ocfs2/cluster: Add new function o2net_fill_node_map()
[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 %pI4:%u"
76 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
77                           &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 inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
157 {
158         nst->st_sock_time = ktime_get();
159 }
160
161 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
162 {
163         nst->st_send_time = ktime_get();
164 }
165
166 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
167 {
168         nst->st_status_time = ktime_get();
169 }
170
171 static inline 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 inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
178                                         u32 msg_id)
179 {
180         nst->st_id = msg_id;
181 }
182
183 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
184 {
185         sc->sc_tv_timer = ktime_get();
186 }
187
188 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
189 {
190         sc->sc_tv_data_ready = ktime_get();
191 }
192
193 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
194 {
195         sc->sc_tv_advance_start = ktime_get();
196 }
197
198 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
199 {
200         sc->sc_tv_advance_stop = ktime_get();
201 }
202
203 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
204 {
205         sc->sc_tv_func_start = ktime_get();
206 }
207
208 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
209 {
210         sc->sc_tv_func_stop = ktime_get();
211 }
212
213 #else  /* CONFIG_DEBUG_FS */
214 # define o2net_init_nst(a, b, c, d, e)
215 # define o2net_set_nst_sock_time(a)
216 # define o2net_set_nst_send_time(a)
217 # define o2net_set_nst_status_time(a)
218 # define o2net_set_nst_sock_container(a, b)
219 # define o2net_set_nst_msg_id(a, b)
220 # define o2net_set_sock_timer(a)
221 # define o2net_set_data_ready_time(a)
222 # define o2net_set_advance_start_time(a)
223 # define o2net_set_advance_stop_time(a)
224 # define o2net_set_func_start_time(a)
225 # define o2net_set_func_stop_time(a)
226 #endif /* CONFIG_DEBUG_FS */
227
228 #ifdef CONFIG_OCFS2_FS_STATS
229 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
230 {
231         return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
232 }
233
234 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
235                                     struct o2net_sock_container *sc)
236 {
237         sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
238                                            ktime_sub(ktime_get(),
239                                                      nst->st_status_time));
240         sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
241                                          ktime_sub(nst->st_status_time,
242                                                    nst->st_send_time));
243         sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
244                                             ktime_sub(nst->st_send_time,
245                                                       nst->st_sock_time));
246         sc->sc_send_count++;
247 }
248
249 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
250 {
251         sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
252                                             o2net_get_func_run_time(sc));
253         sc->sc_recv_count++;
254 }
255
256 #else
257
258 # define o2net_update_send_stats(a, b)
259
260 # define o2net_update_recv_stats(sc)
261
262 #endif /* CONFIG_OCFS2_FS_STATS */
263
264 static inline int o2net_reconnect_delay(void)
265 {
266         return o2nm_single_cluster->cl_reconnect_delay_ms;
267 }
268
269 static inline int o2net_keepalive_delay(void)
270 {
271         return o2nm_single_cluster->cl_keepalive_delay_ms;
272 }
273
274 static inline int o2net_idle_timeout(void)
275 {
276         return o2nm_single_cluster->cl_idle_timeout_ms;
277 }
278
279 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
280 {
281         int trans;
282         BUG_ON(err >= O2NET_ERR_MAX);
283         trans = o2net_sys_err_translations[err];
284
285         /* Just in case we mess up the translation table above */
286         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
287         return trans;
288 }
289
290 static struct o2net_node * o2net_nn_from_num(u8 node_num)
291 {
292         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
293         return &o2net_nodes[node_num];
294 }
295
296 static u8 o2net_num_from_nn(struct o2net_node *nn)
297 {
298         BUG_ON(nn == NULL);
299         return nn - o2net_nodes;
300 }
301
302 /* ------------------------------------------------------------ */
303
304 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
305 {
306         int ret = 0;
307
308         do {
309                 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
310                         ret = -EAGAIN;
311                         break;
312                 }
313                 spin_lock(&nn->nn_lock);
314                 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
315                 if (ret == 0)
316                         list_add_tail(&nsw->ns_node_item,
317                                       &nn->nn_status_list);
318                 spin_unlock(&nn->nn_lock);
319         } while (ret == -EAGAIN);
320
321         if (ret == 0)  {
322                 init_waitqueue_head(&nsw->ns_wq);
323                 nsw->ns_sys_status = O2NET_ERR_NONE;
324                 nsw->ns_status = 0;
325         }
326
327         return ret;
328 }
329
330 static void o2net_complete_nsw_locked(struct o2net_node *nn,
331                                       struct o2net_status_wait *nsw,
332                                       enum o2net_system_error sys_status,
333                                       s32 status)
334 {
335         assert_spin_locked(&nn->nn_lock);
336
337         if (!list_empty(&nsw->ns_node_item)) {
338                 list_del_init(&nsw->ns_node_item);
339                 nsw->ns_sys_status = sys_status;
340                 nsw->ns_status = status;
341                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
342                 wake_up(&nsw->ns_wq);
343         }
344 }
345
346 static void o2net_complete_nsw(struct o2net_node *nn,
347                                struct o2net_status_wait *nsw,
348                                u64 id, enum o2net_system_error sys_status,
349                                s32 status)
350 {
351         spin_lock(&nn->nn_lock);
352         if (nsw == NULL) {
353                 if (id > INT_MAX)
354                         goto out;
355
356                 nsw = idr_find(&nn->nn_status_idr, id);
357                 if (nsw == NULL)
358                         goto out;
359         }
360
361         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
362
363 out:
364         spin_unlock(&nn->nn_lock);
365         return;
366 }
367
368 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
369 {
370         struct o2net_status_wait *nsw, *tmp;
371         unsigned int num_kills = 0;
372
373         assert_spin_locked(&nn->nn_lock);
374
375         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
376                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
377                 num_kills++;
378         }
379
380         mlog(0, "completed %d messages for node %u\n", num_kills,
381              o2net_num_from_nn(nn));
382 }
383
384 static int o2net_nsw_completed(struct o2net_node *nn,
385                                struct o2net_status_wait *nsw)
386 {
387         int completed;
388         spin_lock(&nn->nn_lock);
389         completed = list_empty(&nsw->ns_node_item);
390         spin_unlock(&nn->nn_lock);
391         return completed;
392 }
393
394 /* ------------------------------------------------------------ */
395
396 static void sc_kref_release(struct kref *kref)
397 {
398         struct o2net_sock_container *sc = container_of(kref,
399                                         struct o2net_sock_container, sc_kref);
400         BUG_ON(timer_pending(&sc->sc_idle_timeout));
401
402         sclog(sc, "releasing\n");
403
404         if (sc->sc_sock) {
405                 sock_release(sc->sc_sock);
406                 sc->sc_sock = NULL;
407         }
408
409         o2nm_undepend_item(&sc->sc_node->nd_item);
410         o2nm_node_put(sc->sc_node);
411         sc->sc_node = NULL;
412
413         o2net_debug_del_sc(sc);
414         kfree(sc);
415 }
416
417 static void sc_put(struct o2net_sock_container *sc)
418 {
419         sclog(sc, "put\n");
420         kref_put(&sc->sc_kref, sc_kref_release);
421 }
422 static void sc_get(struct o2net_sock_container *sc)
423 {
424         sclog(sc, "get\n");
425         kref_get(&sc->sc_kref);
426 }
427 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
428 {
429         struct o2net_sock_container *sc, *ret = NULL;
430         struct page *page = NULL;
431         int status = 0;
432
433         page = alloc_page(GFP_NOFS);
434         sc = kzalloc(sizeof(*sc), GFP_NOFS);
435         if (sc == NULL || page == NULL)
436                 goto out;
437
438         kref_init(&sc->sc_kref);
439         o2nm_node_get(node);
440         sc->sc_node = node;
441
442         /* pin the node item of the remote node */
443         status = o2nm_depend_item(&node->nd_item);
444         if (status) {
445                 mlog_errno(status);
446                 o2nm_node_put(node);
447                 goto out;
448         }
449         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
450         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
451         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
452         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
453
454         init_timer(&sc->sc_idle_timeout);
455         sc->sc_idle_timeout.function = o2net_idle_timer;
456         sc->sc_idle_timeout.data = (unsigned long)sc;
457
458         sclog(sc, "alloced\n");
459
460         ret = sc;
461         sc->sc_page = page;
462         o2net_debug_add_sc(sc);
463         sc = NULL;
464         page = NULL;
465
466 out:
467         if (page)
468                 __free_page(page);
469         kfree(sc);
470
471         return ret;
472 }
473
474 /* ------------------------------------------------------------ */
475
476 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
477                                 struct work_struct *work)
478 {
479         sc_get(sc);
480         if (!queue_work(o2net_wq, work))
481                 sc_put(sc);
482 }
483 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
484                                         struct delayed_work *work,
485                                         int delay)
486 {
487         sc_get(sc);
488         if (!queue_delayed_work(o2net_wq, work, delay))
489                 sc_put(sc);
490 }
491 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
492                                          struct delayed_work *work)
493 {
494         if (cancel_delayed_work(work))
495                 sc_put(sc);
496 }
497
498 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
499
500 int o2net_num_connected_peers(void)
501 {
502         return atomic_read(&o2net_connected_peers);
503 }
504
505 static void o2net_set_nn_state(struct o2net_node *nn,
506                                struct o2net_sock_container *sc,
507                                unsigned valid, int err)
508 {
509         int was_valid = nn->nn_sc_valid;
510         int was_err = nn->nn_persistent_error;
511         struct o2net_sock_container *old_sc = nn->nn_sc;
512
513         assert_spin_locked(&nn->nn_lock);
514
515         if (old_sc && !sc)
516                 atomic_dec(&o2net_connected_peers);
517         else if (!old_sc && sc)
518                 atomic_inc(&o2net_connected_peers);
519
520         /* the node num comparison and single connect/accept path should stop
521          * an non-null sc from being overwritten with another */
522         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
523         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
524         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
525
526         if (was_valid && !valid && err == 0)
527                 err = -ENOTCONN;
528
529         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
530              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
531              nn->nn_persistent_error, err);
532
533         nn->nn_sc = sc;
534         nn->nn_sc_valid = valid ? 1 : 0;
535         nn->nn_persistent_error = err;
536
537         /* mirrors o2net_tx_can_proceed() */
538         if (nn->nn_persistent_error || nn->nn_sc_valid)
539                 wake_up(&nn->nn_sc_wq);
540
541         if (!was_err && nn->nn_persistent_error) {
542                 o2quo_conn_err(o2net_num_from_nn(nn));
543                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
544                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
545         }
546
547         if (was_valid && !valid) {
548                 printk(KERN_NOTICE "o2net: No longer connected to "
549                        SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
550                 o2net_complete_nodes_nsw(nn);
551         }
552
553         if (!was_valid && valid) {
554                 o2quo_conn_up(o2net_num_from_nn(nn));
555                 cancel_delayed_work(&nn->nn_connect_expired);
556                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
557                        o2nm_this_node() > sc->sc_node->nd_num ?
558                        "Connected to" : "Accepted connection from",
559                        SC_NODEF_ARGS(sc));
560         }
561
562         /* trigger the connecting worker func as long as we're not valid,
563          * it will back off if it shouldn't connect.  This can be called
564          * from node config teardown and so needs to be careful about
565          * the work queue actually being up. */
566         if (!valid && o2net_wq) {
567                 unsigned long delay;
568                 /* delay if we're within a RECONNECT_DELAY of the
569                  * last attempt */
570                 delay = (nn->nn_last_connect_attempt +
571                          msecs_to_jiffies(o2net_reconnect_delay()))
572                         - jiffies;
573                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
574                         delay = 0;
575                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
576                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
577
578                 /*
579                  * Delay the expired work after idle timeout.
580                  *
581                  * We might have lots of failed connection attempts that run
582                  * through here but we only cancel the connect_expired work when
583                  * a connection attempt succeeds.  So only the first enqueue of
584                  * the connect_expired work will do anything.  The rest will see
585                  * that it's already queued and do nothing.
586                  */
587                 delay += msecs_to_jiffies(o2net_idle_timeout());
588                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
589         }
590
591         /* keep track of the nn's sc ref for the caller */
592         if ((old_sc == NULL) && sc)
593                 sc_get(sc);
594         if (old_sc && (old_sc != sc)) {
595                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
596                 sc_put(old_sc);
597         }
598 }
599
600 /* see o2net_register_callbacks() */
601 static void o2net_data_ready(struct sock *sk, int bytes)
602 {
603         void (*ready)(struct sock *sk, int bytes);
604
605         read_lock(&sk->sk_callback_lock);
606         if (sk->sk_user_data) {
607                 struct o2net_sock_container *sc = sk->sk_user_data;
608                 sclog(sc, "data_ready hit\n");
609                 o2net_set_data_ready_time(sc);
610                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
611                 ready = sc->sc_data_ready;
612         } else {
613                 ready = sk->sk_data_ready;
614         }
615         read_unlock(&sk->sk_callback_lock);
616
617         ready(sk, bytes);
618 }
619
620 /* see o2net_register_callbacks() */
621 static void o2net_state_change(struct sock *sk)
622 {
623         void (*state_change)(struct sock *sk);
624         struct o2net_sock_container *sc;
625
626         read_lock(&sk->sk_callback_lock);
627         sc = sk->sk_user_data;
628         if (sc == NULL) {
629                 state_change = sk->sk_state_change;
630                 goto out;
631         }
632
633         sclog(sc, "state_change to %d\n", sk->sk_state);
634
635         state_change = sc->sc_state_change;
636
637         switch(sk->sk_state) {
638                 /* ignore connecting sockets as they make progress */
639                 case TCP_SYN_SENT:
640                 case TCP_SYN_RECV:
641                         break;
642                 case TCP_ESTABLISHED:
643                         o2net_sc_queue_work(sc, &sc->sc_connect_work);
644                         break;
645                 default:
646                         printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
647                               " shutdown, state %d\n",
648                               SC_NODEF_ARGS(sc), sk->sk_state);
649                         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
650                         break;
651         }
652 out:
653         read_unlock(&sk->sk_callback_lock);
654         state_change(sk);
655 }
656
657 /*
658  * we register callbacks so we can queue work on events before calling
659  * the original callbacks.  our callbacks our careful to test user_data
660  * to discover when they've reaced with o2net_unregister_callbacks().
661  */
662 static void o2net_register_callbacks(struct sock *sk,
663                                      struct o2net_sock_container *sc)
664 {
665         write_lock_bh(&sk->sk_callback_lock);
666
667         /* accepted sockets inherit the old listen socket data ready */
668         if (sk->sk_data_ready == o2net_listen_data_ready) {
669                 sk->sk_data_ready = sk->sk_user_data;
670                 sk->sk_user_data = NULL;
671         }
672
673         BUG_ON(sk->sk_user_data != NULL);
674         sk->sk_user_data = sc;
675         sc_get(sc);
676
677         sc->sc_data_ready = sk->sk_data_ready;
678         sc->sc_state_change = sk->sk_state_change;
679         sk->sk_data_ready = o2net_data_ready;
680         sk->sk_state_change = o2net_state_change;
681
682         mutex_init(&sc->sc_send_lock);
683
684         write_unlock_bh(&sk->sk_callback_lock);
685 }
686
687 static int o2net_unregister_callbacks(struct sock *sk,
688                                    struct o2net_sock_container *sc)
689 {
690         int ret = 0;
691
692         write_lock_bh(&sk->sk_callback_lock);
693         if (sk->sk_user_data == sc) {
694                 ret = 1;
695                 sk->sk_user_data = NULL;
696                 sk->sk_data_ready = sc->sc_data_ready;
697                 sk->sk_state_change = sc->sc_state_change;
698         }
699         write_unlock_bh(&sk->sk_callback_lock);
700
701         return ret;
702 }
703
704 /*
705  * this is a little helper that is called by callers who have seen a problem
706  * with an sc and want to detach it from the nn if someone already hasn't beat
707  * them to it.  if an error is given then the shutdown will be persistent
708  * and pending transmits will be canceled.
709  */
710 static void o2net_ensure_shutdown(struct o2net_node *nn,
711                                    struct o2net_sock_container *sc,
712                                    int err)
713 {
714         spin_lock(&nn->nn_lock);
715         if (nn->nn_sc == sc)
716                 o2net_set_nn_state(nn, NULL, 0, err);
717         spin_unlock(&nn->nn_lock);
718 }
719
720 /*
721  * This work queue function performs the blocking parts of socket shutdown.  A
722  * few paths lead here.  set_nn_state will trigger this callback if it sees an
723  * sc detached from the nn.  state_change will also trigger this callback
724  * directly when it sees errors.  In that case we need to call set_nn_state
725  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
726  * itself.
727  */
728 static void o2net_shutdown_sc(struct work_struct *work)
729 {
730         struct o2net_sock_container *sc =
731                 container_of(work, struct o2net_sock_container,
732                              sc_shutdown_work);
733         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
734
735         sclog(sc, "shutting down\n");
736
737         /* drop the callbacks ref and call shutdown only once */
738         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
739                 /* we shouldn't flush as we're in the thread, the
740                  * races with pending sc work structs are harmless */
741                 del_timer_sync(&sc->sc_idle_timeout);
742                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
743                 sc_put(sc);
744                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
745         }
746
747         /* not fatal so failed connects before the other guy has our
748          * heartbeat can be retried */
749         o2net_ensure_shutdown(nn, sc, 0);
750         sc_put(sc);
751 }
752
753 /* ------------------------------------------------------------ */
754
755 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
756                              u32 key)
757 {
758         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
759
760         if (ret == 0)
761                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
762
763         return ret;
764 }
765
766 static struct o2net_msg_handler *
767 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
768                           struct rb_node **ret_parent)
769 {
770         struct rb_node **p = &o2net_handler_tree.rb_node;
771         struct rb_node *parent = NULL;
772         struct o2net_msg_handler *nmh, *ret = NULL;
773         int cmp;
774
775         while (*p) {
776                 parent = *p;
777                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
778                 cmp = o2net_handler_cmp(nmh, msg_type, key);
779
780                 if (cmp < 0)
781                         p = &(*p)->rb_left;
782                 else if (cmp > 0)
783                         p = &(*p)->rb_right;
784                 else {
785                         ret = nmh;
786                         break;
787                 }
788         }
789
790         if (ret_p != NULL)
791                 *ret_p = p;
792         if (ret_parent != NULL)
793                 *ret_parent = parent;
794
795         return ret;
796 }
797
798 static void o2net_handler_kref_release(struct kref *kref)
799 {
800         struct o2net_msg_handler *nmh;
801         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
802
803         kfree(nmh);
804 }
805
806 static void o2net_handler_put(struct o2net_msg_handler *nmh)
807 {
808         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
809 }
810
811 /* max_len is protection for the handler func.  incoming messages won't
812  * be given to the handler if their payload is longer than the max. */
813 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
814                            o2net_msg_handler_func *func, void *data,
815                            o2net_post_msg_handler_func *post_func,
816                            struct list_head *unreg_list)
817 {
818         struct o2net_msg_handler *nmh = NULL;
819         struct rb_node **p, *parent;
820         int ret = 0;
821
822         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
823                 mlog(0, "max_len for message handler out of range: %u\n",
824                         max_len);
825                 ret = -EINVAL;
826                 goto out;
827         }
828
829         if (!msg_type) {
830                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
831                 ret = -EINVAL;
832                 goto out;
833
834         }
835         if (!func) {
836                 mlog(0, "no message handler provided: %u, %p\n",
837                        msg_type, func);
838                 ret = -EINVAL;
839                 goto out;
840         }
841
842         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
843         if (nmh == NULL) {
844                 ret = -ENOMEM;
845                 goto out;
846         }
847
848         nmh->nh_func = func;
849         nmh->nh_func_data = data;
850         nmh->nh_post_func = post_func;
851         nmh->nh_msg_type = msg_type;
852         nmh->nh_max_len = max_len;
853         nmh->nh_key = key;
854         /* the tree and list get this ref.. they're both removed in
855          * unregister when this ref is dropped */
856         kref_init(&nmh->nh_kref);
857         INIT_LIST_HEAD(&nmh->nh_unregister_item);
858
859         write_lock(&o2net_handler_lock);
860         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
861                 ret = -EEXIST;
862         else {
863                 rb_link_node(&nmh->nh_node, parent, p);
864                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
865                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
866
867                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
868                      func, msg_type, key);
869                 /* we've had some trouble with handlers seemingly vanishing. */
870                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
871                                                           &parent) == NULL,
872                                 "couldn't find handler we *just* registerd "
873                                 "for type %u key %08x\n", msg_type, key);
874         }
875         write_unlock(&o2net_handler_lock);
876         if (ret)
877                 goto out;
878
879 out:
880         if (ret)
881                 kfree(nmh);
882
883         return ret;
884 }
885 EXPORT_SYMBOL_GPL(o2net_register_handler);
886
887 void o2net_unregister_handler_list(struct list_head *list)
888 {
889         struct o2net_msg_handler *nmh, *n;
890
891         write_lock(&o2net_handler_lock);
892         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
893                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
894                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
895                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
896                 list_del_init(&nmh->nh_unregister_item);
897                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
898         }
899         write_unlock(&o2net_handler_lock);
900 }
901 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
902
903 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
904 {
905         struct o2net_msg_handler *nmh;
906
907         read_lock(&o2net_handler_lock);
908         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
909         if (nmh)
910                 kref_get(&nmh->nh_kref);
911         read_unlock(&o2net_handler_lock);
912
913         return nmh;
914 }
915
916 /* ------------------------------------------------------------ */
917
918 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
919 {
920         int ret;
921         mm_segment_t oldfs;
922         struct kvec vec = {
923                 .iov_len = len,
924                 .iov_base = data,
925         };
926         struct msghdr msg = {
927                 .msg_iovlen = 1,
928                 .msg_iov = (struct iovec *)&vec,
929                 .msg_flags = MSG_DONTWAIT,
930         };
931
932         oldfs = get_fs();
933         set_fs(get_ds());
934         ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
935         set_fs(oldfs);
936
937         return ret;
938 }
939
940 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
941                               size_t veclen, size_t total)
942 {
943         int ret;
944         mm_segment_t oldfs;
945         struct msghdr msg = {
946                 .msg_iov = (struct iovec *)vec,
947                 .msg_iovlen = veclen,
948         };
949
950         if (sock == NULL) {
951                 ret = -EINVAL;
952                 goto out;
953         }
954
955         oldfs = get_fs();
956         set_fs(get_ds());
957         ret = sock_sendmsg(sock, &msg, total);
958         set_fs(oldfs);
959         if (ret != total) {
960                 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
961                      total);
962                 if (ret >= 0)
963                         ret = -EPIPE; /* should be smarter, I bet */
964                 goto out;
965         }
966
967         ret = 0;
968 out:
969         if (ret < 0)
970                 mlog(0, "returning error: %d\n", ret);
971         return ret;
972 }
973
974 static void o2net_sendpage(struct o2net_sock_container *sc,
975                            void *kmalloced_virt,
976                            size_t size)
977 {
978         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
979         ssize_t ret;
980
981         while (1) {
982                 mutex_lock(&sc->sc_send_lock);
983                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
984                                                  virt_to_page(kmalloced_virt),
985                                                  (long)kmalloced_virt & ~PAGE_MASK,
986                                                  size, MSG_DONTWAIT);
987                 mutex_unlock(&sc->sc_send_lock);
988                 if (ret == size)
989                         break;
990                 if (ret == (ssize_t)-EAGAIN) {
991                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
992                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
993                         cond_resched();
994                         continue;
995                 }
996                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
997                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
998                 o2net_ensure_shutdown(nn, sc, 0);
999                 break;
1000         }
1001 }
1002
1003 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
1004 {
1005         memset(msg, 0, sizeof(struct o2net_msg));
1006         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
1007         msg->data_len = cpu_to_be16(data_len);
1008         msg->msg_type = cpu_to_be16(msg_type);
1009         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
1010         msg->status = 0;
1011         msg->key = cpu_to_be32(key);
1012 }
1013
1014 static int o2net_tx_can_proceed(struct o2net_node *nn,
1015                                 struct o2net_sock_container **sc_ret,
1016                                 int *error)
1017 {
1018         int ret = 0;
1019
1020         spin_lock(&nn->nn_lock);
1021         if (nn->nn_persistent_error) {
1022                 ret = 1;
1023                 *sc_ret = NULL;
1024                 *error = nn->nn_persistent_error;
1025         } else if (nn->nn_sc_valid) {
1026                 kref_get(&nn->nn_sc->sc_kref);
1027
1028                 ret = 1;
1029                 *sc_ret = nn->nn_sc;
1030                 *error = 0;
1031         }
1032         spin_unlock(&nn->nn_lock);
1033
1034         return ret;
1035 }
1036
1037 /* Get a map of all nodes to which this node is currently connected to */
1038 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1039 {
1040         struct o2net_sock_container *sc;
1041         int node, ret;
1042
1043         BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1044
1045         memset(map, 0, bytes);
1046         for (node = 0; node < O2NM_MAX_NODES; ++node) {
1047                 o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret);
1048                 if (!ret) {
1049                         set_bit(node, map);
1050                         sc_put(sc);
1051                 }
1052         }
1053 }
1054 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1055
1056 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1057                            size_t caller_veclen, u8 target_node, int *status)
1058 {
1059         int ret = 0;
1060         struct o2net_msg *msg = NULL;
1061         size_t veclen, caller_bytes = 0;
1062         struct kvec *vec = NULL;
1063         struct o2net_sock_container *sc = NULL;
1064         struct o2net_node *nn = o2net_nn_from_num(target_node);
1065         struct o2net_status_wait nsw = {
1066                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1067         };
1068         struct o2net_send_tracking nst;
1069
1070         o2net_init_nst(&nst, msg_type, key, current, target_node);
1071
1072         if (o2net_wq == NULL) {
1073                 mlog(0, "attempt to tx without o2netd running\n");
1074                 ret = -ESRCH;
1075                 goto out;
1076         }
1077
1078         if (caller_veclen == 0) {
1079                 mlog(0, "bad kvec array length\n");
1080                 ret = -EINVAL;
1081                 goto out;
1082         }
1083
1084         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1085         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1086                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1087                 ret = -EINVAL;
1088                 goto out;
1089         }
1090
1091         if (target_node == o2nm_this_node()) {
1092                 ret = -ELOOP;
1093                 goto out;
1094         }
1095
1096         o2net_debug_add_nst(&nst);
1097
1098         o2net_set_nst_sock_time(&nst);
1099
1100         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1101         if (ret)
1102                 goto out;
1103
1104         o2net_set_nst_sock_container(&nst, sc);
1105
1106         veclen = caller_veclen + 1;
1107         vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1108         if (vec == NULL) {
1109                 mlog(0, "failed to %zu element kvec!\n", veclen);
1110                 ret = -ENOMEM;
1111                 goto out;
1112         }
1113
1114         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1115         if (!msg) {
1116                 mlog(0, "failed to allocate a o2net_msg!\n");
1117                 ret = -ENOMEM;
1118                 goto out;
1119         }
1120
1121         o2net_init_msg(msg, caller_bytes, msg_type, key);
1122
1123         vec[0].iov_len = sizeof(struct o2net_msg);
1124         vec[0].iov_base = msg;
1125         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1126
1127         ret = o2net_prep_nsw(nn, &nsw);
1128         if (ret)
1129                 goto out;
1130
1131         msg->msg_num = cpu_to_be32(nsw.ns_id);
1132         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1133
1134         o2net_set_nst_send_time(&nst);
1135
1136         /* finally, convert the message header to network byte-order
1137          * and send */
1138         mutex_lock(&sc->sc_send_lock);
1139         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1140                                  sizeof(struct o2net_msg) + caller_bytes);
1141         mutex_unlock(&sc->sc_send_lock);
1142         msglog(msg, "sending returned %d\n", ret);
1143         if (ret < 0) {
1144                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1145                 goto out;
1146         }
1147
1148         /* wait on other node's handler */
1149         o2net_set_nst_status_time(&nst);
1150         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1151
1152         o2net_update_send_stats(&nst, sc);
1153
1154         /* Note that we avoid overwriting the callers status return
1155          * variable if a system error was reported on the other
1156          * side. Callers beware. */
1157         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1158         if (status && !ret)
1159                 *status = nsw.ns_status;
1160
1161         mlog(0, "woken, returning system status %d, user status %d\n",
1162              ret, nsw.ns_status);
1163 out:
1164         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1165         if (sc)
1166                 sc_put(sc);
1167         if (vec)
1168                 kfree(vec);
1169         if (msg)
1170                 kfree(msg);
1171         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1172         return ret;
1173 }
1174 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1175
1176 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1177                        u8 target_node, int *status)
1178 {
1179         struct kvec vec = {
1180                 .iov_base = data,
1181                 .iov_len = len,
1182         };
1183         return o2net_send_message_vec(msg_type, key, &vec, 1,
1184                                       target_node, status);
1185 }
1186 EXPORT_SYMBOL_GPL(o2net_send_message);
1187
1188 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1189                                    enum o2net_system_error syserr, int err)
1190 {
1191         struct kvec vec = {
1192                 .iov_base = hdr,
1193                 .iov_len = sizeof(struct o2net_msg),
1194         };
1195
1196         BUG_ON(syserr >= O2NET_ERR_MAX);
1197
1198         /* leave other fields intact from the incoming message, msg_num
1199          * in particular */
1200         hdr->sys_status = cpu_to_be32(syserr);
1201         hdr->status = cpu_to_be32(err);
1202         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1203         hdr->data_len = 0;
1204
1205         msglog(hdr, "about to send status magic %d\n", err);
1206         /* hdr has been in host byteorder this whole time */
1207         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1208 }
1209
1210 /* this returns -errno if the header was unknown or too large, etc.
1211  * after this is called the buffer us reused for the next message */
1212 static int o2net_process_message(struct o2net_sock_container *sc,
1213                                  struct o2net_msg *hdr)
1214 {
1215         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1216         int ret = 0, handler_status;
1217         enum  o2net_system_error syserr;
1218         struct o2net_msg_handler *nmh = NULL;
1219         void *ret_data = NULL;
1220
1221         msglog(hdr, "processing message\n");
1222
1223         o2net_sc_postpone_idle(sc);
1224
1225         switch(be16_to_cpu(hdr->magic)) {
1226                 case O2NET_MSG_STATUS_MAGIC:
1227                         /* special type for returning message status */
1228                         o2net_complete_nsw(nn, NULL,
1229                                            be32_to_cpu(hdr->msg_num),
1230                                            be32_to_cpu(hdr->sys_status),
1231                                            be32_to_cpu(hdr->status));
1232                         goto out;
1233                 case O2NET_MSG_KEEP_REQ_MAGIC:
1234                         o2net_sendpage(sc, o2net_keep_resp,
1235                                        sizeof(*o2net_keep_resp));
1236                         goto out;
1237                 case O2NET_MSG_KEEP_RESP_MAGIC:
1238                         goto out;
1239                 case O2NET_MSG_MAGIC:
1240                         break;
1241                 default:
1242                         msglog(hdr, "bad magic\n");
1243                         ret = -EINVAL;
1244                         goto out;
1245                         break;
1246         }
1247
1248         /* find a handler for it */
1249         handler_status = 0;
1250         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1251                                 be32_to_cpu(hdr->key));
1252         if (!nmh) {
1253                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1254                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1255                 syserr = O2NET_ERR_NO_HNDLR;
1256                 goto out_respond;
1257         }
1258
1259         syserr = O2NET_ERR_NONE;
1260
1261         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1262                 syserr = O2NET_ERR_OVERFLOW;
1263
1264         if (syserr != O2NET_ERR_NONE)
1265                 goto out_respond;
1266
1267         o2net_set_func_start_time(sc);
1268         sc->sc_msg_key = be32_to_cpu(hdr->key);
1269         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1270         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1271                                              be16_to_cpu(hdr->data_len),
1272                                         nmh->nh_func_data, &ret_data);
1273         o2net_set_func_stop_time(sc);
1274
1275         o2net_update_recv_stats(sc);
1276
1277 out_respond:
1278         /* this destroys the hdr, so don't use it after this */
1279         mutex_lock(&sc->sc_send_lock);
1280         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1281                                       handler_status);
1282         mutex_unlock(&sc->sc_send_lock);
1283         hdr = NULL;
1284         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1285              handler_status, syserr, ret);
1286
1287         if (nmh) {
1288                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1289                 if (nmh->nh_post_func)
1290                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1291                                             ret_data);
1292         }
1293
1294 out:
1295         if (nmh)
1296                 o2net_handler_put(nmh);
1297         return ret;
1298 }
1299
1300 static int o2net_check_handshake(struct o2net_sock_container *sc)
1301 {
1302         struct o2net_handshake *hand = page_address(sc->sc_page);
1303         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1304
1305         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1306                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1307                        "protocol version %llu but %llu is required. "
1308                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1309                        (unsigned long long)be64_to_cpu(hand->protocol_version),
1310                        O2NET_PROTOCOL_VERSION);
1311
1312                 /* don't bother reconnecting if its the wrong version. */
1313                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1314                 return -1;
1315         }
1316
1317         /*
1318          * Ensure timeouts are consistent with other nodes, otherwise
1319          * we can end up with one node thinking that the other must be down,
1320          * but isn't. This can ultimately cause corruption.
1321          */
1322         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1323                                 o2net_idle_timeout()) {
1324                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1325                        "idle timeout of %u ms, but we use %u ms locally. "
1326                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1327                        be32_to_cpu(hand->o2net_idle_timeout_ms),
1328                        o2net_idle_timeout());
1329                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1330                 return -1;
1331         }
1332
1333         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1334                         o2net_keepalive_delay()) {
1335                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1336                        "delay of %u ms, but we use %u ms locally. "
1337                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1338                        be32_to_cpu(hand->o2net_keepalive_delay_ms),
1339                        o2net_keepalive_delay());
1340                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1341                 return -1;
1342         }
1343
1344         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1345                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1346                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1347                        "timeout of %u ms, but we use %u ms locally. "
1348                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1349                        be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1350                        O2HB_MAX_WRITE_TIMEOUT_MS);
1351                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1352                 return -1;
1353         }
1354
1355         sc->sc_handshake_ok = 1;
1356
1357         spin_lock(&nn->nn_lock);
1358         /* set valid and queue the idle timers only if it hasn't been
1359          * shut down already */
1360         if (nn->nn_sc == sc) {
1361                 o2net_sc_reset_idle_timer(sc);
1362                 atomic_set(&nn->nn_timeout, 0);
1363                 o2net_set_nn_state(nn, sc, 1, 0);
1364         }
1365         spin_unlock(&nn->nn_lock);
1366
1367         /* shift everything up as though it wasn't there */
1368         sc->sc_page_off -= sizeof(struct o2net_handshake);
1369         if (sc->sc_page_off)
1370                 memmove(hand, hand + 1, sc->sc_page_off);
1371
1372         return 0;
1373 }
1374
1375 /* this demuxes the queued rx bytes into header or payload bits and calls
1376  * handlers as each full message is read off the socket.  it returns -error,
1377  * == 0 eof, or > 0 for progress made.*/
1378 static int o2net_advance_rx(struct o2net_sock_container *sc)
1379 {
1380         struct o2net_msg *hdr;
1381         int ret = 0;
1382         void *data;
1383         size_t datalen;
1384
1385         sclog(sc, "receiving\n");
1386         o2net_set_advance_start_time(sc);
1387
1388         if (unlikely(sc->sc_handshake_ok == 0)) {
1389                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1390                         data = page_address(sc->sc_page) + sc->sc_page_off;
1391                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1392                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1393                         if (ret > 0)
1394                                 sc->sc_page_off += ret;
1395                 }
1396
1397                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1398                         o2net_check_handshake(sc);
1399                         if (unlikely(sc->sc_handshake_ok == 0))
1400                                 ret = -EPROTO;
1401                 }
1402                 goto out;
1403         }
1404
1405         /* do we need more header? */
1406         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1407                 data = page_address(sc->sc_page) + sc->sc_page_off;
1408                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1409                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1410                 if (ret > 0) {
1411                         sc->sc_page_off += ret;
1412                         /* only swab incoming here.. we can
1413                          * only get here once as we cross from
1414                          * being under to over */
1415                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1416                                 hdr = page_address(sc->sc_page);
1417                                 if (be16_to_cpu(hdr->data_len) >
1418                                     O2NET_MAX_PAYLOAD_BYTES)
1419                                         ret = -EOVERFLOW;
1420                         }
1421                 }
1422                 if (ret <= 0)
1423                         goto out;
1424         }
1425
1426         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1427                 /* oof, still don't have a header */
1428                 goto out;
1429         }
1430
1431         /* this was swabbed above when we first read it */
1432         hdr = page_address(sc->sc_page);
1433
1434         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1435
1436         /* do we need more payload? */
1437         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1438                 /* need more payload */
1439                 data = page_address(sc->sc_page) + sc->sc_page_off;
1440                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1441                           sc->sc_page_off;
1442                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1443                 if (ret > 0)
1444                         sc->sc_page_off += ret;
1445                 if (ret <= 0)
1446                         goto out;
1447         }
1448
1449         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1450                 /* we can only get here once, the first time we read
1451                  * the payload.. so set ret to progress if the handler
1452                  * works out. after calling this the message is toast */
1453                 ret = o2net_process_message(sc, hdr);
1454                 if (ret == 0)
1455                         ret = 1;
1456                 sc->sc_page_off = 0;
1457         }
1458
1459 out:
1460         sclog(sc, "ret = %d\n", ret);
1461         o2net_set_advance_stop_time(sc);
1462         return ret;
1463 }
1464
1465 /* this work func is triggerd by data ready.  it reads until it can read no
1466  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1467  * our work the work struct will be marked and we'll be called again. */
1468 static void o2net_rx_until_empty(struct work_struct *work)
1469 {
1470         struct o2net_sock_container *sc =
1471                 container_of(work, struct o2net_sock_container, sc_rx_work);
1472         int ret;
1473
1474         do {
1475                 ret = o2net_advance_rx(sc);
1476         } while (ret > 0);
1477
1478         if (ret <= 0 && ret != -EAGAIN) {
1479                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1480                 sclog(sc, "saw error %d, closing\n", ret);
1481                 /* not permanent so read failed handshake can retry */
1482                 o2net_ensure_shutdown(nn, sc, 0);
1483         }
1484
1485         sc_put(sc);
1486 }
1487
1488 static int o2net_set_nodelay(struct socket *sock)
1489 {
1490         int ret, val = 1;
1491         mm_segment_t oldfs;
1492
1493         oldfs = get_fs();
1494         set_fs(KERNEL_DS);
1495
1496         /*
1497          * Dear unsuspecting programmer,
1498          *
1499          * Don't use sock_setsockopt() for SOL_TCP.  It doesn't check its level
1500          * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1501          * silently turn into SO_DEBUG.
1502          *
1503          * Yours,
1504          * Keeper of hilariously fragile interfaces.
1505          */
1506         ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1507                                     (char __user *)&val, sizeof(val));
1508
1509         set_fs(oldfs);
1510         return ret;
1511 }
1512
1513 static void o2net_initialize_handshake(void)
1514 {
1515         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1516                 O2HB_MAX_WRITE_TIMEOUT_MS);
1517         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1518         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1519                 o2net_keepalive_delay());
1520         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1521                 o2net_reconnect_delay());
1522 }
1523
1524 /* ------------------------------------------------------------ */
1525
1526 /* called when a connect completes and after a sock is accepted.  the
1527  * rx path will see the response and mark the sc valid */
1528 static void o2net_sc_connect_completed(struct work_struct *work)
1529 {
1530         struct o2net_sock_container *sc =
1531                 container_of(work, struct o2net_sock_container,
1532                              sc_connect_work);
1533
1534         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1535               (unsigned long long)O2NET_PROTOCOL_VERSION,
1536               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1537
1538         o2net_initialize_handshake();
1539         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1540         sc_put(sc);
1541 }
1542
1543 /* this is called as a work_struct func. */
1544 static void o2net_sc_send_keep_req(struct work_struct *work)
1545 {
1546         struct o2net_sock_container *sc =
1547                 container_of(work, struct o2net_sock_container,
1548                              sc_keepalive_work.work);
1549
1550         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1551         sc_put(sc);
1552 }
1553
1554 /* socket shutdown does a del_timer_sync against this as it tears down.
1555  * we can't start this timer until we've got to the point in sc buildup
1556  * where shutdown is going to be involved */
1557 static void o2net_idle_timer(unsigned long data)
1558 {
1559         struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1560         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1561 #ifdef CONFIG_DEBUG_FS
1562         unsigned long msecs = ktime_to_ms(ktime_get()) -
1563                 ktime_to_ms(sc->sc_tv_timer);
1564 #else
1565         unsigned long msecs = o2net_idle_timeout();
1566 #endif
1567
1568         printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1569                "idle for %lu.%lu secs, shutting it down.\n", SC_NODEF_ARGS(sc),
1570                msecs / 1000, msecs % 1000);
1571
1572         /*
1573          * Initialize the nn_timeout so that the next connection attempt
1574          * will continue in o2net_start_connect.
1575          */
1576         atomic_set(&nn->nn_timeout, 1);
1577
1578         o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1579 }
1580
1581 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1582 {
1583         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1584         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1585                       msecs_to_jiffies(o2net_keepalive_delay()));
1586         o2net_set_sock_timer(sc);
1587         mod_timer(&sc->sc_idle_timeout,
1588                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1589 }
1590
1591 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1592 {
1593         /* Only push out an existing timer */
1594         if (timer_pending(&sc->sc_idle_timeout))
1595                 o2net_sc_reset_idle_timer(sc);
1596 }
1597
1598 /* this work func is kicked whenever a path sets the nn state which doesn't
1599  * have valid set.  This includes seeing hb come up, losing a connection,
1600  * having a connect attempt fail, etc. This centralizes the logic which decides
1601  * if a connect attempt should be made or if we should give up and all future
1602  * transmit attempts should fail */
1603 static void o2net_start_connect(struct work_struct *work)
1604 {
1605         struct o2net_node *nn =
1606                 container_of(work, struct o2net_node, nn_connect_work.work);
1607         struct o2net_sock_container *sc = NULL;
1608         struct o2nm_node *node = NULL, *mynode = NULL;
1609         struct socket *sock = NULL;
1610         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1611         int ret = 0, stop;
1612         unsigned int timeout;
1613
1614         /* if we're greater we initiate tx, otherwise we accept */
1615         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1616                 goto out;
1617
1618         /* watch for racing with tearing a node down */
1619         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1620         if (node == NULL) {
1621                 ret = 0;
1622                 goto out;
1623         }
1624
1625         mynode = o2nm_get_node_by_num(o2nm_this_node());
1626         if (mynode == NULL) {
1627                 ret = 0;
1628                 goto out;
1629         }
1630
1631         spin_lock(&nn->nn_lock);
1632         /*
1633          * see if we already have one pending or have given up.
1634          * For nn_timeout, it is set when we close the connection
1635          * because of the idle time out. So it means that we have
1636          * at least connected to that node successfully once,
1637          * now try to connect to it again.
1638          */
1639         timeout = atomic_read(&nn->nn_timeout);
1640         stop = (nn->nn_sc ||
1641                 (nn->nn_persistent_error &&
1642                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1643         spin_unlock(&nn->nn_lock);
1644         if (stop)
1645                 goto out;
1646
1647         nn->nn_last_connect_attempt = jiffies;
1648
1649         sc = sc_alloc(node);
1650         if (sc == NULL) {
1651                 mlog(0, "couldn't allocate sc\n");
1652                 ret = -ENOMEM;
1653                 goto out;
1654         }
1655
1656         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1657         if (ret < 0) {
1658                 mlog(0, "can't create socket: %d\n", ret);
1659                 goto out;
1660         }
1661         sc->sc_sock = sock; /* freed by sc_kref_release */
1662
1663         sock->sk->sk_allocation = GFP_ATOMIC;
1664
1665         myaddr.sin_family = AF_INET;
1666         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1667         myaddr.sin_port = htons(0); /* any port */
1668
1669         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1670                               sizeof(myaddr));
1671         if (ret) {
1672                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1673                      ret, &mynode->nd_ipv4_address);
1674                 goto out;
1675         }
1676
1677         ret = o2net_set_nodelay(sc->sc_sock);
1678         if (ret) {
1679                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1680                 goto out;
1681         }
1682
1683         o2net_register_callbacks(sc->sc_sock->sk, sc);
1684
1685         spin_lock(&nn->nn_lock);
1686         /* handshake completion will set nn->nn_sc_valid */
1687         o2net_set_nn_state(nn, sc, 0, 0);
1688         spin_unlock(&nn->nn_lock);
1689
1690         remoteaddr.sin_family = AF_INET;
1691         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1692         remoteaddr.sin_port = node->nd_ipv4_port;
1693
1694         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1695                                         (struct sockaddr *)&remoteaddr,
1696                                         sizeof(remoteaddr),
1697                                         O_NONBLOCK);
1698         if (ret == -EINPROGRESS)
1699                 ret = 0;
1700
1701 out:
1702         if (ret) {
1703                 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1704                        " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1705                 /* 0 err so that another will be queued and attempted
1706                  * from set_nn_state */
1707                 if (sc)
1708                         o2net_ensure_shutdown(nn, sc, 0);
1709         }
1710         if (sc)
1711                 sc_put(sc);
1712         if (node)
1713                 o2nm_node_put(node);
1714         if (mynode)
1715                 o2nm_node_put(mynode);
1716
1717         return;
1718 }
1719
1720 static void o2net_connect_expired(struct work_struct *work)
1721 {
1722         struct o2net_node *nn =
1723                 container_of(work, struct o2net_node, nn_connect_expired.work);
1724
1725         spin_lock(&nn->nn_lock);
1726         if (!nn->nn_sc_valid) {
1727                 printk(KERN_NOTICE "o2net: No connection established with "
1728                        "node %u after %u.%u seconds, giving up.\n",
1729                      o2net_num_from_nn(nn),
1730                      o2net_idle_timeout() / 1000,
1731                      o2net_idle_timeout() % 1000);
1732
1733                 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1734         }
1735         spin_unlock(&nn->nn_lock);
1736 }
1737
1738 static void o2net_still_up(struct work_struct *work)
1739 {
1740         struct o2net_node *nn =
1741                 container_of(work, struct o2net_node, nn_still_up.work);
1742
1743         o2quo_hb_still_up(o2net_num_from_nn(nn));
1744 }
1745
1746 /* ------------------------------------------------------------ */
1747
1748 void o2net_disconnect_node(struct o2nm_node *node)
1749 {
1750         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1751
1752         /* don't reconnect until it's heartbeating again */
1753         spin_lock(&nn->nn_lock);
1754         atomic_set(&nn->nn_timeout, 0);
1755         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1756         spin_unlock(&nn->nn_lock);
1757
1758         if (o2net_wq) {
1759                 cancel_delayed_work(&nn->nn_connect_expired);
1760                 cancel_delayed_work(&nn->nn_connect_work);
1761                 cancel_delayed_work(&nn->nn_still_up);
1762                 flush_workqueue(o2net_wq);
1763         }
1764 }
1765
1766 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1767                                   void *data)
1768 {
1769         o2quo_hb_down(node_num);
1770
1771         if (!node)
1772                 return;
1773
1774         if (node_num != o2nm_this_node())
1775                 o2net_disconnect_node(node);
1776
1777         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1778 }
1779
1780 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1781                                 void *data)
1782 {
1783         struct o2net_node *nn = o2net_nn_from_num(node_num);
1784
1785         o2quo_hb_up(node_num);
1786
1787         BUG_ON(!node);
1788
1789         /* ensure an immediate connect attempt */
1790         nn->nn_last_connect_attempt = jiffies -
1791                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1792
1793         if (node_num != o2nm_this_node()) {
1794                 /* believe it or not, accept and node hearbeating testing
1795                  * can succeed for this node before we got here.. so
1796                  * only use set_nn_state to clear the persistent error
1797                  * if that hasn't already happened */
1798                 spin_lock(&nn->nn_lock);
1799                 atomic_set(&nn->nn_timeout, 0);
1800                 if (nn->nn_persistent_error)
1801                         o2net_set_nn_state(nn, NULL, 0, 0);
1802                 spin_unlock(&nn->nn_lock);
1803         }
1804 }
1805
1806 void o2net_unregister_hb_callbacks(void)
1807 {
1808         o2hb_unregister_callback(NULL, &o2net_hb_up);
1809         o2hb_unregister_callback(NULL, &o2net_hb_down);
1810 }
1811
1812 int o2net_register_hb_callbacks(void)
1813 {
1814         int ret;
1815
1816         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1817                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1818         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1819                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1820
1821         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1822         if (ret == 0)
1823                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1824
1825         if (ret)
1826                 o2net_unregister_hb_callbacks();
1827
1828         return ret;
1829 }
1830
1831 /* ------------------------------------------------------------ */
1832
1833 static int o2net_accept_one(struct socket *sock)
1834 {
1835         int ret, slen;
1836         struct sockaddr_in sin;
1837         struct socket *new_sock = NULL;
1838         struct o2nm_node *node = NULL;
1839         struct o2nm_node *local_node = NULL;
1840         struct o2net_sock_container *sc = NULL;
1841         struct o2net_node *nn;
1842
1843         BUG_ON(sock == NULL);
1844         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1845                                sock->sk->sk_protocol, &new_sock);
1846         if (ret)
1847                 goto out;
1848
1849         new_sock->type = sock->type;
1850         new_sock->ops = sock->ops;
1851         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1852         if (ret < 0)
1853                 goto out;
1854
1855         new_sock->sk->sk_allocation = GFP_ATOMIC;
1856
1857         ret = o2net_set_nodelay(new_sock);
1858         if (ret) {
1859                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1860                 goto out;
1861         }
1862
1863         slen = sizeof(sin);
1864         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1865                                        &slen, 1);
1866         if (ret < 0)
1867                 goto out;
1868
1869         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1870         if (node == NULL) {
1871                 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1872                        "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1873                        ntohs(sin.sin_port));
1874                 ret = -EINVAL;
1875                 goto out;
1876         }
1877
1878         if (o2nm_this_node() >= node->nd_num) {
1879                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1880                 printk(KERN_NOTICE "o2net: Unexpected connect attempt seen "
1881                        "at node '%s' (%u, %pI4:%d) from node '%s' (%u, "
1882                        "%pI4:%d)\n", local_node->nd_name, local_node->nd_num,
1883                        &(local_node->nd_ipv4_address),
1884                        ntohs(local_node->nd_ipv4_port), node->nd_name,
1885                        node->nd_num, &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1886                 ret = -EINVAL;
1887                 goto out;
1888         }
1889
1890         /* this happens all the time when the other node sees our heartbeat
1891          * and tries to connect before we see their heartbeat */
1892         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1893                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1894                      "%pI4:%d but it isn't heartbeating\n",
1895                      node->nd_name, &sin.sin_addr.s_addr,
1896                      ntohs(sin.sin_port));
1897                 ret = -EINVAL;
1898                 goto out;
1899         }
1900
1901         nn = o2net_nn_from_num(node->nd_num);
1902
1903         spin_lock(&nn->nn_lock);
1904         if (nn->nn_sc)
1905                 ret = -EBUSY;
1906         else
1907                 ret = 0;
1908         spin_unlock(&nn->nn_lock);
1909         if (ret) {
1910                 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1911                        "at %pI4:%d but it already has an open connection\n",
1912                        node->nd_name, &sin.sin_addr.s_addr,
1913                        ntohs(sin.sin_port));
1914                 goto out;
1915         }
1916
1917         sc = sc_alloc(node);
1918         if (sc == NULL) {
1919                 ret = -ENOMEM;
1920                 goto out;
1921         }
1922
1923         sc->sc_sock = new_sock;
1924         new_sock = NULL;
1925
1926         spin_lock(&nn->nn_lock);
1927         atomic_set(&nn->nn_timeout, 0);
1928         o2net_set_nn_state(nn, sc, 0, 0);
1929         spin_unlock(&nn->nn_lock);
1930
1931         o2net_register_callbacks(sc->sc_sock->sk, sc);
1932         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1933
1934         o2net_initialize_handshake();
1935         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1936
1937 out:
1938         if (new_sock)
1939                 sock_release(new_sock);
1940         if (node)
1941                 o2nm_node_put(node);
1942         if (local_node)
1943                 o2nm_node_put(local_node);
1944         if (sc)
1945                 sc_put(sc);
1946         return ret;
1947 }
1948
1949 static void o2net_accept_many(struct work_struct *work)
1950 {
1951         struct socket *sock = o2net_listen_sock;
1952         while (o2net_accept_one(sock) == 0)
1953                 cond_resched();
1954 }
1955
1956 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1957 {
1958         void (*ready)(struct sock *sk, int bytes);
1959
1960         read_lock(&sk->sk_callback_lock);
1961         ready = sk->sk_user_data;
1962         if (ready == NULL) { /* check for teardown race */
1963                 ready = sk->sk_data_ready;
1964                 goto out;
1965         }
1966
1967         /* ->sk_data_ready is also called for a newly established child socket
1968          * before it has been accepted and the acceptor has set up their
1969          * data_ready.. we only want to queue listen work for our listening
1970          * socket */
1971         if (sk->sk_state == TCP_LISTEN) {
1972                 mlog(ML_TCP, "bytes: %d\n", bytes);
1973                 queue_work(o2net_wq, &o2net_listen_work);
1974         }
1975
1976 out:
1977         read_unlock(&sk->sk_callback_lock);
1978         ready(sk, bytes);
1979 }
1980
1981 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1982 {
1983         struct socket *sock = NULL;
1984         int ret;
1985         struct sockaddr_in sin = {
1986                 .sin_family = PF_INET,
1987                 .sin_addr = { .s_addr = addr },
1988                 .sin_port = port,
1989         };
1990
1991         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1992         if (ret < 0) {
1993                 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
1994                 goto out;
1995         }
1996
1997         sock->sk->sk_allocation = GFP_ATOMIC;
1998
1999         write_lock_bh(&sock->sk->sk_callback_lock);
2000         sock->sk->sk_user_data = sock->sk->sk_data_ready;
2001         sock->sk->sk_data_ready = o2net_listen_data_ready;
2002         write_unlock_bh(&sock->sk->sk_callback_lock);
2003
2004         o2net_listen_sock = sock;
2005         INIT_WORK(&o2net_listen_work, o2net_accept_many);
2006
2007         sock->sk->sk_reuse = 1;
2008         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2009         if (ret < 0) {
2010                 printk(KERN_ERR "o2net: Error %d while binding socket at "
2011                        "%pI4:%u\n", ret, &addr, ntohs(port)); 
2012                 goto out;
2013         }
2014
2015         ret = sock->ops->listen(sock, 64);
2016         if (ret < 0)
2017                 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2018                        ret, &addr, ntohs(port));
2019
2020 out:
2021         if (ret) {
2022                 o2net_listen_sock = NULL;
2023                 if (sock)
2024                         sock_release(sock);
2025         }
2026         return ret;
2027 }
2028
2029 /*
2030  * called from node manager when we should bring up our network listening
2031  * socket.  node manager handles all the serialization to only call this
2032  * once and to match it with o2net_stop_listening().  note,
2033  * o2nm_this_node() doesn't work yet as we're being called while it
2034  * is being set up.
2035  */
2036 int o2net_start_listening(struct o2nm_node *node)
2037 {
2038         int ret = 0;
2039
2040         BUG_ON(o2net_wq != NULL);
2041         BUG_ON(o2net_listen_sock != NULL);
2042
2043         mlog(ML_KTHREAD, "starting o2net thread...\n");
2044         o2net_wq = create_singlethread_workqueue("o2net");
2045         if (o2net_wq == NULL) {
2046                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2047                 return -ENOMEM; /* ? */
2048         }
2049
2050         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2051                                         node->nd_ipv4_port);
2052         if (ret) {
2053                 destroy_workqueue(o2net_wq);
2054                 o2net_wq = NULL;
2055         } else
2056                 o2quo_conn_up(node->nd_num);
2057
2058         return ret;
2059 }
2060
2061 /* again, o2nm_this_node() doesn't work here as we're involved in
2062  * tearing it down */
2063 void o2net_stop_listening(struct o2nm_node *node)
2064 {
2065         struct socket *sock = o2net_listen_sock;
2066         size_t i;
2067
2068         BUG_ON(o2net_wq == NULL);
2069         BUG_ON(o2net_listen_sock == NULL);
2070
2071         /* stop the listening socket from generating work */
2072         write_lock_bh(&sock->sk->sk_callback_lock);
2073         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2074         sock->sk->sk_user_data = NULL;
2075         write_unlock_bh(&sock->sk->sk_callback_lock);
2076
2077         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2078                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2079                 if (node) {
2080                         o2net_disconnect_node(node);
2081                         o2nm_node_put(node);
2082                 }
2083         }
2084
2085         /* finish all work and tear down the work queue */
2086         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2087         destroy_workqueue(o2net_wq);
2088         o2net_wq = NULL;
2089
2090         sock_release(o2net_listen_sock);
2091         o2net_listen_sock = NULL;
2092
2093         o2quo_conn_err(node->nd_num);
2094 }
2095
2096 /* ------------------------------------------------------------ */
2097
2098 int o2net_init(void)
2099 {
2100         unsigned long i;
2101
2102         o2quo_init();
2103
2104         if (o2net_debugfs_init())
2105                 return -ENOMEM;
2106
2107         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2108         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2109         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2110         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2111                 kfree(o2net_hand);
2112                 kfree(o2net_keep_req);
2113                 kfree(o2net_keep_resp);
2114                 return -ENOMEM;
2115         }
2116
2117         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2118         o2net_hand->connector_id = cpu_to_be64(1);
2119
2120         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2121         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2122
2123         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2124                 struct o2net_node *nn = o2net_nn_from_num(i);
2125
2126                 atomic_set(&nn->nn_timeout, 0);
2127                 spin_lock_init(&nn->nn_lock);
2128                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2129                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2130                                   o2net_connect_expired);
2131                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2132                 /* until we see hb from a node we'll return einval */
2133                 nn->nn_persistent_error = -ENOTCONN;
2134                 init_waitqueue_head(&nn->nn_sc_wq);
2135                 idr_init(&nn->nn_status_idr);
2136                 INIT_LIST_HEAD(&nn->nn_status_list);
2137         }
2138
2139         return 0;
2140 }
2141
2142 void o2net_exit(void)
2143 {
2144         o2quo_exit();
2145         kfree(o2net_hand);
2146         kfree(o2net_keep_req);
2147         kfree(o2net_keep_resp);
2148         o2net_debugfs_exit();
2149 }