ocfs2/trivial: Remove trailing whitespaces
[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 #ifdef CONFIG_DEBUG_FS
146 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
147                            u32 msgkey, struct task_struct *task, u8 node)
148 {
149         INIT_LIST_HEAD(&nst->st_net_debug_item);
150         nst->st_task = task;
151         nst->st_msg_type = msgtype;
152         nst->st_msg_key = msgkey;
153         nst->st_node = node;
154 }
155
156 static void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
157 {
158         do_gettimeofday(&nst->st_sock_time);
159 }
160
161 static void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
162 {
163         do_gettimeofday(&nst->st_send_time);
164 }
165
166 static void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
167 {
168         do_gettimeofday(&nst->st_status_time);
169 }
170
171 static void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
172                                          struct o2net_sock_container *sc)
173 {
174         nst->st_sc = sc;
175 }
176
177 static void o2net_set_nst_msg_id(struct o2net_send_tracking *nst, u32 msg_id)
178 {
179         nst->st_id = msg_id;
180 }
181
182 #else  /* CONFIG_DEBUG_FS */
183
184 static inline void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
185                                   u32 msgkey, struct task_struct *task, u8 node)
186 {
187 }
188
189 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
190 {
191 }
192
193 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
194 {
195 }
196
197 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
198 {
199 }
200
201 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
202                                                 struct o2net_sock_container *sc)
203 {
204 }
205
206 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
207                                         u32 msg_id)
208 {
209 }
210
211 #endif /* CONFIG_DEBUG_FS */
212
213 static inline int o2net_reconnect_delay(void)
214 {
215         return o2nm_single_cluster->cl_reconnect_delay_ms;
216 }
217
218 static inline int o2net_keepalive_delay(void)
219 {
220         return o2nm_single_cluster->cl_keepalive_delay_ms;
221 }
222
223 static inline int o2net_idle_timeout(void)
224 {
225         return o2nm_single_cluster->cl_idle_timeout_ms;
226 }
227
228 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
229 {
230         int trans;
231         BUG_ON(err >= O2NET_ERR_MAX);
232         trans = o2net_sys_err_translations[err];
233
234         /* Just in case we mess up the translation table above */
235         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
236         return trans;
237 }
238
239 static struct o2net_node * o2net_nn_from_num(u8 node_num)
240 {
241         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
242         return &o2net_nodes[node_num];
243 }
244
245 static u8 o2net_num_from_nn(struct o2net_node *nn)
246 {
247         BUG_ON(nn == NULL);
248         return nn - o2net_nodes;
249 }
250
251 /* ------------------------------------------------------------ */
252
253 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
254 {
255         int ret = 0;
256
257         do {
258                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
259                         ret = -EAGAIN;
260                         break;
261                 }
262                 spin_lock(&nn->nn_lock);
263                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
264                 if (ret == 0)
265                         list_add_tail(&nsw->ns_node_item,
266                                       &nn->nn_status_list);
267                 spin_unlock(&nn->nn_lock);
268         } while (ret == -EAGAIN);
269
270         if (ret == 0)  {
271                 init_waitqueue_head(&nsw->ns_wq);
272                 nsw->ns_sys_status = O2NET_ERR_NONE;
273                 nsw->ns_status = 0;
274         }
275
276         return ret;
277 }
278
279 static void o2net_complete_nsw_locked(struct o2net_node *nn,
280                                       struct o2net_status_wait *nsw,
281                                       enum o2net_system_error sys_status,
282                                       s32 status)
283 {
284         assert_spin_locked(&nn->nn_lock);
285
286         if (!list_empty(&nsw->ns_node_item)) {
287                 list_del_init(&nsw->ns_node_item);
288                 nsw->ns_sys_status = sys_status;
289                 nsw->ns_status = status;
290                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
291                 wake_up(&nsw->ns_wq);
292         }
293 }
294
295 static void o2net_complete_nsw(struct o2net_node *nn,
296                                struct o2net_status_wait *nsw,
297                                u64 id, enum o2net_system_error sys_status,
298                                s32 status)
299 {
300         spin_lock(&nn->nn_lock);
301         if (nsw == NULL) {
302                 if (id > INT_MAX)
303                         goto out;
304
305                 nsw = idr_find(&nn->nn_status_idr, id);
306                 if (nsw == NULL)
307                         goto out;
308         }
309
310         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
311
312 out:
313         spin_unlock(&nn->nn_lock);
314         return;
315 }
316
317 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
318 {
319         struct o2net_status_wait *nsw, *tmp;
320         unsigned int num_kills = 0;
321
322         assert_spin_locked(&nn->nn_lock);
323
324         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
325                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
326                 num_kills++;
327         }
328
329         mlog(0, "completed %d messages for node %u\n", num_kills,
330              o2net_num_from_nn(nn));
331 }
332
333 static int o2net_nsw_completed(struct o2net_node *nn,
334                                struct o2net_status_wait *nsw)
335 {
336         int completed;
337         spin_lock(&nn->nn_lock);
338         completed = list_empty(&nsw->ns_node_item);
339         spin_unlock(&nn->nn_lock);
340         return completed;
341 }
342
343 /* ------------------------------------------------------------ */
344
345 static void sc_kref_release(struct kref *kref)
346 {
347         struct o2net_sock_container *sc = container_of(kref,
348                                         struct o2net_sock_container, sc_kref);
349         BUG_ON(timer_pending(&sc->sc_idle_timeout));
350
351         sclog(sc, "releasing\n");
352
353         if (sc->sc_sock) {
354                 sock_release(sc->sc_sock);
355                 sc->sc_sock = NULL;
356         }
357
358         o2nm_node_put(sc->sc_node);
359         sc->sc_node = NULL;
360
361         o2net_debug_del_sc(sc);
362         kfree(sc);
363 }
364
365 static void sc_put(struct o2net_sock_container *sc)
366 {
367         sclog(sc, "put\n");
368         kref_put(&sc->sc_kref, sc_kref_release);
369 }
370 static void sc_get(struct o2net_sock_container *sc)
371 {
372         sclog(sc, "get\n");
373         kref_get(&sc->sc_kref);
374 }
375 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
376 {
377         struct o2net_sock_container *sc, *ret = NULL;
378         struct page *page = NULL;
379
380         page = alloc_page(GFP_NOFS);
381         sc = kzalloc(sizeof(*sc), GFP_NOFS);
382         if (sc == NULL || page == NULL)
383                 goto out;
384
385         kref_init(&sc->sc_kref);
386         o2nm_node_get(node);
387         sc->sc_node = node;
388
389         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
390         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
391         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
392         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
393
394         init_timer(&sc->sc_idle_timeout);
395         sc->sc_idle_timeout.function = o2net_idle_timer;
396         sc->sc_idle_timeout.data = (unsigned long)sc;
397
398         sclog(sc, "alloced\n");
399
400         ret = sc;
401         sc->sc_page = page;
402         o2net_debug_add_sc(sc);
403         sc = NULL;
404         page = NULL;
405
406 out:
407         if (page)
408                 __free_page(page);
409         kfree(sc);
410
411         return ret;
412 }
413
414 /* ------------------------------------------------------------ */
415
416 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
417                                 struct work_struct *work)
418 {
419         sc_get(sc);
420         if (!queue_work(o2net_wq, work))
421                 sc_put(sc);
422 }
423 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
424                                         struct delayed_work *work,
425                                         int delay)
426 {
427         sc_get(sc);
428         if (!queue_delayed_work(o2net_wq, work, delay))
429                 sc_put(sc);
430 }
431 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
432                                          struct delayed_work *work)
433 {
434         if (cancel_delayed_work(work))
435                 sc_put(sc);
436 }
437
438 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
439
440 int o2net_num_connected_peers(void)
441 {
442         return atomic_read(&o2net_connected_peers);
443 }
444
445 static void o2net_set_nn_state(struct o2net_node *nn,
446                                struct o2net_sock_container *sc,
447                                unsigned valid, int err)
448 {
449         int was_valid = nn->nn_sc_valid;
450         int was_err = nn->nn_persistent_error;
451         struct o2net_sock_container *old_sc = nn->nn_sc;
452
453         assert_spin_locked(&nn->nn_lock);
454
455         if (old_sc && !sc)
456                 atomic_dec(&o2net_connected_peers);
457         else if (!old_sc && sc)
458                 atomic_inc(&o2net_connected_peers);
459
460         /* the node num comparison and single connect/accept path should stop
461          * an non-null sc from being overwritten with another */
462         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
463         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
464         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
465
466         if (was_valid && !valid && err == 0)
467                 err = -ENOTCONN;
468
469         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
470              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
471              nn->nn_persistent_error, err);
472
473         nn->nn_sc = sc;
474         nn->nn_sc_valid = valid ? 1 : 0;
475         nn->nn_persistent_error = err;
476
477         /* mirrors o2net_tx_can_proceed() */
478         if (nn->nn_persistent_error || nn->nn_sc_valid)
479                 wake_up(&nn->nn_sc_wq);
480
481         if (!was_err && nn->nn_persistent_error) {
482                 o2quo_conn_err(o2net_num_from_nn(nn));
483                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
484                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
485         }
486
487         if (was_valid && !valid) {
488                 printk(KERN_INFO "o2net: no longer connected to "
489                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
490                 o2net_complete_nodes_nsw(nn);
491         }
492
493         if (!was_valid && valid) {
494                 o2quo_conn_up(o2net_num_from_nn(nn));
495                 cancel_delayed_work(&nn->nn_connect_expired);
496                 printk(KERN_INFO "o2net: %s " SC_NODEF_FMT "\n",
497                        o2nm_this_node() > sc->sc_node->nd_num ?
498                                 "connected to" : "accepted connection from",
499                        SC_NODEF_ARGS(sc));
500         }
501
502         /* trigger the connecting worker func as long as we're not valid,
503          * it will back off if it shouldn't connect.  This can be called
504          * from node config teardown and so needs to be careful about
505          * the work queue actually being up. */
506         if (!valid && o2net_wq) {
507                 unsigned long delay;
508                 /* delay if we're withing a RECONNECT_DELAY of the
509                  * last attempt */
510                 delay = (nn->nn_last_connect_attempt +
511                          msecs_to_jiffies(o2net_reconnect_delay()))
512                         - jiffies;
513                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
514                         delay = 0;
515                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
516                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
517
518                 /*
519                  * Delay the expired work after idle timeout.
520                  *
521                  * We might have lots of failed connection attempts that run
522                  * through here but we only cancel the connect_expired work when
523                  * a connection attempt succeeds.  So only the first enqueue of
524                  * the connect_expired work will do anything.  The rest will see
525                  * that it's already queued and do nothing.
526                  */
527                 delay += msecs_to_jiffies(o2net_idle_timeout());
528                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
529         }
530
531         /* keep track of the nn's sc ref for the caller */
532         if ((old_sc == NULL) && sc)
533                 sc_get(sc);
534         if (old_sc && (old_sc != sc)) {
535                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
536                 sc_put(old_sc);
537         }
538 }
539
540 /* see o2net_register_callbacks() */
541 static void o2net_data_ready(struct sock *sk, int bytes)
542 {
543         void (*ready)(struct sock *sk, int bytes);
544
545         read_lock(&sk->sk_callback_lock);
546         if (sk->sk_user_data) {
547                 struct o2net_sock_container *sc = sk->sk_user_data;
548                 sclog(sc, "data_ready hit\n");
549                 do_gettimeofday(&sc->sc_tv_data_ready);
550                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
551                 ready = sc->sc_data_ready;
552         } else {
553                 ready = sk->sk_data_ready;
554         }
555         read_unlock(&sk->sk_callback_lock);
556
557         ready(sk, bytes);
558 }
559
560 /* see o2net_register_callbacks() */
561 static void o2net_state_change(struct sock *sk)
562 {
563         void (*state_change)(struct sock *sk);
564         struct o2net_sock_container *sc;
565
566         read_lock(&sk->sk_callback_lock);
567         sc = sk->sk_user_data;
568         if (sc == NULL) {
569                 state_change = sk->sk_state_change;
570                 goto out;
571         }
572
573         sclog(sc, "state_change to %d\n", sk->sk_state);
574
575         state_change = sc->sc_state_change;
576
577         switch(sk->sk_state) {
578                 /* ignore connecting sockets as they make progress */
579                 case TCP_SYN_SENT:
580                 case TCP_SYN_RECV:
581                         break;
582                 case TCP_ESTABLISHED:
583                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
584                         break;
585                 default:
586                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
587                         break;
588         }
589 out:
590         read_unlock(&sk->sk_callback_lock);
591         state_change(sk);
592 }
593
594 /*
595  * we register callbacks so we can queue work on events before calling
596  * the original callbacks.  our callbacks our careful to test user_data
597  * to discover when they've reaced with o2net_unregister_callbacks().
598  */
599 static void o2net_register_callbacks(struct sock *sk,
600                                      struct o2net_sock_container *sc)
601 {
602         write_lock_bh(&sk->sk_callback_lock);
603
604         /* accepted sockets inherit the old listen socket data ready */
605         if (sk->sk_data_ready == o2net_listen_data_ready) {
606                 sk->sk_data_ready = sk->sk_user_data;
607                 sk->sk_user_data = NULL;
608         }
609
610         BUG_ON(sk->sk_user_data != NULL);
611         sk->sk_user_data = sc;
612         sc_get(sc);
613
614         sc->sc_data_ready = sk->sk_data_ready;
615         sc->sc_state_change = sk->sk_state_change;
616         sk->sk_data_ready = o2net_data_ready;
617         sk->sk_state_change = o2net_state_change;
618
619         mutex_init(&sc->sc_send_lock);
620
621         write_unlock_bh(&sk->sk_callback_lock);
622 }
623
624 static int o2net_unregister_callbacks(struct sock *sk,
625                                    struct o2net_sock_container *sc)
626 {
627         int ret = 0;
628
629         write_lock_bh(&sk->sk_callback_lock);
630         if (sk->sk_user_data == sc) {
631                 ret = 1;
632                 sk->sk_user_data = NULL;
633                 sk->sk_data_ready = sc->sc_data_ready;
634                 sk->sk_state_change = sc->sc_state_change;
635         }
636         write_unlock_bh(&sk->sk_callback_lock);
637
638         return ret;
639 }
640
641 /*
642  * this is a little helper that is called by callers who have seen a problem
643  * with an sc and want to detach it from the nn if someone already hasn't beat
644  * them to it.  if an error is given then the shutdown will be persistent
645  * and pending transmits will be canceled.
646  */
647 static void o2net_ensure_shutdown(struct o2net_node *nn,
648                                    struct o2net_sock_container *sc,
649                                    int err)
650 {
651         spin_lock(&nn->nn_lock);
652         if (nn->nn_sc == sc)
653                 o2net_set_nn_state(nn, NULL, 0, err);
654         spin_unlock(&nn->nn_lock);
655 }
656
657 /*
658  * This work queue function performs the blocking parts of socket shutdown.  A
659  * few paths lead here.  set_nn_state will trigger this callback if it sees an
660  * sc detached from the nn.  state_change will also trigger this callback
661  * directly when it sees errors.  In that case we need to call set_nn_state
662  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
663  * itself.
664  */
665 static void o2net_shutdown_sc(struct work_struct *work)
666 {
667         struct o2net_sock_container *sc =
668                 container_of(work, struct o2net_sock_container,
669                              sc_shutdown_work);
670         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
671
672         sclog(sc, "shutting down\n");
673
674         /* drop the callbacks ref and call shutdown only once */
675         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
676                 /* we shouldn't flush as we're in the thread, the
677                  * races with pending sc work structs are harmless */
678                 del_timer_sync(&sc->sc_idle_timeout);
679                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
680                 sc_put(sc);
681                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
682         }
683
684         /* not fatal so failed connects before the other guy has our
685          * heartbeat can be retried */
686         o2net_ensure_shutdown(nn, sc, 0);
687         sc_put(sc);
688 }
689
690 /* ------------------------------------------------------------ */
691
692 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
693                              u32 key)
694 {
695         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
696
697         if (ret == 0)
698                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
699
700         return ret;
701 }
702
703 static struct o2net_msg_handler *
704 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
705                           struct rb_node **ret_parent)
706 {
707         struct rb_node **p = &o2net_handler_tree.rb_node;
708         struct rb_node *parent = NULL;
709         struct o2net_msg_handler *nmh, *ret = NULL;
710         int cmp;
711
712         while (*p) {
713                 parent = *p;
714                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
715                 cmp = o2net_handler_cmp(nmh, msg_type, key);
716
717                 if (cmp < 0)
718                         p = &(*p)->rb_left;
719                 else if (cmp > 0)
720                         p = &(*p)->rb_right;
721                 else {
722                         ret = nmh;
723                         break;
724                 }
725         }
726
727         if (ret_p != NULL)
728                 *ret_p = p;
729         if (ret_parent != NULL)
730                 *ret_parent = parent;
731
732         return ret;
733 }
734
735 static void o2net_handler_kref_release(struct kref *kref)
736 {
737         struct o2net_msg_handler *nmh;
738         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
739
740         kfree(nmh);
741 }
742
743 static void o2net_handler_put(struct o2net_msg_handler *nmh)
744 {
745         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
746 }
747
748 /* max_len is protection for the handler func.  incoming messages won't
749  * be given to the handler if their payload is longer than the max. */
750 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
751                            o2net_msg_handler_func *func, void *data,
752                            o2net_post_msg_handler_func *post_func,
753                            struct list_head *unreg_list)
754 {
755         struct o2net_msg_handler *nmh = NULL;
756         struct rb_node **p, *parent;
757         int ret = 0;
758
759         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
760                 mlog(0, "max_len for message handler out of range: %u\n",
761                         max_len);
762                 ret = -EINVAL;
763                 goto out;
764         }
765
766         if (!msg_type) {
767                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
768                 ret = -EINVAL;
769                 goto out;
770
771         }
772         if (!func) {
773                 mlog(0, "no message handler provided: %u, %p\n",
774                        msg_type, func);
775                 ret = -EINVAL;
776                 goto out;
777         }
778
779         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
780         if (nmh == NULL) {
781                 ret = -ENOMEM;
782                 goto out;
783         }
784
785         nmh->nh_func = func;
786         nmh->nh_func_data = data;
787         nmh->nh_post_func = post_func;
788         nmh->nh_msg_type = msg_type;
789         nmh->nh_max_len = max_len;
790         nmh->nh_key = key;
791         /* the tree and list get this ref.. they're both removed in
792          * unregister when this ref is dropped */
793         kref_init(&nmh->nh_kref);
794         INIT_LIST_HEAD(&nmh->nh_unregister_item);
795
796         write_lock(&o2net_handler_lock);
797         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
798                 ret = -EEXIST;
799         else {
800                 rb_link_node(&nmh->nh_node, parent, p);
801                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
802                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
803
804                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
805                      func, msg_type, key);
806                 /* we've had some trouble with handlers seemingly vanishing. */
807                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
808                                                           &parent) == NULL,
809                                 "couldn't find handler we *just* registerd "
810                                 "for type %u key %08x\n", msg_type, key);
811         }
812         write_unlock(&o2net_handler_lock);
813         if (ret)
814                 goto out;
815
816 out:
817         if (ret)
818                 kfree(nmh);
819
820         return ret;
821 }
822 EXPORT_SYMBOL_GPL(o2net_register_handler);
823
824 void o2net_unregister_handler_list(struct list_head *list)
825 {
826         struct o2net_msg_handler *nmh, *n;
827
828         write_lock(&o2net_handler_lock);
829         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
830                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
831                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
832                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
833                 list_del_init(&nmh->nh_unregister_item);
834                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
835         }
836         write_unlock(&o2net_handler_lock);
837 }
838 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
839
840 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
841 {
842         struct o2net_msg_handler *nmh;
843
844         read_lock(&o2net_handler_lock);
845         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
846         if (nmh)
847                 kref_get(&nmh->nh_kref);
848         read_unlock(&o2net_handler_lock);
849
850         return nmh;
851 }
852
853 /* ------------------------------------------------------------ */
854
855 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
856 {
857         int ret;
858         mm_segment_t oldfs;
859         struct kvec vec = {
860                 .iov_len = len,
861                 .iov_base = data,
862         };
863         struct msghdr msg = {
864                 .msg_iovlen = 1,
865                 .msg_iov = (struct iovec *)&vec,
866                 .msg_flags = MSG_DONTWAIT,
867         };
868
869         oldfs = get_fs();
870         set_fs(get_ds());
871         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
872         set_fs(oldfs);
873
874         return ret;
875 }
876
877 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
878                               size_t veclen, size_t total)
879 {
880         int ret;
881         mm_segment_t oldfs;
882         struct msghdr msg = {
883                 .msg_iov = (struct iovec *)vec,
884                 .msg_iovlen = veclen,
885         };
886
887         if (sock == NULL) {
888                 ret = -EINVAL;
889                 goto out;
890         }
891
892         oldfs = get_fs();
893         set_fs(get_ds());
894         ret = sock_sendmsg(sock, &msg, total);
895         set_fs(oldfs);
896         if (ret != total) {
897                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
898                      total);
899                 if (ret >= 0)
900                         ret = -EPIPE; /* should be smarter, I bet */
901                 goto out;
902         }
903
904         ret = 0;
905 out:
906         if (ret < 0)
907                 mlog(0, "returning error: %d\n", ret);
908         return ret;
909 }
910
911 static void o2net_sendpage(struct o2net_sock_container *sc,
912                            void *kmalloced_virt,
913                            size_t size)
914 {
915         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
916         ssize_t ret;
917
918         while (1) {
919                 mutex_lock(&sc->sc_send_lock);
920                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
921                                                  virt_to_page(kmalloced_virt),
922                                                  (long)kmalloced_virt & ~PAGE_MASK,
923                                                  size, MSG_DONTWAIT);
924                 mutex_unlock(&sc->sc_send_lock);
925                 if (ret == size)
926                         break;
927                 if (ret == (ssize_t)-EAGAIN) {
928                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
929                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
930                         cond_resched();
931                         continue;
932                 }
933                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
934                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
935                 o2net_ensure_shutdown(nn, sc, 0);
936                 break;
937         }
938 }
939
940 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
941 {
942         memset(msg, 0, sizeof(struct o2net_msg));
943         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
944         msg->data_len = cpu_to_be16(data_len);
945         msg->msg_type = cpu_to_be16(msg_type);
946         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
947         msg->status = 0;
948         msg->key = cpu_to_be32(key);
949 }
950
951 static int o2net_tx_can_proceed(struct o2net_node *nn,
952                                 struct o2net_sock_container **sc_ret,
953                                 int *error)
954 {
955         int ret = 0;
956
957         spin_lock(&nn->nn_lock);
958         if (nn->nn_persistent_error) {
959                 ret = 1;
960                 *sc_ret = NULL;
961                 *error = nn->nn_persistent_error;
962         } else if (nn->nn_sc_valid) {
963                 kref_get(&nn->nn_sc->sc_kref);
964
965                 ret = 1;
966                 *sc_ret = nn->nn_sc;
967                 *error = 0;
968         }
969         spin_unlock(&nn->nn_lock);
970
971         return ret;
972 }
973
974 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
975                            size_t caller_veclen, u8 target_node, int *status)
976 {
977         int ret;
978         struct o2net_msg *msg = NULL;
979         size_t veclen, caller_bytes = 0;
980         struct kvec *vec = NULL;
981         struct o2net_sock_container *sc = NULL;
982         struct o2net_node *nn = o2net_nn_from_num(target_node);
983         struct o2net_status_wait nsw = {
984                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
985         };
986         struct o2net_send_tracking nst;
987
988         o2net_init_nst(&nst, msg_type, key, current, target_node);
989
990         if (o2net_wq == NULL) {
991                 mlog(0, "attempt to tx without o2netd running\n");
992                 ret = -ESRCH;
993                 goto out;
994         }
995
996         if (caller_veclen == 0) {
997                 mlog(0, "bad kvec array length\n");
998                 ret = -EINVAL;
999                 goto out;
1000         }
1001
1002         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1003         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1004                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1005                 ret = -EINVAL;
1006                 goto out;
1007         }
1008
1009         if (target_node == o2nm_this_node()) {
1010                 ret = -ELOOP;
1011                 goto out;
1012         }
1013
1014         o2net_debug_add_nst(&nst);
1015
1016         o2net_set_nst_sock_time(&nst);
1017
1018         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1019         if (ret)
1020                 goto out;
1021
1022         o2net_set_nst_sock_container(&nst, sc);
1023
1024         veclen = caller_veclen + 1;
1025         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1026         if (vec == NULL) {
1027                 mlog(0, "failed to %zu element kvec!\n", veclen);
1028                 ret = -ENOMEM;
1029                 goto out;
1030         }
1031
1032         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1033         if (!msg) {
1034                 mlog(0, "failed to allocate a o2net_msg!\n");
1035                 ret = -ENOMEM;
1036                 goto out;
1037         }
1038
1039         o2net_init_msg(msg, caller_bytes, msg_type, key);
1040
1041         vec[0].iov_len = sizeof(struct o2net_msg);
1042         vec[0].iov_base = msg;
1043         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1044
1045         ret = o2net_prep_nsw(nn, &nsw);
1046         if (ret)
1047                 goto out;
1048
1049         msg->msg_num = cpu_to_be32(nsw.ns_id);
1050         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1051
1052         o2net_set_nst_send_time(&nst);
1053
1054         /* finally, convert the message header to network byte-order
1055          * and send */
1056         mutex_lock(&sc->sc_send_lock);
1057         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1058                                  sizeof(struct o2net_msg) + caller_bytes);
1059         mutex_unlock(&sc->sc_send_lock);
1060         msglog(msg, "sending returned %d\n", ret);
1061         if (ret < 0) {
1062                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1063                 goto out;
1064         }
1065
1066         /* wait on other node's handler */
1067         o2net_set_nst_status_time(&nst);
1068         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1069
1070         /* Note that we avoid overwriting the callers status return
1071          * variable if a system error was reported on the other
1072          * side. Callers beware. */
1073         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1074         if (status && !ret)
1075                 *status = nsw.ns_status;
1076
1077         mlog(0, "woken, returning system status %d, user status %d\n",
1078              ret, nsw.ns_status);
1079 out:
1080         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1081         if (sc)
1082                 sc_put(sc);
1083         if (vec)
1084                 kfree(vec);
1085         if (msg)
1086                 kfree(msg);
1087         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1088         return ret;
1089 }
1090 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1091
1092 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1093                        u8 target_node, int *status)
1094 {
1095         struct kvec vec = {
1096                 .iov_base = data,
1097                 .iov_len = len,
1098         };
1099         return o2net_send_message_vec(msg_type, key, &vec, 1,
1100                                       target_node, status);
1101 }
1102 EXPORT_SYMBOL_GPL(o2net_send_message);
1103
1104 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1105                                    enum o2net_system_error syserr, int err)
1106 {
1107         struct kvec vec = {
1108                 .iov_base = hdr,
1109                 .iov_len = sizeof(struct o2net_msg),
1110         };
1111
1112         BUG_ON(syserr >= O2NET_ERR_MAX);
1113
1114         /* leave other fields intact from the incoming message, msg_num
1115          * in particular */
1116         hdr->sys_status = cpu_to_be32(syserr);
1117         hdr->status = cpu_to_be32(err);
1118         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1119         hdr->data_len = 0;
1120
1121         msglog(hdr, "about to send status magic %d\n", err);
1122         /* hdr has been in host byteorder this whole time */
1123         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1124 }
1125
1126 /* this returns -errno if the header was unknown or too large, etc.
1127  * after this is called the buffer us reused for the next message */
1128 static int o2net_process_message(struct o2net_sock_container *sc,
1129                                  struct o2net_msg *hdr)
1130 {
1131         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1132         int ret = 0, handler_status;
1133         enum  o2net_system_error syserr;
1134         struct o2net_msg_handler *nmh = NULL;
1135         void *ret_data = NULL;
1136
1137         msglog(hdr, "processing message\n");
1138
1139         o2net_sc_postpone_idle(sc);
1140
1141         switch(be16_to_cpu(hdr->magic)) {
1142                 case O2NET_MSG_STATUS_MAGIC:
1143                         /* special type for returning message status */
1144                         o2net_complete_nsw(nn, NULL,
1145                                            be32_to_cpu(hdr->msg_num),
1146                                            be32_to_cpu(hdr->sys_status),
1147                                            be32_to_cpu(hdr->status));
1148                         goto out;
1149                 case O2NET_MSG_KEEP_REQ_MAGIC:
1150                         o2net_sendpage(sc, o2net_keep_resp,
1151                                        sizeof(*o2net_keep_resp));
1152                         goto out;
1153                 case O2NET_MSG_KEEP_RESP_MAGIC:
1154                         goto out;
1155                 case O2NET_MSG_MAGIC:
1156                         break;
1157                 default:
1158                         msglog(hdr, "bad magic\n");
1159                         ret = -EINVAL;
1160                         goto out;
1161                         break;
1162         }
1163
1164         /* find a handler for it */
1165         handler_status = 0;
1166         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1167                                 be32_to_cpu(hdr->key));
1168         if (!nmh) {
1169                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1170                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1171                 syserr = O2NET_ERR_NO_HNDLR;
1172                 goto out_respond;
1173         }
1174
1175         syserr = O2NET_ERR_NONE;
1176
1177         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1178                 syserr = O2NET_ERR_OVERFLOW;
1179
1180         if (syserr != O2NET_ERR_NONE)
1181                 goto out_respond;
1182
1183         do_gettimeofday(&sc->sc_tv_func_start);
1184         sc->sc_msg_key = be32_to_cpu(hdr->key);
1185         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1186         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1187                                              be16_to_cpu(hdr->data_len),
1188                                         nmh->nh_func_data, &ret_data);
1189         do_gettimeofday(&sc->sc_tv_func_stop);
1190
1191 out_respond:
1192         /* this destroys the hdr, so don't use it after this */
1193         mutex_lock(&sc->sc_send_lock);
1194         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1195                                       handler_status);
1196         mutex_unlock(&sc->sc_send_lock);
1197         hdr = NULL;
1198         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1199              handler_status, syserr, ret);
1200
1201         if (nmh) {
1202                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1203                 if (nmh->nh_post_func)
1204                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1205                                             ret_data);
1206         }
1207
1208 out:
1209         if (nmh)
1210                 o2net_handler_put(nmh);
1211         return ret;
1212 }
1213
1214 static int o2net_check_handshake(struct o2net_sock_container *sc)
1215 {
1216         struct o2net_handshake *hand = page_address(sc->sc_page);
1217         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1218
1219         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1220                 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1221                      "version %llu but %llu is required, disconnecting\n",
1222                      SC_NODEF_ARGS(sc),
1223                      (unsigned long long)be64_to_cpu(hand->protocol_version),
1224                      O2NET_PROTOCOL_VERSION);
1225
1226                 /* don't bother reconnecting if its the wrong version. */
1227                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1228                 return -1;
1229         }
1230
1231         /*
1232          * Ensure timeouts are consistent with other nodes, otherwise
1233          * we can end up with one node thinking that the other must be down,
1234          * but isn't. This can ultimately cause corruption.
1235          */
1236         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1237                                 o2net_idle_timeout()) {
1238                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1239                      "%u ms, but we use %u ms locally.  disconnecting\n",
1240                      SC_NODEF_ARGS(sc),
1241                      be32_to_cpu(hand->o2net_idle_timeout_ms),
1242                      o2net_idle_timeout());
1243                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1244                 return -1;
1245         }
1246
1247         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1248                         o2net_keepalive_delay()) {
1249                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1250                      "%u ms, but we use %u ms locally.  disconnecting\n",
1251                      SC_NODEF_ARGS(sc),
1252                      be32_to_cpu(hand->o2net_keepalive_delay_ms),
1253                      o2net_keepalive_delay());
1254                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1255                 return -1;
1256         }
1257
1258         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1259                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1260                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1261                      "%u ms, but we use %u ms locally.  disconnecting\n",
1262                      SC_NODEF_ARGS(sc),
1263                      be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1264                      O2HB_MAX_WRITE_TIMEOUT_MS);
1265                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1266                 return -1;
1267         }
1268
1269         sc->sc_handshake_ok = 1;
1270
1271         spin_lock(&nn->nn_lock);
1272         /* set valid and queue the idle timers only if it hasn't been
1273          * shut down already */
1274         if (nn->nn_sc == sc) {
1275                 o2net_sc_reset_idle_timer(sc);
1276                 atomic_set(&nn->nn_timeout, 0);
1277                 o2net_set_nn_state(nn, sc, 1, 0);
1278         }
1279         spin_unlock(&nn->nn_lock);
1280
1281         /* shift everything up as though it wasn't there */
1282         sc->sc_page_off -= sizeof(struct o2net_handshake);
1283         if (sc->sc_page_off)
1284                 memmove(hand, hand + 1, sc->sc_page_off);
1285
1286         return 0;
1287 }
1288
1289 /* this demuxes the queued rx bytes into header or payload bits and calls
1290  * handlers as each full message is read off the socket.  it returns -error,
1291  * == 0 eof, or > 0 for progress made.*/
1292 static int o2net_advance_rx(struct o2net_sock_container *sc)
1293 {
1294         struct o2net_msg *hdr;
1295         int ret = 0;
1296         void *data;
1297         size_t datalen;
1298
1299         sclog(sc, "receiving\n");
1300         do_gettimeofday(&sc->sc_tv_advance_start);
1301
1302         if (unlikely(sc->sc_handshake_ok == 0)) {
1303                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1304                         data = page_address(sc->sc_page) + sc->sc_page_off;
1305                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1306                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1307                         if (ret > 0)
1308                                 sc->sc_page_off += ret;
1309                 }
1310
1311                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1312                         o2net_check_handshake(sc);
1313                         if (unlikely(sc->sc_handshake_ok == 0))
1314                                 ret = -EPROTO;
1315                 }
1316                 goto out;
1317         }
1318
1319         /* do we need more header? */
1320         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1321                 data = page_address(sc->sc_page) + sc->sc_page_off;
1322                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1323                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1324                 if (ret > 0) {
1325                         sc->sc_page_off += ret;
1326                         /* only swab incoming here.. we can
1327                          * only get here once as we cross from
1328                          * being under to over */
1329                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1330                                 hdr = page_address(sc->sc_page);
1331                                 if (be16_to_cpu(hdr->data_len) >
1332                                     O2NET_MAX_PAYLOAD_BYTES)
1333                                         ret = -EOVERFLOW;
1334                         }
1335                 }
1336                 if (ret <= 0)
1337                         goto out;
1338         }
1339
1340         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1341                 /* oof, still don't have a header */
1342                 goto out;
1343         }
1344
1345         /* this was swabbed above when we first read it */
1346         hdr = page_address(sc->sc_page);
1347
1348         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1349
1350         /* do we need more payload? */
1351         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1352                 /* need more payload */
1353                 data = page_address(sc->sc_page) + sc->sc_page_off;
1354                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1355                           sc->sc_page_off;
1356                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1357                 if (ret > 0)
1358                         sc->sc_page_off += ret;
1359                 if (ret <= 0)
1360                         goto out;
1361         }
1362
1363         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1364                 /* we can only get here once, the first time we read
1365                  * the payload.. so set ret to progress if the handler
1366                  * works out. after calling this the message is toast */
1367                 ret = o2net_process_message(sc, hdr);
1368                 if (ret == 0)
1369                         ret = 1;
1370                 sc->sc_page_off = 0;
1371         }
1372
1373 out:
1374         sclog(sc, "ret = %d\n", ret);
1375         do_gettimeofday(&sc->sc_tv_advance_stop);
1376         return ret;
1377 }
1378
1379 /* this work func is triggerd by data ready.  it reads until it can read no
1380  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1381  * our work the work struct will be marked and we'll be called again. */
1382 static void o2net_rx_until_empty(struct work_struct *work)
1383 {
1384         struct o2net_sock_container *sc =
1385                 container_of(work, struct o2net_sock_container, sc_rx_work);
1386         int ret;
1387
1388         do {
1389                 ret = o2net_advance_rx(sc);
1390         } while (ret > 0);
1391
1392         if (ret <= 0 && ret != -EAGAIN) {
1393                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1394                 sclog(sc, "saw error %d, closing\n", ret);
1395                 /* not permanent so read failed handshake can retry */
1396                 o2net_ensure_shutdown(nn, sc, 0);
1397         }
1398
1399         sc_put(sc);
1400 }
1401
1402 static int o2net_set_nodelay(struct socket *sock)
1403 {
1404         int ret, val = 1;
1405         mm_segment_t oldfs;
1406
1407         oldfs = get_fs();
1408         set_fs(KERNEL_DS);
1409
1410         /*
1411          * Dear unsuspecting programmer,
1412          *
1413          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1414          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1415          * silently turn into SO_DEBUG.
1416          *
1417          * Yours,
1418          * Keeper of hilariously fragile interfaces.
1419          */
1420         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1421                                     (char __user *)&val, sizeof(val));
1422
1423         set_fs(oldfs);
1424         return ret;
1425 }
1426
1427 static void o2net_initialize_handshake(void)
1428 {
1429         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1430                 O2HB_MAX_WRITE_TIMEOUT_MS);
1431         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1432         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1433                 o2net_keepalive_delay());
1434         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1435                 o2net_reconnect_delay());
1436 }
1437
1438 /* ------------------------------------------------------------ */
1439
1440 /* called when a connect completes and after a sock is accepted.  the
1441  * rx path will see the response and mark the sc valid */
1442 static void o2net_sc_connect_completed(struct work_struct *work)
1443 {
1444         struct o2net_sock_container *sc =
1445                 container_of(work, struct o2net_sock_container,
1446                              sc_connect_work);
1447
1448         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1449               (unsigned long long)O2NET_PROTOCOL_VERSION,
1450               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1451
1452         o2net_initialize_handshake();
1453         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1454         sc_put(sc);
1455 }
1456
1457 /* this is called as a work_struct func. */
1458 static void o2net_sc_send_keep_req(struct work_struct *work)
1459 {
1460         struct o2net_sock_container *sc =
1461                 container_of(work, struct o2net_sock_container,
1462                              sc_keepalive_work.work);
1463
1464         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1465         sc_put(sc);
1466 }
1467
1468 /* socket shutdown does a del_timer_sync against this as it tears down.
1469  * we can't start this timer until we've got to the point in sc buildup
1470  * where shutdown is going to be involved */
1471 static void o2net_idle_timer(unsigned long data)
1472 {
1473         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1474         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1475         struct timeval now;
1476
1477         do_gettimeofday(&now);
1478
1479         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1480              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1481                      o2net_idle_timeout() / 1000,
1482                      o2net_idle_timeout() % 1000);
1483         mlog(ML_NOTICE, "here are some times that might help debug the "
1484              "situation: (tmr %ld.%ld now %ld.%ld dr %ld.%ld adv "
1485              "%ld.%ld:%ld.%ld func (%08x:%u) %ld.%ld:%ld.%ld)\n",
1486              sc->sc_tv_timer.tv_sec, (long) sc->sc_tv_timer.tv_usec,
1487              now.tv_sec, (long) now.tv_usec,
1488              sc->sc_tv_data_ready.tv_sec, (long) sc->sc_tv_data_ready.tv_usec,
1489              sc->sc_tv_advance_start.tv_sec,
1490              (long) sc->sc_tv_advance_start.tv_usec,
1491              sc->sc_tv_advance_stop.tv_sec,
1492              (long) sc->sc_tv_advance_stop.tv_usec,
1493              sc->sc_msg_key, sc->sc_msg_type,
1494              sc->sc_tv_func_start.tv_sec, (long) sc->sc_tv_func_start.tv_usec,
1495              sc->sc_tv_func_stop.tv_sec, (long) sc->sc_tv_func_stop.tv_usec);
1496
1497         /*
1498          * Initialize the nn_timeout so that the next connection attempt
1499          * will continue in o2net_start_connect.
1500          */
1501         atomic_set(&nn->nn_timeout, 1);
1502
1503         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1504 }
1505
1506 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1507 {
1508         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1509         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1510                       msecs_to_jiffies(o2net_keepalive_delay()));
1511         do_gettimeofday(&sc->sc_tv_timer);
1512         mod_timer(&sc->sc_idle_timeout,
1513                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1514 }
1515
1516 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1517 {
1518         /* Only push out an existing timer */
1519         if (timer_pending(&sc->sc_idle_timeout))
1520                 o2net_sc_reset_idle_timer(sc);
1521 }
1522
1523 /* this work func is kicked whenever a path sets the nn state which doesn't
1524  * have valid set.  This includes seeing hb come up, losing a connection,
1525  * having a connect attempt fail, etc. This centralizes the logic which decides
1526  * if a connect attempt should be made or if we should give up and all future
1527  * transmit attempts should fail */
1528 static void o2net_start_connect(struct work_struct *work)
1529 {
1530         struct o2net_node *nn =
1531                 container_of(work, struct o2net_node, nn_connect_work.work);
1532         struct o2net_sock_container *sc = NULL;
1533         struct o2nm_node *node = NULL, *mynode = NULL;
1534         struct socket *sock = NULL;
1535         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1536         int ret = 0, stop;
1537         unsigned int timeout;
1538
1539         /* if we're greater we initiate tx, otherwise we accept */
1540         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1541                 goto out;
1542
1543         /* watch for racing with tearing a node down */
1544         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1545         if (node == NULL) {
1546                 ret = 0;
1547                 goto out;
1548         }
1549
1550         mynode = o2nm_get_node_by_num(o2nm_this_node());
1551         if (mynode == NULL) {
1552                 ret = 0;
1553                 goto out;
1554         }
1555
1556         spin_lock(&nn->nn_lock);
1557         /*
1558          * see if we already have one pending or have given up.
1559          * For nn_timeout, it is set when we close the connection
1560          * because of the idle time out. So it means that we have
1561          * at least connected to that node successfully once,
1562          * now try to connect to it again.
1563          */
1564         timeout = atomic_read(&nn->nn_timeout);
1565         stop = (nn->nn_sc ||
1566                 (nn->nn_persistent_error &&
1567                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1568         spin_unlock(&nn->nn_lock);
1569         if (stop)
1570                 goto out;
1571
1572         nn->nn_last_connect_attempt = jiffies;
1573
1574         sc = sc_alloc(node);
1575         if (sc == NULL) {
1576                 mlog(0, "couldn't allocate sc\n");
1577                 ret = -ENOMEM;
1578                 goto out;
1579         }
1580
1581         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1582         if (ret < 0) {
1583                 mlog(0, "can't create socket: %d\n", ret);
1584                 goto out;
1585         }
1586         sc->sc_sock = sock; /* freed by sc_kref_release */
1587
1588         sock->sk->sk_allocation = GFP_ATOMIC;
1589
1590         myaddr.sin_family = AF_INET;
1591         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1592         myaddr.sin_port = htons(0); /* any port */
1593
1594         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1595                               sizeof(myaddr));
1596         if (ret) {
1597                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1598                      ret, &mynode->nd_ipv4_address);
1599                 goto out;
1600         }
1601
1602         ret = o2net_set_nodelay(sc->sc_sock);
1603         if (ret) {
1604                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1605                 goto out;
1606         }
1607
1608         o2net_register_callbacks(sc->sc_sock->sk, sc);
1609
1610         spin_lock(&nn->nn_lock);
1611         /* handshake completion will set nn->nn_sc_valid */
1612         o2net_set_nn_state(nn, sc, 0, 0);
1613         spin_unlock(&nn->nn_lock);
1614
1615         remoteaddr.sin_family = AF_INET;
1616         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1617         remoteaddr.sin_port = node->nd_ipv4_port;
1618
1619         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1620                                         (struct sockaddr *)&remoteaddr,
1621                                         sizeof(remoteaddr),
1622                                         O_NONBLOCK);
1623         if (ret == -EINPROGRESS)
1624                 ret = 0;
1625
1626 out:
1627         if (ret) {
1628                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1629                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1630                 /* 0 err so that another will be queued and attempted
1631                  * from set_nn_state */
1632                 if (sc)
1633                         o2net_ensure_shutdown(nn, sc, 0);
1634         }
1635         if (sc)
1636                 sc_put(sc);
1637         if (node)
1638                 o2nm_node_put(node);
1639         if (mynode)
1640                 o2nm_node_put(mynode);
1641
1642         return;
1643 }
1644
1645 static void o2net_connect_expired(struct work_struct *work)
1646 {
1647         struct o2net_node *nn =
1648                 container_of(work, struct o2net_node, nn_connect_expired.work);
1649
1650         spin_lock(&nn->nn_lock);
1651         if (!nn->nn_sc_valid) {
1652                 mlog(ML_ERROR, "no connection established with node %u after "
1653                      "%u.%u seconds, giving up and returning errors.\n",
1654                      o2net_num_from_nn(nn),
1655                      o2net_idle_timeout() / 1000,
1656                      o2net_idle_timeout() % 1000);
1657
1658                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1659         }
1660         spin_unlock(&nn->nn_lock);
1661 }
1662
1663 static void o2net_still_up(struct work_struct *work)
1664 {
1665         struct o2net_node *nn =
1666                 container_of(work, struct o2net_node, nn_still_up.work);
1667
1668         o2quo_hb_still_up(o2net_num_from_nn(nn));
1669 }
1670
1671 /* ------------------------------------------------------------ */
1672
1673 void o2net_disconnect_node(struct o2nm_node *node)
1674 {
1675         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1676
1677         /* don't reconnect until it's heartbeating again */
1678         spin_lock(&nn->nn_lock);
1679         atomic_set(&nn->nn_timeout, 0);
1680         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1681         spin_unlock(&nn->nn_lock);
1682
1683         if (o2net_wq) {
1684                 cancel_delayed_work(&nn->nn_connect_expired);
1685                 cancel_delayed_work(&nn->nn_connect_work);
1686                 cancel_delayed_work(&nn->nn_still_up);
1687                 flush_workqueue(o2net_wq);
1688         }
1689 }
1690
1691 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1692                                   void *data)
1693 {
1694         o2quo_hb_down(node_num);
1695
1696         if (node_num != o2nm_this_node())
1697                 o2net_disconnect_node(node);
1698
1699         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1700 }
1701
1702 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1703                                 void *data)
1704 {
1705         struct o2net_node *nn = o2net_nn_from_num(node_num);
1706
1707         o2quo_hb_up(node_num);
1708
1709         /* ensure an immediate connect attempt */
1710         nn->nn_last_connect_attempt = jiffies -
1711                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1712
1713         if (node_num != o2nm_this_node()) {
1714                 /* believe it or not, accept and node hearbeating testing
1715                  * can succeed for this node before we got here.. so
1716                  * only use set_nn_state to clear the persistent error
1717                  * if that hasn't already happened */
1718                 spin_lock(&nn->nn_lock);
1719                 atomic_set(&nn->nn_timeout, 0);
1720                 if (nn->nn_persistent_error)
1721                         o2net_set_nn_state(nn, NULL, 0, 0);
1722                 spin_unlock(&nn->nn_lock);
1723         }
1724 }
1725
1726 void o2net_unregister_hb_callbacks(void)
1727 {
1728         o2hb_unregister_callback(NULL, &o2net_hb_up);
1729         o2hb_unregister_callback(NULL, &o2net_hb_down);
1730 }
1731
1732 int o2net_register_hb_callbacks(void)
1733 {
1734         int ret;
1735
1736         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1737                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1738         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1739                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1740
1741         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1742         if (ret == 0)
1743                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1744
1745         if (ret)
1746                 o2net_unregister_hb_callbacks();
1747
1748         return ret;
1749 }
1750
1751 /* ------------------------------------------------------------ */
1752
1753 static int o2net_accept_one(struct socket *sock)
1754 {
1755         int ret, slen;
1756         struct sockaddr_in sin;
1757         struct socket *new_sock = NULL;
1758         struct o2nm_node *node = NULL;
1759         struct o2net_sock_container *sc = NULL;
1760         struct o2net_node *nn;
1761
1762         BUG_ON(sock == NULL);
1763         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1764                                sock->sk->sk_protocol, &new_sock);
1765         if (ret)
1766                 goto out;
1767
1768         new_sock->type = sock->type;
1769         new_sock->ops = sock->ops;
1770         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1771         if (ret < 0)
1772                 goto out;
1773
1774         new_sock->sk->sk_allocation = GFP_ATOMIC;
1775
1776         ret = o2net_set_nodelay(new_sock);
1777         if (ret) {
1778                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1779                 goto out;
1780         }
1781
1782         slen = sizeof(sin);
1783         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1784                                        &slen, 1);
1785         if (ret < 0)
1786                 goto out;
1787
1788         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1789         if (node == NULL) {
1790                 mlog(ML_NOTICE, "attempt to connect from unknown node at %pI4:%d\n",
1791                      &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1792                 ret = -EINVAL;
1793                 goto out;
1794         }
1795
1796         if (o2nm_this_node() > node->nd_num) {
1797                 mlog(ML_NOTICE, "unexpected connect attempted from a lower "
1798                      "numbered node '%s' at " "%pI4:%d with num %u\n",
1799                      node->nd_name, &sin.sin_addr.s_addr,
1800                      ntohs(sin.sin_port), node->nd_num);
1801                 ret = -EINVAL;
1802                 goto out;
1803         }
1804
1805         /* this happens all the time when the other node sees our heartbeat
1806          * and tries to connect before we see their heartbeat */
1807         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1808                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1809                      "%pI4:%d but it isn't heartbeating\n",
1810                      node->nd_name, &sin.sin_addr.s_addr,
1811                      ntohs(sin.sin_port));
1812                 ret = -EINVAL;
1813                 goto out;
1814         }
1815
1816         nn = o2net_nn_from_num(node->nd_num);
1817
1818         spin_lock(&nn->nn_lock);
1819         if (nn->nn_sc)
1820                 ret = -EBUSY;
1821         else
1822                 ret = 0;
1823         spin_unlock(&nn->nn_lock);
1824         if (ret) {
1825                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1826                      "%pI4:%d but it already has an open connection\n",
1827                      node->nd_name, &sin.sin_addr.s_addr,
1828                      ntohs(sin.sin_port));
1829                 goto out;
1830         }
1831
1832         sc = sc_alloc(node);
1833         if (sc == NULL) {
1834                 ret = -ENOMEM;
1835                 goto out;
1836         }
1837
1838         sc->sc_sock = new_sock;
1839         new_sock = NULL;
1840
1841         spin_lock(&nn->nn_lock);
1842         atomic_set(&nn->nn_timeout, 0);
1843         o2net_set_nn_state(nn, sc, 0, 0);
1844         spin_unlock(&nn->nn_lock);
1845
1846         o2net_register_callbacks(sc->sc_sock->sk, sc);
1847         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1848
1849         o2net_initialize_handshake();
1850         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1851
1852 out:
1853         if (new_sock)
1854                 sock_release(new_sock);
1855         if (node)
1856                 o2nm_node_put(node);
1857         if (sc)
1858                 sc_put(sc);
1859         return ret;
1860 }
1861
1862 static void o2net_accept_many(struct work_struct *work)
1863 {
1864         struct socket *sock = o2net_listen_sock;
1865         while (o2net_accept_one(sock) == 0)
1866                 cond_resched();
1867 }
1868
1869 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1870 {
1871         void (*ready)(struct sock *sk, int bytes);
1872
1873         read_lock(&sk->sk_callback_lock);
1874         ready = sk->sk_user_data;
1875         if (ready == NULL) { /* check for teardown race */
1876                 ready = sk->sk_data_ready;
1877                 goto out;
1878         }
1879
1880         /* ->sk_data_ready is also called for a newly established child socket
1881          * before it has been accepted and the acceptor has set up their
1882          * data_ready.. we only want to queue listen work for our listening
1883          * socket */
1884         if (sk->sk_state == TCP_LISTEN) {
1885                 mlog(ML_TCP, "bytes: %d\n", bytes);
1886                 queue_work(o2net_wq, &o2net_listen_work);
1887         }
1888
1889 out:
1890         read_unlock(&sk->sk_callback_lock);
1891         ready(sk, bytes);
1892 }
1893
1894 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1895 {
1896         struct socket *sock = NULL;
1897         int ret;
1898         struct sockaddr_in sin = {
1899                 .sin_family = PF_INET,
1900                 .sin_addr = { .s_addr = addr },
1901                 .sin_port = port,
1902         };
1903
1904         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1905         if (ret < 0) {
1906                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1907                 goto out;
1908         }
1909
1910         sock->sk->sk_allocation = GFP_ATOMIC;
1911
1912         write_lock_bh(&sock->sk->sk_callback_lock);
1913         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1914         sock->sk->sk_data_ready = o2net_listen_data_ready;
1915         write_unlock_bh(&sock->sk->sk_callback_lock);
1916
1917         o2net_listen_sock = sock;
1918         INIT_WORK(&o2net_listen_work, o2net_accept_many);
1919
1920         sock->sk->sk_reuse = 1;
1921         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1922         if (ret < 0) {
1923                 mlog(ML_ERROR, "unable to bind socket at %pI4:%u, "
1924                      "ret=%d\n", &addr, ntohs(port), ret);
1925                 goto out;
1926         }
1927
1928         ret = sock->ops->listen(sock, 64);
1929         if (ret < 0) {
1930                 mlog(ML_ERROR, "unable to listen on %pI4:%u, ret=%d\n",
1931                      &addr, ntohs(port), ret);
1932         }
1933
1934 out:
1935         if (ret) {
1936                 o2net_listen_sock = NULL;
1937                 if (sock)
1938                         sock_release(sock);
1939         }
1940         return ret;
1941 }
1942
1943 /*
1944  * called from node manager when we should bring up our network listening
1945  * socket.  node manager handles all the serialization to only call this
1946  * once and to match it with o2net_stop_listening().  note,
1947  * o2nm_this_node() doesn't work yet as we're being called while it
1948  * is being set up.
1949  */
1950 int o2net_start_listening(struct o2nm_node *node)
1951 {
1952         int ret = 0;
1953
1954         BUG_ON(o2net_wq != NULL);
1955         BUG_ON(o2net_listen_sock != NULL);
1956
1957         mlog(ML_KTHREAD, "starting o2net thread...\n");
1958         o2net_wq = create_singlethread_workqueue("o2net");
1959         if (o2net_wq == NULL) {
1960                 mlog(ML_ERROR, "unable to launch o2net thread\n");
1961                 return -ENOMEM; /* ? */
1962         }
1963
1964         ret = o2net_open_listening_sock(node->nd_ipv4_address,
1965                                         node->nd_ipv4_port);
1966         if (ret) {
1967                 destroy_workqueue(o2net_wq);
1968                 o2net_wq = NULL;
1969         } else
1970                 o2quo_conn_up(node->nd_num);
1971
1972         return ret;
1973 }
1974
1975 /* again, o2nm_this_node() doesn't work here as we're involved in
1976  * tearing it down */
1977 void o2net_stop_listening(struct o2nm_node *node)
1978 {
1979         struct socket *sock = o2net_listen_sock;
1980         size_t i;
1981
1982         BUG_ON(o2net_wq == NULL);
1983         BUG_ON(o2net_listen_sock == NULL);
1984
1985         /* stop the listening socket from generating work */
1986         write_lock_bh(&sock->sk->sk_callback_lock);
1987         sock->sk->sk_data_ready = sock->sk->sk_user_data;
1988         sock->sk->sk_user_data = NULL;
1989         write_unlock_bh(&sock->sk->sk_callback_lock);
1990
1991         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
1992                 struct o2nm_node *node = o2nm_get_node_by_num(i);
1993                 if (node) {
1994                         o2net_disconnect_node(node);
1995                         o2nm_node_put(node);
1996                 }
1997         }
1998
1999         /* finish all work and tear down the work queue */
2000         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2001         destroy_workqueue(o2net_wq);
2002         o2net_wq = NULL;
2003
2004         sock_release(o2net_listen_sock);
2005         o2net_listen_sock = NULL;
2006
2007         o2quo_conn_err(node->nd_num);
2008 }
2009
2010 /* ------------------------------------------------------------ */
2011
2012 int o2net_init(void)
2013 {
2014         unsigned long i;
2015
2016         o2quo_init();
2017
2018         if (o2net_debugfs_init())
2019                 return -ENOMEM;
2020
2021         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2022         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2023         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2024         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2025                 kfree(o2net_hand);
2026                 kfree(o2net_keep_req);
2027                 kfree(o2net_keep_resp);
2028                 return -ENOMEM;
2029         }
2030
2031         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2032         o2net_hand->connector_id = cpu_to_be64(1);
2033
2034         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2035         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2036
2037         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2038                 struct o2net_node *nn = o2net_nn_from_num(i);
2039
2040                 atomic_set(&nn->nn_timeout, 0);
2041                 spin_lock_init(&nn->nn_lock);
2042                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2043                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2044                                   o2net_connect_expired);
2045                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2046                 /* until we see hb from a node we'll return einval */
2047                 nn->nn_persistent_error = -ENOTCONN;
2048                 init_waitqueue_head(&nn->nn_sc_wq);
2049                 idr_init(&nn->nn_status_idr);
2050                 INIT_LIST_HEAD(&nn->nn_status_list);
2051         }
2052
2053         return 0;
2054 }
2055
2056 void o2net_exit(void)
2057 {
2058         o2quo_exit();
2059         kfree(o2net_hand);
2060         kfree(o2net_keep_req);
2061         kfree(o2net_keep_resp);
2062         o2net_debugfs_exit();
2063 }