ocfs2: Depend on configfs heartbeat items.
[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 <net/tcp.h>
62
63 #include <asm/uaccess.h>
64
65 #include "heartbeat.h"
66 #include "tcp.h"
67 #include "nodemanager.h"
68 #define MLOG_MASK_PREFIX ML_TCP
69 #include "masklog.h"
70 #include "quorum.h"
71
72 #include "tcp_internal.h"
73
74 /* 
75  * The linux network stack isn't sparse endian clean.. It has macros like
76  * ntohs() which perform the endian checks and structs like sockaddr_in
77  * which aren't annotated.  So __force is found here to get the build
78  * clean.  When they emerge from the dark ages and annotate the code
79  * we can remove these.
80  */
81
82 #define SC_NODEF_FMT "node %s (num %u) at %u.%u.%u.%u:%u"
83 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
84                           NIPQUAD(sc->sc_node->nd_ipv4_address),        \
85                           ntohs(sc->sc_node->nd_ipv4_port)
86
87 /*
88  * In the following two log macros, the whitespace after the ',' just
89  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
90  * previous token if args expands to nothing.
91  */
92 #define msglog(hdr, fmt, args...) do {                                  \
93         typeof(hdr) __hdr = (hdr);                                      \
94         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
95              "key %08x num %u] " fmt,                                   \
96              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
97              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
98              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
99              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
100 } while (0)
101
102 #define sclog(sc, fmt, args...) do {                                    \
103         typeof(sc) __sc = (sc);                                         \
104         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
105              "pg_off %zu] " fmt, __sc,                                  \
106              atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,       \
107             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
108             ##args);                                                    \
109 } while (0)
110
111 static DEFINE_RWLOCK(o2net_handler_lock);
112 static struct rb_root o2net_handler_tree = RB_ROOT;
113
114 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
115
116 /* XXX someday we'll need better accounting */
117 static struct socket *o2net_listen_sock = NULL;
118
119 /*
120  * listen work is only queued by the listening socket callbacks on the
121  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
122  * quorum work is queued as sock containers are shutdown.. stop_listening
123  * tears down all the node's sock containers, preventing future shutdowns
124  * and queued quroum work, before canceling delayed quorum work and
125  * destroying the work queue.
126  */
127 static struct workqueue_struct *o2net_wq;
128 static struct work_struct o2net_listen_work;
129
130 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
131 #define O2NET_HB_PRI 0x1
132
133 static struct o2net_handshake *o2net_hand;
134 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
135
136 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
137                 {[O2NET_ERR_NONE]       = 0,
138                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
139                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
140                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
141
142 /* can't quite avoid *all* internal declarations :/ */
143 static void o2net_sc_connect_completed(struct work_struct *work);
144 static void o2net_rx_until_empty(struct work_struct *work);
145 static void o2net_shutdown_sc(struct work_struct *work);
146 static void o2net_listen_data_ready(struct sock *sk, int bytes);
147 static void o2net_sc_send_keep_req(struct work_struct *work);
148 static void o2net_idle_timer(unsigned long data);
149 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
150 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
151
152 /*
153  * FIXME: These should use to_o2nm_cluster_from_node(), but we end up
154  * losing our parent link to the cluster during shutdown. This can be
155  * solved by adding a pre-removal callback to configfs, or passing
156  * around the cluster with the node. -jeffm
157  */
158 static inline int o2net_reconnect_delay(struct o2nm_node *node)
159 {
160         return o2nm_single_cluster->cl_reconnect_delay_ms;
161 }
162
163 static inline int o2net_keepalive_delay(struct o2nm_node *node)
164 {
165         return o2nm_single_cluster->cl_keepalive_delay_ms;
166 }
167
168 static inline int o2net_idle_timeout(struct o2nm_node *node)
169 {
170         return o2nm_single_cluster->cl_idle_timeout_ms;
171 }
172
173 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
174 {
175         int trans;
176         BUG_ON(err >= O2NET_ERR_MAX);
177         trans = o2net_sys_err_translations[err];
178
179         /* Just in case we mess up the translation table above */
180         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
181         return trans;
182 }
183
184 static struct o2net_node * o2net_nn_from_num(u8 node_num)
185 {
186         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
187         return &o2net_nodes[node_num];
188 }
189
190 static u8 o2net_num_from_nn(struct o2net_node *nn)
191 {
192         BUG_ON(nn == NULL);
193         return nn - o2net_nodes;
194 }
195
196 /* ------------------------------------------------------------ */
197
198 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
199 {
200         int ret = 0;
201
202         do {
203                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
204                         ret = -EAGAIN;
205                         break;
206                 }
207                 spin_lock(&nn->nn_lock);
208                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
209                 if (ret == 0)
210                         list_add_tail(&nsw->ns_node_item,
211                                       &nn->nn_status_list);
212                 spin_unlock(&nn->nn_lock);
213         } while (ret == -EAGAIN);
214
215         if (ret == 0)  {
216                 init_waitqueue_head(&nsw->ns_wq);
217                 nsw->ns_sys_status = O2NET_ERR_NONE;
218                 nsw->ns_status = 0;
219         }
220
221         return ret;
222 }
223
224 static void o2net_complete_nsw_locked(struct o2net_node *nn,
225                                       struct o2net_status_wait *nsw,
226                                       enum o2net_system_error sys_status,
227                                       s32 status)
228 {
229         assert_spin_locked(&nn->nn_lock);
230
231         if (!list_empty(&nsw->ns_node_item)) {
232                 list_del_init(&nsw->ns_node_item);
233                 nsw->ns_sys_status = sys_status;
234                 nsw->ns_status = status;
235                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
236                 wake_up(&nsw->ns_wq);
237         }
238 }
239
240 static void o2net_complete_nsw(struct o2net_node *nn,
241                                struct o2net_status_wait *nsw,
242                                u64 id, enum o2net_system_error sys_status,
243                                s32 status)
244 {
245         spin_lock(&nn->nn_lock);
246         if (nsw == NULL) {
247                 if (id > INT_MAX)
248                         goto out;
249
250                 nsw = idr_find(&nn->nn_status_idr, id);
251                 if (nsw == NULL)
252                         goto out;
253         }
254
255         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
256
257 out:
258         spin_unlock(&nn->nn_lock);
259         return;
260 }
261
262 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
263 {
264         struct list_head *iter, *tmp;
265         unsigned int num_kills = 0;
266         struct o2net_status_wait *nsw;
267
268         assert_spin_locked(&nn->nn_lock);
269
270         list_for_each_safe(iter, tmp, &nn->nn_status_list) {
271                 nsw = list_entry(iter, struct o2net_status_wait, ns_node_item);
272                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
273                 num_kills++;
274         }
275
276         mlog(0, "completed %d messages for node %u\n", num_kills,
277              o2net_num_from_nn(nn));
278 }
279
280 static int o2net_nsw_completed(struct o2net_node *nn,
281                                struct o2net_status_wait *nsw)
282 {
283         int completed;
284         spin_lock(&nn->nn_lock);
285         completed = list_empty(&nsw->ns_node_item);
286         spin_unlock(&nn->nn_lock);
287         return completed;
288 }
289
290 /* ------------------------------------------------------------ */
291
292 static void sc_kref_release(struct kref *kref)
293 {
294         struct o2net_sock_container *sc = container_of(kref,
295                                         struct o2net_sock_container, sc_kref);
296         BUG_ON(timer_pending(&sc->sc_idle_timeout));
297
298         sclog(sc, "releasing\n");
299
300         if (sc->sc_sock) {
301                 sock_release(sc->sc_sock);
302                 sc->sc_sock = NULL;
303         }
304
305         o2nm_node_put(sc->sc_node);
306         sc->sc_node = NULL;
307
308         kfree(sc);
309 }
310
311 static void sc_put(struct o2net_sock_container *sc)
312 {
313         sclog(sc, "put\n");
314         kref_put(&sc->sc_kref, sc_kref_release);
315 }
316 static void sc_get(struct o2net_sock_container *sc)
317 {
318         sclog(sc, "get\n");
319         kref_get(&sc->sc_kref);
320 }
321 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
322 {
323         struct o2net_sock_container *sc, *ret = NULL;
324         struct page *page = NULL;
325
326         page = alloc_page(GFP_NOFS);
327         sc = kzalloc(sizeof(*sc), GFP_NOFS);
328         if (sc == NULL || page == NULL)
329                 goto out;
330
331         kref_init(&sc->sc_kref);
332         o2nm_node_get(node);
333         sc->sc_node = node;
334
335         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
336         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
337         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
338         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
339
340         init_timer(&sc->sc_idle_timeout);
341         sc->sc_idle_timeout.function = o2net_idle_timer;
342         sc->sc_idle_timeout.data = (unsigned long)sc;
343
344         sclog(sc, "alloced\n");
345
346         ret = sc;
347         sc->sc_page = page;
348         sc = NULL;
349         page = NULL;
350
351 out:
352         if (page)
353                 __free_page(page);
354         kfree(sc);
355
356         return ret;
357 }
358
359 /* ------------------------------------------------------------ */
360
361 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
362                                 struct work_struct *work)
363 {
364         sc_get(sc);
365         if (!queue_work(o2net_wq, work))
366                 sc_put(sc);
367 }
368 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
369                                         struct delayed_work *work,
370                                         int delay)
371 {
372         sc_get(sc);
373         if (!queue_delayed_work(o2net_wq, work, delay))
374                 sc_put(sc);
375 }
376 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
377                                          struct delayed_work *work)
378 {
379         if (cancel_delayed_work(work))
380                 sc_put(sc);
381 }
382
383 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
384
385 int o2net_num_connected_peers(void)
386 {
387         return atomic_read(&o2net_connected_peers);
388 }
389
390 static void o2net_set_nn_state(struct o2net_node *nn,
391                                struct o2net_sock_container *sc,
392                                unsigned valid, int err)
393 {
394         int was_valid = nn->nn_sc_valid;
395         int was_err = nn->nn_persistent_error;
396         struct o2net_sock_container *old_sc = nn->nn_sc;
397
398         assert_spin_locked(&nn->nn_lock);
399
400         if (old_sc && !sc)
401                 atomic_dec(&o2net_connected_peers);
402         else if (!old_sc && sc)
403                 atomic_inc(&o2net_connected_peers);
404
405         /* the node num comparison and single connect/accept path should stop
406          * an non-null sc from being overwritten with another */
407         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
408         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
409         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
410
411         /* we won't reconnect after our valid conn goes away for
412          * this hb iteration.. here so it shows up in the logs */
413         if (was_valid && !valid && err == 0)
414                 err = -ENOTCONN;
415
416         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
417              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
418              nn->nn_persistent_error, err);
419
420         nn->nn_sc = sc;
421         nn->nn_sc_valid = valid ? 1 : 0;
422         nn->nn_persistent_error = err;
423
424         /* mirrors o2net_tx_can_proceed() */
425         if (nn->nn_persistent_error || nn->nn_sc_valid)
426                 wake_up(&nn->nn_sc_wq);
427
428         if (!was_err && nn->nn_persistent_error) {
429                 o2quo_conn_err(o2net_num_from_nn(nn));
430                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
431                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
432         }
433
434         if (was_valid && !valid) {
435                 printk(KERN_INFO "o2net: no longer connected to "
436                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
437                 o2net_complete_nodes_nsw(nn);
438         }
439
440         if (!was_valid && valid) {
441                 o2quo_conn_up(o2net_num_from_nn(nn));
442                 /* this is a bit of a hack.  we only try reconnecting
443                  * when heartbeating starts until we get a connection.
444                  * if that connection then dies we don't try reconnecting.
445                  * the only way to start connecting again is to down
446                  * heartbeat and bring it back up. */
447                 cancel_delayed_work(&nn->nn_connect_expired);
448                 printk(KERN_INFO "o2net: %s " SC_NODEF_FMT "\n",
449                        o2nm_this_node() > sc->sc_node->nd_num ?
450                                 "connected to" : "accepted connection from",
451                        SC_NODEF_ARGS(sc));
452         }
453
454         /* trigger the connecting worker func as long as we're not valid,
455          * it will back off if it shouldn't connect.  This can be called
456          * from node config teardown and so needs to be careful about
457          * the work queue actually being up. */
458         if (!valid && o2net_wq) {
459                 unsigned long delay;
460                 /* delay if we're withing a RECONNECT_DELAY of the
461                  * last attempt */
462                 delay = (nn->nn_last_connect_attempt +
463                          msecs_to_jiffies(o2net_reconnect_delay(sc->sc_node)))
464                         - jiffies;
465                 if (delay > msecs_to_jiffies(o2net_reconnect_delay(sc->sc_node)))
466                         delay = 0;
467                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
468                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
469         }
470
471         /* keep track of the nn's sc ref for the caller */
472         if ((old_sc == NULL) && sc)
473                 sc_get(sc);
474         if (old_sc && (old_sc != sc)) {
475                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
476                 sc_put(old_sc);
477         }
478 }
479
480 /* see o2net_register_callbacks() */
481 static void o2net_data_ready(struct sock *sk, int bytes)
482 {
483         void (*ready)(struct sock *sk, int bytes);
484
485         read_lock(&sk->sk_callback_lock);
486         if (sk->sk_user_data) {
487                 struct o2net_sock_container *sc = sk->sk_user_data;
488                 sclog(sc, "data_ready hit\n");
489                 do_gettimeofday(&sc->sc_tv_data_ready);
490                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
491                 ready = sc->sc_data_ready;
492         } else {
493                 ready = sk->sk_data_ready;
494         }
495         read_unlock(&sk->sk_callback_lock);
496
497         ready(sk, bytes);
498 }
499
500 /* see o2net_register_callbacks() */
501 static void o2net_state_change(struct sock *sk)
502 {
503         void (*state_change)(struct sock *sk);
504         struct o2net_sock_container *sc;
505
506         read_lock(&sk->sk_callback_lock);
507         sc = sk->sk_user_data;
508         if (sc == NULL) {
509                 state_change = sk->sk_state_change;
510                 goto out;
511         }
512
513         sclog(sc, "state_change to %d\n", sk->sk_state);
514
515         state_change = sc->sc_state_change;
516
517         switch(sk->sk_state) {
518                 /* ignore connecting sockets as they make progress */
519                 case TCP_SYN_SENT:
520                 case TCP_SYN_RECV:
521                         break;
522                 case TCP_ESTABLISHED:
523                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
524                         break;
525                 default:
526                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
527                         break;
528         }
529 out:
530         read_unlock(&sk->sk_callback_lock);
531         state_change(sk);
532 }
533
534 /*
535  * we register callbacks so we can queue work on events before calling
536  * the original callbacks.  our callbacks our careful to test user_data
537  * to discover when they've reaced with o2net_unregister_callbacks().
538  */
539 static void o2net_register_callbacks(struct sock *sk,
540                                      struct o2net_sock_container *sc)
541 {
542         write_lock_bh(&sk->sk_callback_lock);
543
544         /* accepted sockets inherit the old listen socket data ready */
545         if (sk->sk_data_ready == o2net_listen_data_ready) {
546                 sk->sk_data_ready = sk->sk_user_data;
547                 sk->sk_user_data = NULL;
548         }
549
550         BUG_ON(sk->sk_user_data != NULL);
551         sk->sk_user_data = sc;
552         sc_get(sc);
553
554         sc->sc_data_ready = sk->sk_data_ready;
555         sc->sc_state_change = sk->sk_state_change;
556         sk->sk_data_ready = o2net_data_ready;
557         sk->sk_state_change = o2net_state_change;
558
559         mutex_init(&sc->sc_send_lock);
560
561         write_unlock_bh(&sk->sk_callback_lock);
562 }
563
564 static int o2net_unregister_callbacks(struct sock *sk,
565                                    struct o2net_sock_container *sc)
566 {
567         int ret = 0;
568
569         write_lock_bh(&sk->sk_callback_lock);
570         if (sk->sk_user_data == sc) {
571                 ret = 1;
572                 sk->sk_user_data = NULL;
573                 sk->sk_data_ready = sc->sc_data_ready;
574                 sk->sk_state_change = sc->sc_state_change;
575         }
576         write_unlock_bh(&sk->sk_callback_lock);
577
578         return ret;
579 }
580
581 /*
582  * this is a little helper that is called by callers who have seen a problem
583  * with an sc and want to detach it from the nn if someone already hasn't beat
584  * them to it.  if an error is given then the shutdown will be persistent
585  * and pending transmits will be canceled.
586  */
587 static void o2net_ensure_shutdown(struct o2net_node *nn,
588                                    struct o2net_sock_container *sc,
589                                    int err)
590 {
591         spin_lock(&nn->nn_lock);
592         if (nn->nn_sc == sc)
593                 o2net_set_nn_state(nn, NULL, 0, err);
594         spin_unlock(&nn->nn_lock);
595 }
596
597 /*
598  * This work queue function performs the blocking parts of socket shutdown.  A
599  * few paths lead here.  set_nn_state will trigger this callback if it sees an
600  * sc detached from the nn.  state_change will also trigger this callback
601  * directly when it sees errors.  In that case we need to call set_nn_state
602  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
603  * itself.
604  */
605 static void o2net_shutdown_sc(struct work_struct *work)
606 {
607         struct o2net_sock_container *sc =
608                 container_of(work, struct o2net_sock_container,
609                              sc_shutdown_work);
610         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
611
612         sclog(sc, "shutting down\n");
613
614         /* drop the callbacks ref and call shutdown only once */
615         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
616                 /* we shouldn't flush as we're in the thread, the
617                  * races with pending sc work structs are harmless */
618                 del_timer_sync(&sc->sc_idle_timeout);
619                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
620                 sc_put(sc);
621                 sc->sc_sock->ops->shutdown(sc->sc_sock,
622                                            RCV_SHUTDOWN|SEND_SHUTDOWN);
623         }
624
625         /* not fatal so failed connects before the other guy has our
626          * heartbeat can be retried */
627         o2net_ensure_shutdown(nn, sc, 0);
628         sc_put(sc);
629 }
630
631 /* ------------------------------------------------------------ */
632
633 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
634                              u32 key)
635 {
636         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
637
638         if (ret == 0)
639                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
640
641         return ret;
642 }
643
644 static struct o2net_msg_handler *
645 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
646                           struct rb_node **ret_parent)
647 {
648         struct rb_node **p = &o2net_handler_tree.rb_node;
649         struct rb_node *parent = NULL;
650         struct o2net_msg_handler *nmh, *ret = NULL;
651         int cmp;
652
653         while (*p) {
654                 parent = *p;
655                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
656                 cmp = o2net_handler_cmp(nmh, msg_type, key);
657
658                 if (cmp < 0)
659                         p = &(*p)->rb_left;
660                 else if (cmp > 0)
661                         p = &(*p)->rb_right;
662                 else {
663                         ret = nmh;
664                         break;
665                 }
666         }
667
668         if (ret_p != NULL)
669                 *ret_p = p;
670         if (ret_parent != NULL)
671                 *ret_parent = parent;
672
673         return ret;
674 }
675
676 static void o2net_handler_kref_release(struct kref *kref)
677 {
678         struct o2net_msg_handler *nmh;
679         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
680
681         kfree(nmh);
682 }
683
684 static void o2net_handler_put(struct o2net_msg_handler *nmh)
685 {
686         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
687 }
688
689 /* max_len is protection for the handler func.  incoming messages won't
690  * be given to the handler if their payload is longer than the max. */
691 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
692                            o2net_msg_handler_func *func, void *data,
693                            o2net_post_msg_handler_func *post_func,
694                            struct list_head *unreg_list)
695 {
696         struct o2net_msg_handler *nmh = NULL;
697         struct rb_node **p, *parent;
698         int ret = 0;
699
700         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
701                 mlog(0, "max_len for message handler out of range: %u\n",
702                         max_len);
703                 ret = -EINVAL;
704                 goto out;
705         }
706
707         if (!msg_type) {
708                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
709                 ret = -EINVAL;
710                 goto out;
711
712         }
713         if (!func) {
714                 mlog(0, "no message handler provided: %u, %p\n",
715                        msg_type, func);
716                 ret = -EINVAL;
717                 goto out;
718         }
719
720         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
721         if (nmh == NULL) {
722                 ret = -ENOMEM;
723                 goto out;
724         }
725
726         nmh->nh_func = func;
727         nmh->nh_func_data = data;
728         nmh->nh_post_func = post_func;
729         nmh->nh_msg_type = msg_type;
730         nmh->nh_max_len = max_len;
731         nmh->nh_key = key;
732         /* the tree and list get this ref.. they're both removed in
733          * unregister when this ref is dropped */
734         kref_init(&nmh->nh_kref);
735         INIT_LIST_HEAD(&nmh->nh_unregister_item);
736
737         write_lock(&o2net_handler_lock);
738         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
739                 ret = -EEXIST;
740         else {
741                 rb_link_node(&nmh->nh_node, parent, p);
742                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
743                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
744
745                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
746                      func, msg_type, key);
747                 /* we've had some trouble with handlers seemingly vanishing. */
748                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
749                                                           &parent) == NULL,
750                                 "couldn't find handler we *just* registerd "
751                                 "for type %u key %08x\n", msg_type, key);
752         }
753         write_unlock(&o2net_handler_lock);
754         if (ret)
755                 goto out;
756
757 out:
758         if (ret)
759                 kfree(nmh);
760
761         return ret;
762 }
763 EXPORT_SYMBOL_GPL(o2net_register_handler);
764
765 void o2net_unregister_handler_list(struct list_head *list)
766 {
767         struct list_head *pos, *n;
768         struct o2net_msg_handler *nmh;
769
770         write_lock(&o2net_handler_lock);
771         list_for_each_safe(pos, n, list) {
772                 nmh = list_entry(pos, struct o2net_msg_handler,
773                                  nh_unregister_item);
774                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
775                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
776                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
777                 list_del_init(&nmh->nh_unregister_item);
778                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
779         }
780         write_unlock(&o2net_handler_lock);
781 }
782 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
783
784 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
785 {
786         struct o2net_msg_handler *nmh;
787
788         read_lock(&o2net_handler_lock);
789         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
790         if (nmh)
791                 kref_get(&nmh->nh_kref);
792         read_unlock(&o2net_handler_lock);
793
794         return nmh;
795 }
796
797 /* ------------------------------------------------------------ */
798
799 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
800 {
801         int ret;
802         mm_segment_t oldfs;
803         struct kvec vec = {
804                 .iov_len = len,
805                 .iov_base = data,
806         };
807         struct msghdr msg = {
808                 .msg_iovlen = 1,
809                 .msg_iov = (struct iovec *)&vec,
810                 .msg_flags = MSG_DONTWAIT,
811         };
812
813         oldfs = get_fs();
814         set_fs(get_ds());
815         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
816         set_fs(oldfs);
817
818         return ret;
819 }
820
821 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
822                               size_t veclen, size_t total)
823 {
824         int ret;
825         mm_segment_t oldfs;
826         struct msghdr msg = {
827                 .msg_iov = (struct iovec *)vec,
828                 .msg_iovlen = veclen,
829         };
830
831         if (sock == NULL) {
832                 ret = -EINVAL;
833                 goto out;
834         }
835
836         oldfs = get_fs();
837         set_fs(get_ds());
838         ret = sock_sendmsg(sock, &msg, total);
839         set_fs(oldfs);
840         if (ret != total) {
841                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
842                      total);
843                 if (ret >= 0)
844                         ret = -EPIPE; /* should be smarter, I bet */
845                 goto out;
846         }
847
848         ret = 0;
849 out:
850         if (ret < 0)
851                 mlog(0, "returning error: %d\n", ret);
852         return ret;
853 }
854
855 static void o2net_sendpage(struct o2net_sock_container *sc,
856                            void *kmalloced_virt,
857                            size_t size)
858 {
859         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
860         ssize_t ret;
861
862
863         mutex_lock(&sc->sc_send_lock);
864         ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
865                                          virt_to_page(kmalloced_virt),
866                                          (long)kmalloced_virt & ~PAGE_MASK,
867                                          size, MSG_DONTWAIT);
868         mutex_unlock(&sc->sc_send_lock);
869         if (ret != size) {
870                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT 
871                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
872                 o2net_ensure_shutdown(nn, sc, 0);
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                 o2net_set_nn_state(nn, sc, 1, 0);
1202         }
1203         spin_unlock(&nn->nn_lock);
1204
1205         /* shift everything up as though it wasn't there */
1206         sc->sc_page_off -= sizeof(struct o2net_handshake);
1207         if (sc->sc_page_off)
1208                 memmove(hand, hand + 1, sc->sc_page_off);
1209
1210         return 0;
1211 }
1212
1213 /* this demuxes the queued rx bytes into header or payload bits and calls
1214  * handlers as each full message is read off the socket.  it returns -error,
1215  * == 0 eof, or > 0 for progress made.*/
1216 static int o2net_advance_rx(struct o2net_sock_container *sc)
1217 {
1218         struct o2net_msg *hdr;
1219         int ret = 0;
1220         void *data;
1221         size_t datalen;
1222
1223         sclog(sc, "receiving\n");
1224         do_gettimeofday(&sc->sc_tv_advance_start);
1225
1226         if (unlikely(sc->sc_handshake_ok == 0)) {
1227                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1228                         data = page_address(sc->sc_page) + sc->sc_page_off;
1229                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1230                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1231                         if (ret > 0)
1232                                 sc->sc_page_off += ret;
1233                 }
1234
1235                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1236                         o2net_check_handshake(sc);
1237                         if (unlikely(sc->sc_handshake_ok == 0))
1238                                 ret = -EPROTO;
1239                 }
1240                 goto out;
1241         }
1242
1243         /* do we need more header? */
1244         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1245                 data = page_address(sc->sc_page) + sc->sc_page_off;
1246                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1247                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1248                 if (ret > 0) {
1249                         sc->sc_page_off += ret;
1250                         /* only swab incoming here.. we can
1251                          * only get here once as we cross from
1252                          * being under to over */
1253                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1254                                 hdr = page_address(sc->sc_page);
1255                                 if (be16_to_cpu(hdr->data_len) >
1256                                     O2NET_MAX_PAYLOAD_BYTES)
1257                                         ret = -EOVERFLOW;
1258                         }
1259                 }
1260                 if (ret <= 0)
1261                         goto out;
1262         }
1263
1264         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1265                 /* oof, still don't have a header */
1266                 goto out;
1267         }
1268
1269         /* this was swabbed above when we first read it */
1270         hdr = page_address(sc->sc_page);
1271
1272         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1273
1274         /* do we need more payload? */
1275         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1276                 /* need more payload */
1277                 data = page_address(sc->sc_page) + sc->sc_page_off;
1278                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1279                           sc->sc_page_off;
1280                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1281                 if (ret > 0)
1282                         sc->sc_page_off += ret;
1283                 if (ret <= 0)
1284                         goto out;
1285         }
1286
1287         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1288                 /* we can only get here once, the first time we read
1289                  * the payload.. so set ret to progress if the handler
1290                  * works out. after calling this the message is toast */
1291                 ret = o2net_process_message(sc, hdr);
1292                 if (ret == 0)
1293                         ret = 1;
1294                 sc->sc_page_off = 0;
1295         }
1296
1297 out:
1298         sclog(sc, "ret = %d\n", ret);
1299         do_gettimeofday(&sc->sc_tv_advance_stop);
1300         return ret;
1301 }
1302
1303 /* this work func is triggerd by data ready.  it reads until it can read no
1304  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1305  * our work the work struct will be marked and we'll be called again. */
1306 static void o2net_rx_until_empty(struct work_struct *work)
1307 {
1308         struct o2net_sock_container *sc =
1309                 container_of(work, struct o2net_sock_container, sc_rx_work);
1310         int ret;
1311
1312         do {
1313                 ret = o2net_advance_rx(sc);
1314         } while (ret > 0);
1315
1316         if (ret <= 0 && ret != -EAGAIN) {
1317                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1318                 sclog(sc, "saw error %d, closing\n", ret);
1319                 /* not permanent so read failed handshake can retry */
1320                 o2net_ensure_shutdown(nn, sc, 0);
1321         }
1322
1323         sc_put(sc);
1324 }
1325
1326 static int o2net_set_nodelay(struct socket *sock)
1327 {
1328         int ret, val = 1;
1329         mm_segment_t oldfs;
1330
1331         oldfs = get_fs();
1332         set_fs(KERNEL_DS);
1333
1334         /*
1335          * Dear unsuspecting programmer,
1336          *
1337          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1338          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1339          * silently turn into SO_DEBUG.
1340          *
1341          * Yours,
1342          * Keeper of hilariously fragile interfaces.
1343          */
1344         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1345                                     (char __user *)&val, sizeof(val));
1346
1347         set_fs(oldfs);
1348         return ret;
1349 }
1350
1351 static void o2net_initialize_handshake(void)
1352 {
1353         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1354                 O2HB_MAX_WRITE_TIMEOUT_MS);
1355         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(
1356                 o2net_idle_timeout(NULL));
1357         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1358                 o2net_keepalive_delay(NULL));
1359         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1360                 o2net_reconnect_delay(NULL));
1361 }
1362
1363 /* ------------------------------------------------------------ */
1364
1365 /* called when a connect completes and after a sock is accepted.  the
1366  * rx path will see the response and mark the sc valid */
1367 static void o2net_sc_connect_completed(struct work_struct *work)
1368 {
1369         struct o2net_sock_container *sc =
1370                 container_of(work, struct o2net_sock_container,
1371                              sc_connect_work);
1372
1373         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1374               (unsigned long long)O2NET_PROTOCOL_VERSION,
1375               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1376
1377         o2net_initialize_handshake();
1378         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1379         sc_put(sc);
1380 }
1381
1382 /* this is called as a work_struct func. */
1383 static void o2net_sc_send_keep_req(struct work_struct *work)
1384 {
1385         struct o2net_sock_container *sc =
1386                 container_of(work, struct o2net_sock_container,
1387                              sc_keepalive_work.work);
1388
1389         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1390         sc_put(sc);
1391 }
1392
1393 /* socket shutdown does a del_timer_sync against this as it tears down.
1394  * we can't start this timer until we've got to the point in sc buildup
1395  * where shutdown is going to be involved */
1396 static void o2net_idle_timer(unsigned long data)
1397 {
1398         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1399         struct timeval now;
1400
1401         do_gettimeofday(&now);
1402
1403         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1404              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1405                      o2net_idle_timeout(sc->sc_node) / 1000,
1406                      o2net_idle_timeout(sc->sc_node) % 1000);
1407         mlog(ML_NOTICE, "here are some times that might help debug the "
1408              "situation: (tmr %ld.%ld now %ld.%ld dr %ld.%ld adv "
1409              "%ld.%ld:%ld.%ld func (%08x:%u) %ld.%ld:%ld.%ld)\n",
1410              sc->sc_tv_timer.tv_sec, (long) sc->sc_tv_timer.tv_usec, 
1411              now.tv_sec, (long) now.tv_usec,
1412              sc->sc_tv_data_ready.tv_sec, (long) sc->sc_tv_data_ready.tv_usec,
1413              sc->sc_tv_advance_start.tv_sec,
1414              (long) sc->sc_tv_advance_start.tv_usec,
1415              sc->sc_tv_advance_stop.tv_sec,
1416              (long) sc->sc_tv_advance_stop.tv_usec,
1417              sc->sc_msg_key, sc->sc_msg_type,
1418              sc->sc_tv_func_start.tv_sec, (long) sc->sc_tv_func_start.tv_usec,
1419              sc->sc_tv_func_stop.tv_sec, (long) sc->sc_tv_func_stop.tv_usec);
1420
1421         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1422 }
1423
1424 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1425 {
1426         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1427         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1428                       msecs_to_jiffies(o2net_keepalive_delay(sc->sc_node)));
1429         do_gettimeofday(&sc->sc_tv_timer);
1430         mod_timer(&sc->sc_idle_timeout,
1431                jiffies + msecs_to_jiffies(o2net_idle_timeout(sc->sc_node)));
1432 }
1433
1434 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1435 {
1436         /* Only push out an existing timer */
1437         if (timer_pending(&sc->sc_idle_timeout))
1438                 o2net_sc_reset_idle_timer(sc);
1439 }
1440
1441 /* this work func is kicked whenever a path sets the nn state which doesn't
1442  * have valid set.  This includes seeing hb come up, losing a connection,
1443  * having a connect attempt fail, etc. This centralizes the logic which decides
1444  * if a connect attempt should be made or if we should give up and all future
1445  * transmit attempts should fail */
1446 static void o2net_start_connect(struct work_struct *work)
1447 {
1448         struct o2net_node *nn =
1449                 container_of(work, struct o2net_node, nn_connect_work.work);
1450         struct o2net_sock_container *sc = NULL;
1451         struct o2nm_node *node = NULL, *mynode = NULL;
1452         struct socket *sock = NULL;
1453         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1454         int ret = 0, stop;
1455
1456         /* if we're greater we initiate tx, otherwise we accept */
1457         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1458                 goto out;
1459
1460         /* watch for racing with tearing a node down */
1461         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1462         if (node == NULL) {
1463                 ret = 0;
1464                 goto out;
1465         }
1466
1467         mynode = o2nm_get_node_by_num(o2nm_this_node());
1468         if (mynode == NULL) {
1469                 ret = 0;
1470                 goto out;
1471         }
1472
1473         spin_lock(&nn->nn_lock);
1474         /* see if we already have one pending or have given up */
1475         stop = (nn->nn_sc || nn->nn_persistent_error);
1476         spin_unlock(&nn->nn_lock);
1477         if (stop)
1478                 goto out;
1479
1480         nn->nn_last_connect_attempt = jiffies;
1481
1482         sc = sc_alloc(node);
1483         if (sc == NULL) {
1484                 mlog(0, "couldn't allocate sc\n");
1485                 ret = -ENOMEM;
1486                 goto out;
1487         }
1488
1489         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1490         if (ret < 0) {
1491                 mlog(0, "can't create socket: %d\n", ret);
1492                 goto out;
1493         }
1494         sc->sc_sock = sock; /* freed by sc_kref_release */
1495
1496         sock->sk->sk_allocation = GFP_ATOMIC;
1497
1498         myaddr.sin_family = AF_INET;
1499         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1500         myaddr.sin_port = (__force u16)htons(0); /* any port */
1501
1502         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1503                               sizeof(myaddr));
1504         if (ret) {
1505                 mlog(ML_ERROR, "bind failed with %d at address %u.%u.%u.%u\n",
1506                      ret, NIPQUAD(mynode->nd_ipv4_address));
1507                 goto out;
1508         }
1509
1510         ret = o2net_set_nodelay(sc->sc_sock);
1511         if (ret) {
1512                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1513                 goto out;
1514         }
1515
1516         o2net_register_callbacks(sc->sc_sock->sk, sc);
1517
1518         spin_lock(&nn->nn_lock);
1519         /* handshake completion will set nn->nn_sc_valid */
1520         o2net_set_nn_state(nn, sc, 0, 0);
1521         spin_unlock(&nn->nn_lock);
1522
1523         remoteaddr.sin_family = AF_INET;
1524         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1525         remoteaddr.sin_port = node->nd_ipv4_port;
1526
1527         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1528                                         (struct sockaddr *)&remoteaddr,
1529                                         sizeof(remoteaddr),
1530                                         O_NONBLOCK);
1531         if (ret == -EINPROGRESS)
1532                 ret = 0;
1533
1534 out:
1535         if (ret) {
1536                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1537                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1538                 /* 0 err so that another will be queued and attempted
1539                  * from set_nn_state */
1540                 if (sc)
1541                         o2net_ensure_shutdown(nn, sc, 0);
1542         }
1543         if (sc)
1544                 sc_put(sc);
1545         if (node)
1546                 o2nm_node_put(node);
1547         if (mynode)
1548                 o2nm_node_put(mynode);
1549
1550         return;
1551 }
1552
1553 static void o2net_connect_expired(struct work_struct *work)
1554 {
1555         struct o2net_node *nn =
1556                 container_of(work, struct o2net_node, nn_connect_expired.work);
1557
1558         spin_lock(&nn->nn_lock);
1559         if (!nn->nn_sc_valid) {
1560                 struct o2nm_node *node = nn->nn_sc->sc_node;
1561                 mlog(ML_ERROR, "no connection established with node %u after "
1562                      "%u.%u seconds, giving up and returning errors.\n",
1563                      o2net_num_from_nn(nn),
1564                      o2net_idle_timeout(node) / 1000,
1565                      o2net_idle_timeout(node) % 1000);
1566
1567                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1568         }
1569         spin_unlock(&nn->nn_lock);
1570 }
1571
1572 static void o2net_still_up(struct work_struct *work)
1573 {
1574         struct o2net_node *nn =
1575                 container_of(work, struct o2net_node, nn_still_up.work);
1576
1577         o2quo_hb_still_up(o2net_num_from_nn(nn));
1578 }
1579
1580 /* ------------------------------------------------------------ */
1581
1582 void o2net_disconnect_node(struct o2nm_node *node)
1583 {
1584         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1585
1586         /* don't reconnect until it's heartbeating again */
1587         spin_lock(&nn->nn_lock);
1588         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1589         spin_unlock(&nn->nn_lock);
1590
1591         if (o2net_wq) {
1592                 cancel_delayed_work(&nn->nn_connect_expired);
1593                 cancel_delayed_work(&nn->nn_connect_work);
1594                 cancel_delayed_work(&nn->nn_still_up);
1595                 flush_workqueue(o2net_wq);
1596         }
1597 }
1598
1599 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1600                                   void *data)
1601 {
1602         o2quo_hb_down(node_num);
1603
1604         if (node_num != o2nm_this_node())
1605                 o2net_disconnect_node(node);
1606
1607         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1608 }
1609
1610 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1611                                 void *data)
1612 {
1613         struct o2net_node *nn = o2net_nn_from_num(node_num);
1614
1615         o2quo_hb_up(node_num);
1616
1617         /* ensure an immediate connect attempt */
1618         nn->nn_last_connect_attempt = jiffies -
1619                 (msecs_to_jiffies(o2net_reconnect_delay(node)) + 1);
1620
1621         if (node_num != o2nm_this_node()) {
1622                 /* heartbeat doesn't work unless a local node number is
1623                  * configured and doing so brings up the o2net_wq, so we can
1624                  * use it.. */
1625                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired,
1626                                    msecs_to_jiffies(o2net_idle_timeout(node)));
1627
1628                 /* believe it or not, accept and node hearbeating testing
1629                  * can succeed for this node before we got here.. so
1630                  * only use set_nn_state to clear the persistent error
1631                  * if that hasn't already happened */
1632                 spin_lock(&nn->nn_lock);
1633                 if (nn->nn_persistent_error)
1634                         o2net_set_nn_state(nn, NULL, 0, 0);
1635                 spin_unlock(&nn->nn_lock);
1636         }
1637 }
1638
1639 void o2net_unregister_hb_callbacks(void)
1640 {
1641         o2hb_unregister_callback(NULL, &o2net_hb_up);
1642         o2hb_unregister_callback(NULL, &o2net_hb_down);
1643 }
1644
1645 int o2net_register_hb_callbacks(void)
1646 {
1647         int ret;
1648
1649         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1650                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1651         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1652                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1653
1654         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1655         if (ret == 0)
1656                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1657
1658         if (ret)
1659                 o2net_unregister_hb_callbacks();
1660
1661         return ret;
1662 }
1663
1664 /* ------------------------------------------------------------ */
1665
1666 static int o2net_accept_one(struct socket *sock)
1667 {
1668         int ret, slen;
1669         struct sockaddr_in sin;
1670         struct socket *new_sock = NULL;
1671         struct o2nm_node *node = NULL;
1672         struct o2net_sock_container *sc = NULL;
1673         struct o2net_node *nn;
1674
1675         BUG_ON(sock == NULL);
1676         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1677                                sock->sk->sk_protocol, &new_sock);
1678         if (ret)
1679                 goto out;
1680
1681         new_sock->type = sock->type;
1682         new_sock->ops = sock->ops;
1683         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1684         if (ret < 0)
1685                 goto out;
1686
1687         new_sock->sk->sk_allocation = GFP_ATOMIC;
1688
1689         ret = o2net_set_nodelay(new_sock);
1690         if (ret) {
1691                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1692                 goto out;
1693         }
1694
1695         slen = sizeof(sin);
1696         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1697                                        &slen, 1);
1698         if (ret < 0)
1699                 goto out;
1700
1701         node = o2nm_get_node_by_ip((__force __be32)sin.sin_addr.s_addr);
1702         if (node == NULL) {
1703                 mlog(ML_NOTICE, "attempt to connect from unknown node at "
1704                      "%u.%u.%u.%u:%d\n", NIPQUAD(sin.sin_addr.s_addr),
1705                      ntohs((__force __be16)sin.sin_port));
1706                 ret = -EINVAL;
1707                 goto out;
1708         }
1709
1710         if (o2nm_this_node() > node->nd_num) {
1711                 mlog(ML_NOTICE, "unexpected connect attempted from a lower "
1712                      "numbered node '%s' at " "%u.%u.%u.%u:%d with num %u\n",
1713                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1714                      ntohs((__force __be16)sin.sin_port), node->nd_num);
1715                 ret = -EINVAL;
1716                 goto out;
1717         }
1718
1719         /* this happens all the time when the other node sees our heartbeat
1720          * and tries to connect before we see their heartbeat */
1721         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1722                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1723                      "%u.%u.%u.%u:%d but it isn't heartbeating\n",
1724                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1725                      ntohs((__force __be16)sin.sin_port));
1726                 ret = -EINVAL;
1727                 goto out;
1728         }
1729
1730         nn = o2net_nn_from_num(node->nd_num);
1731
1732         spin_lock(&nn->nn_lock);
1733         if (nn->nn_sc)
1734                 ret = -EBUSY;
1735         else
1736                 ret = 0;
1737         spin_unlock(&nn->nn_lock);
1738         if (ret) {
1739                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1740                      "%u.%u.%u.%u:%d but it already has an open connection\n",
1741                      node->nd_name, NIPQUAD(sin.sin_addr.s_addr),
1742                      ntohs((__force __be16)sin.sin_port));
1743                 goto out;
1744         }
1745
1746         sc = sc_alloc(node);
1747         if (sc == NULL) {
1748                 ret = -ENOMEM;
1749                 goto out;
1750         }
1751
1752         sc->sc_sock = new_sock;
1753         new_sock = NULL;
1754
1755         spin_lock(&nn->nn_lock);
1756         o2net_set_nn_state(nn, sc, 0, 0);
1757         spin_unlock(&nn->nn_lock);
1758
1759         o2net_register_callbacks(sc->sc_sock->sk, sc);
1760         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1761
1762         o2net_initialize_handshake();
1763         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1764
1765 out:
1766         if (new_sock)
1767                 sock_release(new_sock);
1768         if (node)
1769                 o2nm_node_put(node);
1770         if (sc)
1771                 sc_put(sc);
1772         return ret;
1773 }
1774
1775 static void o2net_accept_many(struct work_struct *work)
1776 {
1777         struct socket *sock = o2net_listen_sock;
1778         while (o2net_accept_one(sock) == 0)
1779                 cond_resched();
1780 }
1781
1782 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1783 {
1784         void (*ready)(struct sock *sk, int bytes);
1785
1786         read_lock(&sk->sk_callback_lock);
1787         ready = sk->sk_user_data;
1788         if (ready == NULL) { /* check for teardown race */
1789                 ready = sk->sk_data_ready;
1790                 goto out;
1791         }
1792
1793         /* ->sk_data_ready is also called for a newly established child socket
1794          * before it has been accepted and the acceptor has set up their
1795          * data_ready.. we only want to queue listen work for our listening
1796          * socket */
1797         if (sk->sk_state == TCP_LISTEN) {
1798                 mlog(ML_TCP, "bytes: %d\n", bytes);
1799                 queue_work(o2net_wq, &o2net_listen_work);
1800         }
1801
1802 out:
1803         read_unlock(&sk->sk_callback_lock);
1804         ready(sk, bytes);
1805 }
1806
1807 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1808 {
1809         struct socket *sock = NULL;
1810         int ret;
1811         struct sockaddr_in sin = {
1812                 .sin_family = PF_INET,
1813                 .sin_addr = { .s_addr = addr },
1814                 .sin_port = port,
1815         };
1816
1817         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1818         if (ret < 0) {
1819                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1820                 goto out;
1821         }
1822
1823         sock->sk->sk_allocation = GFP_ATOMIC;
1824
1825         write_lock_bh(&sock->sk->sk_callback_lock);
1826         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1827         sock->sk->sk_data_ready = o2net_listen_data_ready;
1828         write_unlock_bh(&sock->sk->sk_callback_lock);
1829
1830         o2net_listen_sock = sock;
1831         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1832
1833         sock->sk->sk_reuse = 1;
1834         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1835         if (ret < 0) {
1836                 mlog(ML_ERROR, "unable to bind socket at %u.%u.%u.%u:%u, "
1837                      "ret=%d\n", NIPQUAD(addr), ntohs(port), ret);
1838                 goto out;
1839         }
1840
1841         ret = sock->ops->listen(sock, 64);
1842         if (ret < 0) {
1843                 mlog(ML_ERROR, "unable to listen on %u.%u.%u.%u:%u, ret=%d\n",
1844                      NIPQUAD(addr), ntohs(port), ret);
1845         }
1846
1847 out:
1848         if (ret) {
1849                 o2net_listen_sock = NULL;
1850                 if (sock)
1851                         sock_release(sock);
1852         }
1853         return ret;
1854 }
1855
1856 /*
1857  * called from node manager when we should bring up our network listening
1858  * socket.  node manager handles all the serialization to only call this
1859  * once and to match it with o2net_stop_listening().  note,
1860  * o2nm_this_node() doesn't work yet as we're being called while it
1861  * is being set up.
1862  */
1863 int o2net_start_listening(struct o2nm_node *node)
1864 {
1865         int ret = 0;
1866
1867         BUG_ON(o2net_wq != NULL);
1868         BUG_ON(o2net_listen_sock != NULL);
1869
1870         mlog(ML_KTHREAD, "starting o2net thread...\n");
1871         o2net_wq = create_singlethread_workqueue("o2net");
1872         if (o2net_wq == NULL) {
1873                 mlog(ML_ERROR, "unable to launch o2net thread\n");
1874                 return -ENOMEM; /* ? */
1875         }
1876
1877         ret = o2net_open_listening_sock(node->nd_ipv4_address,
1878                                         node->nd_ipv4_port);
1879         if (ret) {
1880                 destroy_workqueue(o2net_wq);
1881                 o2net_wq = NULL;
1882         } else
1883                 o2quo_conn_up(node->nd_num);
1884
1885         return ret;
1886 }
1887
1888 /* again, o2nm_this_node() doesn't work here as we're involved in
1889  * tearing it down */
1890 void o2net_stop_listening(struct o2nm_node *node)
1891 {
1892         struct socket *sock = o2net_listen_sock;
1893         size_t i;
1894
1895         BUG_ON(o2net_wq == NULL);
1896         BUG_ON(o2net_listen_sock == NULL);
1897
1898         /* stop the listening socket from generating work */
1899         write_lock_bh(&sock->sk->sk_callback_lock);
1900         sock->sk->sk_data_ready = sock->sk->sk_user_data;
1901         sock->sk->sk_user_data = NULL;
1902         write_unlock_bh(&sock->sk->sk_callback_lock);
1903
1904         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1905                 struct o2nm_node *node = o2nm_get_node_by_num(i);
1906                 if (node) {
1907                         o2net_disconnect_node(node);
1908                         o2nm_node_put(node);
1909                 }
1910         }
1911
1912         /* finish all work and tear down the work queue */
1913         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
1914         destroy_workqueue(o2net_wq);
1915         o2net_wq = NULL;
1916
1917         sock_release(o2net_listen_sock);
1918         o2net_listen_sock = NULL;
1919
1920         o2quo_conn_err(node->nd_num);
1921 }
1922
1923 /* ------------------------------------------------------------ */
1924
1925 int o2net_init(void)
1926 {
1927         unsigned long i;
1928
1929         o2quo_init();
1930
1931         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
1932         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1933         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
1934         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
1935                 kfree(o2net_hand);
1936                 kfree(o2net_keep_req);
1937                 kfree(o2net_keep_resp);
1938                 return -ENOMEM;
1939         }
1940
1941         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
1942         o2net_hand->connector_id = cpu_to_be64(1);
1943
1944         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
1945         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
1946
1947         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1948                 struct o2net_node *nn = o2net_nn_from_num(i);
1949
1950                 spin_lock_init(&nn->nn_lock);
1951                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
1952                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
1953                                   o2net_connect_expired);
1954                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
1955                 /* until we see hb from a node we'll return einval */
1956                 nn->nn_persistent_error = -ENOTCONN;
1957                 init_waitqueue_head(&nn->nn_sc_wq);
1958                 idr_init(&nn->nn_status_idr);
1959                 INIT_LIST_HEAD(&nn->nn_status_list);
1960         }
1961
1962         return 0;
1963 }
1964
1965 void o2net_exit(void)
1966 {
1967         o2quo_exit();
1968         kfree(o2net_hand);
1969         kfree(o2net_keep_req);
1970         kfree(o2net_keep_resp);
1971 }