fs: add export.h to files using EXPORT_SYMBOL/THIS_MODULE macros
[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 <linux/export.h>
63 #include <net/tcp.h>
64
65 #include <asm/uaccess.h>
66
67 #include "heartbeat.h"
68 #include "tcp.h"
69 #include "nodemanager.h"
70 #define MLOG_MASK_PREFIX ML_TCP
71 #include "masklog.h"
72 #include "quorum.h"
73
74 #include "tcp_internal.h"
75
76 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
77 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
78                           &sc->sc_node->nd_ipv4_address,                \
79                           ntohs(sc->sc_node->nd_ipv4_port)
80
81 /*
82  * In the following two log macros, the whitespace after the ',' just
83  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
84  * previous token if args expands to nothing.
85  */
86 #define msglog(hdr, fmt, args...) do {                                  \
87         typeof(hdr) __hdr = (hdr);                                      \
88         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
89              "key %08x num %u] " fmt,                                   \
90              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
91              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
92              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
93              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
94 } while (0)
95
96 #define sclog(sc, fmt, args...) do {                                    \
97         typeof(sc) __sc = (sc);                                         \
98         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
99              "pg_off %zu] " fmt, __sc,                                  \
100              atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock,       \
101             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
102             ##args);                                                    \
103 } while (0)
104
105 static DEFINE_RWLOCK(o2net_handler_lock);
106 static struct rb_root o2net_handler_tree = RB_ROOT;
107
108 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
109
110 /* XXX someday we'll need better accounting */
111 static struct socket *o2net_listen_sock = NULL;
112
113 /*
114  * listen work is only queued by the listening socket callbacks on the
115  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
116  * quorum work is queued as sock containers are shutdown.. stop_listening
117  * tears down all the node's sock containers, preventing future shutdowns
118  * and queued quroum work, before canceling delayed quorum work and
119  * destroying the work queue.
120  */
121 static struct workqueue_struct *o2net_wq;
122 static struct work_struct o2net_listen_work;
123
124 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
125 #define O2NET_HB_PRI 0x1
126
127 static struct o2net_handshake *o2net_hand;
128 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
129
130 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
131                 {[O2NET_ERR_NONE]       = 0,
132                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
133                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
134                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
135
136 /* can't quite avoid *all* internal declarations :/ */
137 static void o2net_sc_connect_completed(struct work_struct *work);
138 static void o2net_rx_until_empty(struct work_struct *work);
139 static void o2net_shutdown_sc(struct work_struct *work);
140 static void o2net_listen_data_ready(struct sock *sk, int bytes);
141 static void o2net_sc_send_keep_req(struct work_struct *work);
142 static void o2net_idle_timer(unsigned long data);
143 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
144 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
145
146 #ifdef CONFIG_DEBUG_FS
147 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
148                            u32 msgkey, struct task_struct *task, u8 node)
149 {
150         INIT_LIST_HEAD(&nst->st_net_debug_item);
151         nst->st_task = task;
152         nst->st_msg_type = msgtype;
153         nst->st_msg_key = msgkey;
154         nst->st_node = node;
155 }
156
157 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
158 {
159         nst->st_sock_time = ktime_get();
160 }
161
162 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
163 {
164         nst->st_send_time = ktime_get();
165 }
166
167 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
168 {
169         nst->st_status_time = ktime_get();
170 }
171
172 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
173                                                 struct o2net_sock_container *sc)
174 {
175         nst->st_sc = sc;
176 }
177
178 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
179                                         u32 msg_id)
180 {
181         nst->st_id = msg_id;
182 }
183
184 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
185 {
186         sc->sc_tv_timer = ktime_get();
187 }
188
189 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
190 {
191         sc->sc_tv_data_ready = ktime_get();
192 }
193
194 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
195 {
196         sc->sc_tv_advance_start = ktime_get();
197 }
198
199 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
200 {
201         sc->sc_tv_advance_stop = ktime_get();
202 }
203
204 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
205 {
206         sc->sc_tv_func_start = ktime_get();
207 }
208
209 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
210 {
211         sc->sc_tv_func_stop = ktime_get();
212 }
213
214 #else  /* CONFIG_DEBUG_FS */
215 # define o2net_init_nst(a, b, c, d, e)
216 # define o2net_set_nst_sock_time(a)
217 # define o2net_set_nst_send_time(a)
218 # define o2net_set_nst_status_time(a)
219 # define o2net_set_nst_sock_container(a, b)
220 # define o2net_set_nst_msg_id(a, b)
221 # define o2net_set_sock_timer(a)
222 # define o2net_set_data_ready_time(a)
223 # define o2net_set_advance_start_time(a)
224 # define o2net_set_advance_stop_time(a)
225 # define o2net_set_func_start_time(a)
226 # define o2net_set_func_stop_time(a)
227 #endif /* CONFIG_DEBUG_FS */
228
229 #ifdef CONFIG_OCFS2_FS_STATS
230 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
231 {
232         return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
233 }
234
235 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
236                                     struct o2net_sock_container *sc)
237 {
238         sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
239                                            ktime_sub(ktime_get(),
240                                                      nst->st_status_time));
241         sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
242                                          ktime_sub(nst->st_status_time,
243                                                    nst->st_send_time));
244         sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
245                                             ktime_sub(nst->st_send_time,
246                                                       nst->st_sock_time));
247         sc->sc_send_count++;
248 }
249
250 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
251 {
252         sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
253                                             o2net_get_func_run_time(sc));
254         sc->sc_recv_count++;
255 }
256
257 #else
258
259 # define o2net_update_send_stats(a, b)
260
261 # define o2net_update_recv_stats(sc)
262
263 #endif /* CONFIG_OCFS2_FS_STATS */
264
265 static inline int o2net_reconnect_delay(void)
266 {
267         return o2nm_single_cluster->cl_reconnect_delay_ms;
268 }
269
270 static inline int o2net_keepalive_delay(void)
271 {
272         return o2nm_single_cluster->cl_keepalive_delay_ms;
273 }
274
275 static inline int o2net_idle_timeout(void)
276 {
277         return o2nm_single_cluster->cl_idle_timeout_ms;
278 }
279
280 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
281 {
282         int trans;
283         BUG_ON(err >= O2NET_ERR_MAX);
284         trans = o2net_sys_err_translations[err];
285
286         /* Just in case we mess up the translation table above */
287         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
288         return trans;
289 }
290
291 static struct o2net_node * o2net_nn_from_num(u8 node_num)
292 {
293         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
294         return &o2net_nodes[node_num];
295 }
296
297 static u8 o2net_num_from_nn(struct o2net_node *nn)
298 {
299         BUG_ON(nn == NULL);
300         return nn - o2net_nodes;
301 }
302
303 /* ------------------------------------------------------------ */
304
305 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
306 {
307         int ret = 0;
308
309         do {
310                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
311                         ret = -EAGAIN;
312                         break;
313                 }
314                 spin_lock(&nn->nn_lock);
315                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
316                 if (ret == 0)
317                         list_add_tail(&nsw->ns_node_item,
318                                       &nn->nn_status_list);
319                 spin_unlock(&nn->nn_lock);
320         } while (ret == -EAGAIN);
321
322         if (ret == 0)  {
323                 init_waitqueue_head(&nsw->ns_wq);
324                 nsw->ns_sys_status = O2NET_ERR_NONE;
325                 nsw->ns_status = 0;
326         }
327
328         return ret;
329 }
330
331 static void o2net_complete_nsw_locked(struct o2net_node *nn,
332                                       struct o2net_status_wait *nsw,
333                                       enum o2net_system_error sys_status,
334                                       s32 status)
335 {
336         assert_spin_locked(&nn->nn_lock);
337
338         if (!list_empty(&nsw->ns_node_item)) {
339                 list_del_init(&nsw->ns_node_item);
340                 nsw->ns_sys_status = sys_status;
341                 nsw->ns_status = status;
342                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
343                 wake_up(&nsw->ns_wq);
344         }
345 }
346
347 static void o2net_complete_nsw(struct o2net_node *nn,
348                                struct o2net_status_wait *nsw,
349                                u64 id, enum o2net_system_error sys_status,
350                                s32 status)
351 {
352         spin_lock(&nn->nn_lock);
353         if (nsw == NULL) {
354                 if (id > INT_MAX)
355                         goto out;
356
357                 nsw = idr_find(&nn->nn_status_idr, id);
358                 if (nsw == NULL)
359                         goto out;
360         }
361
362         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
363
364 out:
365         spin_unlock(&nn->nn_lock);
366         return;
367 }
368
369 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
370 {
371         struct o2net_status_wait *nsw, *tmp;
372         unsigned int num_kills = 0;
373
374         assert_spin_locked(&nn->nn_lock);
375
376         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
377                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
378                 num_kills++;
379         }
380
381         mlog(0, "completed %d messages for node %u\n", num_kills,
382              o2net_num_from_nn(nn));
383 }
384
385 static int o2net_nsw_completed(struct o2net_node *nn,
386                                struct o2net_status_wait *nsw)
387 {
388         int completed;
389         spin_lock(&nn->nn_lock);
390         completed = list_empty(&nsw->ns_node_item);
391         spin_unlock(&nn->nn_lock);
392         return completed;
393 }
394
395 /* ------------------------------------------------------------ */
396
397 static void sc_kref_release(struct kref *kref)
398 {
399         struct o2net_sock_container *sc = container_of(kref,
400                                         struct o2net_sock_container, sc_kref);
401         BUG_ON(timer_pending(&sc->sc_idle_timeout));
402
403         sclog(sc, "releasing\n");
404
405         if (sc->sc_sock) {
406                 sock_release(sc->sc_sock);
407                 sc->sc_sock = NULL;
408         }
409
410         o2nm_undepend_item(&sc->sc_node->nd_item);
411         o2nm_node_put(sc->sc_node);
412         sc->sc_node = NULL;
413
414         o2net_debug_del_sc(sc);
415         kfree(sc);
416 }
417
418 static void sc_put(struct o2net_sock_container *sc)
419 {
420         sclog(sc, "put\n");
421         kref_put(&sc->sc_kref, sc_kref_release);
422 }
423 static void sc_get(struct o2net_sock_container *sc)
424 {
425         sclog(sc, "get\n");
426         kref_get(&sc->sc_kref);
427 }
428 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
429 {
430         struct o2net_sock_container *sc, *ret = NULL;
431         struct page *page = NULL;
432         int status = 0;
433
434         page = alloc_page(GFP_NOFS);
435         sc = kzalloc(sizeof(*sc), GFP_NOFS);
436         if (sc == NULL || page == NULL)
437                 goto out;
438
439         kref_init(&sc->sc_kref);
440         o2nm_node_get(node);
441         sc->sc_node = node;
442
443         /* pin the node item of the remote node */
444         status = o2nm_depend_item(&node->nd_item);
445         if (status) {
446                 mlog_errno(status);
447                 o2nm_node_put(node);
448                 goto out;
449         }
450         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
451         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
452         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
453         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
454
455         init_timer(&sc->sc_idle_timeout);
456         sc->sc_idle_timeout.function = o2net_idle_timer;
457         sc->sc_idle_timeout.data = (unsigned long)sc;
458
459         sclog(sc, "alloced\n");
460
461         ret = sc;
462         sc->sc_page = page;
463         o2net_debug_add_sc(sc);
464         sc = NULL;
465         page = NULL;
466
467 out:
468         if (page)
469                 __free_page(page);
470         kfree(sc);
471
472         return ret;
473 }
474
475 /* ------------------------------------------------------------ */
476
477 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
478                                 struct work_struct *work)
479 {
480         sc_get(sc);
481         if (!queue_work(o2net_wq, work))
482                 sc_put(sc);
483 }
484 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
485                                         struct delayed_work *work,
486                                         int delay)
487 {
488         sc_get(sc);
489         if (!queue_delayed_work(o2net_wq, work, delay))
490                 sc_put(sc);
491 }
492 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
493                                          struct delayed_work *work)
494 {
495         if (cancel_delayed_work(work))
496                 sc_put(sc);
497 }
498
499 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
500
501 int o2net_num_connected_peers(void)
502 {
503         return atomic_read(&o2net_connected_peers);
504 }
505
506 static void o2net_set_nn_state(struct o2net_node *nn,
507                                struct o2net_sock_container *sc,
508                                unsigned valid, int err)
509 {
510         int was_valid = nn->nn_sc_valid;
511         int was_err = nn->nn_persistent_error;
512         struct o2net_sock_container *old_sc = nn->nn_sc;
513
514         assert_spin_locked(&nn->nn_lock);
515
516         if (old_sc && !sc)
517                 atomic_dec(&o2net_connected_peers);
518         else if (!old_sc && sc)
519                 atomic_inc(&o2net_connected_peers);
520
521         /* the node num comparison and single connect/accept path should stop
522          * an non-null sc from being overwritten with another */
523         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
524         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
525         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
526
527         if (was_valid && !valid && err == 0)
528                 err = -ENOTCONN;
529
530         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
531              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
532              nn->nn_persistent_error, err);
533
534         nn->nn_sc = sc;
535         nn->nn_sc_valid = valid ? 1 : 0;
536         nn->nn_persistent_error = err;
537
538         /* mirrors o2net_tx_can_proceed() */
539         if (nn->nn_persistent_error || nn->nn_sc_valid)
540                 wake_up(&nn->nn_sc_wq);
541
542         if (!was_err && nn->nn_persistent_error) {
543                 o2quo_conn_err(o2net_num_from_nn(nn));
544                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
545                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
546         }
547
548         if (was_valid && !valid) {
549                 printk(KERN_NOTICE "o2net: no longer connected to "
550                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
551                 o2net_complete_nodes_nsw(nn);
552         }
553
554         if (!was_valid && valid) {
555                 o2quo_conn_up(o2net_num_from_nn(nn));
556                 cancel_delayed_work(&nn->nn_connect_expired);
557                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
558                        o2nm_this_node() > sc->sc_node->nd_num ?
559                                 "connected to" : "accepted connection from",
560                        SC_NODEF_ARGS(sc));
561         }
562
563         /* trigger the connecting worker func as long as we're not valid,
564          * it will back off if it shouldn't connect.  This can be called
565          * from node config teardown and so needs to be careful about
566          * the work queue actually being up. */
567         if (!valid && o2net_wq) {
568                 unsigned long delay;
569                 /* delay if we're within a RECONNECT_DELAY of the
570                  * last attempt */
571                 delay = (nn->nn_last_connect_attempt +
572                          msecs_to_jiffies(o2net_reconnect_delay()))
573                         - jiffies;
574                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
575                         delay = 0;
576                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
577                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
578
579                 /*
580                  * Delay the expired work after idle timeout.
581                  *
582                  * We might have lots of failed connection attempts that run
583                  * through here but we only cancel the connect_expired work when
584                  * a connection attempt succeeds.  So only the first enqueue of
585                  * the connect_expired work will do anything.  The rest will see
586                  * that it's already queued and do nothing.
587                  */
588                 delay += msecs_to_jiffies(o2net_idle_timeout());
589                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
590         }
591
592         /* keep track of the nn's sc ref for the caller */
593         if ((old_sc == NULL) && sc)
594                 sc_get(sc);
595         if (old_sc && (old_sc != sc)) {
596                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
597                 sc_put(old_sc);
598         }
599 }
600
601 /* see o2net_register_callbacks() */
602 static void o2net_data_ready(struct sock *sk, int bytes)
603 {
604         void (*ready)(struct sock *sk, int bytes);
605
606         read_lock(&sk->sk_callback_lock);
607         if (sk->sk_user_data) {
608                 struct o2net_sock_container *sc = sk->sk_user_data;
609                 sclog(sc, "data_ready hit\n");
610                 o2net_set_data_ready_time(sc);
611                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
612                 ready = sc->sc_data_ready;
613         } else {
614                 ready = sk->sk_data_ready;
615         }
616         read_unlock(&sk->sk_callback_lock);
617
618         ready(sk, bytes);
619 }
620
621 /* see o2net_register_callbacks() */
622 static void o2net_state_change(struct sock *sk)
623 {
624         void (*state_change)(struct sock *sk);
625         struct o2net_sock_container *sc;
626
627         read_lock(&sk->sk_callback_lock);
628         sc = sk->sk_user_data;
629         if (sc == NULL) {
630                 state_change = sk->sk_state_change;
631                 goto out;
632         }
633
634         sclog(sc, "state_change to %d\n", sk->sk_state);
635
636         state_change = sc->sc_state_change;
637
638         switch(sk->sk_state) {
639                 /* ignore connecting sockets as they make progress */
640                 case TCP_SYN_SENT:
641                 case TCP_SYN_RECV:
642                         break;
643                 case TCP_ESTABLISHED:
644                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
645                         break;
646                 default:
647                         printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT
648                               " shutdown, state %d\n",
649                               SC_NODEF_ARGS(sc), sk->sk_state);
650                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
651                         break;
652         }
653 out:
654         read_unlock(&sk->sk_callback_lock);
655         state_change(sk);
656 }
657
658 /*
659  * we register callbacks so we can queue work on events before calling
660  * the original callbacks.  our callbacks our careful to test user_data
661  * to discover when they've reaced with o2net_unregister_callbacks().
662  */
663 static void o2net_register_callbacks(struct sock *sk,
664                                      struct o2net_sock_container *sc)
665 {
666         write_lock_bh(&sk->sk_callback_lock);
667
668         /* accepted sockets inherit the old listen socket data ready */
669         if (sk->sk_data_ready == o2net_listen_data_ready) {
670                 sk->sk_data_ready = sk->sk_user_data;
671                 sk->sk_user_data = NULL;
672         }
673
674         BUG_ON(sk->sk_user_data != NULL);
675         sk->sk_user_data = sc;
676         sc_get(sc);
677
678         sc->sc_data_ready = sk->sk_data_ready;
679         sc->sc_state_change = sk->sk_state_change;
680         sk->sk_data_ready = o2net_data_ready;
681         sk->sk_state_change = o2net_state_change;
682
683         mutex_init(&sc->sc_send_lock);
684
685         write_unlock_bh(&sk->sk_callback_lock);
686 }
687
688 static int o2net_unregister_callbacks(struct sock *sk,
689                                    struct o2net_sock_container *sc)
690 {
691         int ret = 0;
692
693         write_lock_bh(&sk->sk_callback_lock);
694         if (sk->sk_user_data == sc) {
695                 ret = 1;
696                 sk->sk_user_data = NULL;
697                 sk->sk_data_ready = sc->sc_data_ready;
698                 sk->sk_state_change = sc->sc_state_change;
699         }
700         write_unlock_bh(&sk->sk_callback_lock);
701
702         return ret;
703 }
704
705 /*
706  * this is a little helper that is called by callers who have seen a problem
707  * with an sc and want to detach it from the nn if someone already hasn't beat
708  * them to it.  if an error is given then the shutdown will be persistent
709  * and pending transmits will be canceled.
710  */
711 static void o2net_ensure_shutdown(struct o2net_node *nn,
712                                    struct o2net_sock_container *sc,
713                                    int err)
714 {
715         spin_lock(&nn->nn_lock);
716         if (nn->nn_sc == sc)
717                 o2net_set_nn_state(nn, NULL, 0, err);
718         spin_unlock(&nn->nn_lock);
719 }
720
721 /*
722  * This work queue function performs the blocking parts of socket shutdown.  A
723  * few paths lead here.  set_nn_state will trigger this callback if it sees an
724  * sc detached from the nn.  state_change will also trigger this callback
725  * directly when it sees errors.  In that case we need to call set_nn_state
726  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
727  * itself.
728  */
729 static void o2net_shutdown_sc(struct work_struct *work)
730 {
731         struct o2net_sock_container *sc =
732                 container_of(work, struct o2net_sock_container,
733                              sc_shutdown_work);
734         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
735
736         sclog(sc, "shutting down\n");
737
738         /* drop the callbacks ref and call shutdown only once */
739         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
740                 /* we shouldn't flush as we're in the thread, the
741                  * races with pending sc work structs are harmless */
742                 del_timer_sync(&sc->sc_idle_timeout);
743                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
744                 sc_put(sc);
745                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
746         }
747
748         /* not fatal so failed connects before the other guy has our
749          * heartbeat can be retried */
750         o2net_ensure_shutdown(nn, sc, 0);
751         sc_put(sc);
752 }
753
754 /* ------------------------------------------------------------ */
755
756 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
757                              u32 key)
758 {
759         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
760
761         if (ret == 0)
762                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
763
764         return ret;
765 }
766
767 static struct o2net_msg_handler *
768 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
769                           struct rb_node **ret_parent)
770 {
771         struct rb_node **p = &o2net_handler_tree.rb_node;
772         struct rb_node *parent = NULL;
773         struct o2net_msg_handler *nmh, *ret = NULL;
774         int cmp;
775
776         while (*p) {
777                 parent = *p;
778                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
779                 cmp = o2net_handler_cmp(nmh, msg_type, key);
780
781                 if (cmp < 0)
782                         p = &(*p)->rb_left;
783                 else if (cmp > 0)
784                         p = &(*p)->rb_right;
785                 else {
786                         ret = nmh;
787                         break;
788                 }
789         }
790
791         if (ret_p != NULL)
792                 *ret_p = p;
793         if (ret_parent != NULL)
794                 *ret_parent = parent;
795
796         return ret;
797 }
798
799 static void o2net_handler_kref_release(struct kref *kref)
800 {
801         struct o2net_msg_handler *nmh;
802         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
803
804         kfree(nmh);
805 }
806
807 static void o2net_handler_put(struct o2net_msg_handler *nmh)
808 {
809         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
810 }
811
812 /* max_len is protection for the handler func.  incoming messages won't
813  * be given to the handler if their payload is longer than the max. */
814 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
815                            o2net_msg_handler_func *func, void *data,
816                            o2net_post_msg_handler_func *post_func,
817                            struct list_head *unreg_list)
818 {
819         struct o2net_msg_handler *nmh = NULL;
820         struct rb_node **p, *parent;
821         int ret = 0;
822
823         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
824                 mlog(0, "max_len for message handler out of range: %u\n",
825                         max_len);
826                 ret = -EINVAL;
827                 goto out;
828         }
829
830         if (!msg_type) {
831                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
832                 ret = -EINVAL;
833                 goto out;
834
835         }
836         if (!func) {
837                 mlog(0, "no message handler provided: %u, %p\n",
838                        msg_type, func);
839                 ret = -EINVAL;
840                 goto out;
841         }
842
843         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
844         if (nmh == NULL) {
845                 ret = -ENOMEM;
846                 goto out;
847         }
848
849         nmh->nh_func = func;
850         nmh->nh_func_data = data;
851         nmh->nh_post_func = post_func;
852         nmh->nh_msg_type = msg_type;
853         nmh->nh_max_len = max_len;
854         nmh->nh_key = key;
855         /* the tree and list get this ref.. they're both removed in
856          * unregister when this ref is dropped */
857         kref_init(&nmh->nh_kref);
858         INIT_LIST_HEAD(&nmh->nh_unregister_item);
859
860         write_lock(&o2net_handler_lock);
861         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
862                 ret = -EEXIST;
863         else {
864                 rb_link_node(&nmh->nh_node, parent, p);
865                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
866                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
867
868                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
869                      func, msg_type, key);
870                 /* we've had some trouble with handlers seemingly vanishing. */
871                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
872                                                           &parent) == NULL,
873                                 "couldn't find handler we *just* registerd "
874                                 "for type %u key %08x\n", msg_type, key);
875         }
876         write_unlock(&o2net_handler_lock);
877         if (ret)
878                 goto out;
879
880 out:
881         if (ret)
882                 kfree(nmh);
883
884         return ret;
885 }
886 EXPORT_SYMBOL_GPL(o2net_register_handler);
887
888 void o2net_unregister_handler_list(struct list_head *list)
889 {
890         struct o2net_msg_handler *nmh, *n;
891
892         write_lock(&o2net_handler_lock);
893         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
894                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
895                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
896                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
897                 list_del_init(&nmh->nh_unregister_item);
898                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
899         }
900         write_unlock(&o2net_handler_lock);
901 }
902 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
903
904 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
905 {
906         struct o2net_msg_handler *nmh;
907
908         read_lock(&o2net_handler_lock);
909         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
910         if (nmh)
911                 kref_get(&nmh->nh_kref);
912         read_unlock(&o2net_handler_lock);
913
914         return nmh;
915 }
916
917 /* ------------------------------------------------------------ */
918
919 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
920 {
921         int ret;
922         mm_segment_t oldfs;
923         struct kvec vec = {
924                 .iov_len = len,
925                 .iov_base = data,
926         };
927         struct msghdr msg = {
928                 .msg_iovlen = 1,
929                 .msg_iov = (struct iovec *)&vec,
930                 .msg_flags = MSG_DONTWAIT,
931         };
932
933         oldfs = get_fs();
934         set_fs(get_ds());
935         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
936         set_fs(oldfs);
937
938         return ret;
939 }
940
941 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
942                               size_t veclen, size_t total)
943 {
944         int ret;
945         mm_segment_t oldfs;
946         struct msghdr msg = {
947                 .msg_iov = (struct iovec *)vec,
948                 .msg_iovlen = veclen,
949         };
950
951         if (sock == NULL) {
952                 ret = -EINVAL;
953                 goto out;
954         }
955
956         oldfs = get_fs();
957         set_fs(get_ds());
958         ret = sock_sendmsg(sock, &msg, total);
959         set_fs(oldfs);
960         if (ret != total) {
961                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
962                      total);
963                 if (ret >= 0)
964                         ret = -EPIPE; /* should be smarter, I bet */
965                 goto out;
966         }
967
968         ret = 0;
969 out:
970         if (ret < 0)
971                 mlog(0, "returning error: %d\n", ret);
972         return ret;
973 }
974
975 static void o2net_sendpage(struct o2net_sock_container *sc,
976                            void *kmalloced_virt,
977                            size_t size)
978 {
979         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
980         ssize_t ret;
981
982         while (1) {
983                 mutex_lock(&sc->sc_send_lock);
984                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
985                                                  virt_to_page(kmalloced_virt),
986                                                  (long)kmalloced_virt & ~PAGE_MASK,
987                                                  size, MSG_DONTWAIT);
988                 mutex_unlock(&sc->sc_send_lock);
989                 if (ret == size)
990                         break;
991                 if (ret == (ssize_t)-EAGAIN) {
992                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
993                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
994                         cond_resched();
995                         continue;
996                 }
997                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
998                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
999                 o2net_ensure_shutdown(nn, sc, 0);
1000                 break;
1001         }
1002 }
1003
1004 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
1005 {
1006         memset(msg, 0, sizeof(struct o2net_msg));
1007         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
1008         msg->data_len = cpu_to_be16(data_len);
1009         msg->msg_type = cpu_to_be16(msg_type);
1010         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
1011         msg->status = 0;
1012         msg->key = cpu_to_be32(key);
1013 }
1014
1015 static int o2net_tx_can_proceed(struct o2net_node *nn,
1016                                 struct o2net_sock_container **sc_ret,
1017                                 int *error)
1018 {
1019         int ret = 0;
1020
1021         spin_lock(&nn->nn_lock);
1022         if (nn->nn_persistent_error) {
1023                 ret = 1;
1024                 *sc_ret = NULL;
1025                 *error = nn->nn_persistent_error;
1026         } else if (nn->nn_sc_valid) {
1027                 kref_get(&nn->nn_sc->sc_kref);
1028
1029                 ret = 1;
1030                 *sc_ret = nn->nn_sc;
1031                 *error = 0;
1032         }
1033         spin_unlock(&nn->nn_lock);
1034
1035         return ret;
1036 }
1037
1038 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1039                            size_t caller_veclen, u8 target_node, int *status)
1040 {
1041         int ret = 0;
1042         struct o2net_msg *msg = NULL;
1043         size_t veclen, caller_bytes = 0;
1044         struct kvec *vec = NULL;
1045         struct o2net_sock_container *sc = NULL;
1046         struct o2net_node *nn = o2net_nn_from_num(target_node);
1047         struct o2net_status_wait nsw = {
1048                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1049         };
1050         struct o2net_send_tracking nst;
1051
1052         o2net_init_nst(&nst, msg_type, key, current, target_node);
1053
1054         if (o2net_wq == NULL) {
1055                 mlog(0, "attempt to tx without o2netd running\n");
1056                 ret = -ESRCH;
1057                 goto out;
1058         }
1059
1060         if (caller_veclen == 0) {
1061                 mlog(0, "bad kvec array length\n");
1062                 ret = -EINVAL;
1063                 goto out;
1064         }
1065
1066         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1067         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1068                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1069                 ret = -EINVAL;
1070                 goto out;
1071         }
1072
1073         if (target_node == o2nm_this_node()) {
1074                 ret = -ELOOP;
1075                 goto out;
1076         }
1077
1078         o2net_debug_add_nst(&nst);
1079
1080         o2net_set_nst_sock_time(&nst);
1081
1082         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1083         if (ret)
1084                 goto out;
1085
1086         o2net_set_nst_sock_container(&nst, sc);
1087
1088         veclen = caller_veclen + 1;
1089         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1090         if (vec == NULL) {
1091                 mlog(0, "failed to %zu element kvec!\n", veclen);
1092                 ret = -ENOMEM;
1093                 goto out;
1094         }
1095
1096         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1097         if (!msg) {
1098                 mlog(0, "failed to allocate a o2net_msg!\n");
1099                 ret = -ENOMEM;
1100                 goto out;
1101         }
1102
1103         o2net_init_msg(msg, caller_bytes, msg_type, key);
1104
1105         vec[0].iov_len = sizeof(struct o2net_msg);
1106         vec[0].iov_base = msg;
1107         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1108
1109         ret = o2net_prep_nsw(nn, &nsw);
1110         if (ret)
1111                 goto out;
1112
1113         msg->msg_num = cpu_to_be32(nsw.ns_id);
1114         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1115
1116         o2net_set_nst_send_time(&nst);
1117
1118         /* finally, convert the message header to network byte-order
1119          * and send */
1120         mutex_lock(&sc->sc_send_lock);
1121         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1122                                  sizeof(struct o2net_msg) + caller_bytes);
1123         mutex_unlock(&sc->sc_send_lock);
1124         msglog(msg, "sending returned %d\n", ret);
1125         if (ret < 0) {
1126                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1127                 goto out;
1128         }
1129
1130         /* wait on other node's handler */
1131         o2net_set_nst_status_time(&nst);
1132         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1133
1134         o2net_update_send_stats(&nst, sc);
1135
1136         /* Note that we avoid overwriting the callers status return
1137          * variable if a system error was reported on the other
1138          * side. Callers beware. */
1139         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1140         if (status && !ret)
1141                 *status = nsw.ns_status;
1142
1143         mlog(0, "woken, returning system status %d, user status %d\n",
1144              ret, nsw.ns_status);
1145 out:
1146         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1147         if (sc)
1148                 sc_put(sc);
1149         if (vec)
1150                 kfree(vec);
1151         if (msg)
1152                 kfree(msg);
1153         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1154         return ret;
1155 }
1156 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1157
1158 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1159                        u8 target_node, int *status)
1160 {
1161         struct kvec vec = {
1162                 .iov_base = data,
1163                 .iov_len = len,
1164         };
1165         return o2net_send_message_vec(msg_type, key, &vec, 1,
1166                                       target_node, status);
1167 }
1168 EXPORT_SYMBOL_GPL(o2net_send_message);
1169
1170 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1171                                    enum o2net_system_error syserr, int err)
1172 {
1173         struct kvec vec = {
1174                 .iov_base = hdr,
1175                 .iov_len = sizeof(struct o2net_msg),
1176         };
1177
1178         BUG_ON(syserr >= O2NET_ERR_MAX);
1179
1180         /* leave other fields intact from the incoming message, msg_num
1181          * in particular */
1182         hdr->sys_status = cpu_to_be32(syserr);
1183         hdr->status = cpu_to_be32(err);
1184         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1185         hdr->data_len = 0;
1186
1187         msglog(hdr, "about to send status magic %d\n", err);
1188         /* hdr has been in host byteorder this whole time */
1189         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1190 }
1191
1192 /* this returns -errno if the header was unknown or too large, etc.
1193  * after this is called the buffer us reused for the next message */
1194 static int o2net_process_message(struct o2net_sock_container *sc,
1195                                  struct o2net_msg *hdr)
1196 {
1197         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1198         int ret = 0, handler_status;
1199         enum  o2net_system_error syserr;
1200         struct o2net_msg_handler *nmh = NULL;
1201         void *ret_data = NULL;
1202
1203         msglog(hdr, "processing message\n");
1204
1205         o2net_sc_postpone_idle(sc);
1206
1207         switch(be16_to_cpu(hdr->magic)) {
1208                 case O2NET_MSG_STATUS_MAGIC:
1209                         /* special type for returning message status */
1210                         o2net_complete_nsw(nn, NULL,
1211                                            be32_to_cpu(hdr->msg_num),
1212                                            be32_to_cpu(hdr->sys_status),
1213                                            be32_to_cpu(hdr->status));
1214                         goto out;
1215                 case O2NET_MSG_KEEP_REQ_MAGIC:
1216                         o2net_sendpage(sc, o2net_keep_resp,
1217                                        sizeof(*o2net_keep_resp));
1218                         goto out;
1219                 case O2NET_MSG_KEEP_RESP_MAGIC:
1220                         goto out;
1221                 case O2NET_MSG_MAGIC:
1222                         break;
1223                 default:
1224                         msglog(hdr, "bad magic\n");
1225                         ret = -EINVAL;
1226                         goto out;
1227                         break;
1228         }
1229
1230         /* find a handler for it */
1231         handler_status = 0;
1232         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1233                                 be32_to_cpu(hdr->key));
1234         if (!nmh) {
1235                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1236                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1237                 syserr = O2NET_ERR_NO_HNDLR;
1238                 goto out_respond;
1239         }
1240
1241         syserr = O2NET_ERR_NONE;
1242
1243         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1244                 syserr = O2NET_ERR_OVERFLOW;
1245
1246         if (syserr != O2NET_ERR_NONE)
1247                 goto out_respond;
1248
1249         o2net_set_func_start_time(sc);
1250         sc->sc_msg_key = be32_to_cpu(hdr->key);
1251         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1252         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1253                                              be16_to_cpu(hdr->data_len),
1254                                         nmh->nh_func_data, &ret_data);
1255         o2net_set_func_stop_time(sc);
1256
1257         o2net_update_recv_stats(sc);
1258
1259 out_respond:
1260         /* this destroys the hdr, so don't use it after this */
1261         mutex_lock(&sc->sc_send_lock);
1262         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1263                                       handler_status);
1264         mutex_unlock(&sc->sc_send_lock);
1265         hdr = NULL;
1266         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1267              handler_status, syserr, ret);
1268
1269         if (nmh) {
1270                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1271                 if (nmh->nh_post_func)
1272                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1273                                             ret_data);
1274         }
1275
1276 out:
1277         if (nmh)
1278                 o2net_handler_put(nmh);
1279         return ret;
1280 }
1281
1282 static int o2net_check_handshake(struct o2net_sock_container *sc)
1283 {
1284         struct o2net_handshake *hand = page_address(sc->sc_page);
1285         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1286
1287         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1288                 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1289                      "version %llu but %llu is required, disconnecting\n",
1290                      SC_NODEF_ARGS(sc),
1291                      (unsigned long long)be64_to_cpu(hand->protocol_version),
1292                      O2NET_PROTOCOL_VERSION);
1293
1294                 /* don't bother reconnecting if its the wrong version. */
1295                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1296                 return -1;
1297         }
1298
1299         /*
1300          * Ensure timeouts are consistent with other nodes, otherwise
1301          * we can end up with one node thinking that the other must be down,
1302          * but isn't. This can ultimately cause corruption.
1303          */
1304         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1305                                 o2net_idle_timeout()) {
1306                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1307                      "%u ms, but we use %u ms locally.  disconnecting\n",
1308                      SC_NODEF_ARGS(sc),
1309                      be32_to_cpu(hand->o2net_idle_timeout_ms),
1310                      o2net_idle_timeout());
1311                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1312                 return -1;
1313         }
1314
1315         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1316                         o2net_keepalive_delay()) {
1317                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1318                      "%u ms, but we use %u ms locally.  disconnecting\n",
1319                      SC_NODEF_ARGS(sc),
1320                      be32_to_cpu(hand->o2net_keepalive_delay_ms),
1321                      o2net_keepalive_delay());
1322                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1323                 return -1;
1324         }
1325
1326         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1327                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1328                 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1329                      "%u ms, but we use %u ms locally.  disconnecting\n",
1330                      SC_NODEF_ARGS(sc),
1331                      be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1332                      O2HB_MAX_WRITE_TIMEOUT_MS);
1333                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1334                 return -1;
1335         }
1336
1337         sc->sc_handshake_ok = 1;
1338
1339         spin_lock(&nn->nn_lock);
1340         /* set valid and queue the idle timers only if it hasn't been
1341          * shut down already */
1342         if (nn->nn_sc == sc) {
1343                 o2net_sc_reset_idle_timer(sc);
1344                 atomic_set(&nn->nn_timeout, 0);
1345                 o2net_set_nn_state(nn, sc, 1, 0);
1346         }
1347         spin_unlock(&nn->nn_lock);
1348
1349         /* shift everything up as though it wasn't there */
1350         sc->sc_page_off -= sizeof(struct o2net_handshake);
1351         if (sc->sc_page_off)
1352                 memmove(hand, hand + 1, sc->sc_page_off);
1353
1354         return 0;
1355 }
1356
1357 /* this demuxes the queued rx bytes into header or payload bits and calls
1358  * handlers as each full message is read off the socket.  it returns -error,
1359  * == 0 eof, or > 0 for progress made.*/
1360 static int o2net_advance_rx(struct o2net_sock_container *sc)
1361 {
1362         struct o2net_msg *hdr;
1363         int ret = 0;
1364         void *data;
1365         size_t datalen;
1366
1367         sclog(sc, "receiving\n");
1368         o2net_set_advance_start_time(sc);
1369
1370         if (unlikely(sc->sc_handshake_ok == 0)) {
1371                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1372                         data = page_address(sc->sc_page) + sc->sc_page_off;
1373                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1374                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1375                         if (ret > 0)
1376                                 sc->sc_page_off += ret;
1377                 }
1378
1379                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1380                         o2net_check_handshake(sc);
1381                         if (unlikely(sc->sc_handshake_ok == 0))
1382                                 ret = -EPROTO;
1383                 }
1384                 goto out;
1385         }
1386
1387         /* do we need more header? */
1388         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1389                 data = page_address(sc->sc_page) + sc->sc_page_off;
1390                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1391                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1392                 if (ret > 0) {
1393                         sc->sc_page_off += ret;
1394                         /* only swab incoming here.. we can
1395                          * only get here once as we cross from
1396                          * being under to over */
1397                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1398                                 hdr = page_address(sc->sc_page);
1399                                 if (be16_to_cpu(hdr->data_len) >
1400                                     O2NET_MAX_PAYLOAD_BYTES)
1401                                         ret = -EOVERFLOW;
1402                         }
1403                 }
1404                 if (ret <= 0)
1405                         goto out;
1406         }
1407
1408         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1409                 /* oof, still don't have a header */
1410                 goto out;
1411         }
1412
1413         /* this was swabbed above when we first read it */
1414         hdr = page_address(sc->sc_page);
1415
1416         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1417
1418         /* do we need more payload? */
1419         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1420                 /* need more payload */
1421                 data = page_address(sc->sc_page) + sc->sc_page_off;
1422                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1423                           sc->sc_page_off;
1424                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1425                 if (ret > 0)
1426                         sc->sc_page_off += ret;
1427                 if (ret <= 0)
1428                         goto out;
1429         }
1430
1431         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1432                 /* we can only get here once, the first time we read
1433                  * the payload.. so set ret to progress if the handler
1434                  * works out. after calling this the message is toast */
1435                 ret = o2net_process_message(sc, hdr);
1436                 if (ret == 0)
1437                         ret = 1;
1438                 sc->sc_page_off = 0;
1439         }
1440
1441 out:
1442         sclog(sc, "ret = %d\n", ret);
1443         o2net_set_advance_stop_time(sc);
1444         return ret;
1445 }
1446
1447 /* this work func is triggerd by data ready.  it reads until it can read no
1448  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1449  * our work the work struct will be marked and we'll be called again. */
1450 static void o2net_rx_until_empty(struct work_struct *work)
1451 {
1452         struct o2net_sock_container *sc =
1453                 container_of(work, struct o2net_sock_container, sc_rx_work);
1454         int ret;
1455
1456         do {
1457                 ret = o2net_advance_rx(sc);
1458         } while (ret > 0);
1459
1460         if (ret <= 0 && ret != -EAGAIN) {
1461                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1462                 sclog(sc, "saw error %d, closing\n", ret);
1463                 /* not permanent so read failed handshake can retry */
1464                 o2net_ensure_shutdown(nn, sc, 0);
1465         }
1466
1467         sc_put(sc);
1468 }
1469
1470 static int o2net_set_nodelay(struct socket *sock)
1471 {
1472         int ret, val = 1;
1473         mm_segment_t oldfs;
1474
1475         oldfs = get_fs();
1476         set_fs(KERNEL_DS);
1477
1478         /*
1479          * Dear unsuspecting programmer,
1480          *
1481          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1482          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1483          * silently turn into SO_DEBUG.
1484          *
1485          * Yours,
1486          * Keeper of hilariously fragile interfaces.
1487          */
1488         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1489                                     (char __user *)&val, sizeof(val));
1490
1491         set_fs(oldfs);
1492         return ret;
1493 }
1494
1495 static void o2net_initialize_handshake(void)
1496 {
1497         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1498                 O2HB_MAX_WRITE_TIMEOUT_MS);
1499         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1500         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1501                 o2net_keepalive_delay());
1502         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1503                 o2net_reconnect_delay());
1504 }
1505
1506 /* ------------------------------------------------------------ */
1507
1508 /* called when a connect completes and after a sock is accepted.  the
1509  * rx path will see the response and mark the sc valid */
1510 static void o2net_sc_connect_completed(struct work_struct *work)
1511 {
1512         struct o2net_sock_container *sc =
1513                 container_of(work, struct o2net_sock_container,
1514                              sc_connect_work);
1515
1516         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1517               (unsigned long long)O2NET_PROTOCOL_VERSION,
1518               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1519
1520         o2net_initialize_handshake();
1521         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1522         sc_put(sc);
1523 }
1524
1525 /* this is called as a work_struct func. */
1526 static void o2net_sc_send_keep_req(struct work_struct *work)
1527 {
1528         struct o2net_sock_container *sc =
1529                 container_of(work, struct o2net_sock_container,
1530                              sc_keepalive_work.work);
1531
1532         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1533         sc_put(sc);
1534 }
1535
1536 /* socket shutdown does a del_timer_sync against this as it tears down.
1537  * we can't start this timer until we've got to the point in sc buildup
1538  * where shutdown is going to be involved */
1539 static void o2net_idle_timer(unsigned long data)
1540 {
1541         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1542         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1543
1544 #ifdef CONFIG_DEBUG_FS
1545         ktime_t now = ktime_get();
1546 #endif
1547
1548         printk(KERN_NOTICE "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1549              "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1550                      o2net_idle_timeout() / 1000,
1551                      o2net_idle_timeout() % 1000);
1552
1553 #ifdef CONFIG_DEBUG_FS
1554         mlog(ML_NOTICE, "Here are some times that might help debug the "
1555              "situation: (Timer: %lld, Now %lld, DataReady %lld, Advance %lld-%lld, "
1556              "Key 0x%08x, Func %u, FuncTime %lld-%lld)\n",
1557              (long long)ktime_to_us(sc->sc_tv_timer), (long long)ktime_to_us(now),
1558              (long long)ktime_to_us(sc->sc_tv_data_ready),
1559              (long long)ktime_to_us(sc->sc_tv_advance_start),
1560              (long long)ktime_to_us(sc->sc_tv_advance_stop),
1561              sc->sc_msg_key, sc->sc_msg_type,
1562              (long long)ktime_to_us(sc->sc_tv_func_start),
1563              (long long)ktime_to_us(sc->sc_tv_func_stop));
1564 #endif
1565
1566         /*
1567          * Initialize the nn_timeout so that the next connection attempt
1568          * will continue in o2net_start_connect.
1569          */
1570         atomic_set(&nn->nn_timeout, 1);
1571
1572         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1573 }
1574
1575 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1576 {
1577         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1578         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1579                       msecs_to_jiffies(o2net_keepalive_delay()));
1580         o2net_set_sock_timer(sc);
1581         mod_timer(&sc->sc_idle_timeout,
1582                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1583 }
1584
1585 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1586 {
1587         /* Only push out an existing timer */
1588         if (timer_pending(&sc->sc_idle_timeout))
1589                 o2net_sc_reset_idle_timer(sc);
1590 }
1591
1592 /* this work func is kicked whenever a path sets the nn state which doesn't
1593  * have valid set.  This includes seeing hb come up, losing a connection,
1594  * having a connect attempt fail, etc. This centralizes the logic which decides
1595  * if a connect attempt should be made or if we should give up and all future
1596  * transmit attempts should fail */
1597 static void o2net_start_connect(struct work_struct *work)
1598 {
1599         struct o2net_node *nn =
1600                 container_of(work, struct o2net_node, nn_connect_work.work);
1601         struct o2net_sock_container *sc = NULL;
1602         struct o2nm_node *node = NULL, *mynode = NULL;
1603         struct socket *sock = NULL;
1604         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1605         int ret = 0, stop;
1606         unsigned int timeout;
1607
1608         /* if we're greater we initiate tx, otherwise we accept */
1609         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1610                 goto out;
1611
1612         /* watch for racing with tearing a node down */
1613         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1614         if (node == NULL) {
1615                 ret = 0;
1616                 goto out;
1617         }
1618
1619         mynode = o2nm_get_node_by_num(o2nm_this_node());
1620         if (mynode == NULL) {
1621                 ret = 0;
1622                 goto out;
1623         }
1624
1625         spin_lock(&nn->nn_lock);
1626         /*
1627          * see if we already have one pending or have given up.
1628          * For nn_timeout, it is set when we close the connection
1629          * because of the idle time out. So it means that we have
1630          * at least connected to that node successfully once,
1631          * now try to connect to it again.
1632          */
1633         timeout = atomic_read(&nn->nn_timeout);
1634         stop = (nn->nn_sc ||
1635                 (nn->nn_persistent_error &&
1636                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1637         spin_unlock(&nn->nn_lock);
1638         if (stop)
1639                 goto out;
1640
1641         nn->nn_last_connect_attempt = jiffies;
1642
1643         sc = sc_alloc(node);
1644         if (sc == NULL) {
1645                 mlog(0, "couldn't allocate sc\n");
1646                 ret = -ENOMEM;
1647                 goto out;
1648         }
1649
1650         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1651         if (ret < 0) {
1652                 mlog(0, "can't create socket: %d\n", ret);
1653                 goto out;
1654         }
1655         sc->sc_sock = sock; /* freed by sc_kref_release */
1656
1657         sock->sk->sk_allocation = GFP_ATOMIC;
1658
1659         myaddr.sin_family = AF_INET;
1660         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1661         myaddr.sin_port = htons(0); /* any port */
1662
1663         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1664                               sizeof(myaddr));
1665         if (ret) {
1666                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1667                      ret, &mynode->nd_ipv4_address);
1668                 goto out;
1669         }
1670
1671         ret = o2net_set_nodelay(sc->sc_sock);
1672         if (ret) {
1673                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1674                 goto out;
1675         }
1676
1677         o2net_register_callbacks(sc->sc_sock->sk, sc);
1678
1679         spin_lock(&nn->nn_lock);
1680         /* handshake completion will set nn->nn_sc_valid */
1681         o2net_set_nn_state(nn, sc, 0, 0);
1682         spin_unlock(&nn->nn_lock);
1683
1684         remoteaddr.sin_family = AF_INET;
1685         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1686         remoteaddr.sin_port = node->nd_ipv4_port;
1687
1688         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1689                                         (struct sockaddr *)&remoteaddr,
1690                                         sizeof(remoteaddr),
1691                                         O_NONBLOCK);
1692         if (ret == -EINPROGRESS)
1693                 ret = 0;
1694
1695 out:
1696         if (ret) {
1697                 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1698                      "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1699                 /* 0 err so that another will be queued and attempted
1700                  * from set_nn_state */
1701                 if (sc)
1702                         o2net_ensure_shutdown(nn, sc, 0);
1703         }
1704         if (sc)
1705                 sc_put(sc);
1706         if (node)
1707                 o2nm_node_put(node);
1708         if (mynode)
1709                 o2nm_node_put(mynode);
1710
1711         return;
1712 }
1713
1714 static void o2net_connect_expired(struct work_struct *work)
1715 {
1716         struct o2net_node *nn =
1717                 container_of(work, struct o2net_node, nn_connect_expired.work);
1718
1719         spin_lock(&nn->nn_lock);
1720         if (!nn->nn_sc_valid) {
1721                 mlog(ML_ERROR, "no connection established with node %u after "
1722                      "%u.%u seconds, giving up and returning errors.\n",
1723                      o2net_num_from_nn(nn),
1724                      o2net_idle_timeout() / 1000,
1725                      o2net_idle_timeout() % 1000);
1726
1727                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1728         }
1729         spin_unlock(&nn->nn_lock);
1730 }
1731
1732 static void o2net_still_up(struct work_struct *work)
1733 {
1734         struct o2net_node *nn =
1735                 container_of(work, struct o2net_node, nn_still_up.work);
1736
1737         o2quo_hb_still_up(o2net_num_from_nn(nn));
1738 }
1739
1740 /* ------------------------------------------------------------ */
1741
1742 void o2net_disconnect_node(struct o2nm_node *node)
1743 {
1744         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1745
1746         /* don't reconnect until it's heartbeating again */
1747         spin_lock(&nn->nn_lock);
1748         atomic_set(&nn->nn_timeout, 0);
1749         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1750         spin_unlock(&nn->nn_lock);
1751
1752         if (o2net_wq) {
1753                 cancel_delayed_work(&nn->nn_connect_expired);
1754                 cancel_delayed_work(&nn->nn_connect_work);
1755                 cancel_delayed_work(&nn->nn_still_up);
1756                 flush_workqueue(o2net_wq);
1757         }
1758 }
1759
1760 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1761                                   void *data)
1762 {
1763         o2quo_hb_down(node_num);
1764
1765         if (!node)
1766                 return;
1767
1768         if (node_num != o2nm_this_node())
1769                 o2net_disconnect_node(node);
1770
1771         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1772 }
1773
1774 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1775                                 void *data)
1776 {
1777         struct o2net_node *nn = o2net_nn_from_num(node_num);
1778
1779         o2quo_hb_up(node_num);
1780
1781         BUG_ON(!node);
1782
1783         /* ensure an immediate connect attempt */
1784         nn->nn_last_connect_attempt = jiffies -
1785                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1786
1787         if (node_num != o2nm_this_node()) {
1788                 /* believe it or not, accept and node hearbeating testing
1789                  * can succeed for this node before we got here.. so
1790                  * only use set_nn_state to clear the persistent error
1791                  * if that hasn't already happened */
1792                 spin_lock(&nn->nn_lock);
1793                 atomic_set(&nn->nn_timeout, 0);
1794                 if (nn->nn_persistent_error)
1795                         o2net_set_nn_state(nn, NULL, 0, 0);
1796                 spin_unlock(&nn->nn_lock);
1797         }
1798 }
1799
1800 void o2net_unregister_hb_callbacks(void)
1801 {
1802         o2hb_unregister_callback(NULL, &o2net_hb_up);
1803         o2hb_unregister_callback(NULL, &o2net_hb_down);
1804 }
1805
1806 int o2net_register_hb_callbacks(void)
1807 {
1808         int ret;
1809
1810         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1811                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1812         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1813                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1814
1815         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1816         if (ret == 0)
1817                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1818
1819         if (ret)
1820                 o2net_unregister_hb_callbacks();
1821
1822         return ret;
1823 }
1824
1825 /* ------------------------------------------------------------ */
1826
1827 static int o2net_accept_one(struct socket *sock)
1828 {
1829         int ret, slen;
1830         struct sockaddr_in sin;
1831         struct socket *new_sock = NULL;
1832         struct o2nm_node *node = NULL;
1833         struct o2nm_node *local_node = NULL;
1834         struct o2net_sock_container *sc = NULL;
1835         struct o2net_node *nn;
1836
1837         BUG_ON(sock == NULL);
1838         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1839                                sock->sk->sk_protocol, &new_sock);
1840         if (ret)
1841                 goto out;
1842
1843         new_sock->type = sock->type;
1844         new_sock->ops = sock->ops;
1845         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1846         if (ret < 0)
1847                 goto out;
1848
1849         new_sock->sk->sk_allocation = GFP_ATOMIC;
1850
1851         ret = o2net_set_nodelay(new_sock);
1852         if (ret) {
1853                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1854                 goto out;
1855         }
1856
1857         slen = sizeof(sin);
1858         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1859                                        &slen, 1);
1860         if (ret < 0)
1861                 goto out;
1862
1863         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1864         if (node == NULL) {
1865                 mlog(ML_NOTICE, "attempt to connect from unknown node at %pI4:%d\n",
1866                      &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1867                 ret = -EINVAL;
1868                 goto out;
1869         }
1870
1871         if (o2nm_this_node() >= node->nd_num) {
1872                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1873                 mlog(ML_NOTICE, "unexpected connect attempt seen at node '%s' ("
1874                      "%u, %pI4:%d) from node '%s' (%u, %pI4:%d)\n",
1875                      local_node->nd_name, local_node->nd_num,
1876                      &(local_node->nd_ipv4_address),
1877                      ntohs(local_node->nd_ipv4_port),
1878                      node->nd_name, node->nd_num, &sin.sin_addr.s_addr,
1879                      ntohs(sin.sin_port));
1880                 ret = -EINVAL;
1881                 goto out;
1882         }
1883
1884         /* this happens all the time when the other node sees our heartbeat
1885          * and tries to connect before we see their heartbeat */
1886         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1887                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1888                      "%pI4:%d but it isn't heartbeating\n",
1889                      node->nd_name, &sin.sin_addr.s_addr,
1890                      ntohs(sin.sin_port));
1891                 ret = -EINVAL;
1892                 goto out;
1893         }
1894
1895         nn = o2net_nn_from_num(node->nd_num);
1896
1897         spin_lock(&nn->nn_lock);
1898         if (nn->nn_sc)
1899                 ret = -EBUSY;
1900         else
1901                 ret = 0;
1902         spin_unlock(&nn->nn_lock);
1903         if (ret) {
1904                 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1905                      "%pI4:%d but it already has an open connection\n",
1906                      node->nd_name, &sin.sin_addr.s_addr,
1907                      ntohs(sin.sin_port));
1908                 goto out;
1909         }
1910
1911         sc = sc_alloc(node);
1912         if (sc == NULL) {
1913                 ret = -ENOMEM;
1914                 goto out;
1915         }
1916
1917         sc->sc_sock = new_sock;
1918         new_sock = NULL;
1919
1920         spin_lock(&nn->nn_lock);
1921         atomic_set(&nn->nn_timeout, 0);
1922         o2net_set_nn_state(nn, sc, 0, 0);
1923         spin_unlock(&nn->nn_lock);
1924
1925         o2net_register_callbacks(sc->sc_sock->sk, sc);
1926         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1927
1928         o2net_initialize_handshake();
1929         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1930
1931 out:
1932         if (new_sock)
1933                 sock_release(new_sock);
1934         if (node)
1935                 o2nm_node_put(node);
1936         if (local_node)
1937                 o2nm_node_put(local_node);
1938         if (sc)
1939                 sc_put(sc);
1940         return ret;
1941 }
1942
1943 static void o2net_accept_many(struct work_struct *work)
1944 {
1945         struct socket *sock = o2net_listen_sock;
1946         while (o2net_accept_one(sock) == 0)
1947                 cond_resched();
1948 }
1949
1950 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1951 {
1952         void (*ready)(struct sock *sk, int bytes);
1953
1954         read_lock(&sk->sk_callback_lock);
1955         ready = sk->sk_user_data;
1956         if (ready == NULL) { /* check for teardown race */
1957                 ready = sk->sk_data_ready;
1958                 goto out;
1959         }
1960
1961         /* ->sk_data_ready is also called for a newly established child socket
1962          * before it has been accepted and the acceptor has set up their
1963          * data_ready.. we only want to queue listen work for our listening
1964          * socket */
1965         if (sk->sk_state == TCP_LISTEN) {
1966                 mlog(ML_TCP, "bytes: %d\n", bytes);
1967                 queue_work(o2net_wq, &o2net_listen_work);
1968         }
1969
1970 out:
1971         read_unlock(&sk->sk_callback_lock);
1972         ready(sk, bytes);
1973 }
1974
1975 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1976 {
1977         struct socket *sock = NULL;
1978         int ret;
1979         struct sockaddr_in sin = {
1980                 .sin_family = PF_INET,
1981                 .sin_addr = { .s_addr = addr },
1982                 .sin_port = port,
1983         };
1984
1985         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1986         if (ret < 0) {
1987                 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1988                 goto out;
1989         }
1990
1991         sock->sk->sk_allocation = GFP_ATOMIC;
1992
1993         write_lock_bh(&sock->sk->sk_callback_lock);
1994         sock->sk->sk_user_data = sock->sk->sk_data_ready;
1995         sock->sk->sk_data_ready = o2net_listen_data_ready;
1996         write_unlock_bh(&sock->sk->sk_callback_lock);
1997
1998         o2net_listen_sock = sock;
1999         INIT_WORK(&o2net_listen_work, o2net_accept_many);
2000
2001         sock->sk->sk_reuse = 1;
2002         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2003         if (ret < 0) {
2004                 mlog(ML_ERROR, "unable to bind socket at %pI4:%u, "
2005                      "ret=%d\n", &addr, ntohs(port), ret);
2006                 goto out;
2007         }
2008
2009         ret = sock->ops->listen(sock, 64);
2010         if (ret < 0) {
2011                 mlog(ML_ERROR, "unable to listen on %pI4:%u, ret=%d\n",
2012                      &addr, ntohs(port), ret);
2013         }
2014
2015 out:
2016         if (ret) {
2017                 o2net_listen_sock = NULL;
2018                 if (sock)
2019                         sock_release(sock);
2020         }
2021         return ret;
2022 }
2023
2024 /*
2025  * called from node manager when we should bring up our network listening
2026  * socket.  node manager handles all the serialization to only call this
2027  * once and to match it with o2net_stop_listening().  note,
2028  * o2nm_this_node() doesn't work yet as we're being called while it
2029  * is being set up.
2030  */
2031 int o2net_start_listening(struct o2nm_node *node)
2032 {
2033         int ret = 0;
2034
2035         BUG_ON(o2net_wq != NULL);
2036         BUG_ON(o2net_listen_sock != NULL);
2037
2038         mlog(ML_KTHREAD, "starting o2net thread...\n");
2039         o2net_wq = create_singlethread_workqueue("o2net");
2040         if (o2net_wq == NULL) {
2041                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2042                 return -ENOMEM; /* ? */
2043         }
2044
2045         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2046                                         node->nd_ipv4_port);
2047         if (ret) {
2048                 destroy_workqueue(o2net_wq);
2049                 o2net_wq = NULL;
2050         } else
2051                 o2quo_conn_up(node->nd_num);
2052
2053         return ret;
2054 }
2055
2056 /* again, o2nm_this_node() doesn't work here as we're involved in
2057  * tearing it down */
2058 void o2net_stop_listening(struct o2nm_node *node)
2059 {
2060         struct socket *sock = o2net_listen_sock;
2061         size_t i;
2062
2063         BUG_ON(o2net_wq == NULL);
2064         BUG_ON(o2net_listen_sock == NULL);
2065
2066         /* stop the listening socket from generating work */
2067         write_lock_bh(&sock->sk->sk_callback_lock);
2068         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2069         sock->sk->sk_user_data = NULL;
2070         write_unlock_bh(&sock->sk->sk_callback_lock);
2071
2072         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2073                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2074                 if (node) {
2075                         o2net_disconnect_node(node);
2076                         o2nm_node_put(node);
2077                 }
2078         }
2079
2080         /* finish all work and tear down the work queue */
2081         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2082         destroy_workqueue(o2net_wq);
2083         o2net_wq = NULL;
2084
2085         sock_release(o2net_listen_sock);
2086         o2net_listen_sock = NULL;
2087
2088         o2quo_conn_err(node->nd_num);
2089 }
2090
2091 /* ------------------------------------------------------------ */
2092
2093 int o2net_init(void)
2094 {
2095         unsigned long i;
2096
2097         o2quo_init();
2098
2099         if (o2net_debugfs_init())
2100                 return -ENOMEM;
2101
2102         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2103         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2104         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2105         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2106                 kfree(o2net_hand);
2107                 kfree(o2net_keep_req);
2108                 kfree(o2net_keep_resp);
2109                 return -ENOMEM;
2110         }
2111
2112         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2113         o2net_hand->connector_id = cpu_to_be64(1);
2114
2115         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2116         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2117
2118         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2119                 struct o2net_node *nn = o2net_nn_from_num(i);
2120
2121                 atomic_set(&nn->nn_timeout, 0);
2122                 spin_lock_init(&nn->nn_lock);
2123                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2124                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2125                                   o2net_connect_expired);
2126                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2127                 /* until we see hb from a node we'll return einval */
2128                 nn->nn_persistent_error = -ENOTCONN;
2129                 init_waitqueue_head(&nn->nn_sc_wq);
2130                 idr_init(&nn->nn_status_idr);
2131                 INIT_LIST_HEAD(&nn->nn_status_list);
2132         }
2133
2134         return 0;
2135 }
2136
2137 void o2net_exit(void)
2138 {
2139         o2quo_exit();
2140         kfree(o2net_hand);
2141         kfree(o2net_keep_req);
2142         kfree(o2net_keep_resp);
2143         o2net_debugfs_exit();
2144 }