xref: /openbmc/linux/fs/ocfs2/cluster/tcp.c (revision 732a675a)
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 <net/tcp.h>
63 
64 #include <asm/uaccess.h>
65 
66 #include "heartbeat.h"
67 #include "tcp.h"
68 #include "nodemanager.h"
69 #define MLOG_MASK_PREFIX ML_TCP
70 #include "masklog.h"
71 #include "quorum.h"
72 
73 #include "tcp_internal.h"
74 
75 #define SC_NODEF_FMT "node %s (num %u) at %u.%u.%u.%u:%u"
76 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,	\
77 			  NIPQUAD(sc->sc_node->nd_ipv4_address),	\
78 			  ntohs(sc->sc_node->nd_ipv4_port)
79 
80 /*
81  * In the following two log macros, the whitespace after the ',' just
82  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
83  * previous token if args expands to nothing.
84  */
85 #define msglog(hdr, fmt, args...) do {					\
86 	typeof(hdr) __hdr = (hdr);					\
87 	mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "	\
88 	     "key %08x num %u] " fmt,					\
89 	     be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len), 	\
90 	     be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),	\
91 	     be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),	\
92 	     be32_to_cpu(__hdr->msg_num) ,  ##args);			\
93 } while (0)
94 
95 #define sclog(sc, fmt, args...) do {					\
96 	typeof(sc) __sc = (sc);						\
97 	mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "	\
98 	     "pg_off %zu] " fmt, __sc,					\
99 	     atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,	\
100 	    __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,	\
101 	    ##args);							\
102 } while (0)
103 
104 static DEFINE_RWLOCK(o2net_handler_lock);
105 static struct rb_root o2net_handler_tree = RB_ROOT;
106 
107 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
108 
109 /* XXX someday we'll need better accounting */
110 static struct socket *o2net_listen_sock = NULL;
111 
112 /*
113  * listen work is only queued by the listening socket callbacks on the
114  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
115  * quorum work is queued as sock containers are shutdown.. stop_listening
116  * tears down all the node's sock containers, preventing future shutdowns
117  * and queued quroum work, before canceling delayed quorum work and
118  * destroying the work queue.
119  */
120 static struct workqueue_struct *o2net_wq;
121 static struct work_struct o2net_listen_work;
122 
123 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
124 #define O2NET_HB_PRI 0x1
125 
126 static struct o2net_handshake *o2net_hand;
127 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
128 
129 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
130 		{[O2NET_ERR_NONE]	= 0,
131 		 [O2NET_ERR_NO_HNDLR]	= -ENOPROTOOPT,
132 		 [O2NET_ERR_OVERFLOW]	= -EOVERFLOW,
133 		 [O2NET_ERR_DIED]	= -EHOSTDOWN,};
134 
135 /* can't quite avoid *all* internal declarations :/ */
136 static void o2net_sc_connect_completed(struct work_struct *work);
137 static void o2net_rx_until_empty(struct work_struct *work);
138 static void o2net_shutdown_sc(struct work_struct *work);
139 static void o2net_listen_data_ready(struct sock *sk, int bytes);
140 static void o2net_sc_send_keep_req(struct work_struct *work);
141 static void o2net_idle_timer(unsigned long data);
142 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
143 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
144 
145 #ifdef CONFIG_DEBUG_FS
146 void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
147 		    u32 msgkey, struct task_struct *task, u8 node)
148 {
149 	INIT_LIST_HEAD(&nst->st_net_debug_item);
150 	nst->st_task = task;
151 	nst->st_msg_type = msgtype;
152 	nst->st_msg_key = msgkey;
153 	nst->st_node = node;
154 }
155 
156 void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
157 {
158 	do_gettimeofday(&nst->st_sock_time);
159 }
160 
161 void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
162 {
163 	do_gettimeofday(&nst->st_send_time);
164 }
165 
166 void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
167 {
168 	do_gettimeofday(&nst->st_status_time);
169 }
170 
171 void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
172 					 struct o2net_sock_container *sc)
173 {
174 	nst->st_sc = sc;
175 }
176 
177 void o2net_set_nst_msg_id(struct o2net_send_tracking *nst, u32 msg_id)
178 {
179 	nst->st_id = msg_id;
180 }
181 #endif /* CONFIG_DEBUG_FS */
182 
183 static inline int o2net_reconnect_delay(void)
184 {
185 	return o2nm_single_cluster->cl_reconnect_delay_ms;
186 }
187 
188 static inline int o2net_keepalive_delay(void)
189 {
190 	return o2nm_single_cluster->cl_keepalive_delay_ms;
191 }
192 
193 static inline int o2net_idle_timeout(void)
194 {
195 	return o2nm_single_cluster->cl_idle_timeout_ms;
196 }
197 
198 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
199 {
200 	int trans;
201 	BUG_ON(err >= O2NET_ERR_MAX);
202 	trans = o2net_sys_err_translations[err];
203 
204 	/* Just in case we mess up the translation table above */
205 	BUG_ON(err != O2NET_ERR_NONE && trans == 0);
206 	return trans;
207 }
208 
209 static struct o2net_node * o2net_nn_from_num(u8 node_num)
210 {
211 	BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
212 	return &o2net_nodes[node_num];
213 }
214 
215 static u8 o2net_num_from_nn(struct o2net_node *nn)
216 {
217 	BUG_ON(nn == NULL);
218 	return nn - o2net_nodes;
219 }
220 
221 /* ------------------------------------------------------------ */
222 
223 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
224 {
225 	int ret = 0;
226 
227 	do {
228 		if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
229 			ret = -EAGAIN;
230 			break;
231 		}
232 		spin_lock(&nn->nn_lock);
233 		ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
234 		if (ret == 0)
235 			list_add_tail(&nsw->ns_node_item,
236 				      &nn->nn_status_list);
237 		spin_unlock(&nn->nn_lock);
238 	} while (ret == -EAGAIN);
239 
240 	if (ret == 0)  {
241 		init_waitqueue_head(&nsw->ns_wq);
242 		nsw->ns_sys_status = O2NET_ERR_NONE;
243 		nsw->ns_status = 0;
244 	}
245 
246 	return ret;
247 }
248 
249 static void o2net_complete_nsw_locked(struct o2net_node *nn,
250 				      struct o2net_status_wait *nsw,
251 				      enum o2net_system_error sys_status,
252 				      s32 status)
253 {
254 	assert_spin_locked(&nn->nn_lock);
255 
256 	if (!list_empty(&nsw->ns_node_item)) {
257 		list_del_init(&nsw->ns_node_item);
258 		nsw->ns_sys_status = sys_status;
259 		nsw->ns_status = status;
260 		idr_remove(&nn->nn_status_idr, nsw->ns_id);
261 		wake_up(&nsw->ns_wq);
262 	}
263 }
264 
265 static void o2net_complete_nsw(struct o2net_node *nn,
266 			       struct o2net_status_wait *nsw,
267 			       u64 id, enum o2net_system_error sys_status,
268 			       s32 status)
269 {
270 	spin_lock(&nn->nn_lock);
271 	if (nsw == NULL) {
272 		if (id > INT_MAX)
273 			goto out;
274 
275 		nsw = idr_find(&nn->nn_status_idr, id);
276 		if (nsw == NULL)
277 			goto out;
278 	}
279 
280 	o2net_complete_nsw_locked(nn, nsw, sys_status, status);
281 
282 out:
283 	spin_unlock(&nn->nn_lock);
284 	return;
285 }
286 
287 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
288 {
289 	struct o2net_status_wait *nsw, *tmp;
290 	unsigned int num_kills = 0;
291 
292 	assert_spin_locked(&nn->nn_lock);
293 
294 	list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
295 		o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
296 		num_kills++;
297 	}
298 
299 	mlog(0, "completed %d messages for node %u\n", num_kills,
300 	     o2net_num_from_nn(nn));
301 }
302 
303 static int o2net_nsw_completed(struct o2net_node *nn,
304 			       struct o2net_status_wait *nsw)
305 {
306 	int completed;
307 	spin_lock(&nn->nn_lock);
308 	completed = list_empty(&nsw->ns_node_item);
309 	spin_unlock(&nn->nn_lock);
310 	return completed;
311 }
312 
313 /* ------------------------------------------------------------ */
314 
315 static void sc_kref_release(struct kref *kref)
316 {
317 	struct o2net_sock_container *sc = container_of(kref,
318 					struct o2net_sock_container, sc_kref);
319 	BUG_ON(timer_pending(&sc->sc_idle_timeout));
320 
321 	sclog(sc, "releasing\n");
322 
323 	if (sc->sc_sock) {
324 		sock_release(sc->sc_sock);
325 		sc->sc_sock = NULL;
326 	}
327 
328 	o2nm_node_put(sc->sc_node);
329 	sc->sc_node = NULL;
330 
331 	o2net_debug_del_sc(sc);
332 	kfree(sc);
333 }
334 
335 static void sc_put(struct o2net_sock_container *sc)
336 {
337 	sclog(sc, "put\n");
338 	kref_put(&sc->sc_kref, sc_kref_release);
339 }
340 static void sc_get(struct o2net_sock_container *sc)
341 {
342 	sclog(sc, "get\n");
343 	kref_get(&sc->sc_kref);
344 }
345 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
346 {
347 	struct o2net_sock_container *sc, *ret = NULL;
348 	struct page *page = NULL;
349 
350 	page = alloc_page(GFP_NOFS);
351 	sc = kzalloc(sizeof(*sc), GFP_NOFS);
352 	if (sc == NULL || page == NULL)
353 		goto out;
354 
355 	kref_init(&sc->sc_kref);
356 	o2nm_node_get(node);
357 	sc->sc_node = node;
358 
359 	INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
360 	INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
361 	INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
362 	INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
363 
364 	init_timer(&sc->sc_idle_timeout);
365 	sc->sc_idle_timeout.function = o2net_idle_timer;
366 	sc->sc_idle_timeout.data = (unsigned long)sc;
367 
368 	sclog(sc, "alloced\n");
369 
370 	ret = sc;
371 	sc->sc_page = page;
372 	o2net_debug_add_sc(sc);
373 	sc = NULL;
374 	page = NULL;
375 
376 out:
377 	if (page)
378 		__free_page(page);
379 	kfree(sc);
380 
381 	return ret;
382 }
383 
384 /* ------------------------------------------------------------ */
385 
386 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
387 				struct work_struct *work)
388 {
389 	sc_get(sc);
390 	if (!queue_work(o2net_wq, work))
391 		sc_put(sc);
392 }
393 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
394 					struct delayed_work *work,
395 					int delay)
396 {
397 	sc_get(sc);
398 	if (!queue_delayed_work(o2net_wq, work, delay))
399 		sc_put(sc);
400 }
401 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
402 					 struct delayed_work *work)
403 {
404 	if (cancel_delayed_work(work))
405 		sc_put(sc);
406 }
407 
408 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
409 
410 int o2net_num_connected_peers(void)
411 {
412 	return atomic_read(&o2net_connected_peers);
413 }
414 
415 static void o2net_set_nn_state(struct o2net_node *nn,
416 			       struct o2net_sock_container *sc,
417 			       unsigned valid, int err)
418 {
419 	int was_valid = nn->nn_sc_valid;
420 	int was_err = nn->nn_persistent_error;
421 	struct o2net_sock_container *old_sc = nn->nn_sc;
422 
423 	assert_spin_locked(&nn->nn_lock);
424 
425 	if (old_sc && !sc)
426 		atomic_dec(&o2net_connected_peers);
427 	else if (!old_sc && sc)
428 		atomic_inc(&o2net_connected_peers);
429 
430 	/* the node num comparison and single connect/accept path should stop
431 	 * an non-null sc from being overwritten with another */
432 	BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
433 	mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
434 	mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
435 
436 	if (was_valid && !valid && err == 0)
437 		err = -ENOTCONN;
438 
439 	mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
440 	     o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
441 	     nn->nn_persistent_error, err);
442 
443 	nn->nn_sc = sc;
444 	nn->nn_sc_valid = valid ? 1 : 0;
445 	nn->nn_persistent_error = err;
446 
447 	/* mirrors o2net_tx_can_proceed() */
448 	if (nn->nn_persistent_error || nn->nn_sc_valid)
449 		wake_up(&nn->nn_sc_wq);
450 
451 	if (!was_err && nn->nn_persistent_error) {
452 		o2quo_conn_err(o2net_num_from_nn(nn));
453 		queue_delayed_work(o2net_wq, &nn->nn_still_up,
454 				   msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
455 	}
456 
457 	if (was_valid && !valid) {
458 		printk(KERN_INFO "o2net: no longer connected to "
459 		       SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
460 		o2net_complete_nodes_nsw(nn);
461 	}
462 
463 	if (!was_valid && valid) {
464 		o2quo_conn_up(o2net_num_from_nn(nn));
465 		cancel_delayed_work(&nn->nn_connect_expired);
466 		printk(KERN_INFO "o2net: %s " SC_NODEF_FMT "\n",
467 		       o2nm_this_node() > sc->sc_node->nd_num ?
468 		       		"connected to" : "accepted connection from",
469 		       SC_NODEF_ARGS(sc));
470 	}
471 
472 	/* trigger the connecting worker func as long as we're not valid,
473 	 * it will back off if it shouldn't connect.  This can be called
474 	 * from node config teardown and so needs to be careful about
475 	 * the work queue actually being up. */
476 	if (!valid && o2net_wq) {
477 		unsigned long delay;
478 		/* delay if we're withing a RECONNECT_DELAY of the
479 		 * last attempt */
480 		delay = (nn->nn_last_connect_attempt +
481 			 msecs_to_jiffies(o2net_reconnect_delay()))
482 			- jiffies;
483 		if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
484 			delay = 0;
485 		mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
486 		queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
487 
488 		/*
489 		 * Delay the expired work after idle timeout.
490 		 *
491 		 * We might have lots of failed connection attempts that run
492 		 * through here but we only cancel the connect_expired work when
493 		 * a connection attempt succeeds.  So only the first enqueue of
494 		 * the connect_expired work will do anything.  The rest will see
495 		 * that it's already queued and do nothing.
496 		 */
497 		delay += msecs_to_jiffies(o2net_idle_timeout());
498 		queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
499 	}
500 
501 	/* keep track of the nn's sc ref for the caller */
502 	if ((old_sc == NULL) && sc)
503 		sc_get(sc);
504 	if (old_sc && (old_sc != sc)) {
505 		o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
506 		sc_put(old_sc);
507 	}
508 }
509 
510 /* see o2net_register_callbacks() */
511 static void o2net_data_ready(struct sock *sk, int bytes)
512 {
513 	void (*ready)(struct sock *sk, int bytes);
514 
515 	read_lock(&sk->sk_callback_lock);
516 	if (sk->sk_user_data) {
517 		struct o2net_sock_container *sc = sk->sk_user_data;
518 		sclog(sc, "data_ready hit\n");
519 		do_gettimeofday(&sc->sc_tv_data_ready);
520 		o2net_sc_queue_work(sc, &sc->sc_rx_work);
521 		ready = sc->sc_data_ready;
522 	} else {
523 		ready = sk->sk_data_ready;
524 	}
525 	read_unlock(&sk->sk_callback_lock);
526 
527 	ready(sk, bytes);
528 }
529 
530 /* see o2net_register_callbacks() */
531 static void o2net_state_change(struct sock *sk)
532 {
533 	void (*state_change)(struct sock *sk);
534 	struct o2net_sock_container *sc;
535 
536 	read_lock(&sk->sk_callback_lock);
537 	sc = sk->sk_user_data;
538 	if (sc == NULL) {
539 		state_change = sk->sk_state_change;
540 		goto out;
541 	}
542 
543 	sclog(sc, "state_change to %d\n", sk->sk_state);
544 
545 	state_change = sc->sc_state_change;
546 
547 	switch(sk->sk_state) {
548 		/* ignore connecting sockets as they make progress */
549 		case TCP_SYN_SENT:
550 		case TCP_SYN_RECV:
551 			break;
552 		case TCP_ESTABLISHED:
553 			o2net_sc_queue_work(sc, &sc->sc_connect_work);
554 			break;
555 		default:
556 			o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
557 			break;
558 	}
559 out:
560 	read_unlock(&sk->sk_callback_lock);
561 	state_change(sk);
562 }
563 
564 /*
565  * we register callbacks so we can queue work on events before calling
566  * the original callbacks.  our callbacks our careful to test user_data
567  * to discover when they've reaced with o2net_unregister_callbacks().
568  */
569 static void o2net_register_callbacks(struct sock *sk,
570 				     struct o2net_sock_container *sc)
571 {
572 	write_lock_bh(&sk->sk_callback_lock);
573 
574 	/* accepted sockets inherit the old listen socket data ready */
575 	if (sk->sk_data_ready == o2net_listen_data_ready) {
576 		sk->sk_data_ready = sk->sk_user_data;
577 		sk->sk_user_data = NULL;
578 	}
579 
580 	BUG_ON(sk->sk_user_data != NULL);
581 	sk->sk_user_data = sc;
582 	sc_get(sc);
583 
584 	sc->sc_data_ready = sk->sk_data_ready;
585 	sc->sc_state_change = sk->sk_state_change;
586 	sk->sk_data_ready = o2net_data_ready;
587 	sk->sk_state_change = o2net_state_change;
588 
589 	mutex_init(&sc->sc_send_lock);
590 
591 	write_unlock_bh(&sk->sk_callback_lock);
592 }
593 
594 static int o2net_unregister_callbacks(struct sock *sk,
595 			           struct o2net_sock_container *sc)
596 {
597 	int ret = 0;
598 
599 	write_lock_bh(&sk->sk_callback_lock);
600 	if (sk->sk_user_data == sc) {
601 		ret = 1;
602 		sk->sk_user_data = NULL;
603 		sk->sk_data_ready = sc->sc_data_ready;
604 		sk->sk_state_change = sc->sc_state_change;
605 	}
606 	write_unlock_bh(&sk->sk_callback_lock);
607 
608 	return ret;
609 }
610 
611 /*
612  * this is a little helper that is called by callers who have seen a problem
613  * with an sc and want to detach it from the nn if someone already hasn't beat
614  * them to it.  if an error is given then the shutdown will be persistent
615  * and pending transmits will be canceled.
616  */
617 static void o2net_ensure_shutdown(struct o2net_node *nn,
618 			           struct o2net_sock_container *sc,
619 				   int err)
620 {
621 	spin_lock(&nn->nn_lock);
622 	if (nn->nn_sc == sc)
623 		o2net_set_nn_state(nn, NULL, 0, err);
624 	spin_unlock(&nn->nn_lock);
625 }
626 
627 /*
628  * This work queue function performs the blocking parts of socket shutdown.  A
629  * few paths lead here.  set_nn_state will trigger this callback if it sees an
630  * sc detached from the nn.  state_change will also trigger this callback
631  * directly when it sees errors.  In that case we need to call set_nn_state
632  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
633  * itself.
634  */
635 static void o2net_shutdown_sc(struct work_struct *work)
636 {
637 	struct o2net_sock_container *sc =
638 		container_of(work, struct o2net_sock_container,
639 			     sc_shutdown_work);
640 	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
641 
642 	sclog(sc, "shutting down\n");
643 
644 	/* drop the callbacks ref and call shutdown only once */
645 	if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
646 		/* we shouldn't flush as we're in the thread, the
647 		 * races with pending sc work structs are harmless */
648 		del_timer_sync(&sc->sc_idle_timeout);
649 		o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
650 		sc_put(sc);
651 		kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
652 	}
653 
654 	/* not fatal so failed connects before the other guy has our
655 	 * heartbeat can be retried */
656 	o2net_ensure_shutdown(nn, sc, 0);
657 	sc_put(sc);
658 }
659 
660 /* ------------------------------------------------------------ */
661 
662 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
663 			     u32 key)
664 {
665 	int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
666 
667 	if (ret == 0)
668 		ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
669 
670 	return ret;
671 }
672 
673 static struct o2net_msg_handler *
674 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
675 			  struct rb_node **ret_parent)
676 {
677         struct rb_node **p = &o2net_handler_tree.rb_node;
678         struct rb_node *parent = NULL;
679 	struct o2net_msg_handler *nmh, *ret = NULL;
680 	int cmp;
681 
682         while (*p) {
683                 parent = *p;
684                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
685 		cmp = o2net_handler_cmp(nmh, msg_type, key);
686 
687                 if (cmp < 0)
688                         p = &(*p)->rb_left;
689                 else if (cmp > 0)
690                         p = &(*p)->rb_right;
691                 else {
692 			ret = nmh;
693                         break;
694 		}
695         }
696 
697         if (ret_p != NULL)
698                 *ret_p = p;
699         if (ret_parent != NULL)
700                 *ret_parent = parent;
701 
702         return ret;
703 }
704 
705 static void o2net_handler_kref_release(struct kref *kref)
706 {
707 	struct o2net_msg_handler *nmh;
708 	nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
709 
710 	kfree(nmh);
711 }
712 
713 static void o2net_handler_put(struct o2net_msg_handler *nmh)
714 {
715 	kref_put(&nmh->nh_kref, o2net_handler_kref_release);
716 }
717 
718 /* max_len is protection for the handler func.  incoming messages won't
719  * be given to the handler if their payload is longer than the max. */
720 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
721 			   o2net_msg_handler_func *func, void *data,
722 			   o2net_post_msg_handler_func *post_func,
723 			   struct list_head *unreg_list)
724 {
725 	struct o2net_msg_handler *nmh = NULL;
726 	struct rb_node **p, *parent;
727 	int ret = 0;
728 
729 	if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
730 		mlog(0, "max_len for message handler out of range: %u\n",
731 			max_len);
732 		ret = -EINVAL;
733 		goto out;
734 	}
735 
736 	if (!msg_type) {
737 		mlog(0, "no message type provided: %u, %p\n", msg_type, func);
738 		ret = -EINVAL;
739 		goto out;
740 
741 	}
742 	if (!func) {
743 		mlog(0, "no message handler provided: %u, %p\n",
744 		       msg_type, func);
745 		ret = -EINVAL;
746 		goto out;
747 	}
748 
749        	nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
750 	if (nmh == NULL) {
751 		ret = -ENOMEM;
752 		goto out;
753 	}
754 
755 	nmh->nh_func = func;
756 	nmh->nh_func_data = data;
757 	nmh->nh_post_func = post_func;
758 	nmh->nh_msg_type = msg_type;
759 	nmh->nh_max_len = max_len;
760 	nmh->nh_key = key;
761 	/* the tree and list get this ref.. they're both removed in
762 	 * unregister when this ref is dropped */
763 	kref_init(&nmh->nh_kref);
764 	INIT_LIST_HEAD(&nmh->nh_unregister_item);
765 
766 	write_lock(&o2net_handler_lock);
767 	if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
768 		ret = -EEXIST;
769 	else {
770 	        rb_link_node(&nmh->nh_node, parent, p);
771 		rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
772 		list_add_tail(&nmh->nh_unregister_item, unreg_list);
773 
774 		mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
775 		     func, msg_type, key);
776 		/* we've had some trouble with handlers seemingly vanishing. */
777 		mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
778 							  &parent) == NULL,
779 			        "couldn't find handler we *just* registerd "
780 				"for type %u key %08x\n", msg_type, key);
781 	}
782 	write_unlock(&o2net_handler_lock);
783 	if (ret)
784 		goto out;
785 
786 out:
787 	if (ret)
788 		kfree(nmh);
789 
790 	return ret;
791 }
792 EXPORT_SYMBOL_GPL(o2net_register_handler);
793 
794 void o2net_unregister_handler_list(struct list_head *list)
795 {
796 	struct o2net_msg_handler *nmh, *n;
797 
798 	write_lock(&o2net_handler_lock);
799 	list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
800 		mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
801 		     nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
802 		rb_erase(&nmh->nh_node, &o2net_handler_tree);
803 		list_del_init(&nmh->nh_unregister_item);
804 		kref_put(&nmh->nh_kref, o2net_handler_kref_release);
805 	}
806 	write_unlock(&o2net_handler_lock);
807 }
808 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
809 
810 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
811 {
812 	struct o2net_msg_handler *nmh;
813 
814 	read_lock(&o2net_handler_lock);
815 	nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
816 	if (nmh)
817 		kref_get(&nmh->nh_kref);
818 	read_unlock(&o2net_handler_lock);
819 
820 	return nmh;
821 }
822 
823 /* ------------------------------------------------------------ */
824 
825 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
826 {
827 	int ret;
828 	mm_segment_t oldfs;
829 	struct kvec vec = {
830 		.iov_len = len,
831 		.iov_base = data,
832 	};
833 	struct msghdr msg = {
834 		.msg_iovlen = 1,
835 		.msg_iov = (struct iovec *)&vec,
836        		.msg_flags = MSG_DONTWAIT,
837 	};
838 
839 	oldfs = get_fs();
840 	set_fs(get_ds());
841 	ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
842 	set_fs(oldfs);
843 
844 	return ret;
845 }
846 
847 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
848 			      size_t veclen, size_t total)
849 {
850 	int ret;
851 	mm_segment_t oldfs;
852 	struct msghdr msg = {
853 		.msg_iov = (struct iovec *)vec,
854 		.msg_iovlen = veclen,
855 	};
856 
857 	if (sock == NULL) {
858 		ret = -EINVAL;
859 		goto out;
860 	}
861 
862 	oldfs = get_fs();
863 	set_fs(get_ds());
864 	ret = sock_sendmsg(sock, &msg, total);
865 	set_fs(oldfs);
866 	if (ret != total) {
867 		mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
868 		     total);
869 		if (ret >= 0)
870 			ret = -EPIPE; /* should be smarter, I bet */
871 		goto out;
872 	}
873 
874 	ret = 0;
875 out:
876 	if (ret < 0)
877 		mlog(0, "returning error: %d\n", ret);
878 	return ret;
879 }
880 
881 static void o2net_sendpage(struct o2net_sock_container *sc,
882 			   void *kmalloced_virt,
883 			   size_t size)
884 {
885 	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
886 	ssize_t ret;
887 
888 	while (1) {
889 		mutex_lock(&sc->sc_send_lock);
890 		ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
891 						 virt_to_page(kmalloced_virt),
892 						 (long)kmalloced_virt & ~PAGE_MASK,
893 						 size, MSG_DONTWAIT);
894 		mutex_unlock(&sc->sc_send_lock);
895 		if (ret == size)
896 			break;
897 		if (ret == (ssize_t)-EAGAIN) {
898 			mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
899 			     " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
900 			cond_resched();
901 			continue;
902 		}
903 		mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
904 		     " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
905 		o2net_ensure_shutdown(nn, sc, 0);
906 		break;
907 	}
908 }
909 
910 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
911 {
912 	memset(msg, 0, sizeof(struct o2net_msg));
913 	msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
914 	msg->data_len = cpu_to_be16(data_len);
915 	msg->msg_type = cpu_to_be16(msg_type);
916 	msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
917 	msg->status = 0;
918 	msg->key = cpu_to_be32(key);
919 }
920 
921 static int o2net_tx_can_proceed(struct o2net_node *nn,
922 			        struct o2net_sock_container **sc_ret,
923 				int *error)
924 {
925 	int ret = 0;
926 
927 	spin_lock(&nn->nn_lock);
928 	if (nn->nn_persistent_error) {
929 		ret = 1;
930 		*sc_ret = NULL;
931 		*error = nn->nn_persistent_error;
932 	} else if (nn->nn_sc_valid) {
933 		kref_get(&nn->nn_sc->sc_kref);
934 
935 		ret = 1;
936 		*sc_ret = nn->nn_sc;
937 		*error = 0;
938 	}
939 	spin_unlock(&nn->nn_lock);
940 
941 	return ret;
942 }
943 
944 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
945 			   size_t caller_veclen, u8 target_node, int *status)
946 {
947 	int ret, error = 0;
948 	struct o2net_msg *msg = NULL;
949 	size_t veclen, caller_bytes = 0;
950 	struct kvec *vec = NULL;
951 	struct o2net_sock_container *sc = NULL;
952 	struct o2net_node *nn = o2net_nn_from_num(target_node);
953 	struct o2net_status_wait nsw = {
954 		.ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
955 	};
956 	struct o2net_send_tracking nst;
957 
958 	o2net_init_nst(&nst, msg_type, key, current, target_node);
959 
960 	if (o2net_wq == NULL) {
961 		mlog(0, "attempt to tx without o2netd running\n");
962 		ret = -ESRCH;
963 		goto out;
964 	}
965 
966 	if (caller_veclen == 0) {
967 		mlog(0, "bad kvec array length\n");
968 		ret = -EINVAL;
969 		goto out;
970 	}
971 
972 	caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
973 	if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
974 		mlog(0, "total payload len %zu too large\n", caller_bytes);
975 		ret = -EINVAL;
976 		goto out;
977 	}
978 
979 	if (target_node == o2nm_this_node()) {
980 		ret = -ELOOP;
981 		goto out;
982 	}
983 
984 	o2net_debug_add_nst(&nst);
985 
986 	o2net_set_nst_sock_time(&nst);
987 
988 	ret = wait_event_interruptible(nn->nn_sc_wq,
989 				       o2net_tx_can_proceed(nn, &sc, &error));
990 	if (!ret && error)
991 		ret = error;
992 	if (ret)
993 		goto out;
994 
995 	o2net_set_nst_sock_container(&nst, sc);
996 
997 	veclen = caller_veclen + 1;
998 	vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
999 	if (vec == NULL) {
1000 		mlog(0, "failed to %zu element kvec!\n", veclen);
1001 		ret = -ENOMEM;
1002 		goto out;
1003 	}
1004 
1005 	msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1006 	if (!msg) {
1007 		mlog(0, "failed to allocate a o2net_msg!\n");
1008 		ret = -ENOMEM;
1009 		goto out;
1010 	}
1011 
1012 	o2net_init_msg(msg, caller_bytes, msg_type, key);
1013 
1014 	vec[0].iov_len = sizeof(struct o2net_msg);
1015 	vec[0].iov_base = msg;
1016 	memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1017 
1018 	ret = o2net_prep_nsw(nn, &nsw);
1019 	if (ret)
1020 		goto out;
1021 
1022 	msg->msg_num = cpu_to_be32(nsw.ns_id);
1023 	o2net_set_nst_msg_id(&nst, nsw.ns_id);
1024 
1025 	o2net_set_nst_send_time(&nst);
1026 
1027 	/* finally, convert the message header to network byte-order
1028 	 * and send */
1029 	mutex_lock(&sc->sc_send_lock);
1030 	ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1031 				 sizeof(struct o2net_msg) + caller_bytes);
1032 	mutex_unlock(&sc->sc_send_lock);
1033 	msglog(msg, "sending returned %d\n", ret);
1034 	if (ret < 0) {
1035 		mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1036 		goto out;
1037 	}
1038 
1039 	/* wait on other node's handler */
1040 	o2net_set_nst_status_time(&nst);
1041 	wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1042 
1043 	/* Note that we avoid overwriting the callers status return
1044 	 * variable if a system error was reported on the other
1045 	 * side. Callers beware. */
1046 	ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1047 	if (status && !ret)
1048 		*status = nsw.ns_status;
1049 
1050 	mlog(0, "woken, returning system status %d, user status %d\n",
1051 	     ret, nsw.ns_status);
1052 out:
1053 	o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1054 	if (sc)
1055 		sc_put(sc);
1056 	if (vec)
1057 		kfree(vec);
1058 	if (msg)
1059 		kfree(msg);
1060 	o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1061 	return ret;
1062 }
1063 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1064 
1065 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1066 		       u8 target_node, int *status)
1067 {
1068 	struct kvec vec = {
1069 		.iov_base = data,
1070 		.iov_len = len,
1071 	};
1072 	return o2net_send_message_vec(msg_type, key, &vec, 1,
1073 				      target_node, status);
1074 }
1075 EXPORT_SYMBOL_GPL(o2net_send_message);
1076 
1077 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1078 				   enum o2net_system_error syserr, int err)
1079 {
1080 	struct kvec vec = {
1081 		.iov_base = hdr,
1082 		.iov_len = sizeof(struct o2net_msg),
1083 	};
1084 
1085 	BUG_ON(syserr >= O2NET_ERR_MAX);
1086 
1087 	/* leave other fields intact from the incoming message, msg_num
1088 	 * in particular */
1089 	hdr->sys_status = cpu_to_be32(syserr);
1090 	hdr->status = cpu_to_be32(err);
1091 	hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1092 	hdr->data_len = 0;
1093 
1094 	msglog(hdr, "about to send status magic %d\n", err);
1095 	/* hdr has been in host byteorder this whole time */
1096 	return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1097 }
1098 
1099 /* this returns -errno if the header was unknown or too large, etc.
1100  * after this is called the buffer us reused for the next message */
1101 static int o2net_process_message(struct o2net_sock_container *sc,
1102 				 struct o2net_msg *hdr)
1103 {
1104 	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1105 	int ret = 0, handler_status;
1106 	enum  o2net_system_error syserr;
1107 	struct o2net_msg_handler *nmh = NULL;
1108 	void *ret_data = NULL;
1109 
1110 	msglog(hdr, "processing message\n");
1111 
1112 	o2net_sc_postpone_idle(sc);
1113 
1114 	switch(be16_to_cpu(hdr->magic)) {
1115 		case O2NET_MSG_STATUS_MAGIC:
1116 			/* special type for returning message status */
1117 			o2net_complete_nsw(nn, NULL,
1118 					   be32_to_cpu(hdr->msg_num),
1119 					   be32_to_cpu(hdr->sys_status),
1120 					   be32_to_cpu(hdr->status));
1121 			goto out;
1122 		case O2NET_MSG_KEEP_REQ_MAGIC:
1123 			o2net_sendpage(sc, o2net_keep_resp,
1124 				       sizeof(*o2net_keep_resp));
1125 			goto out;
1126 		case O2NET_MSG_KEEP_RESP_MAGIC:
1127 			goto out;
1128 		case O2NET_MSG_MAGIC:
1129 			break;
1130 		default:
1131 			msglog(hdr, "bad magic\n");
1132 			ret = -EINVAL;
1133 			goto out;
1134 			break;
1135 	}
1136 
1137 	/* find a handler for it */
1138 	handler_status = 0;
1139 	nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1140 				be32_to_cpu(hdr->key));
1141 	if (!nmh) {
1142 		mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1143 		     be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1144 		syserr = O2NET_ERR_NO_HNDLR;
1145 		goto out_respond;
1146 	}
1147 
1148 	syserr = O2NET_ERR_NONE;
1149 
1150 	if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1151 		syserr = O2NET_ERR_OVERFLOW;
1152 
1153 	if (syserr != O2NET_ERR_NONE)
1154 		goto out_respond;
1155 
1156 	do_gettimeofday(&sc->sc_tv_func_start);
1157 	sc->sc_msg_key = be32_to_cpu(hdr->key);
1158 	sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1159 	handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1160 					     be16_to_cpu(hdr->data_len),
1161 					nmh->nh_func_data, &ret_data);
1162 	do_gettimeofday(&sc->sc_tv_func_stop);
1163 
1164 out_respond:
1165 	/* this destroys the hdr, so don't use it after this */
1166 	mutex_lock(&sc->sc_send_lock);
1167 	ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1168 				      handler_status);
1169 	mutex_unlock(&sc->sc_send_lock);
1170 	hdr = NULL;
1171 	mlog(0, "sending handler status %d, syserr %d returned %d\n",
1172 	     handler_status, syserr, ret);
1173 
1174 	if (nmh) {
1175 		BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1176 		if (nmh->nh_post_func)
1177 			(nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1178 					    ret_data);
1179 	}
1180 
1181 out:
1182 	if (nmh)
1183 		o2net_handler_put(nmh);
1184 	return ret;
1185 }
1186 
1187 static int o2net_check_handshake(struct o2net_sock_container *sc)
1188 {
1189 	struct o2net_handshake *hand = page_address(sc->sc_page);
1190 	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1191 
1192 	if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1193 		mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1194 		     "version %llu but %llu is required, disconnecting\n",
1195 		     SC_NODEF_ARGS(sc),
1196 		     (unsigned long long)be64_to_cpu(hand->protocol_version),
1197 		     O2NET_PROTOCOL_VERSION);
1198 
1199 		/* don't bother reconnecting if its the wrong version. */
1200 		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1201 		return -1;
1202 	}
1203 
1204 	/*
1205 	 * Ensure timeouts are consistent with other nodes, otherwise
1206 	 * we can end up with one node thinking that the other must be down,
1207 	 * but isn't. This can ultimately cause corruption.
1208 	 */
1209 	if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1210 				o2net_idle_timeout()) {
1211 		mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1212 		     "%u ms, but we use %u ms locally.  disconnecting\n",
1213 		     SC_NODEF_ARGS(sc),
1214 		     be32_to_cpu(hand->o2net_idle_timeout_ms),
1215 		     o2net_idle_timeout());
1216 		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1217 		return -1;
1218 	}
1219 
1220 	if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1221 			o2net_keepalive_delay()) {
1222 		mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1223 		     "%u ms, but we use %u ms locally.  disconnecting\n",
1224 		     SC_NODEF_ARGS(sc),
1225 		     be32_to_cpu(hand->o2net_keepalive_delay_ms),
1226 		     o2net_keepalive_delay());
1227 		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1228 		return -1;
1229 	}
1230 
1231 	if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1232 			O2HB_MAX_WRITE_TIMEOUT_MS) {
1233 		mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1234 		     "%u ms, but we use %u ms locally.  disconnecting\n",
1235 		     SC_NODEF_ARGS(sc),
1236 		     be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1237 		     O2HB_MAX_WRITE_TIMEOUT_MS);
1238 		o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1239 		return -1;
1240 	}
1241 
1242 	sc->sc_handshake_ok = 1;
1243 
1244 	spin_lock(&nn->nn_lock);
1245 	/* set valid and queue the idle timers only if it hasn't been
1246 	 * shut down already */
1247 	if (nn->nn_sc == sc) {
1248 		o2net_sc_reset_idle_timer(sc);
1249 		atomic_set(&nn->nn_timeout, 0);
1250 		o2net_set_nn_state(nn, sc, 1, 0);
1251 	}
1252 	spin_unlock(&nn->nn_lock);
1253 
1254 	/* shift everything up as though it wasn't there */
1255 	sc->sc_page_off -= sizeof(struct o2net_handshake);
1256 	if (sc->sc_page_off)
1257 		memmove(hand, hand + 1, sc->sc_page_off);
1258 
1259 	return 0;
1260 }
1261 
1262 /* this demuxes the queued rx bytes into header or payload bits and calls
1263  * handlers as each full message is read off the socket.  it returns -error,
1264  * == 0 eof, or > 0 for progress made.*/
1265 static int o2net_advance_rx(struct o2net_sock_container *sc)
1266 {
1267 	struct o2net_msg *hdr;
1268 	int ret = 0;
1269 	void *data;
1270 	size_t datalen;
1271 
1272 	sclog(sc, "receiving\n");
1273 	do_gettimeofday(&sc->sc_tv_advance_start);
1274 
1275 	if (unlikely(sc->sc_handshake_ok == 0)) {
1276 		if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1277 			data = page_address(sc->sc_page) + sc->sc_page_off;
1278 			datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1279 			ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1280 			if (ret > 0)
1281 				sc->sc_page_off += ret;
1282 		}
1283 
1284 		if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1285 			o2net_check_handshake(sc);
1286 			if (unlikely(sc->sc_handshake_ok == 0))
1287 				ret = -EPROTO;
1288 		}
1289 		goto out;
1290 	}
1291 
1292 	/* do we need more header? */
1293 	if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1294 		data = page_address(sc->sc_page) + sc->sc_page_off;
1295 		datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1296 		ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1297 		if (ret > 0) {
1298 			sc->sc_page_off += ret;
1299 			/* only swab incoming here.. we can
1300 			 * only get here once as we cross from
1301 			 * being under to over */
1302 			if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1303 				hdr = page_address(sc->sc_page);
1304 				if (be16_to_cpu(hdr->data_len) >
1305 				    O2NET_MAX_PAYLOAD_BYTES)
1306 					ret = -EOVERFLOW;
1307 			}
1308 		}
1309 		if (ret <= 0)
1310 			goto out;
1311 	}
1312 
1313 	if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1314 		/* oof, still don't have a header */
1315 		goto out;
1316 	}
1317 
1318 	/* this was swabbed above when we first read it */
1319 	hdr = page_address(sc->sc_page);
1320 
1321 	msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1322 
1323 	/* do we need more payload? */
1324 	if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1325 		/* need more payload */
1326 		data = page_address(sc->sc_page) + sc->sc_page_off;
1327 		datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1328 			  sc->sc_page_off;
1329 		ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1330 		if (ret > 0)
1331 			sc->sc_page_off += ret;
1332 		if (ret <= 0)
1333 			goto out;
1334 	}
1335 
1336 	if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1337 		/* we can only get here once, the first time we read
1338 		 * the payload.. so set ret to progress if the handler
1339 		 * works out. after calling this the message is toast */
1340 		ret = o2net_process_message(sc, hdr);
1341 		if (ret == 0)
1342 			ret = 1;
1343 		sc->sc_page_off = 0;
1344 	}
1345 
1346 out:
1347 	sclog(sc, "ret = %d\n", ret);
1348 	do_gettimeofday(&sc->sc_tv_advance_stop);
1349 	return ret;
1350 }
1351 
1352 /* this work func is triggerd by data ready.  it reads until it can read no
1353  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1354  * our work the work struct will be marked and we'll be called again. */
1355 static void o2net_rx_until_empty(struct work_struct *work)
1356 {
1357 	struct o2net_sock_container *sc =
1358 		container_of(work, struct o2net_sock_container, sc_rx_work);
1359 	int ret;
1360 
1361 	do {
1362 		ret = o2net_advance_rx(sc);
1363 	} while (ret > 0);
1364 
1365 	if (ret <= 0 && ret != -EAGAIN) {
1366 		struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1367 		sclog(sc, "saw error %d, closing\n", ret);
1368 		/* not permanent so read failed handshake can retry */
1369 		o2net_ensure_shutdown(nn, sc, 0);
1370 	}
1371 
1372 	sc_put(sc);
1373 }
1374 
1375 static int o2net_set_nodelay(struct socket *sock)
1376 {
1377 	int ret, val = 1;
1378 	mm_segment_t oldfs;
1379 
1380 	oldfs = get_fs();
1381 	set_fs(KERNEL_DS);
1382 
1383 	/*
1384 	 * Dear unsuspecting programmer,
1385 	 *
1386 	 * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1387 	 * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1388 	 * silently turn into SO_DEBUG.
1389 	 *
1390 	 * Yours,
1391 	 * Keeper of hilariously fragile interfaces.
1392 	 */
1393 	ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1394 				    (char __user *)&val, sizeof(val));
1395 
1396 	set_fs(oldfs);
1397 	return ret;
1398 }
1399 
1400 static void o2net_initialize_handshake(void)
1401 {
1402 	o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1403 		O2HB_MAX_WRITE_TIMEOUT_MS);
1404 	o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1405 	o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1406 		o2net_keepalive_delay());
1407 	o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1408 		o2net_reconnect_delay());
1409 }
1410 
1411 /* ------------------------------------------------------------ */
1412 
1413 /* called when a connect completes and after a sock is accepted.  the
1414  * rx path will see the response and mark the sc valid */
1415 static void o2net_sc_connect_completed(struct work_struct *work)
1416 {
1417 	struct o2net_sock_container *sc =
1418 		container_of(work, struct o2net_sock_container,
1419 			     sc_connect_work);
1420 
1421 	mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1422               (unsigned long long)O2NET_PROTOCOL_VERSION,
1423 	      (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1424 
1425 	o2net_initialize_handshake();
1426 	o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1427 	sc_put(sc);
1428 }
1429 
1430 /* this is called as a work_struct func. */
1431 static void o2net_sc_send_keep_req(struct work_struct *work)
1432 {
1433 	struct o2net_sock_container *sc =
1434 		container_of(work, struct o2net_sock_container,
1435 			     sc_keepalive_work.work);
1436 
1437 	o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1438 	sc_put(sc);
1439 }
1440 
1441 /* socket shutdown does a del_timer_sync against this as it tears down.
1442  * we can't start this timer until we've got to the point in sc buildup
1443  * where shutdown is going to be involved */
1444 static void o2net_idle_timer(unsigned long data)
1445 {
1446 	struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1447 	struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1448 	struct timeval now;
1449 
1450 	do_gettimeofday(&now);
1451 
1452 	printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1453 	     "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1454 		     o2net_idle_timeout() / 1000,
1455 		     o2net_idle_timeout() % 1000);
1456 	mlog(ML_NOTICE, "here are some times that might help debug the "
1457 	     "situation: (tmr %ld.%ld now %ld.%ld dr %ld.%ld adv "
1458 	     "%ld.%ld:%ld.%ld func (%08x:%u) %ld.%ld:%ld.%ld)\n",
1459 	     sc->sc_tv_timer.tv_sec, (long) sc->sc_tv_timer.tv_usec,
1460 	     now.tv_sec, (long) now.tv_usec,
1461 	     sc->sc_tv_data_ready.tv_sec, (long) sc->sc_tv_data_ready.tv_usec,
1462 	     sc->sc_tv_advance_start.tv_sec,
1463 	     (long) sc->sc_tv_advance_start.tv_usec,
1464 	     sc->sc_tv_advance_stop.tv_sec,
1465 	     (long) sc->sc_tv_advance_stop.tv_usec,
1466 	     sc->sc_msg_key, sc->sc_msg_type,
1467 	     sc->sc_tv_func_start.tv_sec, (long) sc->sc_tv_func_start.tv_usec,
1468 	     sc->sc_tv_func_stop.tv_sec, (long) sc->sc_tv_func_stop.tv_usec);
1469 
1470 	/*
1471 	 * Initialize the nn_timeout so that the next connection attempt
1472 	 * will continue in o2net_start_connect.
1473 	 */
1474 	atomic_set(&nn->nn_timeout, 1);
1475 
1476 	o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1477 }
1478 
1479 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1480 {
1481 	o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1482 	o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1483 		      msecs_to_jiffies(o2net_keepalive_delay()));
1484 	do_gettimeofday(&sc->sc_tv_timer);
1485 	mod_timer(&sc->sc_idle_timeout,
1486 	       jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1487 }
1488 
1489 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1490 {
1491 	/* Only push out an existing timer */
1492 	if (timer_pending(&sc->sc_idle_timeout))
1493 		o2net_sc_reset_idle_timer(sc);
1494 }
1495 
1496 /* this work func is kicked whenever a path sets the nn state which doesn't
1497  * have valid set.  This includes seeing hb come up, losing a connection,
1498  * having a connect attempt fail, etc. This centralizes the logic which decides
1499  * if a connect attempt should be made or if we should give up and all future
1500  * transmit attempts should fail */
1501 static void o2net_start_connect(struct work_struct *work)
1502 {
1503 	struct o2net_node *nn =
1504 		container_of(work, struct o2net_node, nn_connect_work.work);
1505 	struct o2net_sock_container *sc = NULL;
1506 	struct o2nm_node *node = NULL, *mynode = NULL;
1507 	struct socket *sock = NULL;
1508 	struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1509 	int ret = 0, stop;
1510 	unsigned int timeout;
1511 
1512 	/* if we're greater we initiate tx, otherwise we accept */
1513 	if (o2nm_this_node() <= o2net_num_from_nn(nn))
1514 		goto out;
1515 
1516 	/* watch for racing with tearing a node down */
1517 	node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1518 	if (node == NULL) {
1519 		ret = 0;
1520 		goto out;
1521 	}
1522 
1523 	mynode = o2nm_get_node_by_num(o2nm_this_node());
1524 	if (mynode == NULL) {
1525 		ret = 0;
1526 		goto out;
1527 	}
1528 
1529 	spin_lock(&nn->nn_lock);
1530 	/*
1531 	 * see if we already have one pending or have given up.
1532 	 * For nn_timeout, it is set when we close the connection
1533 	 * because of the idle time out. So it means that we have
1534 	 * at least connected to that node successfully once,
1535 	 * now try to connect to it again.
1536 	 */
1537 	timeout = atomic_read(&nn->nn_timeout);
1538 	stop = (nn->nn_sc ||
1539 		(nn->nn_persistent_error &&
1540 		(nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1541 	spin_unlock(&nn->nn_lock);
1542 	if (stop)
1543 		goto out;
1544 
1545 	nn->nn_last_connect_attempt = jiffies;
1546 
1547 	sc = sc_alloc(node);
1548 	if (sc == NULL) {
1549 		mlog(0, "couldn't allocate sc\n");
1550 		ret = -ENOMEM;
1551 		goto out;
1552 	}
1553 
1554 	ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1555 	if (ret < 0) {
1556 		mlog(0, "can't create socket: %d\n", ret);
1557 		goto out;
1558 	}
1559 	sc->sc_sock = sock; /* freed by sc_kref_release */
1560 
1561 	sock->sk->sk_allocation = GFP_ATOMIC;
1562 
1563 	myaddr.sin_family = AF_INET;
1564 	myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1565 	myaddr.sin_port = htons(0); /* any port */
1566 
1567 	ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1568 			      sizeof(myaddr));
1569 	if (ret) {
1570 		mlog(ML_ERROR, "bind failed with %d at address %u.%u.%u.%u\n",
1571 		     ret, NIPQUAD(mynode->nd_ipv4_address));
1572 		goto out;
1573 	}
1574 
1575 	ret = o2net_set_nodelay(sc->sc_sock);
1576 	if (ret) {
1577 		mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1578 		goto out;
1579 	}
1580 
1581 	o2net_register_callbacks(sc->sc_sock->sk, sc);
1582 
1583 	spin_lock(&nn->nn_lock);
1584 	/* handshake completion will set nn->nn_sc_valid */
1585 	o2net_set_nn_state(nn, sc, 0, 0);
1586 	spin_unlock(&nn->nn_lock);
1587 
1588 	remoteaddr.sin_family = AF_INET;
1589 	remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1590 	remoteaddr.sin_port = node->nd_ipv4_port;
1591 
1592 	ret = sc->sc_sock->ops->connect(sc->sc_sock,
1593 					(struct sockaddr *)&remoteaddr,
1594 					sizeof(remoteaddr),
1595 					O_NONBLOCK);
1596 	if (ret == -EINPROGRESS)
1597 		ret = 0;
1598 
1599 out:
1600 	if (ret) {
1601 		mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1602 		     "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1603 		/* 0 err so that another will be queued and attempted
1604 		 * from set_nn_state */
1605 		if (sc)
1606 			o2net_ensure_shutdown(nn, sc, 0);
1607 	}
1608 	if (sc)
1609 		sc_put(sc);
1610 	if (node)
1611 		o2nm_node_put(node);
1612 	if (mynode)
1613 		o2nm_node_put(mynode);
1614 
1615 	return;
1616 }
1617 
1618 static void o2net_connect_expired(struct work_struct *work)
1619 {
1620 	struct o2net_node *nn =
1621 		container_of(work, struct o2net_node, nn_connect_expired.work);
1622 
1623 	spin_lock(&nn->nn_lock);
1624 	if (!nn->nn_sc_valid) {
1625 		mlog(ML_ERROR, "no connection established with node %u after "
1626 		     "%u.%u seconds, giving up and returning errors.\n",
1627 		     o2net_num_from_nn(nn),
1628 		     o2net_idle_timeout() / 1000,
1629 		     o2net_idle_timeout() % 1000);
1630 
1631 		o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1632 	}
1633 	spin_unlock(&nn->nn_lock);
1634 }
1635 
1636 static void o2net_still_up(struct work_struct *work)
1637 {
1638 	struct o2net_node *nn =
1639 		container_of(work, struct o2net_node, nn_still_up.work);
1640 
1641 	o2quo_hb_still_up(o2net_num_from_nn(nn));
1642 }
1643 
1644 /* ------------------------------------------------------------ */
1645 
1646 void o2net_disconnect_node(struct o2nm_node *node)
1647 {
1648 	struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1649 
1650 	/* don't reconnect until it's heartbeating again */
1651 	spin_lock(&nn->nn_lock);
1652 	atomic_set(&nn->nn_timeout, 0);
1653 	o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1654 	spin_unlock(&nn->nn_lock);
1655 
1656 	if (o2net_wq) {
1657 		cancel_delayed_work(&nn->nn_connect_expired);
1658 		cancel_delayed_work(&nn->nn_connect_work);
1659 		cancel_delayed_work(&nn->nn_still_up);
1660 		flush_workqueue(o2net_wq);
1661 	}
1662 }
1663 
1664 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1665 				  void *data)
1666 {
1667 	o2quo_hb_down(node_num);
1668 
1669 	if (node_num != o2nm_this_node())
1670 		o2net_disconnect_node(node);
1671 
1672 	BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1673 }
1674 
1675 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1676 				void *data)
1677 {
1678 	struct o2net_node *nn = o2net_nn_from_num(node_num);
1679 
1680 	o2quo_hb_up(node_num);
1681 
1682 	/* ensure an immediate connect attempt */
1683 	nn->nn_last_connect_attempt = jiffies -
1684 		(msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1685 
1686 	if (node_num != o2nm_this_node()) {
1687 		/* believe it or not, accept and node hearbeating testing
1688 		 * can succeed for this node before we got here.. so
1689 		 * only use set_nn_state to clear the persistent error
1690 		 * if that hasn't already happened */
1691 		spin_lock(&nn->nn_lock);
1692 		atomic_set(&nn->nn_timeout, 0);
1693 		if (nn->nn_persistent_error)
1694 			o2net_set_nn_state(nn, NULL, 0, 0);
1695 		spin_unlock(&nn->nn_lock);
1696 	}
1697 }
1698 
1699 void o2net_unregister_hb_callbacks(void)
1700 {
1701 	o2hb_unregister_callback(NULL, &o2net_hb_up);
1702 	o2hb_unregister_callback(NULL, &o2net_hb_down);
1703 }
1704 
1705 int o2net_register_hb_callbacks(void)
1706 {
1707 	int ret;
1708 
1709 	o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1710 			    o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1711 	o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1712 			    o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1713 
1714 	ret = o2hb_register_callback(NULL, &o2net_hb_up);
1715 	if (ret == 0)
1716 		ret = o2hb_register_callback(NULL, &o2net_hb_down);
1717 
1718 	if (ret)
1719 		o2net_unregister_hb_callbacks();
1720 
1721 	return ret;
1722 }
1723 
1724 /* ------------------------------------------------------------ */
1725 
1726 static int o2net_accept_one(struct socket *sock)
1727 {
1728 	int ret, slen;
1729 	struct sockaddr_in sin;
1730 	struct socket *new_sock = NULL;
1731 	struct o2nm_node *node = NULL;
1732 	struct o2net_sock_container *sc = NULL;
1733 	struct o2net_node *nn;
1734 
1735 	BUG_ON(sock == NULL);
1736 	ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1737 			       sock->sk->sk_protocol, &new_sock);
1738 	if (ret)
1739 		goto out;
1740 
1741 	new_sock->type = sock->type;
1742 	new_sock->ops = sock->ops;
1743 	ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1744 	if (ret < 0)
1745 		goto out;
1746 
1747 	new_sock->sk->sk_allocation = GFP_ATOMIC;
1748 
1749 	ret = o2net_set_nodelay(new_sock);
1750 	if (ret) {
1751 		mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1752 		goto out;
1753 	}
1754 
1755 	slen = sizeof(sin);
1756 	ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1757 				       &slen, 1);
1758 	if (ret < 0)
1759 		goto out;
1760 
1761 	node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1762 	if (node == NULL) {
1763 		mlog(ML_NOTICE, "attempt to connect from unknown node at "
1764 		     "%u.%u.%u.%u:%d\n", NIPQUAD(sin.sin_addr.s_addr),
1765 		     ntohs(sin.sin_port));
1766 		ret = -EINVAL;
1767 		goto out;
1768 	}
1769 
1770 	if (o2nm_this_node() > node->nd_num) {
1771 		mlog(ML_NOTICE, "unexpected connect attempted from a lower "
1772 		     "numbered node '%s' at " "%u.%u.%u.%u:%d with num %u\n",
1773 		     node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1774 		     ntohs(sin.sin_port), node->nd_num);
1775 		ret = -EINVAL;
1776 		goto out;
1777 	}
1778 
1779 	/* this happens all the time when the other node sees our heartbeat
1780 	 * and tries to connect before we see their heartbeat */
1781 	if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1782 		mlog(ML_CONN, "attempt to connect from node '%s' at "
1783 		     "%u.%u.%u.%u:%d but it isn't heartbeating\n",
1784 		     node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1785 		     ntohs(sin.sin_port));
1786 		ret = -EINVAL;
1787 		goto out;
1788 	}
1789 
1790 	nn = o2net_nn_from_num(node->nd_num);
1791 
1792 	spin_lock(&nn->nn_lock);
1793 	if (nn->nn_sc)
1794 		ret = -EBUSY;
1795 	else
1796 		ret = 0;
1797 	spin_unlock(&nn->nn_lock);
1798 	if (ret) {
1799 		mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1800 		     "%u.%u.%u.%u:%d but it already has an open connection\n",
1801 		     node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1802 		     ntohs(sin.sin_port));
1803 		goto out;
1804 	}
1805 
1806 	sc = sc_alloc(node);
1807 	if (sc == NULL) {
1808 		ret = -ENOMEM;
1809 		goto out;
1810 	}
1811 
1812 	sc->sc_sock = new_sock;
1813 	new_sock = NULL;
1814 
1815 	spin_lock(&nn->nn_lock);
1816 	atomic_set(&nn->nn_timeout, 0);
1817 	o2net_set_nn_state(nn, sc, 0, 0);
1818 	spin_unlock(&nn->nn_lock);
1819 
1820 	o2net_register_callbacks(sc->sc_sock->sk, sc);
1821 	o2net_sc_queue_work(sc, &sc->sc_rx_work);
1822 
1823 	o2net_initialize_handshake();
1824 	o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1825 
1826 out:
1827 	if (new_sock)
1828 		sock_release(new_sock);
1829 	if (node)
1830 		o2nm_node_put(node);
1831 	if (sc)
1832 		sc_put(sc);
1833 	return ret;
1834 }
1835 
1836 static void o2net_accept_many(struct work_struct *work)
1837 {
1838 	struct socket *sock = o2net_listen_sock;
1839 	while (o2net_accept_one(sock) == 0)
1840 		cond_resched();
1841 }
1842 
1843 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1844 {
1845 	void (*ready)(struct sock *sk, int bytes);
1846 
1847 	read_lock(&sk->sk_callback_lock);
1848 	ready = sk->sk_user_data;
1849 	if (ready == NULL) { /* check for teardown race */
1850 		ready = sk->sk_data_ready;
1851 		goto out;
1852 	}
1853 
1854 	/* ->sk_data_ready is also called for a newly established child socket
1855 	 * before it has been accepted and the acceptor has set up their
1856 	 * data_ready.. we only want to queue listen work for our listening
1857 	 * socket */
1858 	if (sk->sk_state == TCP_LISTEN) {
1859 		mlog(ML_TCP, "bytes: %d\n", bytes);
1860 		queue_work(o2net_wq, &o2net_listen_work);
1861 	}
1862 
1863 out:
1864 	read_unlock(&sk->sk_callback_lock);
1865 	ready(sk, bytes);
1866 }
1867 
1868 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1869 {
1870 	struct socket *sock = NULL;
1871 	int ret;
1872 	struct sockaddr_in sin = {
1873 		.sin_family = PF_INET,
1874 		.sin_addr = { .s_addr = addr },
1875 		.sin_port = port,
1876 	};
1877 
1878 	ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1879 	if (ret < 0) {
1880 		mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1881 		goto out;
1882 	}
1883 
1884 	sock->sk->sk_allocation = GFP_ATOMIC;
1885 
1886 	write_lock_bh(&sock->sk->sk_callback_lock);
1887 	sock->sk->sk_user_data = sock->sk->sk_data_ready;
1888 	sock->sk->sk_data_ready = o2net_listen_data_ready;
1889 	write_unlock_bh(&sock->sk->sk_callback_lock);
1890 
1891 	o2net_listen_sock = sock;
1892 	INIT_WORK(&o2net_listen_work, o2net_accept_many);
1893 
1894 	sock->sk->sk_reuse = 1;
1895 	ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1896 	if (ret < 0) {
1897 		mlog(ML_ERROR, "unable to bind socket at %u.%u.%u.%u:%u, "
1898 		     "ret=%d\n", NIPQUAD(addr), ntohs(port), ret);
1899 		goto out;
1900 	}
1901 
1902 	ret = sock->ops->listen(sock, 64);
1903 	if (ret < 0) {
1904 		mlog(ML_ERROR, "unable to listen on %u.%u.%u.%u:%u, ret=%d\n",
1905 		     NIPQUAD(addr), ntohs(port), ret);
1906 	}
1907 
1908 out:
1909 	if (ret) {
1910 		o2net_listen_sock = NULL;
1911 		if (sock)
1912 			sock_release(sock);
1913 	}
1914 	return ret;
1915 }
1916 
1917 /*
1918  * called from node manager when we should bring up our network listening
1919  * socket.  node manager handles all the serialization to only call this
1920  * once and to match it with o2net_stop_listening().  note,
1921  * o2nm_this_node() doesn't work yet as we're being called while it
1922  * is being set up.
1923  */
1924 int o2net_start_listening(struct o2nm_node *node)
1925 {
1926 	int ret = 0;
1927 
1928 	BUG_ON(o2net_wq != NULL);
1929 	BUG_ON(o2net_listen_sock != NULL);
1930 
1931 	mlog(ML_KTHREAD, "starting o2net thread...\n");
1932 	o2net_wq = create_singlethread_workqueue("o2net");
1933 	if (o2net_wq == NULL) {
1934 		mlog(ML_ERROR, "unable to launch o2net thread\n");
1935 		return -ENOMEM; /* ? */
1936 	}
1937 
1938 	ret = o2net_open_listening_sock(node->nd_ipv4_address,
1939 					node->nd_ipv4_port);
1940 	if (ret) {
1941 		destroy_workqueue(o2net_wq);
1942 		o2net_wq = NULL;
1943 	} else
1944 		o2quo_conn_up(node->nd_num);
1945 
1946 	return ret;
1947 }
1948 
1949 /* again, o2nm_this_node() doesn't work here as we're involved in
1950  * tearing it down */
1951 void o2net_stop_listening(struct o2nm_node *node)
1952 {
1953 	struct socket *sock = o2net_listen_sock;
1954 	size_t i;
1955 
1956 	BUG_ON(o2net_wq == NULL);
1957 	BUG_ON(o2net_listen_sock == NULL);
1958 
1959 	/* stop the listening socket from generating work */
1960 	write_lock_bh(&sock->sk->sk_callback_lock);
1961 	sock->sk->sk_data_ready = sock->sk->sk_user_data;
1962 	sock->sk->sk_user_data = NULL;
1963 	write_unlock_bh(&sock->sk->sk_callback_lock);
1964 
1965 	for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1966 		struct o2nm_node *node = o2nm_get_node_by_num(i);
1967 		if (node) {
1968 			o2net_disconnect_node(node);
1969 			o2nm_node_put(node);
1970 		}
1971 	}
1972 
1973 	/* finish all work and tear down the work queue */
1974 	mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
1975 	destroy_workqueue(o2net_wq);
1976 	o2net_wq = NULL;
1977 
1978 	sock_release(o2net_listen_sock);
1979 	o2net_listen_sock = NULL;
1980 
1981 	o2quo_conn_err(node->nd_num);
1982 }
1983 
1984 /* ------------------------------------------------------------ */
1985 
1986 int o2net_init(void)
1987 {
1988 	unsigned long i;
1989 
1990 	o2quo_init();
1991 
1992 	if (o2net_debugfs_init())
1993 		return -ENOMEM;
1994 
1995 	o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
1996 	o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1997 	o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1998 	if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
1999 		kfree(o2net_hand);
2000 		kfree(o2net_keep_req);
2001 		kfree(o2net_keep_resp);
2002 		return -ENOMEM;
2003 	}
2004 
2005 	o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2006 	o2net_hand->connector_id = cpu_to_be64(1);
2007 
2008 	o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2009 	o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2010 
2011 	for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2012 		struct o2net_node *nn = o2net_nn_from_num(i);
2013 
2014 		atomic_set(&nn->nn_timeout, 0);
2015 		spin_lock_init(&nn->nn_lock);
2016 		INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2017 		INIT_DELAYED_WORK(&nn->nn_connect_expired,
2018 				  o2net_connect_expired);
2019 		INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2020 		/* until we see hb from a node we'll return einval */
2021 		nn->nn_persistent_error = -ENOTCONN;
2022 		init_waitqueue_head(&nn->nn_sc_wq);
2023 		idr_init(&nn->nn_status_idr);
2024 		INIT_LIST_HEAD(&nn->nn_status_list);
2025 	}
2026 
2027 	return 0;
2028 }
2029 
2030 void o2net_exit(void)
2031 {
2032 	o2quo_exit();
2033 	kfree(o2net_hand);
2034 	kfree(o2net_keep_req);
2035 	kfree(o2net_keep_resp);
2036 	o2net_debugfs_exit();
2037 }
2038