1013	int node, ret;
1014
1015	BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1016
1017	memset(map, 0, bytes);
1018	for (node = 0; node < O2NM_MAX_NODES; ++node) {
1019		if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1020			continue;
1021		if (!ret) {
1022			set_bit(node, map);
1023			sc_put(sc);
1024		}
1025	}
1026}
1027EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1028
1029int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1030			   size_t caller_veclen, u8 target_node, int *status)
1031{
1032	int ret = 0;
1033	struct o2net_msg *msg = NULL;
1034	size_t veclen, caller_bytes = 0;
1035	struct kvec *vec = NULL;
1036	struct o2net_sock_container *sc = NULL;
1037	struct o2net_node *nn = o2net_nn_from_num(target_node);
1038	struct o2net_status_wait nsw = {
1039		.ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1040	};
1041	struct o2net_send_tracking nst;
1042
1043	o2net_init_nst(&nst, msg_type, key, current, target_node);
1044
1045	if (o2net_wq == NULL) {
1046		mlog(0, "attempt to tx without o2netd running\n");
1047		ret = -ESRCH;
1048		goto out;
1049	}
1050
1051	if (caller_veclen == 0) {
1052		mlog(0, "bad kvec array length\n");
1053		ret = -EINVAL;
1054		goto out;
1055	}
1056
1057	caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1058	if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1059		mlog(0, "total payload len %zu too large\n", caller_bytes);
1060		ret = -EINVAL;
1061		goto out;
1062	}
1063
1064	if (target_node == o2nm_this_node()) {
1065		ret = -ELOOP;
1066		goto out;
1067	}
1068
1069	o2net_debug_add_nst(&nst);
1070
1071	o2net_set_nst_sock_time(&nst);
1072
1073	wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1074	if (ret)
1075		goto out;
1076
1077	o2net_set_nst_sock_container(&nst, sc);
1078
1079	veclen = caller_veclen + 1;
1080	vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1081	if (vec == NULL) {
1082		mlog(0, "failed to %zu element kvec!\n", veclen);
1083		ret = -ENOMEM;
1084		goto out;
1085	}
1086
1087	msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1088	if (!msg) {
1089		mlog(0, "failed to allocate a o2net_msg!\n");
1090		ret = -ENOMEM;
1091		goto out;
1092	}
1093
1094	o2net_init_msg(msg, caller_bytes, msg_type, key);
1095
1096	vec[0].iov_len = sizeof(struct o2net_msg);
1097	vec[0].iov_base = msg;
1098	memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1099
1100	ret = o2net_prep_nsw(nn, &nsw);
1101	if (ret)
1102		goto out;
1103
1104	msg->msg_num = cpu_to_be32(nsw.ns_id);
1105	o2net_set_nst_msg_id(&nst, nsw.ns_id);
1106
1107	o2net_set_nst_send_time(&nst);
1108
1109	/* finally, convert the message header to network byte-order
1110	 * and send */
1111	mutex_lock(&sc->sc_send_lock);
1112	ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1113				 sizeof(struct o2net_msg) + caller_bytes);
1114	mutex_unlock(&sc->sc_send_lock);
1115	msglog(msg, "sending returned %d\n", ret);
1116	if (ret < 0) {
1117		mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1118		goto out;
1119	}
1120
1121	/* wait on other node's handler */
1122	o2net_set_nst_status_time(&nst);
1123	wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1124
1125	o2net_update_send_stats(&nst, sc);
1126
1127	/* Note that we avoid overwriting the callers status return
1128	 * variable if a system error was reported on the other
1129	 * side. Callers beware. */
1130	ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1131	if (status && !ret)
1132		*status = nsw.ns_status;
1133
1134	mlog(0, "woken, returning system status %d, user status %d\n",
1135	     ret, nsw.ns_status);
1136out:
1137	o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1138	if (sc)
1139		sc_put(sc);
1140	kfree(vec);
1141	kfree(msg);
1142	o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1143	return ret;
1144}
1145EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1146
1147int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1148		       u8 target_node, int *status)
1149{
1150	struct kvec vec = {
1151		.iov_base = data,
1152		.iov_len = len,
1153	};
1154	return o2net_send_message_vec(msg_type, key, &vec, 1,
1155				      target_node, status);
1156}
1157EXPORT_SYMBOL_GPL(o2net_send_message);
1158
1159static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1160				   enum o2net_system_error syserr, int err)
1161{
1162	struct kvec vec = {
1163		.iov_base = hdr,
1164		.iov_len = sizeof(struct o2net_msg),
1165	};
1166
1167	BUG_ON(syserr >= O2NET_ERR_MAX);
1168
1169	/* leave other fields intact from the incoming message, msg_num
1170	 * in particular */
1171	hdr->sys_status = cpu_to_be32(syserr);
1172	hdr->status = cpu_to_be32(err);
1173	hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1174	hdr->data_len = 0;
1175
1176	msglog(hdr, "about to send status magic %d\n", err);
1177	/* hdr has been in host byteorder this whole time */
1178	return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1179}
1180
1181/* this returns -errno if the header was unknown or too large, etc.
1182 * after this is called the buffer us reused for the next message */
1183static int o2net_process_message(struct o2net_sock_container *sc,
1184				 struct o2net_msg *hdr)
1185{
1186	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1187	int ret = 0, handler_status;
1188	enum  o2net_system_error syserr;
1189	struct o2net_msg_handler *nmh = NULL;
1190	void *ret_data = NULL;
1191
1192	msglog(hdr, "processing message\n");
1193
1194	o2net_sc_postpone_idle(sc);
1195
1196	switch(be16_to_cpu(hdr->magic)) {
1197		case O2NET_MSG_STATUS_MAGIC:
1198			/* special type for returning message status */
1199			o2net_complete_nsw(nn, NULL,
1200					   be32_to_cpu(hdr->msg_num),
1201					   be32_to_cpu(hdr->sys_status),
1202					   be32_to_cpu(hdr->status));
1203			goto out;
1204		case O2NET_MSG_KEEP_REQ_MAGIC:
1205			o2net_sendpage(sc, o2net_keep_resp,
1206				       sizeof(*o2net_keep_resp));
1207			goto out;
1208		case O2NET_MSG_KEEP_RESP_MAGIC:
1209			goto out;
1210		case O2NET_MSG_MAGIC:
1211			break;
1212		default:
1213			msglog(hdr, "bad magic\n");
1214			ret = -EINVAL;
1215			goto out;
1216			break;
1217	}
1218
1219	/* find a handler for it */
1220	handler_status = 0;
1221	nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1222				be32_to_cpu(hdr->key));
1223	if (!nmh) {
1224		mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1225		     be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1226		syserr = O2NET_ERR_NO_HNDLR;
1227		goto out_respond;
1228	}
1229
1230	syserr = O2NET_ERR_NONE;
1231
1232	if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1233		syserr = O2NET_ERR_OVERFLOW;
1234
1235	if (syserr != O2NET_ERR_NONE)
1236		goto out_respond;
1237
1238	o2net_set_func_start_time(sc);
1239	sc->sc_msg_key = be32_to_cpu(hdr->key);
1240	sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1241	handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1242					     be16_to_cpu(hdr->data_len),
1243					nmh->nh_func_data, &ret_data);
1244	o2net_set_func_stop_time(sc);
1245
1246	o2net_update_recv_stats(sc);
1247
1248out_respond:
1249	/* this destroys the hdr, so don't use it after this */
1250	mutex_lock(&sc->sc_send_lock);
1251	ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1252				      handler_status);
1253	mutex_unlock(&sc->sc_send_lock);
1254	hdr = NULL;
1255	mlog(0, "sending handler status %d, syserr %d returned %d\n",
1256	     handler_status, syserr, ret);
1257
1258	if (nmh) {
1259		BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1260		if (nmh->nh_post_func)
1261			(nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1262					    ret_data);
1263	}
1264
1265out:
1266	if (nmh)
1267		o2net_handler_put(nmh);
1268	return ret;
1269}
1270
1271static int o2net_check_handshake(struct o2net_sock_container *sc)
1272{
1273	struct o2net_handshake *hand = page_address(sc->sc_page);
1274	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1275
1276	if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1277		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1278		       "protocol version %llu but %llu is required. "
1279		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1280		       (unsigned long long)be64_to_cpu(hand->protocol_version),
1281		       O2NET_PROTOCOL_VERSION);
1282
1283		/* don't bother reconnecting if its the wrong version. */
1284		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1285		return -1;
1286	}
1287
1288	/*
1289	 * Ensure timeouts are consistent with other nodes, otherwise
1290	 * we can end up with one node thinking that the other must be down,
1291	 * but isn't. This can ultimately cause corruption.
1292	 */
1293	if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1294				o2net_idle_timeout()) {
1295		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1296		       "idle timeout of %u ms, but we use %u ms locally. "
1297		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1298		       be32_to_cpu(hand->o2net_idle_timeout_ms),
1299		       o2net_idle_timeout());
1300		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1301		return -1;
1302	}
1303
1304	if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1305			o2net_keepalive_delay()) {
1306		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1307		       "delay of %u ms, but we use %u ms locally. "
1308		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1309		       be32_to_cpu(hand->o2net_keepalive_delay_ms),
1310		       o2net_keepalive_delay());
1311		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1312		return -1;
1313	}
1314
1315	if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1316			O2HB_MAX_WRITE_TIMEOUT_MS) {
1317		printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1318		       "timeout of %u ms, but we use %u ms locally. "
1319		       "Disconnecting.\n", SC_NODEF_ARGS(sc),
1320		       be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1321		       O2HB_MAX_WRITE_TIMEOUT_MS);
1322		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1323		return -1;
1324	}
1325
1326	sc->sc_handshake_ok = 1;
1327
1328	spin_lock(&nn->nn_lock);
1329	/* set valid and queue the idle timers only if it hasn't been
1330	 * shut down already */
1331	if (nn->nn_sc == sc) {
1332		o2net_sc_reset_idle_timer(sc);
1333		atomic_set(&nn->nn_timeout, 0);
1334		o2net_set_nn_state(nn, sc, 1, 0);
1335	}
1336	spin_unlock(&nn->nn_lock);
1337
1338	/* shift everything up as though it wasn't there */
1339	sc->sc_page_off -= sizeof(struct o2net_handshake);
1340	if (sc->sc_page_off)
1341		memmove(hand, hand + 1, sc->sc_page_off);
1342
1343	return 0;
1344}
1345
1346/* this demuxes the queued rx bytes into header or payload bits and calls
1347 * handlers as each full message is read off the socket.  it returns -error,
1348 * == 0 eof, or > 0 for progress made.*/
1349static int o2net_advance_rx(struct o2net_sock_container *sc)
1350{
1351	struct o2net_msg *hdr;
1352	int ret = 0;
1353	void *data;
1354	size_t datalen;
1355
1356	sclog(sc, "receiving\n");
1357	o2net_set_advance_start_time(sc);
1358
1359	if (unlikely(sc->sc_handshake_ok == 0)) {
1360		if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1361			data = page_address(sc->sc_page) + sc->sc_page_off;
1362			datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1363			ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1364			if (ret > 0)
1365				sc->sc_page_off += ret;
1366		}
1367
1368		if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1369			o2net_check_handshake(sc);
1370			if (unlikely(sc->sc_handshake_ok == 0))
1371				ret = -EPROTO;
1372		}
1373		goto out;
1374	}
1375
1376	/* do we need more header? */
1377	if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1378		data = page_address(sc->sc_page) + sc->sc_page_off;
1379		datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1380		ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1381		if (ret > 0) {
1382			sc->sc_page_off += ret;
1383			/* only swab incoming here.. we can
1384			 * only get here once as we cross from
1385			 * being under to over */
1386			if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1387				hdr = page_address(sc->sc_page);
1388				if (be16_to_cpu(hdr->data_len) >
1389				    O2NET_MAX_PAYLOAD_BYTES)
1390					ret = -EOVERFLOW;
1391			}
1392		}
1393		if (ret <= 0)
1394			goto out;
1395	}
1396
1397	if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1398		/* oof, still don't have a header */
1399		goto out;
1400	}
1401
1402	/* this was swabbed above when we first read it */
1403	hdr = page_address(sc->sc_page);
1404
1405	msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1406
1407	/* do we need more payload? */
1408	if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1409		/* need more payload */
1410		data = page_address(sc->sc_page) + sc->sc_page_off;
1411		datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1412			  sc->sc_page_off;
1413		ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1414		if (ret > 0)
1415			sc->sc_page_off += ret;
1416		if (ret <= 0)
1417			goto out;
1418	}
1419
1420	if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1421		/* we can only get here once, the first time we read
1422		 * the payload.. so set ret to progress if the handler
1423		 * works out. after calling this the message is toast */
1424		ret = o2net_process_message(sc, hdr);
1425		if (ret == 0)
1426			ret = 1;
1427		sc->sc_page_off = 0;
1428	}
1429
1430out:
1431	sclog(sc, "ret = %d\n", ret);
1432	o2net_set_advance_stop_time(sc);
1433	return ret;
1434}
1435
1436/* this work func is triggerd by data ready.  it reads until it can read no
1437 * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1438 * our work the work struct will be marked and we'll be called again. */
1439static void o2net_rx_until_empty(struct work_struct *work)
1440{
1441	struct o2net_sock_container *sc =
1442		container_of(work, struct o2net_sock_container, sc_rx_work);
1443	int ret;
1444
1445	do {
1446		ret = o2net_advance_rx(sc);
1447	} while (ret > 0);
1448
1449	if (ret <= 0 && ret != -EAGAIN) {
1450		struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1451		sclog(sc, "saw error %d, closing\n", ret);
1452		/* not permanent so read failed handshake can retry */
1453		o2net_ensure_shutdown(nn, sc, 0);
1454	}
1455
1456	sc_put(sc);
1457}
1458
1459static int o2net_set_nodelay(struct socket *sock)
1460{
1461	int ret, val = 1;
1462	mm_segment_t oldfs;
1463
1464	oldfs = get_fs();
1465	set_fs(KERNEL_DS);
1466
1467	/*
1468	 * Dear unsuspecting programmer,
1469	 *
1470	 * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1471	 * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1472	 * silently turn into SO_DEBUG.
1473	 *
1474	 * Yours,
1475	 * Keeper of hilariously fragile interfaces.
1476	 */
1477	ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1478				    (char __user *)&val, sizeof(val));
1479
1480	set_fs(oldfs);
1481	return ret;
1482}
1483
1484static int o2net_set_usertimeout(struct socket *sock)
1485{
1486	int user_timeout = O2NET_TCP_USER_TIMEOUT;
1487
1488	return kernel_setsockopt(sock, SOL_TCP, TCP_USER_TIMEOUT,
1489				(char *)&user_timeout, sizeof(user_timeout));
1490}
1491
1492static void o2net_initialize_handshake(void)
1493{
1494	o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1495		O2HB_MAX_WRITE_TIMEOUT_MS);
1496	o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1497	o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1498		o2net_keepalive_delay());
1499	o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1500		o2net_reconnect_delay());
1501}
1502
1503/* ------------------------------------------------------------ */
1504
1505/* called when a connect completes and after a sock is accepted.  the
1506 * rx path will see the response and mark the sc valid */
1507static void o2net_sc_connect_completed(struct work_struct *work)
1508{
1509	struct o2net_sock_container *sc =
1510		container_of(work, struct o2net_sock_container,
1511			     sc_connect_work);
1512
1513	mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1514              (unsigned long long)O2NET_PROTOCOL_VERSION,
1515	      (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1516
1517	o2net_initialize_handshake();
1518	o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1519	sc_put(sc);
1520}
1521
1522/* this is called as a work_struct func. */
1523static void o2net_sc_send_keep_req(struct work_struct *work)
1524{
1525	struct o2net_sock_container *sc =
1526		container_of(work, struct o2net_sock_container,
1527			     sc_keepalive_work.work);
1528
1529	o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1530	sc_put(sc);
1531}
1532
1533/* socket shutdown does a del_timer_sync against this as it tears down.
1534 * we can't start this timer until we've got to the point in sc buildup
1535 * where shutdown is going to be involved */
1536static void o2net_idle_timer(unsigned long data)
1537{
1538	struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1539	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1540#ifdef CONFIG_DEBUG_FS
1541	unsigned long msecs = ktime_to_ms(ktime_get()) -
1542		ktime_to_ms(sc->sc_tv_timer);
1543#else
1544	unsigned long msecs = o2net_idle_timeout();
1545#endif
1546
1547	printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1548	       "idle for %lu.%lu secs.\n",
1549	       SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1550
1551	/* idle timerout happen, don't shutdown the connection, but
1552	 * make fence decision. Maybe the connection can recover before
1553	 * the decision is made.
1554	 */
1555	atomic_set(&nn->nn_timeout, 1);
1556	o2quo_conn_err(o2net_num_from_nn(nn));
1557	queue_delayed_work(o2net_wq, &nn->nn_still_up,
1558			msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1559
1560	o2net_sc_reset_idle_timer(sc);
1561
1562}
1563
1564static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1565{
1566	o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1567	o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1568		      msecs_to_jiffies(o2net_keepalive_delay()));
1569	o2net_set_sock_timer(sc);
1570	mod_timer(&sc->sc_idle_timeout,
1571	       jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1572}
1573
1574static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1575{
1576	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1577
1578	/* clear fence decision since the connection recover from timeout*/
1579	if (atomic_read(&nn->nn_timeout)) {
1580		o2quo_conn_up(o2net_num_from_nn(nn));
1581		cancel_delayed_work(&nn->nn_still_up);
1582		atomic_set(&nn->nn_timeout, 0);
1583	}
1584
1585	/* Only push out an existing timer */
1586	if (timer_pending(&sc->sc_idle_timeout))
1587		o2net_sc_reset_idle_timer(sc);
1588}
1589
1590/* this work func is kicked whenever a path sets the nn state which doesn't
1591 * have valid set.  This includes seeing hb come up, losing a connection,
1592 * having a connect attempt fail, etc. This centralizes the logic which decides
1593 * if a connect attempt should be made or if we should give up and all future
1594 * transmit attempts should fail */
1595static void o2net_start_connect(struct work_struct *work)
1596{
1597	struct o2net_node *nn =
1598		container_of(work, struct o2net_node, nn_connect_work.work);
1599	struct o2net_sock_container *sc = NULL;
1600	struct o2nm_node *node = NULL, *mynode = NULL;
1601	struct socket *sock = NULL;
1602	struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1603	int ret = 0, stop;
1604	unsigned int timeout;
1605	unsigned int noio_flag;
1606
1607	/*
1608	 * sock_create allocates the sock with GFP_KERNEL. We must set
1609	 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1610	 * by this process are done as if GFP_NOIO was specified. So we
1611	 * are not reentering filesystem while doing memory reclaim.
1612	 */
1613	noio_flag = memalloc_noio_save();
1614	/* if we're greater we initiate tx, otherwise we accept */
1615	if (o2nm_this_node() <= o2net_num_from_nn(nn))
1616		goto out;
1617
1618	/* watch for racing with tearing a node down */
1619	node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1620	if (node == NULL) {
1621		ret = 0;
1622		goto out;
1623	}
1624
1625	mynode = o2nm_get_node_by_num(o2nm_this_node());
1626	if (mynode == NULL) {
1627		ret = 0;
1628		goto out;
1629	}
1630
1631	spin_lock(&nn->nn_lock);
1632	/*
1633	 * see if we already have one pending or have given up.
1634	 * For nn_timeout, it is set when we close the connection
1635	 * because of the idle time out. So it means that we have
1636	 * at least connected to that node successfully once,
1637	 * now try to connect to it again.
1638	 */
1639	timeout = atomic_read(&nn->nn_timeout);
1640	stop = (nn->nn_sc ||
1641		(nn->nn_persistent_error &&
1642		(nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1643	spin_unlock(&nn->nn_lock);
1644	if (stop)
1645		goto out;
1646
1647	nn->nn_last_connect_attempt = jiffies;
1648
1649	sc = sc_alloc(node);
1650	if (sc == NULL) {
1651		mlog(0, "couldn't allocate sc\n");
1652		ret = -ENOMEM;
1653		goto out;
1654	}
1655
1656	ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1657	if (ret < 0) {
1658		mlog(0, "can't create socket: %d\n", ret);
1659		goto out;
1660	}
1661	sc->sc_sock = sock; /* freed by sc_kref_release */
1662
1663	sock->sk->sk_allocation = GFP_ATOMIC;
 
1664
1665	myaddr.sin_family = AF_INET;
1666	myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1667	myaddr.sin_port = htons(0); /* any port */
1668
1669	ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1670			      sizeof(myaddr));
1671	if (ret) {
1672		mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1673		     ret, &mynode->nd_ipv4_address);
1674		goto out;
1675	}
1676
1677	ret = o2net_set_nodelay(sc->sc_sock);
1678	if (ret) {
1679		mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1680		goto out;
1681	}
1682
1683	ret = o2net_set_usertimeout(sock);
1684	if (ret) {
1685		mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1686		goto out;
1687	}
1688
1689	o2net_register_callbacks(sc->sc_sock->sk, sc);
1690
1691	spin_lock(&nn->nn_lock);
1692	/* handshake completion will set nn->nn_sc_valid */
1693	o2net_set_nn_state(nn, sc, 0, 0);
1694	spin_unlock(&nn->nn_lock);
1695
1696	remoteaddr.sin_family = AF_INET;
1697	remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1698	remoteaddr.sin_port = node->nd_ipv4_port;
1699
1700	ret = sc->sc_sock->ops->connect(sc->sc_sock,
1701					(struct sockaddr *)&remoteaddr,
1702					sizeof(remoteaddr),
1703					O_NONBLOCK);
1704	if (ret == -EINPROGRESS)
1705		ret = 0;
1706
1707out:
1708	if (ret && sc) {
1709		printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1710		       " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1711		/* 0 err so that another will be queued and attempted
1712		 * from set_nn_state */
1713		o2net_ensure_shutdown(nn, sc, 0);
1714	}
1715	if (sc)
1716		sc_put(sc);
1717	if (node)
1718		o2nm_node_put(node);
1719	if (mynode)
1720		o2nm_node_put(mynode);
1721
1722	memalloc_noio_restore(noio_flag);
1723	return;
1724}
1725
1726static void o2net_connect_expired(struct work_struct *work)
1727{
1728	struct o2net_node *nn =
1729		container_of(work, struct o2net_node, nn_connect_expired.work);
1730
1731	spin_lock(&nn->nn_lock);
1732	if (!nn->nn_sc_valid) {
1733		printk(KERN_NOTICE "o2net: No connection established with "
1734		       "node %u after %u.%u seconds, check network and"
1735		       " cluster configuration.\n",
1736		     o2net_num_from_nn(nn),
1737		     o2net_idle_timeout() / 1000,
1738		     o2net_idle_timeout() % 1000);
1739
1740		o2net_set_nn_state(nn, NULL, 0, 0);
1741	}
1742	spin_unlock(&nn->nn_lock);
1743}
1744
1745static void o2net_still_up(struct work_struct *work)
1746{
1747	struct o2net_node *nn =
1748		container_of(work, struct o2net_node, nn_still_up.work);
1749
1750	o2quo_hb_still_up(o2net_num_from_nn(nn));
1751}
1752
1753/* ------------------------------------------------------------ */
1754
1755void o2net_disconnect_node(struct o2nm_node *node)
1756{
1757	struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1758
1759	/* don't reconnect until it's heartbeating again */
1760	spin_lock(&nn->nn_lock);
1761	atomic_set(&nn->nn_timeout, 0);
1762	o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1763	spin_unlock(&nn->nn_lock);
1764
1765	if (o2net_wq) {
1766		cancel_delayed_work(&nn->nn_connect_expired);
1767		cancel_delayed_work(&nn->nn_connect_work);
1768		cancel_delayed_work(&nn->nn_still_up);
1769		flush_workqueue(o2net_wq);
1770	}
1771}
1772
1773static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1774				  void *data)
1775{
1776	o2quo_hb_down(node_num);
1777
1778	if (!node)
1779		return;
1780
1781	if (node_num != o2nm_this_node())
1782		o2net_disconnect_node(node);
1783
1784	BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1785}
1786
1787static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1788				void *data)
1789{
1790	struct o2net_node *nn = o2net_nn_from_num(node_num);
1791
1792	o2quo_hb_up(node_num);
1793
1794	BUG_ON(!node);
1795
1796	/* ensure an immediate connect attempt */
1797	nn->nn_last_connect_attempt = jiffies -
1798		(msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1799
1800	if (node_num != o2nm_this_node()) {
1801		/* believe it or not, accept and node hearbeating testing
1802		 * can succeed for this node before we got here.. so
1803		 * only use set_nn_state to clear the persistent error
1804		 * if that hasn't already happened */
1805		spin_lock(&nn->nn_lock);
1806		atomic_set(&nn->nn_timeout, 0);
1807		if (nn->nn_persistent_error)
1808			o2net_set_nn_state(nn, NULL, 0, 0);
1809		spin_unlock(&nn->nn_lock);
1810	}
1811}
1812
1813void o2net_unregister_hb_callbacks(void)
1814{
1815	o2hb_unregister_callback(NULL, &o2net_hb_up);
1816	o2hb_unregister_callback(NULL, &o2net_hb_down);
1817}
1818
1819int o2net_register_hb_callbacks(void)
1820{
1821	int ret;
1822
1823	o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1824			    o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1825	o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1826			    o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1827
1828	ret = o2hb_register_callback(NULL, &o2net_hb_up);
1829	if (ret == 0)
1830		ret = o2hb_register_callback(NULL, &o2net_hb_down);
1831
1832	if (ret)
1833		o2net_unregister_hb_callbacks();
1834
1835	return ret;
1836}
1837
1838/* ------------------------------------------------------------ */
1839
1840static int o2net_accept_one(struct socket *sock, int *more)
1841{
1842	int ret, slen;
1843	struct sockaddr_in sin;
1844	struct socket *new_sock = NULL;
1845	struct o2nm_node *node = NULL;
1846	struct o2nm_node *local_node = NULL;
1847	struct o2net_sock_container *sc = NULL;
 
 
 
1848	struct o2net_node *nn;
1849	unsigned int noio_flag;
1850
1851	/*
1852	 * sock_create_lite allocates the sock with GFP_KERNEL. We must set
1853	 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1854	 * by this process are done as if GFP_NOIO was specified. So we
1855	 * are not reentering filesystem while doing memory reclaim.
1856	 */
1857	noio_flag = memalloc_noio_save();
1858
1859	BUG_ON(sock == NULL);
1860	*more = 0;
1861	ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1862			       sock->sk->sk_protocol, &new_sock);
1863	if (ret)
1864		goto out;
1865
1866	new_sock->type = sock->type;
1867	new_sock->ops = sock->ops;
1868	ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1869	if (ret < 0)
1870		goto out;
1871
1872	*more = 1;
1873	new_sock->sk->sk_allocation = GFP_ATOMIC;
1874
1875	ret = o2net_set_nodelay(new_sock);
1876	if (ret) {
1877		mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1878		goto out;
1879	}
1880
1881	ret = o2net_set_usertimeout(new_sock);
1882	if (ret) {
1883		mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1884		goto out;
1885	}
1886
1887	slen = sizeof(sin);
1888	ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1889				       &slen, 1);
1890	if (ret < 0)
1891		goto out;
1892
1893	node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1894	if (node == NULL) {
1895		printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1896		       "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1897		       ntohs(sin.sin_port));
1898		ret = -EINVAL;
1899		goto out;
1900	}
1901
1902	if (o2nm_this_node() >= node->nd_num) {
1903		local_node = o2nm_get_node_by_num(o2nm_this_node());
1904		if (local_node)
1905			printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1906					"seen at node '%s' (%u, %pI4:%d) from "
1907					"node '%s' (%u, %pI4:%d)\n",
1908					local_node->nd_name, local_node->nd_num,
1909					&(local_node->nd_ipv4_address),
1910					ntohs(local_node->nd_ipv4_port),
1911					node->nd_name,
1912					node->nd_num, &sin.sin_addr.s_addr,
1913					ntohs(sin.sin_port));
1914		ret = -EINVAL;
1915		goto out;
1916	}
1917
1918	/* this happens all the time when the other node sees our heartbeat
1919	 * and tries to connect before we see their heartbeat */
1920	if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1921		mlog(ML_CONN, "attempt to connect from node '%s' at "
1922		     "%pI4:%d but it isn't heartbeating\n",
1923		     node->nd_name, &sin.sin_addr.s_addr,
1924		     ntohs(sin.sin_port));
1925		ret = -EINVAL;
1926		goto out;
1927	}
1928
1929	nn = o2net_nn_from_num(node->nd_num);
1930
1931	spin_lock(&nn->nn_lock);
1932	if (nn->nn_sc)
1933		ret = -EBUSY;
1934	else
1935		ret = 0;
1936	spin_unlock(&nn->nn_lock);
1937	if (ret) {
1938		printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1939		       "at %pI4:%d but it already has an open connection\n",
1940		       node->nd_name, &sin.sin_addr.s_addr,
1941		       ntohs(sin.sin_port));
1942		goto out;
1943	}
1944
1945	sc = sc_alloc(node);
1946	if (sc == NULL) {
1947		ret = -ENOMEM;
1948		goto out;
1949	}
1950
1951	sc->sc_sock = new_sock;
1952	new_sock = NULL;
1953
1954	spin_lock(&nn->nn_lock);
1955	atomic_set(&nn->nn_timeout, 0);
1956	o2net_set_nn_state(nn, sc, 0, 0);
1957	spin_unlock(&nn->nn_lock);
1958
1959	o2net_register_callbacks(sc->sc_sock->sk, sc);
1960	o2net_sc_queue_work(sc, &sc->sc_rx_work);
1961
1962	o2net_initialize_handshake();
1963	o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1964
1965out:
1966	if (new_sock)
1967		sock_release(new_sock);
1968	if (node)
1969		o2nm_node_put(node);
1970	if (local_node)
1971		o2nm_node_put(local_node);
1972	if (sc)
1973		sc_put(sc);
1974
1975	memalloc_noio_restore(noio_flag);
1976	return ret;
1977}
1978
1979/*
1980 * This function is invoked in response to one or more
1981 * pending accepts at softIRQ level. We must drain the
1982 * entire que before returning.
1983 */
1984
1985static void o2net_accept_many(struct work_struct *work)
1986{
1987	struct socket *sock = o2net_listen_sock;
1988	int	more;
1989	int	err;
1990
1991	/*
1992	 * It is critical to note that due to interrupt moderation
1993	 * at the network driver level, we can't assume to get a
1994	 * softIRQ for every single conn since tcp SYN packets
1995	 * can arrive back-to-back, and therefore many pending
1996	 * accepts may result in just 1 softIRQ. If we terminate
1997	 * the o2net_accept_one() loop upon seeing an err, what happens
1998	 * to the rest of the conns in the queue? If no new SYN
1999	 * arrives for hours, no softIRQ  will be delivered,
2000	 * and the connections will just sit in the queue.
2001	 */
2002
2003	for (;;) {
2004		err = o2net_accept_one(sock, &more);
2005		if (!more)
2006			break;
2007		cond_resched();
2008	}
2009}
2010
2011static void o2net_listen_data_ready(struct sock *sk)
2012{
2013	void (*ready)(struct sock *sk);
2014
2015	read_lock(&sk->sk_callback_lock);
 
 
2016	ready = sk->sk_user_data;
2017	if (ready == NULL) { /* check for teardown race */
2018		ready = sk->sk_data_ready;
2019		goto out;
2020	}
2021
2022	/* This callback may called twice when a new connection
2023	 * is  being established as a child socket inherits everything
2024	 * from a parent LISTEN socket, including the data_ready cb of
2025	 * the parent. This leads to a hazard. In o2net_accept_one()
2026	 * we are still initializing the child socket but have not
2027	 * changed the inherited data_ready callback yet when
2028	 * data starts arriving.
2029	 * We avoid this hazard by checking the state.
2030	 * For the listening socket,  the state will be TCP_LISTEN; for the new
2031	 * socket, will be  TCP_ESTABLISHED. Also, in this case,
2032	 * sk->sk_user_data is not a valid function pointer.
2033	 */
2034
2035	if (sk->sk_state == TCP_LISTEN) {
2036		queue_work(o2net_wq, &o2net_listen_work);
2037	} else {
2038		ready = NULL;
2039	}
2040
2041out:
2042	read_unlock(&sk->sk_callback_lock);
2043	if (ready != NULL)
2044		ready(sk);
2045}
2046
2047static int o2net_open_listening_sock(__be32 addr, __be16 port)
2048{
2049	struct socket *sock = NULL;
2050	int ret;
2051	struct sockaddr_in sin = {
2052		.sin_family = PF_INET,
2053		.sin_addr = { .s_addr = addr },
2054		.sin_port = port,
2055	};
2056
2057	ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2058	if (ret < 0) {
2059		printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
2060		goto out;
2061	}
2062
2063	sock->sk->sk_allocation = GFP_ATOMIC;
2064
2065	write_lock_bh(&sock->sk->sk_callback_lock);
2066	sock->sk->sk_user_data = sock->sk->sk_data_ready;
2067	sock->sk->sk_data_ready = o2net_listen_data_ready;
2068	write_unlock_bh(&sock->sk->sk_callback_lock);
2069
2070	o2net_listen_sock = sock;
2071	INIT_WORK(&o2net_listen_work, o2net_accept_many);
2072
2073	sock->sk->sk_reuse = SK_CAN_REUSE;
2074	ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2075	if (ret < 0) {
2076		printk(KERN_ERR "o2net: Error %d while binding socket at "
2077		       "%pI4:%u\n", ret, &addr, ntohs(port)); 
2078		goto out;
2079	}
2080
2081	ret = sock->ops->listen(sock, 64);
2082	if (ret < 0)
2083		printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2084		       ret, &addr, ntohs(port));
2085
2086out:
2087	if (ret) {
2088		o2net_listen_sock = NULL;
2089		if (sock)
2090			sock_release(sock);
2091	}
2092	return ret;
2093}
2094
2095/*
2096 * called from node manager when we should bring up our network listening
2097 * socket.  node manager handles all the serialization to only call this
2098 * once and to match it with o2net_stop_listening().  note,
2099 * o2nm_this_node() doesn't work yet as we're being called while it
2100 * is being set up.
2101 */
2102int o2net_start_listening(struct o2nm_node *node)
2103{
2104	int ret = 0;
2105
2106	BUG_ON(o2net_wq != NULL);
2107	BUG_ON(o2net_listen_sock != NULL);
2108
2109	mlog(ML_KTHREAD, "starting o2net thread...\n");
2110	o2net_wq = create_singlethread_workqueue("o2net");
2111	if (o2net_wq == NULL) {
2112		mlog(ML_ERROR, "unable to launch o2net thread\n");
2113		return -ENOMEM; /* ? */
2114	}
2115
2116	ret = o2net_open_listening_sock(node->nd_ipv4_address,
2117					node->nd_ipv4_port);
2118	if (ret) {
2119		destroy_workqueue(o2net_wq);
2120		o2net_wq = NULL;
2121	} else
2122		o2quo_conn_up(node->nd_num);
2123
2124	return ret;
2125}
2126
2127/* again, o2nm_this_node() doesn't work here as we're involved in
2128 * tearing it down */
2129void o2net_stop_listening(struct o2nm_node *node)
2130{
2131	struct socket *sock = o2net_listen_sock;
2132	size_t i;
2133
2134	BUG_ON(o2net_wq == NULL);
2135	BUG_ON(o2net_listen_sock == NULL);
2136
2137	/* stop the listening socket from generating work */
2138	write_lock_bh(&sock->sk->sk_callback_lock);
2139	sock->sk->sk_data_ready = sock->sk->sk_user_data;
2140	sock->sk->sk_user_data = NULL;
2141	write_unlock_bh(&sock->sk->sk_callback_lock);
2142
2143	for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2144		struct o2nm_node *node = o2nm_get_node_by_num(i);
2145		if (node) {
2146			o2net_disconnect_node(node);
2147			o2nm_node_put(node);
2148		}
2149	}
2150
2151	/* finish all work and tear down the work queue */
2152	mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2153	destroy_workqueue(o2net_wq);
2154	o2net_wq = NULL;
2155
2156	sock_release(o2net_listen_sock);
2157	o2net_listen_sock = NULL;
2158
2159	o2quo_conn_err(node->nd_num);
2160}
2161
2162/* ------------------------------------------------------------ */
2163
2164int o2net_init(void)
2165{
 
 
2166	unsigned long i;
2167
2168	o2quo_init();
 
2169
2170	if (o2net_debugfs_init())
 
2171		goto out;
2172
2173	o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2174	o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2175	o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2176	if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp)
2177		goto out;
 
2178
2179	o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2180	o2net_hand->connector_id = cpu_to_be64(1);
2181
2182	o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2183	o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2184
2185	for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2186		struct o2net_node *nn = o2net_nn_from_num(i);
2187
2188		atomic_set(&nn->nn_timeout, 0);
2189		spin_lock_init(&nn->nn_lock);
2190		INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2191		INIT_DELAYED_WORK(&nn->nn_connect_expired,
2192				  o2net_connect_expired);
2193		INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2194		/* until we see hb from a node we'll return einval */
2195		nn->nn_persistent_error = -ENOTCONN;
2196		init_waitqueue_head(&nn->nn_sc_wq);
2197		idr_init(&nn->nn_status_idr);
2198		INIT_LIST_HEAD(&nn->nn_status_list);
2199	}
2200
2201	return 0;
2202
2203out:
2204	kfree(o2net_hand);
2205	kfree(o2net_keep_req);
2206	kfree(o2net_keep_resp);
2207	o2net_debugfs_exit();
2208	o2quo_exit();
2209	return -ENOMEM;
2210}
2211
2212void o2net_exit(void)
2213{
2214	o2quo_exit();
2215	kfree(o2net_hand);
2216	kfree(o2net_keep_req);
2217	kfree(o2net_keep_resp);
2218	o2net_debugfs_exit();
 
2219}