ocfs2: Reconnect after idle time out.
[linux-2.6.git] / fs / ocfs2 / cluster / tcp.c
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 /*
146  * FIXME: These should use to_o2nm_cluster_from_node(), but we end up
147  * losing our parent link to the cluster during shutdown. This can be
148  * solved by adding a pre-removal callback to configfs, or passing
149  * around the cluster with the node. -jeffm
150  */
151 static inline int o2net_reconnect_delay(struct o2nm_node *node)
152 {
153         return o2nm_single_cluster->cl_reconnect_delay_ms;
154 }
155
156 static inline int o2net_keepalive_delay(struct o2nm_node *node)
157 {
158         return o2nm_single_cluster->cl_keepalive_delay_ms;
159 }
160
161 static inline int o2net_idle_timeout(struct o2nm_node *node)
162 {
163         return o2nm_single_cluster->cl_idle_timeout_ms;
164 }
165
166 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
167 {
168         int trans;
169         BUG_ON(err >= O2NET_ERR_MAX);
170         trans = o2net_sys_err_translations[err];
171
172         /* Just in case we mess up the translation table above */
173         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
174         return trans;
175 }
176
177 static struct o2net_node * o2net_nn_from_num(u8 node_num)
178 {
179         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
180         return &o2net_nodes[node_num];
181 }
182
183 static u8 o2net_num_from_nn(struct o2net_node *nn)
184 {
185         BUG_ON(nn == NULL);
186         return nn - o2net_nodes;
187 }
188
189 /* ------------------------------------------------------------ */
190
191 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
192 {
193         int ret = 0;
194
195         do {
196                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
197                         ret = -EAGAIN;
198                         break;
199                 }
200                 spin_lock(&nn->nn_lock);
201                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
202                 if (ret == 0)
203                         list_add_tail(&nsw->ns_node_item,
204                                       &nn->nn_status_list);
205                 spin_unlock(&nn->nn_lock);
206         } while (ret == -EAGAIN);
207
208         if (ret == 0)  {
209                 init_waitqueue_head(&nsw->ns_wq);
210                 nsw->ns_sys_status = O2NET_ERR_NONE;
211                 nsw->ns_status = 0;
212         }
213
214         return ret;
215 }
216
217 static void o2net_complete_nsw_locked(struct o2net_node *nn,
218                                       struct o2net_status_wait *nsw,
219                                       enum o2net_system_error sys_status,
220                                       s32 status)
221 {
222         assert_spin_locked(&nn->nn_lock);
223
224         if (!list_empty(&nsw->ns_node_item)) {
225                 list_del_init(&nsw->ns_node_item);
226                 nsw->ns_sys_status = sys_status;
227                 nsw->ns_status = status;
228                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
229                 wake_up(&nsw->ns_wq);
230         }
231 }
232
233 static void o2net_complete_nsw(struct o2net_node *nn,
234                                struct o2net_status_wait *nsw,
235                                u64 id, enum o2net_system_error sys_status,
236                                s32 status)
237 {
238         spin_lock(&nn->nn_lock);
239         if (nsw == NULL) {
240                 if (id > INT_MAX)
241                         goto out;
242
243                 nsw = idr_find(&nn->nn_status_idr, id);
244                 if (nsw == NULL)
245                         goto out;
246         }
247
248         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
249
250 out:
251         spin_unlock(&nn->nn_lock);
252         return;
253 }
254
255 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
256 {
257         struct o2net_status_wait *nsw, *tmp;
258         unsigned int num_kills = 0;
259
260         assert_spin_locked(&nn->nn_lock);
261
262         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
263                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
264                 num_kills++;
265         }
266
267         mlog(0, "completed %d messages for node %u\n", num_kills,
268              o2net_num_from_nn(nn));
269 }
270
271 static int o2net_nsw_completed(struct o2net_node *nn,
272                                struct o2net_status_wait *nsw)
273 {
274         int completed;
275         spin_lock(&nn->nn_lock);
276         completed = list_empty(&nsw->ns_node_item);
277         spin_unlock(&nn->nn_lock);
278         return completed;
279 }
280
281 /* ------------------------------------------------------------ */
282
283 static void sc_kref_release(struct kref *kref)
284 {
285         struct o2net_sock_container *sc = container_of(kref,
286                                         struct o2net_sock_container, sc_kref);
287         BUG_ON(timer_pending(&sc->sc_idle_timeout));
288
289         sclog(sc, "releasing\n");
290
291         if (sc->sc_sock) {
292                 sock_release(sc->sc_sock);
293                 sc->sc_sock = NULL;
294         }
295
296         o2nm_node_put(sc->sc_node);
297         sc->sc_node = NULL;
298
299         kfree(sc);
300 }
301
302 static void sc_put(struct o2net_sock_container *sc)
303 {
304         sclog(sc, "put\n");
305         kref_put(&sc->sc_kref, sc_kref_release);
306 }
307 static void sc_get(struct o2net_sock_container *sc)
308 {
309         sclog(sc, "get\n");
310         kref_get(&sc->sc_kref);
311 }
312 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
313 {
314         struct o2net_sock_container *sc, *ret = NULL;
315         struct page *page = NULL;
316
317         page = alloc_page(GFP_NOFS);
318         sc = kzalloc(sizeof(*sc), GFP_NOFS);
319         if (sc == NULL || page == NULL)
320                 goto out;
321
322         kref_init(&sc->sc_kref);
323         o2nm_node_get(node);
324         sc->sc_node = node;
325
326         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
327         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
328         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
329         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
330
331         init_timer(&sc->sc_idle_timeout);
332         sc->sc_idle_timeout.function = o2net_idle_timer;
333         sc->sc_idle_timeout.data = (unsigned long)sc;
334
335         sclog(sc, "alloced\n");
336
337         ret = sc;
338         sc->sc_page = page;
339         sc = NULL;
340         page = NULL;
341
342 out:
343         if (page)
344                 __free_page(page);
345         kfree(sc);
346
347         return ret;
348 }
349
350 /* ------------------------------------------------------------ */
351
352 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
353                                 struct work_struct *work)
354 {
355         sc_get(sc);
356         if (!queue_work(o2net_wq, work))
357                 sc_put(sc);
358 }
359 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
360                                         struct delayed_work *work,
361                                         int delay)
362 {
363         sc_get(sc);
364         if (!queue_delayed_work(o2net_wq, work, delay))
365                 sc_put(sc);
366 }
367 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
368                                          struct delayed_work *work)
369 {
370         if (cancel_delayed_work(work))
371                 sc_put(sc);
372 }
373
374 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
375
376 int o2net_num_connected_peers(void)
377 {
378         return atomic_read(&o2net_connected_peers);
379 }
380
381 static void o2net_set_nn_state(struct o2net_node *nn,
382                                struct o2net_sock_container *sc,
383                                unsigned valid, int err)
384 {
385         int was_valid = nn->nn_sc_valid;
386         int was_err = nn->nn_persistent_error;
387         struct o2net_sock_container *old_sc = nn->nn_sc;
388
389         assert_spin_locked(&nn->nn_lock);
390
391         if (old_sc && !sc)
392                 atomic_dec(&o2net_connected_peers);
393         else if (!old_sc && sc)
394                 atomic_inc(&o2net_connected_peers);
395
396         /* the node num comparison and single connect/accept path should stop
397          * an non-null sc from being overwritten with another */
398         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
399         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
400         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
401
402         if (was_valid && !valid && err == 0)
403                 err = -ENOTCONN;
404
405         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
406              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
407              nn->nn_persistent_error, err);
408
409         nn->nn_sc = sc;
410         nn->nn_sc_valid = valid ? 1 : 0;
411         nn->nn_persistent_error = err;
412
413         /* mirrors o2net_tx_can_proceed() */
414         if (nn->nn_persistent_error || nn->nn_sc_valid)
415                 wake_up(&nn->nn_sc_wq);
416
417         if (!was_err && nn->nn_persistent_error) {
418                 o2quo_conn_err(o2net_num_from_nn(nn));
419                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
420                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
421         }
422
423         if (was_valid && !valid) {
424                 printk(KERN_INFO "o2net: no longer connected to "
425                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
426                 o2net_complete_nodes_nsw(nn);
427         }
428
429         if (!was_valid && valid) {
430                 o2quo_conn_up(o2net_num_from_nn(nn));
431                 cancel_delayed_work(&nn->nn_connect_expired);
432                 printk(KERN_INFO "o2net: %s " SC_NODEF_FMT "\n",
433                        o2nm_this_node() > sc->sc_node->nd_num ?
434                                 "connected to" : "accepted connection from",
435                        SC_NODEF_ARGS(sc));
436         }
437
438         /* trigger the connecting worker func as long as we're not valid,
439          * it will back off if it shouldn't connect.  This can be called
440          * from node config teardown and so needs to be careful about
441          * the work queue actually being up. */
442         if (!valid && o2net_wq) {
443                 unsigned long delay;
444                 /* delay if we're withing a RECONNECT_DELAY of the
445                  * last attempt */
446                 delay = (nn->nn_last_connect_attempt +
447                          msecs_to_jiffies(o2net_reconnect_delay(NULL)))
448                         - jiffies;
449                 if (delay > msecs_to_jiffies(o2net_reconnect_delay(NULL)))
450                         delay = 0;
451                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
452                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
453
454                 /*
455                  * Delay the expired work after idle timeout.
456                  *
457                  * We might have lots of failed connection attempts that run
458                  * through here but we only cancel the connect_expired work when
459                  * a connection attempt succeeds.  So only the first enqueue of
460                  * the connect_expired work will do anything.  The rest will see
461                  * that it's already queued and do nothing.
462                  */
463                 delay += msecs_to_jiffies(o2net_idle_timeout(NULL));
464                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
465         }
466
467         /* keep track of the nn's sc ref for the caller */
468         if ((old_sc == NULL) && sc)
469                 sc_get(sc);
470         if (old_sc && (old_sc != sc)) {
471                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
472                 sc_put(old_sc);
473         }
474 }
475
476 /* see o2net_register_callbacks() */
477 static void o2net_data_ready(struct sock *sk, int bytes)
478 {
479         void (*ready)(struct sock *sk, int bytes);
480
481         read_lock(&sk->sk_callback_lock);
482         if (sk->sk_user_data) {
483                 struct o2net_sock_container *sc = sk->sk_user_data;
484                 sclog(sc, "data_ready hit\n");
485                 do_gettimeofday(&sc->sc_tv_data_ready);
486                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
487                 ready = sc->sc_data_ready;
488         } else {
489                 ready = sk->sk_data_ready;
490         }
491         read_unlock(&sk->sk_callback_lock);
492
493         ready(sk, bytes);
494 }
495
496 /* see o2net_register_callbacks() */
497 static void o2net_state_change(struct sock *sk)
498 {
499         void (*state_change)(struct sock *sk);
500         struct o2net_sock_container *sc;
501
502         read_lock(&sk->sk_callback_lock);
503         sc = sk->sk_user_data;
504         if (sc == NULL) {
505                 state_change = sk->sk_state_change;
506                 goto out;
507         }
508
509         sclog(sc, "state_change to %d\n", sk->sk_state);
510
511         state_change = sc->sc_state_change;
512
513         switch(sk->sk_state) {
514                 /* ignore connecting sockets as they make progress */
515                 case TCP_SYN_SENT:
516                 case TCP_SYN_RECV:
517                         break;
518                 case TCP_ESTABLISHED:
519                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
520                         break;
521                 default:
522                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
523                         break;
524         }
525 out:
526         read_unlock(&sk->sk_callback_lock);
527         state_change(sk);
528 }
529
530 /*
531  * we register callbacks so we can queue work on events before calling
532  * the original callbacks.  our callbacks our careful to test user_data
533  * to discover when they've reaced with o2net_unregister_callbacks().
534  */
535 static void o2net_register_callbacks(struct sock *sk,
536                                      struct o2net_sock_container *sc)
537 {
538         write_lock_bh(&sk->sk_callback_lock);
539
540         /* accepted sockets inherit the old listen socket data ready */
541         if (sk->sk_data_ready == o2net_listen_data_ready) {
542                 sk->sk_data_ready = sk->sk_user_data;
543                 sk->sk_user_data = NULL;
544         }
545
546         BUG_ON(sk->sk_user_data != NULL);
547         sk->sk_user_data = sc;
548         sc_get(sc);
549
550         sc->sc_data_ready = sk->sk_data_ready;
551         sc->sc_state_change = sk->sk_state_change;
552         sk->sk_data_ready = o2net_data_ready;
553         sk->sk_state_change = o2net_state_change;
554
555         mutex_init(&sc->sc_send_lock);
556
557         write_unlock_bh(&sk->sk_callback_lock);
558 }
559
560 static int o2net_unregister_callbacks(struct sock *sk,
561                                    struct o2net_sock_container *sc)
562 {
563         int ret = 0;
564
565         write_lock_bh(&sk->sk_callback_lock);
566         if (sk->sk_user_data == sc) {
567                 ret = 1;
568                 sk->sk_user_data = NULL;
569                 sk->sk_data_ready = sc->sc_data_ready;
570                 sk->sk_state_change = sc->sc_state_change;
571         }
572         write_unlock_bh(&sk->sk_callback_lock);
573
574         return ret;
575 }
576
577 /*
578  * this is a little helper that is called by callers who have seen a problem
579  * with an sc and want to detach it from the nn if someone already hasn't beat
580  * them to it.  if an error is given then the shutdown will be persistent
581  * and pending transmits will be canceled.
582  */
583 static void o2net_ensure_shutdown(struct o2net_node *nn,
584                                    struct o2net_sock_container *sc,
585                                    int err)
586 {
587         spin_lock(&nn->nn_lock);
588         if (nn->nn_sc == sc)
589                 o2net_set_nn_state(nn, NULL, 0, err);
590         spin_unlock(&nn->nn_lock);
591 }
592
593 /*
594  * This work queue function performs the blocking parts of socket shutdown.  A
595  * few paths lead here.  set_nn_state will trigger this callback if it sees an
596  * sc detached from the nn.  state_change will also trigger this callback
597  * directly when it sees errors.  In that case we need to call set_nn_state
598  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
599  * itself.
600  */
601 static void o2net_shutdown_sc(struct work_struct *work)
602 {
603         struct o2net_sock_container *sc =
604                 container_of(work, struct o2net_sock_container,
605                              sc_shutdown_work);
606         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
607
608         sclog(sc, "shutting down\n");
609
610         /* drop the callbacks ref and call shutdown only once */
611         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
612                 /* we shouldn't flush as we're in the thread, the
613                  * races with pending sc work structs are harmless */
614                 del_timer_sync(&sc->sc_idle_timeout);
615                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
616                 sc_put(sc);
617                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
618         }
619
620         /* not fatal so failed connects before the other guy has our
621          * heartbeat can be retried */
622         o2net_ensure_shutdown(nn, sc, 0);
623         sc_put(sc);
624 }
625
626 /* ------------------------------------------------------------ */
627
628 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
629                              u32 key)
630 {
631         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
632
633         if (ret == 0)
634                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
635
636         return ret;
637 }
638
639 static struct o2net_msg_handler *
640 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
641                           struct rb_node **ret_parent)
642 {
643         struct rb_node **p = &o2net_handler_tree.rb_node;
644         struct rb_node *parent = NULL;
645         struct o2net_msg_handler *nmh, *ret = NULL;
646         int cmp;
647
648         while (*p) {
649                 parent = *p;
650                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
651                 cmp = o2net_handler_cmp(nmh, msg_type, key);
652
653                 if (cmp < 0)
654                         p = &(*p)->rb_left;
655                 else if (cmp > 0)
656                         p = &(*p)->rb_right;
657                 else {
658                         ret = nmh;
659                         break;
660                 }
661         }
662
663         if (ret_p != NULL)
664                 *ret_p = p;
665         if (ret_parent != NULL)
666                 *ret_parent = parent;
667
668         return ret;
669 }
670
671 static void o2net_handler_kref_release(struct kref *kref)
672 {
673         struct o2net_msg_handler *nmh;
674         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
675
676         kfree(nmh);
677 }
678
679 static void o2net_handler_put(struct o2net_msg_handler *nmh)
680 {
681         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
682 }
683
684 /* max_len is protection for the handler func.  incoming messages won't
685  * be given to the handler if their payload is longer than the max. */
686 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
687                            o2net_msg_handler_func *func, void *data,
688                            o2net_post_msg_handler_func *post_func,
689                            struct list_head *unreg_list)
690 {
691         struct o2net_msg_handler *nmh = NULL;
692         struct rb_node **p, *parent;
693         int ret = 0;
694
695         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
696                 mlog(0, "max_len for message handler out of range: %u\n",
697                         max_len);
698                 ret = -EINVAL;
699                 goto out;
700         }
701
702         if (!msg_type) {
703                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
704                 ret = -EINVAL;
705                 goto out;
706
707         }
708         if (!func) {
709                 mlog(0, "no message handler provided: %u, %p\n",
710                        msg_type, func);
711                 ret = -EINVAL;
712                 goto out;
713         }
714
715         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
716         if (nmh == NULL) {
717                 ret = -ENOMEM;
718                 goto out;
719         }
720
721         nmh->nh_func = func;
722         nmh->nh_func_data = data;
723         nmh->nh_post_func = post_func;
724         nmh->nh_msg_type = msg_type;
725         nmh->nh_max_len = max_len;
726         nmh->nh_key = key;
727         /* the tree and list get this ref.. they're both removed in
728          * unregister when this ref is dropped */
729         kref_init(&nmh->nh_kref);
730         INIT_LIST_HEAD(&nmh->nh_unregister_item);
731
732         write_lock(&o2net_handler_lock);
733         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
734                 ret = -EEXIST;
735         else {
736                 rb_link_node(&nmh->nh_node, parent, p);
737                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
738                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
739
740                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
741                      func, msg_type, key);
742                 /* we've had some trouble with handlers seemingly vanishing. */
743                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
744                                                           &parent) == NULL,
745                                 "couldn't find handler we *just* registerd "
746                                 "for type %u key %08x\n", msg_type, key);
747         }
748         write_unlock(&o2net_handler_lock);
749         if (ret)
750                 goto out;
751
752 out:
753         if (ret)
754                 kfree(nmh);
755
756         return ret;
757 }
758 EXPORT_SYMBOL_GPL(o2net_register_handler);
759
760 void o2net_unregister_handler_list(struct list_head *list)
761 {
762         struct o2net_msg_handler *nmh, *n;
763
764         write_lock(&o2net_handler_lock);
765         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
766                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
767                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
768                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
769                 list_del_init(&nmh->nh_unregister_item);
770                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
771         }
772         write_unlock(&o2net_handler_lock);
773 }
774 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
775
776 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
777 {
778         struct o2net_msg_handler *nmh;
779
780         read_lock(&o2net_handler_lock);
781         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
782         if (nmh)
783                 kref_get(&nmh->nh_kref);
784         read_unlock(&o2net_handler_lock);
785
786         return nmh;
787 }
788
789 /* ------------------------------------------------------------ */
790
791 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
792 {
793         int ret;
794         mm_segment_t oldfs;
795         struct kvec vec = {
796                 .iov_len = len,
797                 .iov_base = data,
798         };
799         struct msghdr msg = {
800                 .msg_iovlen = 1,
801                 .msg_iov = (struct iovec *)&vec,
802                 .msg_flags = MSG_DONTWAIT,
803         };
804
805         oldfs = get_fs();
806         set_fs(get_ds());
807         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
808         set_fs(oldfs);
809
810         return ret;
811 }
812
813 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
814                               size_t veclen, size_t total)
815 {
816         int ret;
817         mm_segment_t oldfs;
818         struct msghdr msg = {
819                 .msg_iov = (struct iovec *)vec,
820                 .msg_iovlen = veclen,
821         };
822
823         if (sock == NULL) {
824                 ret = -EINVAL;
825                 goto out;
826         }
827
828         oldfs = get_fs();
829         set_fs(get_ds());
830         ret = sock_sendmsg(sock, &msg, total);
831         set_fs(oldfs);
832         if (ret != total) {
833                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
834                      total);
835                 if (ret >= 0)
836                         ret = -EPIPE; /* should be smarter, I bet */
837                 goto out;
838         }
839
840         ret = 0;
841 out:
842         if (ret < 0)
843                 mlog(0, "returning error: %d\n", ret);
844         return ret;
845 }
846
847 static void o2net_sendpage(struct o2net_sock_container *sc,
848                            void *kmalloced_virt,
849                            size_t size)
850 {
851         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
852         ssize_t ret;
853
854         while (1) {
855                 mutex_lock(&sc->sc_send_lock);
856                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
857                                                  virt_to_page(kmalloced_virt),
858                                                  (long)kmalloced_virt & ~PAGE_MASK,
859                                                  size, MSG_DONTWAIT);
860                 mutex_unlock(&sc->sc_send_lock);
861                 if (ret == size)
862                         break;
863                 if (ret == (ssize_t)-EAGAIN) {
864                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
865                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
866                         cond_resched();
867                         continue;
868                 }
869                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT 
870                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
871                 o2net_ensure_shutdown(nn, sc, 0);
872                 break;
873         }
874 }
875
876 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
877 {
878         memset(msg, 0, sizeof(struct o2net_msg));
879         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
880         msg->data_len = cpu_to_be16(data_len);
881         msg->msg_type = cpu_to_be16(msg_type);
882         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
883         msg->status = 0;
884         msg->key = cpu_to_be32(key);
885 }
886
887 static int o2net_tx_can_proceed(struct o2net_node *nn,
888                                 struct o2net_sock_container **sc_ret,
889                                 int *error)
890 {
891         int ret = 0;
892
893         spin_lock(&nn->nn_lock);
894         if (nn->nn_persistent_error) {
895                 ret = 1;
896                 *sc_ret = NULL;
897                 *error = nn->nn_persistent_error;
898         } else if (nn->nn_sc_valid) {
899                 kref_get(&nn->nn_sc->sc_kref);
900
901                 ret = 1;
902                 *sc_ret = nn->nn_sc;
903                 *error = 0;
904         }
905         spin_unlock(&nn->nn_lock);
906
907         return ret;
908 }
909
910 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
911                            size_t caller_veclen, u8 target_node, int *status)
912 {
913         int ret, error = 0;
914         struct o2net_msg *msg = NULL;
915         size_t veclen, caller_bytes = 0;
916         struct kvec *vec = NULL;
917         struct o2net_sock_container *sc = NULL;
918         struct o2net_node *nn = o2net_nn_from_num(target_node);
919         struct o2net_status_wait nsw = {
920                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
921         };
922
923         if (o2net_wq == NULL) {
924                 mlog(0, "attempt to tx without o2netd running\n");
925                 ret = -ESRCH;
926                 goto out;
927         }
928
929         if (caller_veclen == 0) {
930                 mlog(0, "bad kvec array length\n");
931                 ret = -EINVAL;
932                 goto out;
933         }
934
935         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
936         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
937                 mlog(0, "total payload len %zu too large\n", caller_bytes);
938                 ret = -EINVAL;
939                 goto out;
940         }
941
942         if (target_node == o2nm_this_node()) {
943                 ret = -ELOOP;
944                 goto out;
945         }
946
947         ret = wait_event_interruptible(nn->nn_sc_wq,
948                                        o2net_tx_can_proceed(nn, &sc, &error));
949         if (!ret && error)
950                 ret = error;
951         if (ret)
952                 goto out;
953
954         veclen = caller_veclen + 1;
955         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
956         if (vec == NULL) {
957                 mlog(0, "failed to %zu element kvec!\n", veclen);
958                 ret = -ENOMEM;
959                 goto out;
960         }
961
962         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
963         if (!msg) {
964                 mlog(0, "failed to allocate a o2net_msg!\n");
965                 ret = -ENOMEM;
966                 goto out;
967         }
968
969         o2net_init_msg(msg, caller_bytes, msg_type, key);
970
971         vec[0].iov_len = sizeof(struct o2net_msg);
972         vec[0].iov_base = msg;
973         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
974
975         ret = o2net_prep_nsw(nn, &nsw);
976         if (ret)
977                 goto out;
978
979         msg->msg_num = cpu_to_be32(nsw.ns_id);
980
981         /* finally, convert the message header to network byte-order
982          * and send */
983         mutex_lock(&sc->sc_send_lock);
984         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
985                                  sizeof(struct o2net_msg) + caller_bytes);
986         mutex_unlock(&sc->sc_send_lock);
987         msglog(msg, "sending returned %d\n", ret);
988         if (ret < 0) {
989                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
990                 goto out;
991         }
992
993         /* wait on other node's handler */
994         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
995
996         /* Note that we avoid overwriting the callers status return
997          * variable if a system error was reported on the other
998          * side. Callers beware. */
999         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1000         if (status && !ret)
1001                 *status = nsw.ns_status;
1002
1003         mlog(0, "woken, returning system status %d, user status %d\n",
1004              ret, nsw.ns_status);
1005 out:
1006         if (sc)
1007                 sc_put(sc);
1008         if (vec)
1009                 kfree(vec);
1010         if (msg)
1011                 kfree(msg);
1012         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1013         return ret;
1014 }
1015 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1016
1017 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1018                        u8 target_node, int *status)
1019 {
1020         struct kvec vec = {
1021                 .iov_base = data,
1022                 .iov_len = len,
1023         };
1024         return o2net_send_message_vec(msg_type, key, &vec, 1,
1025                                       target_node, status);
1026 }
1027 EXPORT_SYMBOL_GPL(o2net_send_message);
1028
1029 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1030                                    enum o2net_system_error syserr, int err)
1031 {
1032         struct kvec vec = {
1033                 .iov_base = hdr,
1034                 .iov_len = sizeof(struct o2net_msg),
1035         };
1036
1037         BUG_ON(syserr >= O2NET_ERR_MAX);
1038
1039         /* leave other fields intact from the incoming message, msg_num
1040          * in particular */
1041         hdr->sys_status = cpu_to_be32(syserr);
1042         hdr->status = cpu_to_be32(err);
1043         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1044         hdr->data_len = 0;
1045
1046         msglog(hdr, "about to send status magic %d\n", err);
1047         /* hdr has been in host byteorder this whole time */
1048         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1049 }
1050
1051 /* this returns -errno if the header was unknown or too large, etc.
1052  * after this is called the buffer us reused for the next message */
1053 static int o2net_process_message(struct o2net_sock_container *sc,
1054                                  struct o2net_msg *hdr)
1055 {
1056         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1057         int ret = 0, handler_status;
1058         enum  o2net_system_error syserr;
1059         struct o2net_msg_handler *nmh = NULL;
1060         void *ret_data = NULL;
1061
1062         msglog(hdr, "processing message\n");
1063
1064         o2net_sc_postpone_idle(sc);
1065
1066         switch(be16_to_cpu(hdr->magic)) {
1067                 case O2NET_MSG_STATUS_MAGIC:
1068                         /* special type for returning message status */
1069                         o2net_complete_nsw(nn, NULL,
1070                                            be32_to_cpu(hdr->msg_num),
1071                                            be32_to_cpu(hdr->sys_status),
1072                                            be32_to_cpu(hdr->status));
1073                         goto out;
1074                 case O2NET_MSG_KEEP_REQ_MAGIC:
1075                         o2net_sendpage(sc, o2net_keep_resp,
1076                                        sizeof(*o2net_keep_resp));
1077                         goto out;
1078                 case O2NET_MSG_KEEP_RESP_MAGIC:
1079                         goto out;
1080                 case O2NET_MSG_MAGIC:
1081                         break;
1082                 default:
1083                         msglog(hdr, "bad magic\n");
1084                         ret = -EINVAL;
1085                         goto out;
1086                         break;
1087         }
1088
1089         /* find a handler for it */
1090         handler_status = 0;
1091         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1092                                 be32_to_cpu(hdr->key));
1093         if (!nmh) {
1094                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1095                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1096                 syserr = O2NET_ERR_NO_HNDLR;
1097                 goto out_respond;
1098         }
1099
1100         syserr = O2NET_ERR_NONE;
1101
1102         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1103                 syserr = O2NET_ERR_OVERFLOW;
1104
1105         if (syserr != O2NET_ERR_NONE)
1106                 goto out_respond;
1107
1108         do_gettimeofday(&sc->sc_tv_func_start);
1109         sc->sc_msg_key = be32_to_cpu(hdr->key);
1110         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1111         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1112                                              be16_to_cpu(hdr->data_len),
1113                                         nmh->nh_func_data, &ret_data);
1114         do_gettimeofday(&sc->sc_tv_func_stop);
1115
1116 out_respond:
1117         /* this destroys the hdr, so don't use it after this */
1118         mutex_lock(&sc->sc_send_lock);
1119         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1120                                       handler_status);
1121         mutex_unlock(&sc->sc_send_lock);
1122         hdr = NULL;
1123         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1124              handler_status, syserr, ret);
1125
1126         if (nmh) {
1127                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1128                 if (nmh->nh_post_func)
1129                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1130                                             ret_data);
1131         }
1132
1133 out:
1134         if (nmh)
1135                 o2net_handler_put(nmh);
1136         return ret;
1137 }
1138
1139 static int o2net_check_handshake(struct o2net_sock_container *sc)
1140 {
1141         struct o2net_handshake *hand = page_address(sc->sc_page);
1142         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1143
1144         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1145                 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1146                      "version %llu but %llu is required, disconnecting\n",
1147                      SC_NODEF_ARGS(sc),
1148                      (unsigned long long)be64_to_cpu(hand->protocol_version),
1149                      O2NET_PROTOCOL_VERSION);
1150
1151                 /* don't bother reconnecting if its the wrong version. */
1152                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1153                 return -1;
1154         }
1155
1156         /*
1157          * Ensure timeouts are consistent with other nodes, otherwise
1158          * we can end up with one node thinking that the other must be down,
1159          * but isn't. This can ultimately cause corruption.
1160          */
1161         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1162                                 o2net_idle_timeout(sc->sc_node)) {
1163                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1164                      "%u ms, but we use %u ms locally.  disconnecting\n",
1165                      SC_NODEF_ARGS(sc),
1166                      be32_to_cpu(hand->o2net_idle_timeout_ms),
1167                      o2net_idle_timeout(sc->sc_node));
1168                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1169                 return -1;
1170         }
1171
1172         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1173                         o2net_keepalive_delay(sc->sc_node)) {
1174                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1175                      "%u ms, but we use %u ms locally.  disconnecting\n",
1176                      SC_NODEF_ARGS(sc),
1177                      be32_to_cpu(hand->o2net_keepalive_delay_ms),
1178                      o2net_keepalive_delay(sc->sc_node));
1179                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1180                 return -1;
1181         }
1182
1183         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1184                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1185                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1186                      "%u ms, but we use %u ms locally.  disconnecting\n",
1187                      SC_NODEF_ARGS(sc),
1188                      be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1189                      O2HB_MAX_WRITE_TIMEOUT_MS);
1190                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1191                 return -1;
1192         }
1193
1194         sc->sc_handshake_ok = 1;
1195
1196         spin_lock(&nn->nn_lock);
1197         /* set valid and queue the idle timers only if it hasn't been
1198          * shut down already */
1199         if (nn->nn_sc == sc) {
1200                 o2net_sc_reset_idle_timer(sc);
1201                 atomic_set(&nn->nn_timeout, 0);
1202                 o2net_set_nn_state(nn, sc, 1, 0);
1203         }
1204         spin_unlock(&nn->nn_lock);
1205
1206         /* shift everything up as though it wasn't there */
1207         sc->sc_page_off -= sizeof(struct o2net_handshake);
1208         if (sc->sc_page_off)
1209                 memmove(hand, hand + 1, sc->sc_page_off);
1210
1211         return 0;
1212 }
1213
1214 /* this demuxes the queued rx bytes into header or payload bits and calls
1215  * handlers as each full message is read off the socket.  it returns -error,
1216  * == 0 eof, or > 0 for progress made.*/
1217 static int o2net_advance_rx(struct o2net_sock_container *sc)
1218 {
1219         struct o2net_msg *hdr;
1220         int ret = 0;
1221         void *data;
1222         size_t datalen;
1223
1224         sclog(sc, "receiving\n");
1225         do_gettimeofday(&sc->sc_tv_advance_start);
1226
1227         if (unlikely(sc->sc_handshake_ok == 0)) {
1228                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1229                         data = page_address(sc->sc_page) + sc->sc_page_off;
1230                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1231                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1232                         if (ret > 0)
1233                                 sc->sc_page_off += ret;
1234                 }
1235
1236                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1237                         o2net_check_handshake(sc);
1238                         if (unlikely(sc->sc_handshake_ok == 0))
1239                                 ret = -EPROTO;
1240                 }
1241                 goto out;
1242         }
1243
1244         /* do we need more header? */
1245         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1246                 data = page_address(sc->sc_page) + sc->sc_page_off;
1247                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1248                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1249                 if (ret > 0) {
1250                         sc->sc_page_off += ret;
1251                         /* only swab incoming here.. we can
1252                          * only get here once as we cross from
1253                          * being under to over */
1254                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1255                                 hdr = page_address(sc->sc_page);
1256                                 if (be16_to_cpu(hdr->data_len) >
1257                                     O2NET_MAX_PAYLOAD_BYTES)
1258                                         ret = -EOVERFLOW;
1259                         }
1260                 }
1261                 if (ret <= 0)
1262                         goto out;
1263         }
1264
1265         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1266                 /* oof, still don't have a header */
1267                 goto out;
1268         }
1269
1270         /* this was swabbed above when we first read it */
1271         hdr = page_address(sc->sc_page);
1272
1273         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1274
1275         /* do we need more payload? */
1276         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1277                 /* need more payload */
1278                 data = page_address(sc->sc_page) + sc->sc_page_off;
1279                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1280                           sc->sc_page_off;
1281                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1282                 if (ret > 0)
1283                         sc->sc_page_off += ret;
1284                 if (ret <= 0)
1285                         goto out;
1286         }
1287
1288         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1289                 /* we can only get here once, the first time we read
1290                  * the payload.. so set ret to progress if the handler
1291                  * works out. after calling this the message is toast */
1292                 ret = o2net_process_message(sc, hdr);
1293                 if (ret == 0)
1294                         ret = 1;
1295                 sc->sc_page_off = 0;
1296         }
1297
1298 out:
1299         sclog(sc, "ret = %d\n", ret);
1300         do_gettimeofday(&sc->sc_tv_advance_stop);
1301         return ret;
1302 }
1303
1304 /* this work func is triggerd by data ready.  it reads until it can read no
1305  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1306  * our work the work struct will be marked and we'll be called again. */
1307 static void o2net_rx_until_empty(struct work_struct *work)
1308 {
1309         struct o2net_sock_container *sc =
1310                 container_of(work, struct o2net_sock_container, sc_rx_work);
1311         int ret;
1312
1313         do {
1314                 ret = o2net_advance_rx(sc);
1315         } while (ret > 0);
1316
1317         if (ret <= 0 && ret != -EAGAIN) {
1318                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1319                 sclog(sc, "saw error %d, closing\n", ret);
1320                 /* not permanent so read failed handshake can retry */
1321                 o2net_ensure_shutdown(nn, sc, 0);
1322         }
1323
1324         sc_put(sc);
1325 }
1326
1327 static int o2net_set_nodelay(struct socket *sock)
1328 {
1329         int ret, val = 1;
1330         mm_segment_t oldfs;
1331
1332         oldfs = get_fs();
1333         set_fs(KERNEL_DS);
1334
1335         /*
1336          * Dear unsuspecting programmer,
1337          *
1338          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1339          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1340          * silently turn into SO_DEBUG.
1341          *
1342          * Yours,
1343          * Keeper of hilariously fragile interfaces.
1344          */
1345         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1346                                     (char __user *)&val, sizeof(val));
1347
1348         set_fs(oldfs);
1349         return ret;
1350 }
1351
1352 static void o2net_initialize_handshake(void)
1353 {
1354         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1355                 O2HB_MAX_WRITE_TIMEOUT_MS);
1356         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(
1357                 o2net_idle_timeout(NULL));
1358         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1359                 o2net_keepalive_delay(NULL));
1360         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1361                 o2net_reconnect_delay(NULL));
1362 }
1363
1364 /* ------------------------------------------------------------ */
1365
1366 /* called when a connect completes and after a sock is accepted.  the
1367  * rx path will see the response and mark the sc valid */
1368 static void o2net_sc_connect_completed(struct work_struct *work)
1369 {
1370         struct o2net_sock_container *sc =
1371                 container_of(work, struct o2net_sock_container,
1372                              sc_connect_work);
1373
1374         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1375               (unsigned long long)O2NET_PROTOCOL_VERSION,
1376               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1377
1378         o2net_initialize_handshake();
1379         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1380         sc_put(sc);
1381 }
1382
1383 /* this is called as a work_struct func. */
1384 static void o2net_sc_send_keep_req(struct work_struct *work)
1385 {
1386         struct o2net_sock_container *sc =
1387                 container_of(work, struct o2net_sock_container,
1388                              sc_keepalive_work.work);
1389
1390         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1391         sc_put(sc);
1392 }
1393
1394 /* socket shutdown does a del_timer_sync against this as it tears down.
1395  * we can't start this timer until we've got to the point in sc buildup
1396  * where shutdown is going to be involved */
1397 static void o2net_idle_timer(unsigned long data)
1398 {
1399         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1400         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1401         struct timeval now;
1402
1403         do_gettimeofday(&now);
1404
1405         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1406              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1407                      o2net_idle_timeout(sc->sc_node) / 1000,
1408                      o2net_idle_timeout(sc->sc_node) % 1000);
1409         mlog(ML_NOTICE, "here are some times that might help debug the "
1410              "situation: (tmr %ld.%ld now %ld.%ld dr %ld.%ld adv "
1411              "%ld.%ld:%ld.%ld func (%08x:%u) %ld.%ld:%ld.%ld)\n",
1412              sc->sc_tv_timer.tv_sec, (long) sc->sc_tv_timer.tv_usec, 
1413              now.tv_sec, (long) now.tv_usec,
1414              sc->sc_tv_data_ready.tv_sec, (long) sc->sc_tv_data_ready.tv_usec,
1415              sc->sc_tv_advance_start.tv_sec,
1416              (long) sc->sc_tv_advance_start.tv_usec,
1417              sc->sc_tv_advance_stop.tv_sec,
1418              (long) sc->sc_tv_advance_stop.tv_usec,
1419              sc->sc_msg_key, sc->sc_msg_type,
1420              sc->sc_tv_func_start.tv_sec, (long) sc->sc_tv_func_start.tv_usec,
1421              sc->sc_tv_func_stop.tv_sec, (long) sc->sc_tv_func_stop.tv_usec);
1422
1423         /*
1424          * Initialize the nn_timeout so that the next connection attempt
1425          * will continue in o2net_start_connect.
1426          */
1427         atomic_set(&nn->nn_timeout, 1);
1428
1429         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1430 }
1431
1432 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1433 {
1434         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1435         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1436                       msecs_to_jiffies(o2net_keepalive_delay(sc->sc_node)));
1437         do_gettimeofday(&sc->sc_tv_timer);
1438         mod_timer(&sc->sc_idle_timeout,
1439                jiffies + msecs_to_jiffies(o2net_idle_timeout(sc->sc_node)));
1440 }
1441
1442 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1443 {
1444         /* Only push out an existing timer */
1445         if (timer_pending(&sc->sc_idle_timeout))
1446                 o2net_sc_reset_idle_timer(sc);
1447 }
1448
1449 /* this work func is kicked whenever a path sets the nn state which doesn't
1450  * have valid set.  This includes seeing hb come up, losing a connection,
1451  * having a connect attempt fail, etc. This centralizes the logic which decides
1452  * if a connect attempt should be made or if we should give up and all future
1453  * transmit attempts should fail */
1454 static void o2net_start_connect(struct work_struct *work)
1455 {
1456         struct o2net_node *nn =
1457                 container_of(work, struct o2net_node, nn_connect_work.work);
1458         struct o2net_sock_container *sc = NULL;
1459         struct o2nm_node *node = NULL, *mynode = NULL;
1460         struct socket *sock = NULL;
1461         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1462         int ret = 0, stop;
1463         unsigned int timeout;
1464
1465         /* if we're greater we initiate tx, otherwise we accept */
1466         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1467                 goto out;
1468
1469         /* watch for racing with tearing a node down */
1470         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1471         if (node == NULL) {
1472                 ret = 0;
1473                 goto out;
1474         }
1475
1476         mynode = o2nm_get_node_by_num(o2nm_this_node());
1477         if (mynode == NULL) {
1478                 ret = 0;
1479                 goto out;
1480         }
1481
1482         spin_lock(&nn->nn_lock);
1483         /*
1484          * see if we already have one pending or have given up.
1485          * For nn_timeout, it is set when we close the connection
1486          * because of the idle time out. So it means that we have
1487          * at least connected to that node successfully once,
1488          * now try to connect to it again.
1489          */
1490         timeout = atomic_read(&nn->nn_timeout);
1491         stop = (nn->nn_sc ||
1492                 (nn->nn_persistent_error &&
1493                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1494         spin_unlock(&nn->nn_lock);
1495         if (stop)
1496                 goto out;
1497
1498         nn->nn_last_connect_attempt = jiffies;
1499
1500         sc = sc_alloc(node);
1501         if (sc == NULL) {
1502                 mlog(0, "couldn't allocate sc\n");
1503                 ret = -ENOMEM;
1504                 goto out;
1505         }
1506
1507         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1508         if (ret < 0) {
1509                 mlog(0, "can't create socket: %d\n", ret);
1510                 goto out;
1511         }
1512         sc->sc_sock = sock; /* freed by sc_kref_release */
1513
1514         sock->sk->sk_allocation = GFP_ATOMIC;
1515
1516         myaddr.sin_family = AF_INET;
1517         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1518         myaddr.sin_port = htons(0); /* any port */
1519
1520         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1521                               sizeof(myaddr));
1522         if (ret) {
1523                 mlog(ML_ERROR, "bind failed with %d at address %u.%u.%u.%u\n",
1524                      ret, NIPQUAD(mynode->nd_ipv4_address));
1525                 goto out;
1526         }
1527
1528         ret = o2net_set_nodelay(sc->sc_sock);
1529         if (ret) {
1530                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1531                 goto out;
1532         }
1533
1534         o2net_register_callbacks(sc->sc_sock->sk, sc);
1535
1536         spin_lock(&nn->nn_lock);
1537         /* handshake completion will set nn->nn_sc_valid */
1538         o2net_set_nn_state(nn, sc, 0, 0);
1539         spin_unlock(&nn->nn_lock);
1540
1541         remoteaddr.sin_family = AF_INET;
1542         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1543         remoteaddr.sin_port = node->nd_ipv4_port;
1544
1545         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1546                                         (struct sockaddr *)&remoteaddr,
1547                                         sizeof(remoteaddr),
1548                                         O_NONBLOCK);
1549         if (ret == -EINPROGRESS)
1550                 ret = 0;
1551
1552 out:
1553         if (ret) {
1554                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1555                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1556                 /* 0 err so that another will be queued and attempted
1557                  * from set_nn_state */
1558                 if (sc)
1559                         o2net_ensure_shutdown(nn, sc, 0);
1560         }
1561         if (sc)
1562                 sc_put(sc);
1563         if (node)
1564                 o2nm_node_put(node);
1565         if (mynode)
1566                 o2nm_node_put(mynode);
1567
1568         return;
1569 }
1570
1571 static void o2net_connect_expired(struct work_struct *work)
1572 {
1573         struct o2net_node *nn =
1574                 container_of(work, struct o2net_node, nn_connect_expired.work);
1575
1576         spin_lock(&nn->nn_lock);
1577         if (!nn->nn_sc_valid) {
1578                 mlog(ML_ERROR, "no connection established with node %u after "
1579                      "%u.%u seconds, giving up and returning errors.\n",
1580                      o2net_num_from_nn(nn),
1581                      o2net_idle_timeout(NULL) / 1000,
1582                      o2net_idle_timeout(NULL) % 1000);
1583
1584                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1585         }
1586         spin_unlock(&nn->nn_lock);
1587 }
1588
1589 static void o2net_still_up(struct work_struct *work)
1590 {
1591         struct o2net_node *nn =
1592                 container_of(work, struct o2net_node, nn_still_up.work);
1593
1594         o2quo_hb_still_up(o2net_num_from_nn(nn));
1595 }
1596
1597 /* ------------------------------------------------------------ */
1598
1599 void o2net_disconnect_node(struct o2nm_node *node)
1600 {
1601         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1602
1603         /* don't reconnect until it's heartbeating again */
1604         spin_lock(&nn->nn_lock);
1605         atomic_set(&nn->nn_timeout, 0);
1606         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1607         spin_unlock(&nn->nn_lock);
1608
1609         if (o2net_wq) {
1610                 cancel_delayed_work(&nn->nn_connect_expired);
1611                 cancel_delayed_work(&nn->nn_connect_work);
1612                 cancel_delayed_work(&nn->nn_still_up);
1613                 flush_workqueue(o2net_wq);
1614         }
1615 }
1616
1617 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1618                                   void *data)
1619 {
1620         o2quo_hb_down(node_num);
1621
1622         if (node_num != o2nm_this_node())
1623                 o2net_disconnect_node(node);
1624
1625         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1626 }
1627
1628 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1629                                 void *data)
1630 {
1631         struct o2net_node *nn = o2net_nn_from_num(node_num);
1632
1633         o2quo_hb_up(node_num);
1634
1635         /* ensure an immediate connect attempt */
1636         nn->nn_last_connect_attempt = jiffies -
1637                 (msecs_to_jiffies(o2net_reconnect_delay(node)) + 1);
1638
1639         if (node_num != o2nm_this_node()) {
1640                 /* believe it or not, accept and node hearbeating testing
1641                  * can succeed for this node before we got here.. so
1642                  * only use set_nn_state to clear the persistent error
1643                  * if that hasn't already happened */
1644                 spin_lock(&nn->nn_lock);
1645                 atomic_set(&nn->nn_timeout, 0);
1646                 if (nn->nn_persistent_error)
1647                         o2net_set_nn_state(nn, NULL, 0, 0);
1648                 spin_unlock(&nn->nn_lock);
1649         }
1650 }
1651
1652 void o2net_unregister_hb_callbacks(void)
1653 {
1654         o2hb_unregister_callback(NULL, &o2net_hb_up);
1655         o2hb_unregister_callback(NULL, &o2net_hb_down);
1656 }
1657
1658 int o2net_register_hb_callbacks(void)
1659 {
1660         int ret;
1661
1662         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1663                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1664         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1665                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1666
1667         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1668         if (ret == 0)
1669                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1670
1671         if (ret)
1672                 o2net_unregister_hb_callbacks();
1673
1674         return ret;
1675 }
1676
1677 /* ------------------------------------------------------------ */
1678
1679 static int o2net_accept_one(struct socket *sock)
1680 {
1681         int ret, slen;
1682         struct sockaddr_in sin;
1683         struct socket *new_sock = NULL;
1684         struct o2nm_node *node = NULL;
1685         struct o2net_sock_container *sc = NULL;
1686         struct o2net_node *nn;
1687
1688         BUG_ON(sock == NULL);
1689         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1690                                sock->sk->sk_protocol, &new_sock);
1691         if (ret)
1692                 goto out;
1693
1694         new_sock->type = sock->type;
1695         new_sock->ops = sock->ops;
1696         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1697         if (ret < 0)
1698                 goto out;
1699
1700         new_sock->sk->sk_allocation = GFP_ATOMIC;
1701
1702         ret = o2net_set_nodelay(new_sock);
1703         if (ret) {
1704                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1705                 goto out;
1706         }
1707
1708         slen = sizeof(sin);
1709         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1710                                        &slen, 1);
1711         if (ret < 0)
1712                 goto out;
1713
1714         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1715         if (node == NULL) {
1716                 mlog(ML_NOTICE, "attempt to connect from unknown node at "
1717                      "%u.%u.%u.%u:%d\n", NIPQUAD(sin.sin_addr.s_addr),
1718                      ntohs(sin.sin_port));
1719                 ret = -EINVAL;
1720                 goto out;
1721         }
1722
1723         if (o2nm_this_node() > node->nd_num) {
1724                 mlog(ML_NOTICE, "unexpected connect attempted from a lower "
1725                      "numbered node '%s' at " "%u.%u.%u.%u:%d with num %u\n",
1726                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1727                      ntohs(sin.sin_port), node->nd_num);
1728                 ret = -EINVAL;
1729                 goto out;
1730         }
1731
1732         /* this happens all the time when the other node sees our heartbeat
1733          * and tries to connect before we see their heartbeat */
1734         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1735                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1736                      "%u.%u.%u.%u:%d but it isn't heartbeating\n",
1737                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1738                      ntohs(sin.sin_port));
1739                 ret = -EINVAL;
1740                 goto out;
1741         }
1742
1743         nn = o2net_nn_from_num(node->nd_num);
1744
1745         spin_lock(&nn->nn_lock);
1746         if (nn->nn_sc)
1747                 ret = -EBUSY;
1748         else
1749                 ret = 0;
1750         spin_unlock(&nn->nn_lock);
1751         if (ret) {
1752                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1753                      "%u.%u.%u.%u:%d but it already has an open connection\n",
1754                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1755                      ntohs(sin.sin_port));
1756                 goto out;
1757         }
1758
1759         sc = sc_alloc(node);
1760         if (sc == NULL) {
1761                 ret = -ENOMEM;
1762                 goto out;
1763         }
1764
1765         sc->sc_sock = new_sock;
1766         new_sock = NULL;
1767
1768         spin_lock(&nn->nn_lock);
1769         atomic_set(&nn->nn_timeout, 0);
1770         o2net_set_nn_state(nn, sc, 0, 0);
1771         spin_unlock(&nn->nn_lock);
1772
1773         o2net_register_callbacks(sc->sc_sock->sk, sc);
1774         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1775
1776         o2net_initialize_handshake();
1777         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1778
1779 out:
1780         if (new_sock)
1781                 sock_release(new_sock);
1782         if (node)
1783                 o2nm_node_put(node);
1784         if (sc)
1785                 sc_put(sc);
1786         return ret;
1787 }
1788
1789 static void o2net_accept_many(struct work_struct *work)
1790 {
1791         struct socket *sock = o2net_listen_sock;
1792         while (o2net_accept_one(sock) == 0)
1793                 cond_resched();
1794 }
1795
1796 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1797 {
1798         void (*ready)(struct sock *sk, int bytes);
1799
1800         read_lock(&sk->sk_callback_lock);
1801         ready = sk->sk_user_data;
1802         if (ready == NULL) { /* check for teardown race */
1803                 ready = sk->sk_data_ready;
1804                 goto out;
1805         }
1806
1807         /* ->sk_data_ready is also called for a newly established child socket
1808          * before it has been accepted and the acceptor has set up their
1809          * data_ready.. we only want to queue listen work for our listening
1810          * socket */
1811         if (sk->sk_state == TCP_LISTEN) {
1812                 mlog(ML_TCP, "bytes: %d\n", bytes);
1813                 queue_work(o2net_wq, &o2net_listen_work);
1814         }
1815
1816 out:
1817         read_unlock(&sk->sk_callback_lock);
1818         ready(sk, bytes);
1819 }
1820
1821 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1822 {
1823         struct socket *sock = NULL;
1824         int ret;
1825         struct sockaddr_in sin = {
1826                 .sin_family = PF_INET,
1827                 .sin_addr = { .s_addr = addr },
1828                 .sin_port = port,
1829         };
1830
1831         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1832         if (ret < 0) {
1833                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1834                 goto out;
1835         }
1836
1837         sock->sk->sk_allocation = GFP_ATOMIC;
1838
1839         write_lock_bh(&sock->sk->sk_callback_lock);
1840         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1841         sock->sk->sk_data_ready = o2net_listen_data_ready;
1842         write_unlock_bh(&sock->sk->sk_callback_lock);
1843
1844         o2net_listen_sock = sock;
1845         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1846
1847         sock->sk->sk_reuse = 1;
1848         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1849         if (ret < 0) {
1850                 mlog(ML_ERROR, "unable to bind socket at %u.%u.%u.%u:%u, "
1851                      "ret=%d\n", NIPQUAD(addr), ntohs(port), ret);
1852                 goto out;
1853         }
1854
1855         ret = sock->ops->listen(sock, 64);
1856         if (ret < 0) {
1857                 mlog(ML_ERROR, "unable to listen on %u.%u.%u.%u:%u, ret=%d\n",
1858                      NIPQUAD(addr), ntohs(port), ret);
1859         }
1860
1861 out:
1862         if (ret) {
1863                 o2net_listen_sock = NULL;
1864                 if (sock)
1865                         sock_release(sock);
1866         }
1867         return ret;
1868 }
1869
1870 /*
1871  * called from node manager when we should bring up our network listening
1872  * socket.  node manager handles all the serialization to only call this
1873  * once and to match it with o2net_stop_listening().  note,
1874  * o2nm_this_node() doesn't work yet as we're being called while it
1875  * is being set up.
1876  */
1877 int o2net_start_listening(struct o2nm_node *node)
1878 {
1879         int ret = 0;
1880
1881         BUG_ON(o2net_wq != NULL);
1882         BUG_ON(o2net_listen_sock != NULL);
1883
1884         mlog(ML_KTHREAD, "starting o2net thread...\n");
1885         o2net_wq = create_singlethread_workqueue("o2net");
1886         if (o2net_wq == NULL) {
1887                 mlog(ML_ERROR, "unable to launch o2net thread\n");
1888                 return -ENOMEM; /* ? */
1889         }
1890
1891         ret = o2net_open_listening_sock(node->nd_ipv4_address,
1892                                         node->nd_ipv4_port);
1893         if (ret) {
1894                 destroy_workqueue(o2net_wq);
1895                 o2net_wq = NULL;
1896         } else
1897                 o2quo_conn_up(node->nd_num);
1898
1899         return ret;
1900 }
1901
1902 /* again, o2nm_this_node() doesn't work here as we're involved in
1903  * tearing it down */
1904 void o2net_stop_listening(struct o2nm_node *node)
1905 {
1906         struct socket *sock = o2net_listen_sock;
1907         size_t i;
1908
1909         BUG_ON(o2net_wq == NULL);
1910         BUG_ON(o2net_listen_sock == NULL);
1911
1912         /* stop the listening socket from generating work */
1913         write_lock_bh(&sock->sk->sk_callback_lock);
1914         sock->sk->sk_data_ready = sock->sk->sk_user_data;
1915         sock->sk->sk_user_data = NULL;
1916         write_unlock_bh(&sock->sk->sk_callback_lock);
1917
1918         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1919                 struct o2nm_node *node = o2nm_get_node_by_num(i);
1920                 if (node) {
1921                         o2net_disconnect_node(node);
1922                         o2nm_node_put(node);
1923                 }
1924         }
1925
1926         /* finish all work and tear down the work queue */
1927         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
1928         destroy_workqueue(o2net_wq);
1929         o2net_wq = NULL;
1930
1931         sock_release(o2net_listen_sock);
1932         o2net_listen_sock = NULL;
1933
1934         o2quo_conn_err(node->nd_num);
1935 }
1936
1937 /* ------------------------------------------------------------ */
1938
1939 int o2net_init(void)
1940 {
1941         unsigned long i;
1942
1943         o2quo_init();
1944
1945         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
1946         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1947         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1948         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
1949                 kfree(o2net_hand);
1950                 kfree(o2net_keep_req);
1951                 kfree(o2net_keep_resp);
1952                 return -ENOMEM;
1953         }
1954
1955         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
1956         o2net_hand->connector_id = cpu_to_be64(1);
1957
1958         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
1959         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
1960
1961         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1962                 struct o2net_node *nn = o2net_nn_from_num(i);
1963
1964                 atomic_set(&nn->nn_timeout, 0);
1965                 spin_lock_init(&nn->nn_lock);
1966                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
1967                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
1968                                   o2net_connect_expired);
1969                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
1970                 /* until we see hb from a node we'll return einval */
1971                 nn->nn_persistent_error = -ENOTCONN;
1972                 init_waitqueue_head(&nn->nn_sc_wq);
1973                 idr_init(&nn->nn_status_idr);
1974                 INIT_LIST_HEAD(&nn->nn_status_list);
1975         }
1976
1977         return 0;
1978 }
1979
1980 void o2net_exit(void)
1981 {
1982         o2quo_exit();
1983         kfree(o2net_hand);
1984         kfree(o2net_keep_req);
1985         kfree(o2net_keep_resp);
1986 }