tcp.c

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