~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/ocfs2/cluster/tcp.c

Version: ~ [ linux-5.16-rc1 ] ~ [ linux-5.15.2 ] ~ [ linux-5.14.18 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.79 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.159 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.217 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.255 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.290 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.292 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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/sched/mm.h>
 58 #include <linux/jiffies.h>
 59 #include <linux/slab.h>
 60 #include <linux/idr.h>
 61 #include <linux/kref.h>
 62 #include <linux/net.h>
 63 #include <linux/export.h>
 64 #include <net/tcp.h>
 65 
 66 #include <linux/uaccess.h>
 67 
 68 #include "heartbeat.h"
 69 #include "tcp.h"
 70 #include "nodemanager.h"
 71 #define MLOG_MASK_PREFIX ML_TCP
 72 #include "masklog.h"
 73 #include "quorum.h"
 74 
 75 #include "tcp_internal.h"
 76 
 77 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
 78 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num,    \
 79                           &sc->sc_node->nd_ipv4_address,                \
 80                           ntohs(sc->sc_node->nd_ipv4_port)
 81 
 82 /*
 83  * In the following two log macros, the whitespace after the ',' just
 84  * before ##args is intentional. Otherwise, gcc 2.95 will eat the
 85  * previous token if args expands to nothing.
 86  */
 87 #define msglog(hdr, fmt, args...) do {                                  \
 88         typeof(hdr) __hdr = (hdr);                                      \
 89         mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d "       \
 90              "key %08x num %u] " fmt,                                   \
 91              be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len),   \
 92              be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status),  \
 93              be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key),   \
 94              be32_to_cpu(__hdr->msg_num) ,  ##args);                    \
 95 } while (0)
 96 
 97 #define sclog(sc, fmt, args...) do {                                    \
 98         typeof(sc) __sc = (sc);                                         \
 99         mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p "       \
100              "pg_off %zu] " fmt, __sc,                                  \
101              kref_read(&__sc->sc_kref), __sc->sc_sock,  \
102             __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off ,   \
103             ##args);                                                    \
104 } while (0)
105 
106 static DEFINE_RWLOCK(o2net_handler_lock);
107 static struct rb_root o2net_handler_tree = RB_ROOT;
108 
109 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
110 
111 /* XXX someday we'll need better accounting */
112 static struct socket *o2net_listen_sock;
113 
114 /*
115  * listen work is only queued by the listening socket callbacks on the
116  * o2net_wq.  teardown detaches the callbacks before destroying the workqueue.
117  * quorum work is queued as sock containers are shutdown.. stop_listening
118  * tears down all the node's sock containers, preventing future shutdowns
119  * and queued quroum work, before canceling delayed quorum work and
120  * destroying the work queue.
121  */
122 static struct workqueue_struct *o2net_wq;
123 static struct work_struct o2net_listen_work;
124 
125 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
126 #define O2NET_HB_PRI 0x1
127 
128 static struct o2net_handshake *o2net_hand;
129 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
130 
131 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
132                 {[O2NET_ERR_NONE]       = 0,
133                  [O2NET_ERR_NO_HNDLR]   = -ENOPROTOOPT,
134                  [O2NET_ERR_OVERFLOW]   = -EOVERFLOW,
135                  [O2NET_ERR_DIED]       = -EHOSTDOWN,};
136 
137 /* can't quite avoid *all* internal declarations :/ */
138 static void o2net_sc_connect_completed(struct work_struct *work);
139 static void o2net_rx_until_empty(struct work_struct *work);
140 static void o2net_shutdown_sc(struct work_struct *work);
141 static void o2net_listen_data_ready(struct sock *sk);
142 static void o2net_sc_send_keep_req(struct work_struct *work);
143 static void o2net_idle_timer(struct timer_list *t);
144 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
145 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
146 
147 #ifdef CONFIG_DEBUG_FS
148 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
149                            u32 msgkey, struct task_struct *task, u8 node)
150 {
151         INIT_LIST_HEAD(&nst->st_net_debug_item);
152         nst->st_task = task;
153         nst->st_msg_type = msgtype;
154         nst->st_msg_key = msgkey;
155         nst->st_node = node;
156 }
157 
158 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
159 {
160         nst->st_sock_time = ktime_get();
161 }
162 
163 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
164 {
165         nst->st_send_time = ktime_get();
166 }
167 
168 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
169 {
170         nst->st_status_time = ktime_get();
171 }
172 
173 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
174                                                 struct o2net_sock_container *sc)
175 {
176         nst->st_sc = sc;
177 }
178 
179 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
180                                         u32 msg_id)
181 {
182         nst->st_id = msg_id;
183 }
184 
185 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
186 {
187         sc->sc_tv_timer = ktime_get();
188 }
189 
190 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
191 {
192         sc->sc_tv_data_ready = ktime_get();
193 }
194 
195 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
196 {
197         sc->sc_tv_advance_start = ktime_get();
198 }
199 
200 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
201 {
202         sc->sc_tv_advance_stop = ktime_get();
203 }
204 
205 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
206 {
207         sc->sc_tv_func_start = ktime_get();
208 }
209 
210 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
211 {
212         sc->sc_tv_func_stop = ktime_get();
213 }
214 
215 #else  /* CONFIG_DEBUG_FS */
216 # define o2net_init_nst(a, b, c, d, e)
217 # define o2net_set_nst_sock_time(a)
218 # define o2net_set_nst_send_time(a)
219 # define o2net_set_nst_status_time(a)
220 # define o2net_set_nst_sock_container(a, b)
221 # define o2net_set_nst_msg_id(a, b)
222 # define o2net_set_sock_timer(a)
223 # define o2net_set_data_ready_time(a)
224 # define o2net_set_advance_start_time(a)
225 # define o2net_set_advance_stop_time(a)
226 # define o2net_set_func_start_time(a)
227 # define o2net_set_func_stop_time(a)
228 #endif /* CONFIG_DEBUG_FS */
229 
230 #ifdef CONFIG_OCFS2_FS_STATS
231 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
232 {
233         return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
234 }
235 
236 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
237                                     struct o2net_sock_container *sc)
238 {
239         sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
240                                            ktime_sub(ktime_get(),
241                                                      nst->st_status_time));
242         sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
243                                          ktime_sub(nst->st_status_time,
244                                                    nst->st_send_time));
245         sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
246                                             ktime_sub(nst->st_send_time,
247                                                       nst->st_sock_time));
248         sc->sc_send_count++;
249 }
250 
251 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
252 {
253         sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
254                                             o2net_get_func_run_time(sc));
255         sc->sc_recv_count++;
256 }
257 
258 #else
259 
260 # define o2net_update_send_stats(a, b)
261 
262 # define o2net_update_recv_stats(sc)
263 
264 #endif /* CONFIG_OCFS2_FS_STATS */
265 
266 static inline unsigned int o2net_reconnect_delay(void)
267 {
268         return o2nm_single_cluster->cl_reconnect_delay_ms;
269 }
270 
271 static inline unsigned int o2net_keepalive_delay(void)
272 {
273         return o2nm_single_cluster->cl_keepalive_delay_ms;
274 }
275 
276 static inline unsigned int o2net_idle_timeout(void)
277 {
278         return o2nm_single_cluster->cl_idle_timeout_ms;
279 }
280 
281 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
282 {
283         int trans;
284         BUG_ON(err >= O2NET_ERR_MAX);
285         trans = o2net_sys_err_translations[err];
286 
287         /* Just in case we mess up the translation table above */
288         BUG_ON(err != O2NET_ERR_NONE && trans == 0);
289         return trans;
290 }
291 
292 static struct o2net_node * o2net_nn_from_num(u8 node_num)
293 {
294         BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
295         return &o2net_nodes[node_num];
296 }
297 
298 static u8 o2net_num_from_nn(struct o2net_node *nn)
299 {
300         BUG_ON(nn == NULL);
301         return nn - o2net_nodes;
302 }
303 
304 /* ------------------------------------------------------------ */
305 
306 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
307 {
308         int ret;
309 
310         spin_lock(&nn->nn_lock);
311         ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
312         if (ret >= 0) {
313                 nsw->ns_id = ret;
314                 list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
315         }
316         spin_unlock(&nn->nn_lock);
317         if (ret < 0)
318                 return ret;
319 
320         init_waitqueue_head(&nsw->ns_wq);
321         nsw->ns_sys_status = O2NET_ERR_NONE;
322         nsw->ns_status = 0;
323         return 0;
324 }
325 
326 static void o2net_complete_nsw_locked(struct o2net_node *nn,
327                                       struct o2net_status_wait *nsw,
328                                       enum o2net_system_error sys_status,
329                                       s32 status)
330 {
331         assert_spin_locked(&nn->nn_lock);
332 
333         if (!list_empty(&nsw->ns_node_item)) {
334                 list_del_init(&nsw->ns_node_item);
335                 nsw->ns_sys_status = sys_status;
336                 nsw->ns_status = status;
337                 idr_remove(&nn->nn_status_idr, nsw->ns_id);
338                 wake_up(&nsw->ns_wq);
339         }
340 }
341 
342 static void o2net_complete_nsw(struct o2net_node *nn,
343                                struct o2net_status_wait *nsw,
344                                u64 id, enum o2net_system_error sys_status,
345                                s32 status)
346 {
347         spin_lock(&nn->nn_lock);
348         if (nsw == NULL) {
349                 if (id > INT_MAX)
350                         goto out;
351 
352                 nsw = idr_find(&nn->nn_status_idr, id);
353                 if (nsw == NULL)
354                         goto out;
355         }
356 
357         o2net_complete_nsw_locked(nn, nsw, sys_status, status);
358 
359 out:
360         spin_unlock(&nn->nn_lock);
361         return;
362 }
363 
364 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
365 {
366         struct o2net_status_wait *nsw, *tmp;
367         unsigned int num_kills = 0;
368 
369         assert_spin_locked(&nn->nn_lock);
370 
371         list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
372                 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
373                 num_kills++;
374         }
375 
376         mlog(0, "completed %d messages for node %u\n", num_kills,
377              o2net_num_from_nn(nn));
378 }
379 
380 static int o2net_nsw_completed(struct o2net_node *nn,
381                                struct o2net_status_wait *nsw)
382 {
383         int completed;
384         spin_lock(&nn->nn_lock);
385         completed = list_empty(&nsw->ns_node_item);
386         spin_unlock(&nn->nn_lock);
387         return completed;
388 }
389 
390 /* ------------------------------------------------------------ */
391 
392 static void sc_kref_release(struct kref *kref)
393 {
394         struct o2net_sock_container *sc = container_of(kref,
395                                         struct o2net_sock_container, sc_kref);
396         BUG_ON(timer_pending(&sc->sc_idle_timeout));
397 
398         sclog(sc, "releasing\n");
399 
400         if (sc->sc_sock) {
401                 sock_release(sc->sc_sock);
402                 sc->sc_sock = NULL;
403         }
404 
405         o2nm_undepend_item(&sc->sc_node->nd_item);
406         o2nm_node_put(sc->sc_node);
407         sc->sc_node = NULL;
408 
409         o2net_debug_del_sc(sc);
410 
411         if (sc->sc_page)
412                 __free_page(sc->sc_page);
413         kfree(sc);
414 }
415 
416 static void sc_put(struct o2net_sock_container *sc)
417 {
418         sclog(sc, "put\n");
419         kref_put(&sc->sc_kref, sc_kref_release);
420 }
421 static void sc_get(struct o2net_sock_container *sc)
422 {
423         sclog(sc, "get\n");
424         kref_get(&sc->sc_kref);
425 }
426 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
427 {
428         struct o2net_sock_container *sc, *ret = NULL;
429         struct page *page = NULL;
430         int status = 0;
431 
432         page = alloc_page(GFP_NOFS);
433         sc = kzalloc(sizeof(*sc), GFP_NOFS);
434         if (sc == NULL || page == NULL)
435                 goto out;
436 
437         kref_init(&sc->sc_kref);
438         o2nm_node_get(node);
439         sc->sc_node = node;
440 
441         /* pin the node item of the remote node */
442         status = o2nm_depend_item(&node->nd_item);
443         if (status) {
444                 mlog_errno(status);
445                 o2nm_node_put(node);
446                 goto out;
447         }
448         INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
449         INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
450         INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
451         INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
452 
453         timer_setup(&sc->sc_idle_timeout, o2net_idle_timer, 0);
454 
455         sclog(sc, "alloced\n");
456 
457         ret = sc;
458         sc->sc_page = page;
459         o2net_debug_add_sc(sc);
460         sc = NULL;
461         page = NULL;
462 
463 out:
464         if (page)
465                 __free_page(page);
466         kfree(sc);
467 
468         return ret;
469 }
470 
471 /* ------------------------------------------------------------ */
472 
473 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
474                                 struct work_struct *work)
475 {
476         sc_get(sc);
477         if (!queue_work(o2net_wq, work))
478                 sc_put(sc);
479 }
480 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
481                                         struct delayed_work *work,
482                                         int delay)
483 {
484         sc_get(sc);
485         if (!queue_delayed_work(o2net_wq, work, delay))
486                 sc_put(sc);
487 }
488 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
489                                          struct delayed_work *work)
490 {
491         if (cancel_delayed_work(work))
492                 sc_put(sc);
493 }
494 
495 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
496 
497 int o2net_num_connected_peers(void)
498 {
499         return atomic_read(&o2net_connected_peers);
500 }
501 
502 static void o2net_set_nn_state(struct o2net_node *nn,
503                                struct o2net_sock_container *sc,
504                                unsigned valid, int err)
505 {
506         int was_valid = nn->nn_sc_valid;
507         int was_err = nn->nn_persistent_error;
508         struct o2net_sock_container *old_sc = nn->nn_sc;
509 
510         assert_spin_locked(&nn->nn_lock);
511 
512         if (old_sc && !sc)
513                 atomic_dec(&o2net_connected_peers);
514         else if (!old_sc && sc)
515                 atomic_inc(&o2net_connected_peers);
516 
517         /* the node num comparison and single connect/accept path should stop
518          * an non-null sc from being overwritten with another */
519         BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
520         mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
521         mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
522 
523         if (was_valid && !valid && err == 0)
524                 err = -ENOTCONN;
525 
526         mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
527              o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
528              nn->nn_persistent_error, err);
529 
530         nn->nn_sc = sc;
531         nn->nn_sc_valid = valid ? 1 : 0;
532         nn->nn_persistent_error = err;
533 
534         /* mirrors o2net_tx_can_proceed() */
535         if (nn->nn_persistent_error || nn->nn_sc_valid)
536                 wake_up(&nn->nn_sc_wq);
537 
538         if (was_valid && !was_err && nn->nn_persistent_error) {
539                 o2quo_conn_err(o2net_num_from_nn(nn));
540                 queue_delayed_work(o2net_wq, &nn->nn_still_up,
541                                    msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
542         }
543 
544         if (was_valid && !valid) {
545                 if (old_sc)
546                         printk(KERN_NOTICE "o2net: No longer connected to "
547                                 SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
548                 o2net_complete_nodes_nsw(nn);
549         }
550 
551         if (!was_valid && valid) {
552                 o2quo_conn_up(o2net_num_from_nn(nn));
553                 cancel_delayed_work(&nn->nn_connect_expired);
554                 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
555                        o2nm_this_node() > sc->sc_node->nd_num ?
556                        "Connected to" : "Accepted connection from",
557                        SC_NODEF_ARGS(sc));
558         }
559 
560         /* trigger the connecting worker func as long as we're not valid,
561          * it will back off if it shouldn't connect.  This can be called
562          * from node config teardown and so needs to be careful about
563          * the work queue actually being up. */
564         if (!valid && o2net_wq) {
565                 unsigned long delay;
566                 /* delay if we're within a RECONNECT_DELAY of the
567                  * last attempt */
568                 delay = (nn->nn_last_connect_attempt +
569                          msecs_to_jiffies(o2net_reconnect_delay()))
570                         - jiffies;
571                 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
572                         delay = 0;
573                 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
574                 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
575 
576                 /*
577                  * Delay the expired work after idle timeout.
578                  *
579                  * We might have lots of failed connection attempts that run
580                  * through here but we only cancel the connect_expired work when
581                  * a connection attempt succeeds.  So only the first enqueue of
582                  * the connect_expired work will do anything.  The rest will see
583                  * that it's already queued and do nothing.
584                  */
585                 delay += msecs_to_jiffies(o2net_idle_timeout());
586                 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
587         }
588 
589         /* keep track of the nn's sc ref for the caller */
590         if ((old_sc == NULL) && sc)
591                 sc_get(sc);
592         if (old_sc && (old_sc != sc)) {
593                 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
594                 sc_put(old_sc);
595         }
596 }
597 
598 /* see o2net_register_callbacks() */
599 static void o2net_data_ready(struct sock *sk)
600 {
601         void (*ready)(struct sock *sk);
602         struct o2net_sock_container *sc;
603 
604         read_lock_bh(&sk->sk_callback_lock);
605         sc = sk->sk_user_data;
606         if (sc) {
607                 sclog(sc, "data_ready hit\n");
608                 o2net_set_data_ready_time(sc);
609                 o2net_sc_queue_work(sc, &sc->sc_rx_work);
610                 ready = sc->sc_data_ready;
611         } else {
612                 ready = sk->sk_data_ready;
613         }
614         read_unlock_bh(&sk->sk_callback_lock);
615 
616         ready(sk);
617 }
618 
619 /* see o2net_register_callbacks() */
620 static void o2net_state_change(struct sock *sk)
621 {
622         void (*state_change)(struct sock *sk);
623         struct o2net_sock_container *sc;
624 
625         read_lock_bh(&sk->sk_callback_lock);
626         sc = sk->sk_user_data;
627         if (sc == NULL) {
628                 state_change = sk->sk_state_change;
629                 goto out;
630         }
631 
632         sclog(sc, "state_change to %d\n", sk->sk_state);
633 
634         state_change = sc->sc_state_change;
635 
636         switch(sk->sk_state) {
637         /* ignore connecting sockets as they make progress */
638         case TCP_SYN_SENT:
639         case TCP_SYN_RECV:
640                 break;
641         case TCP_ESTABLISHED:
642                 o2net_sc_queue_work(sc, &sc->sc_connect_work);
643                 break;
644         default:
645                 printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
646                         " shutdown, state %d\n",
647                         SC_NODEF_ARGS(sc), sk->sk_state);
648                 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
649                 break;
650         }
651 out:
652         read_unlock_bh(&sk->sk_callback_lock);
653         state_change(sk);
654 }
655 
656 /*
657  * we register callbacks so we can queue work on events before calling
658  * the original callbacks.  our callbacks our careful to test user_data
659  * to discover when they've reaced with o2net_unregister_callbacks().
660  */
661 static void o2net_register_callbacks(struct sock *sk,
662                                      struct o2net_sock_container *sc)
663 {
664         write_lock_bh(&sk->sk_callback_lock);
665 
666         /* accepted sockets inherit the old listen socket data ready */
667         if (sk->sk_data_ready == o2net_listen_data_ready) {
668                 sk->sk_data_ready = sk->sk_user_data;
669                 sk->sk_user_data = NULL;
670         }
671 
672         BUG_ON(sk->sk_user_data != NULL);
673         sk->sk_user_data = sc;
674         sc_get(sc);
675 
676         sc->sc_data_ready = sk->sk_data_ready;
677         sc->sc_state_change = sk->sk_state_change;
678         sk->sk_data_ready = o2net_data_ready;
679         sk->sk_state_change = o2net_state_change;
680 
681         mutex_init(&sc->sc_send_lock);
682 
683         write_unlock_bh(&sk->sk_callback_lock);
684 }
685 
686 static int o2net_unregister_callbacks(struct sock *sk,
687                                    struct o2net_sock_container *sc)
688 {
689         int ret = 0;
690 
691         write_lock_bh(&sk->sk_callback_lock);
692         if (sk->sk_user_data == sc) {
693                 ret = 1;
694                 sk->sk_user_data = NULL;
695                 sk->sk_data_ready = sc->sc_data_ready;
696                 sk->sk_state_change = sc->sc_state_change;
697         }
698         write_unlock_bh(&sk->sk_callback_lock);
699 
700         return ret;
701 }
702 
703 /*
704  * this is a little helper that is called by callers who have seen a problem
705  * with an sc and want to detach it from the nn if someone already hasn't beat
706  * them to it.  if an error is given then the shutdown will be persistent
707  * and pending transmits will be canceled.
708  */
709 static void o2net_ensure_shutdown(struct o2net_node *nn,
710                                    struct o2net_sock_container *sc,
711                                    int err)
712 {
713         spin_lock(&nn->nn_lock);
714         if (nn->nn_sc == sc)
715                 o2net_set_nn_state(nn, NULL, 0, err);
716         spin_unlock(&nn->nn_lock);
717 }
718 
719 /*
720  * This work queue function performs the blocking parts of socket shutdown.  A
721  * few paths lead here.  set_nn_state will trigger this callback if it sees an
722  * sc detached from the nn.  state_change will also trigger this callback
723  * directly when it sees errors.  In that case we need to call set_nn_state
724  * ourselves as state_change couldn't get the nn_lock and call set_nn_state
725  * itself.
726  */
727 static void o2net_shutdown_sc(struct work_struct *work)
728 {
729         struct o2net_sock_container *sc =
730                 container_of(work, struct o2net_sock_container,
731                              sc_shutdown_work);
732         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
733 
734         sclog(sc, "shutting down\n");
735 
736         /* drop the callbacks ref and call shutdown only once */
737         if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
738                 /* we shouldn't flush as we're in the thread, the
739                  * races with pending sc work structs are harmless */
740                 del_timer_sync(&sc->sc_idle_timeout);
741                 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
742                 sc_put(sc);
743                 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
744         }
745 
746         /* not fatal so failed connects before the other guy has our
747          * heartbeat can be retried */
748         o2net_ensure_shutdown(nn, sc, 0);
749         sc_put(sc);
750 }
751 
752 /* ------------------------------------------------------------ */
753 
754 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
755                              u32 key)
756 {
757         int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
758 
759         if (ret == 0)
760                 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
761 
762         return ret;
763 }
764 
765 static struct o2net_msg_handler *
766 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
767                           struct rb_node **ret_parent)
768 {
769         struct rb_node **p = &o2net_handler_tree.rb_node;
770         struct rb_node *parent = NULL;
771         struct o2net_msg_handler *nmh, *ret = NULL;
772         int cmp;
773 
774         while (*p) {
775                 parent = *p;
776                 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
777                 cmp = o2net_handler_cmp(nmh, msg_type, key);
778 
779                 if (cmp < 0)
780                         p = &(*p)->rb_left;
781                 else if (cmp > 0)
782                         p = &(*p)->rb_right;
783                 else {
784                         ret = nmh;
785                         break;
786                 }
787         }
788 
789         if (ret_p != NULL)
790                 *ret_p = p;
791         if (ret_parent != NULL)
792                 *ret_parent = parent;
793 
794         return ret;
795 }
796 
797 static void o2net_handler_kref_release(struct kref *kref)
798 {
799         struct o2net_msg_handler *nmh;
800         nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
801 
802         kfree(nmh);
803 }
804 
805 static void o2net_handler_put(struct o2net_msg_handler *nmh)
806 {
807         kref_put(&nmh->nh_kref, o2net_handler_kref_release);
808 }
809 
810 /* max_len is protection for the handler func.  incoming messages won't
811  * be given to the handler if their payload is longer than the max. */
812 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
813                            o2net_msg_handler_func *func, void *data,
814                            o2net_post_msg_handler_func *post_func,
815                            struct list_head *unreg_list)
816 {
817         struct o2net_msg_handler *nmh = NULL;
818         struct rb_node **p, *parent;
819         int ret = 0;
820 
821         if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
822                 mlog(0, "max_len for message handler out of range: %u\n",
823                         max_len);
824                 ret = -EINVAL;
825                 goto out;
826         }
827 
828         if (!msg_type) {
829                 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
830                 ret = -EINVAL;
831                 goto out;
832 
833         }
834         if (!func) {
835                 mlog(0, "no message handler provided: %u, %p\n",
836                        msg_type, func);
837                 ret = -EINVAL;
838                 goto out;
839         }
840 
841         nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
842         if (nmh == NULL) {
843                 ret = -ENOMEM;
844                 goto out;
845         }
846 
847         nmh->nh_func = func;
848         nmh->nh_func_data = data;
849         nmh->nh_post_func = post_func;
850         nmh->nh_msg_type = msg_type;
851         nmh->nh_max_len = max_len;
852         nmh->nh_key = key;
853         /* the tree and list get this ref.. they're both removed in
854          * unregister when this ref is dropped */
855         kref_init(&nmh->nh_kref);
856         INIT_LIST_HEAD(&nmh->nh_unregister_item);
857 
858         write_lock(&o2net_handler_lock);
859         if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
860                 ret = -EEXIST;
861         else {
862                 rb_link_node(&nmh->nh_node, parent, p);
863                 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
864                 list_add_tail(&nmh->nh_unregister_item, unreg_list);
865 
866                 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
867                      func, msg_type, key);
868                 /* we've had some trouble with handlers seemingly vanishing. */
869                 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
870                                                           &parent) == NULL,
871                                 "couldn't find handler we *just* registered "
872                                 "for type %u key %08x\n", msg_type, key);
873         }
874         write_unlock(&o2net_handler_lock);
875         if (ret)
876                 goto out;
877 
878 out:
879         if (ret)
880                 kfree(nmh);
881 
882         return ret;
883 }
884 EXPORT_SYMBOL_GPL(o2net_register_handler);
885 
886 void o2net_unregister_handler_list(struct list_head *list)
887 {
888         struct o2net_msg_handler *nmh, *n;
889 
890         write_lock(&o2net_handler_lock);
891         list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
892                 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
893                      nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
894                 rb_erase(&nmh->nh_node, &o2net_handler_tree);
895                 list_del_init(&nmh->nh_unregister_item);
896                 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
897         }
898         write_unlock(&o2net_handler_lock);
899 }
900 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
901 
902 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
903 {
904         struct o2net_msg_handler *nmh;
905 
906         read_lock(&o2net_handler_lock);
907         nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
908         if (nmh)
909                 kref_get(&nmh->nh_kref);
910         read_unlock(&o2net_handler_lock);
911 
912         return nmh;
913 }
914 
915 /* ------------------------------------------------------------ */
916 
917 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
918 {
919         struct kvec vec = { .iov_len = len, .iov_base = data, };
920         struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
921         iov_iter_kvec(&msg.msg_iter, READ | ITER_KVEC, &vec, 1, len);
922         return sock_recvmsg(sock, &msg, MSG_DONTWAIT);
923 }
924 
925 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
926                               size_t veclen, size_t total)
927 {
928         int ret;
929         struct msghdr msg = {.msg_flags = 0,};
930 
931         if (sock == NULL) {
932                 ret = -EINVAL;
933                 goto out;
934         }
935 
936         ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
937         if (likely(ret == total))
938                 return 0;
939         mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
940         if (ret >= 0)
941                 ret = -EPIPE; /* should be smarter, I bet */
942 out:
943         mlog(0, "returning error: %d\n", ret);
944         return ret;
945 }
946 
947 static void o2net_sendpage(struct o2net_sock_container *sc,
948                            void *kmalloced_virt,
949                            size_t size)
950 {
951         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
952         ssize_t ret;
953 
954         while (1) {
955                 mutex_lock(&sc->sc_send_lock);
956                 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
957                                                  virt_to_page(kmalloced_virt),
958                                                  offset_in_page(kmalloced_virt),
959                                                  size, MSG_DONTWAIT);
960                 mutex_unlock(&sc->sc_send_lock);
961                 if (ret == size)
962                         break;
963                 if (ret == (ssize_t)-EAGAIN) {
964                         mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
965                              " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
966                         cond_resched();
967                         continue;
968                 }
969                 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
970                      " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
971                 o2net_ensure_shutdown(nn, sc, 0);
972                 break;
973         }
974 }
975 
976 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
977 {
978         memset(msg, 0, sizeof(struct o2net_msg));
979         msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
980         msg->data_len = cpu_to_be16(data_len);
981         msg->msg_type = cpu_to_be16(msg_type);
982         msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
983         msg->status = 0;
984         msg->key = cpu_to_be32(key);
985 }
986 
987 static int o2net_tx_can_proceed(struct o2net_node *nn,
988                                 struct o2net_sock_container **sc_ret,
989                                 int *error)
990 {
991         int ret = 0;
992 
993         spin_lock(&nn->nn_lock);
994         if (nn->nn_persistent_error) {
995                 ret = 1;
996                 *sc_ret = NULL;
997                 *error = nn->nn_persistent_error;
998         } else if (nn->nn_sc_valid) {
999                 kref_get(&nn->nn_sc->sc_kref);
1000 
1001                 ret = 1;
1002                 *sc_ret = nn->nn_sc;
1003                 *error = 0;
1004         }
1005         spin_unlock(&nn->nn_lock);
1006 
1007         return ret;
1008 }
1009 
1010 /* Get a map of all nodes to which this node is currently connected to */
1011 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1012 {
1013         struct o2net_sock_container *sc;
1014         int node, ret;
1015 
1016         BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1017 
1018         memset(map, 0, bytes);
1019         for (node = 0; node < O2NM_MAX_NODES; ++node) {
1020                 if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1021                         continue;
1022                 if (!ret) {
1023                         set_bit(node, map);
1024                         sc_put(sc);
1025                 }
1026         }
1027 }
1028 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1029 
1030 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1031                            size_t caller_veclen, u8 target_node, int *status)
1032 {
1033         int ret = 0;
1034         struct o2net_msg *msg = NULL;
1035         size_t veclen, caller_bytes = 0;
1036         struct kvec *vec = NULL;
1037         struct o2net_sock_container *sc = NULL;
1038         struct o2net_node *nn = o2net_nn_from_num(target_node);
1039         struct o2net_status_wait nsw = {
1040                 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1041         };
1042         struct o2net_send_tracking nst;
1043 
1044         o2net_init_nst(&nst, msg_type, key, current, target_node);
1045 
1046         if (o2net_wq == NULL) {
1047                 mlog(0, "attempt to tx without o2netd running\n");
1048                 ret = -ESRCH;
1049                 goto out;
1050         }
1051 
1052         if (caller_veclen == 0) {
1053                 mlog(0, "bad kvec array length\n");
1054                 ret = -EINVAL;
1055                 goto out;
1056         }
1057 
1058         caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1059         if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1060                 mlog(0, "total payload len %zu too large\n", caller_bytes);
1061                 ret = -EINVAL;
1062                 goto out;
1063         }
1064 
1065         if (target_node == o2nm_this_node()) {
1066                 ret = -ELOOP;
1067                 goto out;
1068         }
1069 
1070         o2net_debug_add_nst(&nst);
1071 
1072         o2net_set_nst_sock_time(&nst);
1073 
1074         wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1075         if (ret)
1076                 goto out;
1077 
1078         o2net_set_nst_sock_container(&nst, sc);
1079 
1080         veclen = caller_veclen + 1;
1081         vec = kmalloc_array(veclen, sizeof(struct kvec), GFP_ATOMIC);
1082         if (vec == NULL) {
1083                 mlog(0, "failed to %zu element kvec!\n", veclen);
1084                 ret = -ENOMEM;
1085                 goto out;
1086         }
1087 
1088         msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1089         if (!msg) {
1090                 mlog(0, "failed to allocate a o2net_msg!\n");
1091                 ret = -ENOMEM;
1092                 goto out;
1093         }
1094 
1095         o2net_init_msg(msg, caller_bytes, msg_type, key);
1096 
1097         vec[0].iov_len = sizeof(struct o2net_msg);
1098         vec[0].iov_base = msg;
1099         memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1100 
1101         ret = o2net_prep_nsw(nn, &nsw);
1102         if (ret)
1103                 goto out;
1104 
1105         msg->msg_num = cpu_to_be32(nsw.ns_id);
1106         o2net_set_nst_msg_id(&nst, nsw.ns_id);
1107 
1108         o2net_set_nst_send_time(&nst);
1109 
1110         /* finally, convert the message header to network byte-order
1111          * and send */
1112         mutex_lock(&sc->sc_send_lock);
1113         ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1114                                  sizeof(struct o2net_msg) + caller_bytes);
1115         mutex_unlock(&sc->sc_send_lock);
1116         msglog(msg, "sending returned %d\n", ret);
1117         if (ret < 0) {
1118                 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1119                 goto out;
1120         }
1121 
1122         /* wait on other node's handler */
1123         o2net_set_nst_status_time(&nst);
1124         wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1125 
1126         o2net_update_send_stats(&nst, sc);
1127 
1128         /* Note that we avoid overwriting the callers status return
1129          * variable if a system error was reported on the other
1130          * side. Callers beware. */
1131         ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1132         if (status && !ret)
1133                 *status = nsw.ns_status;
1134 
1135         mlog(0, "woken, returning system status %d, user status %d\n",
1136              ret, nsw.ns_status);
1137 out:
1138         o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1139         if (sc)
1140                 sc_put(sc);
1141         kfree(vec);
1142         kfree(msg);
1143         o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1144         return ret;
1145 }
1146 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1147 
1148 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1149                        u8 target_node, int *status)
1150 {
1151         struct kvec vec = {
1152                 .iov_base = data,
1153                 .iov_len = len,
1154         };
1155         return o2net_send_message_vec(msg_type, key, &vec, 1,
1156                                       target_node, status);
1157 }
1158 EXPORT_SYMBOL_GPL(o2net_send_message);
1159 
1160 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1161                                    enum o2net_system_error syserr, int err)
1162 {
1163         struct kvec vec = {
1164                 .iov_base = hdr,
1165                 .iov_len = sizeof(struct o2net_msg),
1166         };
1167 
1168         BUG_ON(syserr >= O2NET_ERR_MAX);
1169 
1170         /* leave other fields intact from the incoming message, msg_num
1171          * in particular */
1172         hdr->sys_status = cpu_to_be32(syserr);
1173         hdr->status = cpu_to_be32(err);
1174         hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC);  // twiddle the magic
1175         hdr->data_len = 0;
1176 
1177         msglog(hdr, "about to send status magic %d\n", err);
1178         /* hdr has been in host byteorder this whole time */
1179         return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1180 }
1181 
1182 /* this returns -errno if the header was unknown or too large, etc.
1183  * after this is called the buffer us reused for the next message */
1184 static int o2net_process_message(struct o2net_sock_container *sc,
1185                                  struct o2net_msg *hdr)
1186 {
1187         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1188         int ret = 0, handler_status;
1189         enum  o2net_system_error syserr;
1190         struct o2net_msg_handler *nmh = NULL;
1191         void *ret_data = NULL;
1192 
1193         msglog(hdr, "processing message\n");
1194 
1195         o2net_sc_postpone_idle(sc);
1196 
1197         switch(be16_to_cpu(hdr->magic)) {
1198                 case O2NET_MSG_STATUS_MAGIC:
1199                         /* special type for returning message status */
1200                         o2net_complete_nsw(nn, NULL,
1201                                            be32_to_cpu(hdr->msg_num),
1202                                            be32_to_cpu(hdr->sys_status),
1203                                            be32_to_cpu(hdr->status));
1204                         goto out;
1205                 case O2NET_MSG_KEEP_REQ_MAGIC:
1206                         o2net_sendpage(sc, o2net_keep_resp,
1207                                        sizeof(*o2net_keep_resp));
1208                         goto out;
1209                 case O2NET_MSG_KEEP_RESP_MAGIC:
1210                         goto out;
1211                 case O2NET_MSG_MAGIC:
1212                         break;
1213                 default:
1214                         msglog(hdr, "bad magic\n");
1215                         ret = -EINVAL;
1216                         goto out;
1217                         break;
1218         }
1219 
1220         /* find a handler for it */
1221         handler_status = 0;
1222         nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1223                                 be32_to_cpu(hdr->key));
1224         if (!nmh) {
1225                 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1226                      be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1227                 syserr = O2NET_ERR_NO_HNDLR;
1228                 goto out_respond;
1229         }
1230 
1231         syserr = O2NET_ERR_NONE;
1232 
1233         if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1234                 syserr = O2NET_ERR_OVERFLOW;
1235 
1236         if (syserr != O2NET_ERR_NONE)
1237                 goto out_respond;
1238 
1239         o2net_set_func_start_time(sc);
1240         sc->sc_msg_key = be32_to_cpu(hdr->key);
1241         sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1242         handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1243                                              be16_to_cpu(hdr->data_len),
1244                                         nmh->nh_func_data, &ret_data);
1245         o2net_set_func_stop_time(sc);
1246 
1247         o2net_update_recv_stats(sc);
1248 
1249 out_respond:
1250         /* this destroys the hdr, so don't use it after this */
1251         mutex_lock(&sc->sc_send_lock);
1252         ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1253                                       handler_status);
1254         mutex_unlock(&sc->sc_send_lock);
1255         hdr = NULL;
1256         mlog(0, "sending handler status %d, syserr %d returned %d\n",
1257              handler_status, syserr, ret);
1258 
1259         if (nmh) {
1260                 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1261                 if (nmh->nh_post_func)
1262                         (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1263                                             ret_data);
1264         }
1265 
1266 out:
1267         if (nmh)
1268                 o2net_handler_put(nmh);
1269         return ret;
1270 }
1271 
1272 static int o2net_check_handshake(struct o2net_sock_container *sc)
1273 {
1274         struct o2net_handshake *hand = page_address(sc->sc_page);
1275         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1276 
1277         if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1278                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1279                        "protocol version %llu but %llu is required. "
1280                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1281                        (unsigned long long)be64_to_cpu(hand->protocol_version),
1282                        O2NET_PROTOCOL_VERSION);
1283 
1284                 /* don't bother reconnecting if its the wrong version. */
1285                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1286                 return -1;
1287         }
1288 
1289         /*
1290          * Ensure timeouts are consistent with other nodes, otherwise
1291          * we can end up with one node thinking that the other must be down,
1292          * but isn't. This can ultimately cause corruption.
1293          */
1294         if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1295                                 o2net_idle_timeout()) {
1296                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1297                        "idle timeout of %u ms, but we use %u ms locally. "
1298                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1299                        be32_to_cpu(hand->o2net_idle_timeout_ms),
1300                        o2net_idle_timeout());
1301                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1302                 return -1;
1303         }
1304 
1305         if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1306                         o2net_keepalive_delay()) {
1307                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1308                        "delay of %u ms, but we use %u ms locally. "
1309                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1310                        be32_to_cpu(hand->o2net_keepalive_delay_ms),
1311                        o2net_keepalive_delay());
1312                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1313                 return -1;
1314         }
1315 
1316         if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1317                         O2HB_MAX_WRITE_TIMEOUT_MS) {
1318                 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1319                        "timeout of %u ms, but we use %u ms locally. "
1320                        "Disconnecting.\n", SC_NODEF_ARGS(sc),
1321                        be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1322                        O2HB_MAX_WRITE_TIMEOUT_MS);
1323                 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1324                 return -1;
1325         }
1326 
1327         sc->sc_handshake_ok = 1;
1328 
1329         spin_lock(&nn->nn_lock);
1330         /* set valid and queue the idle timers only if it hasn't been
1331          * shut down already */
1332         if (nn->nn_sc == sc) {
1333                 o2net_sc_reset_idle_timer(sc);
1334                 atomic_set(&nn->nn_timeout, 0);
1335                 o2net_set_nn_state(nn, sc, 1, 0);
1336         }
1337         spin_unlock(&nn->nn_lock);
1338 
1339         /* shift everything up as though it wasn't there */
1340         sc->sc_page_off -= sizeof(struct o2net_handshake);
1341         if (sc->sc_page_off)
1342                 memmove(hand, hand + 1, sc->sc_page_off);
1343 
1344         return 0;
1345 }
1346 
1347 /* this demuxes the queued rx bytes into header or payload bits and calls
1348  * handlers as each full message is read off the socket.  it returns -error,
1349  * == 0 eof, or > 0 for progress made.*/
1350 static int o2net_advance_rx(struct o2net_sock_container *sc)
1351 {
1352         struct o2net_msg *hdr;
1353         int ret = 0;
1354         void *data;
1355         size_t datalen;
1356 
1357         sclog(sc, "receiving\n");
1358         o2net_set_advance_start_time(sc);
1359 
1360         if (unlikely(sc->sc_handshake_ok == 0)) {
1361                 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1362                         data = page_address(sc->sc_page) + sc->sc_page_off;
1363                         datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1364                         ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1365                         if (ret > 0)
1366                                 sc->sc_page_off += ret;
1367                 }
1368 
1369                 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1370                         o2net_check_handshake(sc);
1371                         if (unlikely(sc->sc_handshake_ok == 0))
1372                                 ret = -EPROTO;
1373                 }
1374                 goto out;
1375         }
1376 
1377         /* do we need more header? */
1378         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1379                 data = page_address(sc->sc_page) + sc->sc_page_off;
1380                 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1381                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1382                 if (ret > 0) {
1383                         sc->sc_page_off += ret;
1384                         /* only swab incoming here.. we can
1385                          * only get here once as we cross from
1386                          * being under to over */
1387                         if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1388                                 hdr = page_address(sc->sc_page);
1389                                 if (be16_to_cpu(hdr->data_len) >
1390                                     O2NET_MAX_PAYLOAD_BYTES)
1391                                         ret = -EOVERFLOW;
1392                         }
1393                 }
1394                 if (ret <= 0)
1395                         goto out;
1396         }
1397 
1398         if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1399                 /* oof, still don't have a header */
1400                 goto out;
1401         }
1402 
1403         /* this was swabbed above when we first read it */
1404         hdr = page_address(sc->sc_page);
1405 
1406         msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1407 
1408         /* do we need more payload? */
1409         if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1410                 /* need more payload */
1411                 data = page_address(sc->sc_page) + sc->sc_page_off;
1412                 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1413                           sc->sc_page_off;
1414                 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1415                 if (ret > 0)
1416                         sc->sc_page_off += ret;
1417                 if (ret <= 0)
1418                         goto out;
1419         }
1420 
1421         if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1422                 /* we can only get here once, the first time we read
1423                  * the payload.. so set ret to progress if the handler
1424                  * works out. after calling this the message is toast */
1425                 ret = o2net_process_message(sc, hdr);
1426                 if (ret == 0)
1427                         ret = 1;
1428                 sc->sc_page_off = 0;
1429         }
1430 
1431 out:
1432         sclog(sc, "ret = %d\n", ret);
1433         o2net_set_advance_stop_time(sc);
1434         return ret;
1435 }
1436 
1437 /* this work func is triggerd by data ready.  it reads until it can read no
1438  * more.  it interprets 0, eof, as fatal.  if data_ready hits while we're doing
1439  * our work the work struct will be marked and we'll be called again. */
1440 static void o2net_rx_until_empty(struct work_struct *work)
1441 {
1442         struct o2net_sock_container *sc =
1443                 container_of(work, struct o2net_sock_container, sc_rx_work);
1444         int ret;
1445 
1446         do {
1447                 ret = o2net_advance_rx(sc);
1448         } while (ret > 0);
1449 
1450         if (ret <= 0 && ret != -EAGAIN) {
1451                 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1452                 sclog(sc, "saw error %d, closing\n", ret);
1453                 /* not permanent so read failed handshake can retry */
1454                 o2net_ensure_shutdown(nn, sc, 0);
1455         }
1456 
1457         sc_put(sc);
1458 }
1459 
1460 static int o2net_set_nodelay(struct socket *sock)
1461 {
1462         int val = 1;
1463 
1464         return kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
1465                                     (void *)&val, sizeof(val));
1466 }
1467 
1468 static int o2net_set_usertimeout(struct socket *sock)
1469 {
1470         int user_timeout = O2NET_TCP_USER_TIMEOUT;
1471 
1472         return kernel_setsockopt(sock, SOL_TCP, TCP_USER_TIMEOUT,
1473                                 (void *)&user_timeout, sizeof(user_timeout));
1474 }
1475 
1476 static void o2net_initialize_handshake(void)
1477 {
1478         o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1479                 O2HB_MAX_WRITE_TIMEOUT_MS);
1480         o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1481         o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1482                 o2net_keepalive_delay());
1483         o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1484                 o2net_reconnect_delay());
1485 }
1486 
1487 /* ------------------------------------------------------------ */
1488 
1489 /* called when a connect completes and after a sock is accepted.  the
1490  * rx path will see the response and mark the sc valid */
1491 static void o2net_sc_connect_completed(struct work_struct *work)
1492 {
1493         struct o2net_sock_container *sc =
1494                 container_of(work, struct o2net_sock_container,
1495                              sc_connect_work);
1496 
1497         mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1498               (unsigned long long)O2NET_PROTOCOL_VERSION,
1499               (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1500 
1501         o2net_initialize_handshake();
1502         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1503         sc_put(sc);
1504 }
1505 
1506 /* this is called as a work_struct func. */
1507 static void o2net_sc_send_keep_req(struct work_struct *work)
1508 {
1509         struct o2net_sock_container *sc =
1510                 container_of(work, struct o2net_sock_container,
1511                              sc_keepalive_work.work);
1512 
1513         o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1514         sc_put(sc);
1515 }
1516 
1517 /* socket shutdown does a del_timer_sync against this as it tears down.
1518  * we can't start this timer until we've got to the point in sc buildup
1519  * where shutdown is going to be involved */
1520 static void o2net_idle_timer(struct timer_list *t)
1521 {
1522         struct o2net_sock_container *sc = from_timer(sc, t, sc_idle_timeout);
1523         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1524 #ifdef CONFIG_DEBUG_FS
1525         unsigned long msecs = ktime_to_ms(ktime_get()) -
1526                 ktime_to_ms(sc->sc_tv_timer);
1527 #else
1528         unsigned long msecs = o2net_idle_timeout();
1529 #endif
1530 
1531         printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1532                "idle for %lu.%lu secs.\n",
1533                SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1534 
1535         /* idle timerout happen, don't shutdown the connection, but
1536          * make fence decision. Maybe the connection can recover before
1537          * the decision is made.
1538          */
1539         atomic_set(&nn->nn_timeout, 1);
1540         o2quo_conn_err(o2net_num_from_nn(nn));
1541         queue_delayed_work(o2net_wq, &nn->nn_still_up,
1542                         msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1543 
1544         o2net_sc_reset_idle_timer(sc);
1545 
1546 }
1547 
1548 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1549 {
1550         o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1551         o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1552                       msecs_to_jiffies(o2net_keepalive_delay()));
1553         o2net_set_sock_timer(sc);
1554         mod_timer(&sc->sc_idle_timeout,
1555                jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1556 }
1557 
1558 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1559 {
1560         struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1561 
1562         /* clear fence decision since the connection recover from timeout*/
1563         if (atomic_read(&nn->nn_timeout)) {
1564                 o2quo_conn_up(o2net_num_from_nn(nn));
1565                 cancel_delayed_work(&nn->nn_still_up);
1566                 atomic_set(&nn->nn_timeout, 0);
1567         }
1568 
1569         /* Only push out an existing timer */
1570         if (timer_pending(&sc->sc_idle_timeout))
1571                 o2net_sc_reset_idle_timer(sc);
1572 }
1573 
1574 /* this work func is kicked whenever a path sets the nn state which doesn't
1575  * have valid set.  This includes seeing hb come up, losing a connection,
1576  * having a connect attempt fail, etc. This centralizes the logic which decides
1577  * if a connect attempt should be made or if we should give up and all future
1578  * transmit attempts should fail */
1579 static void o2net_start_connect(struct work_struct *work)
1580 {
1581         struct o2net_node *nn =
1582                 container_of(work, struct o2net_node, nn_connect_work.work);
1583         struct o2net_sock_container *sc = NULL;
1584         struct o2nm_node *node = NULL, *mynode = NULL;
1585         struct socket *sock = NULL;
1586         struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1587         int ret = 0, stop;
1588         unsigned int timeout;
1589         unsigned int noio_flag;
1590 
1591         /*
1592          * sock_create allocates the sock with GFP_KERNEL. We must set
1593          * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1594          * by this process are done as if GFP_NOIO was specified. So we
1595          * are not reentering filesystem while doing memory reclaim.
1596          */
1597         noio_flag = memalloc_noio_save();
1598         /* if we're greater we initiate tx, otherwise we accept */
1599         if (o2nm_this_node() <= o2net_num_from_nn(nn))
1600                 goto out;
1601 
1602         /* watch for racing with tearing a node down */
1603         node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1604         if (node == NULL)
1605                 goto out;
1606 
1607         mynode = o2nm_get_node_by_num(o2nm_this_node());
1608         if (mynode == NULL)
1609                 goto out;
1610 
1611         spin_lock(&nn->nn_lock);
1612         /*
1613          * see if we already have one pending or have given up.
1614          * For nn_timeout, it is set when we close the connection
1615          * because of the idle time out. So it means that we have
1616          * at least connected to that node successfully once,
1617          * now try to connect to it again.
1618          */
1619         timeout = atomic_read(&nn->nn_timeout);
1620         stop = (nn->nn_sc ||
1621                 (nn->nn_persistent_error &&
1622                 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1623         spin_unlock(&nn->nn_lock);
1624         if (stop)
1625                 goto out;
1626 
1627         nn->nn_last_connect_attempt = jiffies;
1628 
1629         sc = sc_alloc(node);
1630         if (sc == NULL) {
1631                 mlog(0, "couldn't allocate sc\n");
1632                 ret = -ENOMEM;
1633                 goto out;
1634         }
1635 
1636         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1637         if (ret < 0) {
1638                 mlog(0, "can't create socket: %d\n", ret);
1639                 goto out;
1640         }
1641         sc->sc_sock = sock; /* freed by sc_kref_release */
1642 
1643         sock->sk->sk_allocation = GFP_ATOMIC;
1644 
1645         myaddr.sin_family = AF_INET;
1646         myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1647         myaddr.sin_port = htons(0); /* any port */
1648 
1649         ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1650                               sizeof(myaddr));
1651         if (ret) {
1652                 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1653                      ret, &mynode->nd_ipv4_address);
1654                 goto out;
1655         }
1656 
1657         ret = o2net_set_nodelay(sc->sc_sock);
1658         if (ret) {
1659                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1660                 goto out;
1661         }
1662 
1663         ret = o2net_set_usertimeout(sock);
1664         if (ret) {
1665                 mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1666                 goto out;
1667         }
1668 
1669         o2net_register_callbacks(sc->sc_sock->sk, sc);
1670 
1671         spin_lock(&nn->nn_lock);
1672         /* handshake completion will set nn->nn_sc_valid */
1673         o2net_set_nn_state(nn, sc, 0, 0);
1674         spin_unlock(&nn->nn_lock);
1675 
1676         remoteaddr.sin_family = AF_INET;
1677         remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1678         remoteaddr.sin_port = node->nd_ipv4_port;
1679 
1680         ret = sc->sc_sock->ops->connect(sc->sc_sock,
1681                                         (struct sockaddr *)&remoteaddr,
1682                                         sizeof(remoteaddr),
1683                                         O_NONBLOCK);
1684         if (ret == -EINPROGRESS)
1685                 ret = 0;
1686 
1687 out:
1688         if (ret && sc) {
1689                 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1690                        " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1691                 /* 0 err so that another will be queued and attempted
1692                  * from set_nn_state */
1693                 o2net_ensure_shutdown(nn, sc, 0);
1694         }
1695         if (sc)
1696                 sc_put(sc);
1697         if (node)
1698                 o2nm_node_put(node);
1699         if (mynode)
1700                 o2nm_node_put(mynode);
1701 
1702         memalloc_noio_restore(noio_flag);
1703         return;
1704 }
1705 
1706 static void o2net_connect_expired(struct work_struct *work)
1707 {
1708         struct o2net_node *nn =
1709                 container_of(work, struct o2net_node, nn_connect_expired.work);
1710 
1711         spin_lock(&nn->nn_lock);
1712         if (!nn->nn_sc_valid) {
1713                 printk(KERN_NOTICE "o2net: No connection established with "
1714                        "node %u after %u.%u seconds, check network and"
1715                        " cluster configuration.\n",
1716                      o2net_num_from_nn(nn),
1717                      o2net_idle_timeout() / 1000,
1718                      o2net_idle_timeout() % 1000);
1719 
1720                 o2net_set_nn_state(nn, NULL, 0, 0);
1721         }
1722         spin_unlock(&nn->nn_lock);
1723 }
1724 
1725 static void o2net_still_up(struct work_struct *work)
1726 {
1727         struct o2net_node *nn =
1728                 container_of(work, struct o2net_node, nn_still_up.work);
1729 
1730         o2quo_hb_still_up(o2net_num_from_nn(nn));
1731 }
1732 
1733 /* ------------------------------------------------------------ */
1734 
1735 void o2net_disconnect_node(struct o2nm_node *node)
1736 {
1737         struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1738 
1739         /* don't reconnect until it's heartbeating again */
1740         spin_lock(&nn->nn_lock);
1741         atomic_set(&nn->nn_timeout, 0);
1742         o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1743         spin_unlock(&nn->nn_lock);
1744 
1745         if (o2net_wq) {
1746                 cancel_delayed_work(&nn->nn_connect_expired);
1747                 cancel_delayed_work(&nn->nn_connect_work);
1748                 cancel_delayed_work(&nn->nn_still_up);
1749                 flush_workqueue(o2net_wq);
1750         }
1751 }
1752 
1753 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1754                                   void *data)
1755 {
1756         o2quo_hb_down(node_num);
1757 
1758         if (!node)
1759                 return;
1760 
1761         if (node_num != o2nm_this_node())
1762                 o2net_disconnect_node(node);
1763 
1764         BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1765 }
1766 
1767 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1768                                 void *data)
1769 {
1770         struct o2net_node *nn = o2net_nn_from_num(node_num);
1771 
1772         o2quo_hb_up(node_num);
1773 
1774         BUG_ON(!node);
1775 
1776         /* ensure an immediate connect attempt */
1777         nn->nn_last_connect_attempt = jiffies -
1778                 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1779 
1780         if (node_num != o2nm_this_node()) {
1781                 /* believe it or not, accept and node hearbeating testing
1782                  * can succeed for this node before we got here.. so
1783                  * only use set_nn_state to clear the persistent error
1784                  * if that hasn't already happened */
1785                 spin_lock(&nn->nn_lock);
1786                 atomic_set(&nn->nn_timeout, 0);
1787                 if (nn->nn_persistent_error)
1788                         o2net_set_nn_state(nn, NULL, 0, 0);
1789                 spin_unlock(&nn->nn_lock);
1790         }
1791 }
1792 
1793 void o2net_unregister_hb_callbacks(void)
1794 {
1795         o2hb_unregister_callback(NULL, &o2net_hb_up);
1796         o2hb_unregister_callback(NULL, &o2net_hb_down);
1797 }
1798 
1799 int o2net_register_hb_callbacks(void)
1800 {
1801         int ret;
1802 
1803         o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1804                             o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1805         o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1806                             o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1807 
1808         ret = o2hb_register_callback(NULL, &o2net_hb_up);
1809         if (ret == 0)
1810                 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1811 
1812         if (ret)
1813                 o2net_unregister_hb_callbacks();
1814 
1815         return ret;
1816 }
1817 
1818 /* ------------------------------------------------------------ */
1819 
1820 static int o2net_accept_one(struct socket *sock, int *more)
1821 {
1822         int ret;
1823         struct sockaddr_in sin;
1824         struct socket *new_sock = NULL;
1825         struct o2nm_node *node = NULL;
1826         struct o2nm_node *local_node = NULL;
1827         struct o2net_sock_container *sc = NULL;
1828         struct o2net_node *nn;
1829         unsigned int noio_flag;
1830 
1831         /*
1832          * sock_create_lite allocates the sock with GFP_KERNEL. We must set
1833          * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1834          * by this process are done as if GFP_NOIO was specified. So we
1835          * are not reentering filesystem while doing memory reclaim.
1836          */
1837         noio_flag = memalloc_noio_save();
1838 
1839         BUG_ON(sock == NULL);
1840         *more = 0;
1841         ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1842                                sock->sk->sk_protocol, &new_sock);
1843         if (ret)
1844                 goto out;
1845 
1846         new_sock->type = sock->type;
1847         new_sock->ops = sock->ops;
1848         ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, false);
1849         if (ret < 0)
1850                 goto out;
1851 
1852         *more = 1;
1853         new_sock->sk->sk_allocation = GFP_ATOMIC;
1854 
1855         ret = o2net_set_nodelay(new_sock);
1856         if (ret) {
1857                 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1858                 goto out;
1859         }
1860 
1861         ret = o2net_set_usertimeout(new_sock);
1862         if (ret) {
1863                 mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1864                 goto out;
1865         }
1866 
1867         ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin, 1);
1868         if (ret < 0)
1869                 goto out;
1870 
1871         node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1872         if (node == NULL) {
1873                 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1874                        "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1875                        ntohs(sin.sin_port));
1876                 ret = -EINVAL;
1877                 goto out;
1878         }
1879 
1880         if (o2nm_this_node() >= node->nd_num) {
1881                 local_node = o2nm_get_node_by_num(o2nm_this_node());
1882                 if (local_node)
1883                         printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1884                                         "seen at node '%s' (%u, %pI4:%d) from "
1885                                         "node '%s' (%u, %pI4:%d)\n",
1886                                         local_node->nd_name, local_node->nd_num,
1887                                         &(local_node->nd_ipv4_address),
1888                                         ntohs(local_node->nd_ipv4_port),
1889                                         node->nd_name,
1890                                         node->nd_num, &sin.sin_addr.s_addr,
1891                                         ntohs(sin.sin_port));
1892                 ret = -EINVAL;
1893                 goto out;
1894         }
1895 
1896         /* this happens all the time when the other node sees our heartbeat
1897          * and tries to connect before we see their heartbeat */
1898         if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1899                 mlog(ML_CONN, "attempt to connect from node '%s' at "
1900                      "%pI4:%d but it isn't heartbeating\n",
1901                      node->nd_name, &sin.sin_addr.s_addr,
1902                      ntohs(sin.sin_port));
1903                 ret = -EINVAL;
1904                 goto out;
1905         }
1906 
1907         nn = o2net_nn_from_num(node->nd_num);
1908 
1909         spin_lock(&nn->nn_lock);
1910         if (nn->nn_sc)
1911                 ret = -EBUSY;
1912         else
1913                 ret = 0;
1914         spin_unlock(&nn->nn_lock);
1915         if (ret) {
1916                 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1917                        "at %pI4:%d but it already has an open connection\n",
1918                        node->nd_name, &sin.sin_addr.s_addr,
1919                        ntohs(sin.sin_port));
1920                 goto out;
1921         }
1922 
1923         sc = sc_alloc(node);
1924         if (sc == NULL) {
1925                 ret = -ENOMEM;
1926                 goto out;
1927         }
1928 
1929         sc->sc_sock = new_sock;
1930         new_sock = NULL;
1931 
1932         spin_lock(&nn->nn_lock);
1933         atomic_set(&nn->nn_timeout, 0);
1934         o2net_set_nn_state(nn, sc, 0, 0);
1935         spin_unlock(&nn->nn_lock);
1936 
1937         o2net_register_callbacks(sc->sc_sock->sk, sc);
1938         o2net_sc_queue_work(sc, &sc->sc_rx_work);
1939 
1940         o2net_initialize_handshake();
1941         o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1942 
1943 out:
1944         if (new_sock)
1945                 sock_release(new_sock);
1946         if (node)
1947                 o2nm_node_put(node);
1948         if (local_node)
1949                 o2nm_node_put(local_node);
1950         if (sc)
1951                 sc_put(sc);
1952 
1953         memalloc_noio_restore(noio_flag);
1954         return ret;
1955 }
1956 
1957 /*
1958  * This function is invoked in response to one or more
1959  * pending accepts at softIRQ level. We must drain the
1960  * entire que before returning.
1961  */
1962 
1963 static void o2net_accept_many(struct work_struct *work)
1964 {
1965         struct socket *sock = o2net_listen_sock;
1966         int     more;
1967         int     err;
1968 
1969         /*
1970          * It is critical to note that due to interrupt moderation
1971          * at the network driver level, we can't assume to get a
1972          * softIRQ for every single conn since tcp SYN packets
1973          * can arrive back-to-back, and therefore many pending
1974          * accepts may result in just 1 softIRQ. If we terminate
1975          * the o2net_accept_one() loop upon seeing an err, what happens
1976          * to the rest of the conns in the queue? If no new SYN
1977          * arrives for hours, no softIRQ  will be delivered,
1978          * and the connections will just sit in the queue.
1979          */
1980 
1981         for (;;) {
1982                 err = o2net_accept_one(sock, &more);
1983                 if (!more)
1984                         break;
1985                 cond_resched();
1986         }
1987 }
1988 
1989 static void o2net_listen_data_ready(struct sock *sk)
1990 {
1991         void (*ready)(struct sock *sk);
1992 
1993         read_lock_bh(&sk->sk_callback_lock);
1994         ready = sk->sk_user_data;
1995         if (ready == NULL) { /* check for teardown race */
1996                 ready = sk->sk_data_ready;
1997                 goto out;
1998         }
1999 
2000         /* This callback may called twice when a new connection
2001          * is  being established as a child socket inherits everything
2002          * from a parent LISTEN socket, including the data_ready cb of
2003          * the parent. This leads to a hazard. In o2net_accept_one()
2004          * we are still initializing the child socket but have not
2005          * changed the inherited data_ready callback yet when
2006          * data starts arriving.
2007          * We avoid this hazard by checking the state.
2008          * For the listening socket,  the state will be TCP_LISTEN; for the new
2009          * socket, will be  TCP_ESTABLISHED. Also, in this case,
2010          * sk->sk_user_data is not a valid function pointer.
2011          */
2012 
2013         if (sk->sk_state == TCP_LISTEN) {
2014                 queue_work(o2net_wq, &o2net_listen_work);
2015         } else {
2016                 ready = NULL;
2017         }
2018 
2019 out:
2020         read_unlock_bh(&sk->sk_callback_lock);
2021         if (ready != NULL)
2022                 ready(sk);
2023 }
2024 
2025 static int o2net_open_listening_sock(__be32 addr, __be16 port)
2026 {
2027         struct socket *sock = NULL;
2028         int ret;
2029         struct sockaddr_in sin = {
2030                 .sin_family = PF_INET,
2031                 .sin_addr = { .s_addr = addr },
2032                 .sin_port = port,
2033         };
2034 
2035         ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2036         if (ret < 0) {
2037                 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
2038                 goto out;
2039         }
2040 
2041         sock->sk->sk_allocation = GFP_ATOMIC;
2042 
2043         write_lock_bh(&sock->sk->sk_callback_lock);
2044         sock->sk->sk_user_data = sock->sk->sk_data_ready;
2045         sock->sk->sk_data_ready = o2net_listen_data_ready;
2046         write_unlock_bh(&sock->sk->sk_callback_lock);
2047 
2048         o2net_listen_sock = sock;
2049         INIT_WORK(&o2net_listen_work, o2net_accept_many);
2050 
2051         sock->sk->sk_reuse = SK_CAN_REUSE;
2052         ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2053         if (ret < 0) {
2054                 printk(KERN_ERR "o2net: Error %d while binding socket at "
2055                        "%pI4:%u\n", ret, &addr, ntohs(port)); 
2056                 goto out;
2057         }
2058 
2059         ret = sock->ops->listen(sock, 64);
2060         if (ret < 0)
2061                 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2062                        ret, &addr, ntohs(port));
2063 
2064 out:
2065         if (ret) {
2066                 o2net_listen_sock = NULL;
2067                 if (sock)
2068                         sock_release(sock);
2069         }
2070         return ret;
2071 }
2072 
2073 /*
2074  * called from node manager when we should bring up our network listening
2075  * socket.  node manager handles all the serialization to only call this
2076  * once and to match it with o2net_stop_listening().  note,
2077  * o2nm_this_node() doesn't work yet as we're being called while it
2078  * is being set up.
2079  */
2080 int o2net_start_listening(struct o2nm_node *node)
2081 {
2082         int ret = 0;
2083 
2084         BUG_ON(o2net_wq != NULL);
2085         BUG_ON(o2net_listen_sock != NULL);
2086 
2087         mlog(ML_KTHREAD, "starting o2net thread...\n");
2088         o2net_wq = alloc_ordered_workqueue("o2net", WQ_MEM_RECLAIM);
2089         if (o2net_wq == NULL) {
2090                 mlog(ML_ERROR, "unable to launch o2net thread\n");
2091                 return -ENOMEM; /* ? */
2092         }
2093 
2094         ret = o2net_open_listening_sock(node->nd_ipv4_address,
2095                                         node->nd_ipv4_port);
2096         if (ret) {
2097                 destroy_workqueue(o2net_wq);
2098                 o2net_wq = NULL;
2099         } else
2100                 o2quo_conn_up(node->nd_num);
2101 
2102         return ret;
2103 }
2104 
2105 /* again, o2nm_this_node() doesn't work here as we're involved in
2106  * tearing it down */
2107 void o2net_stop_listening(struct o2nm_node *node)
2108 {
2109         struct socket *sock = o2net_listen_sock;
2110         size_t i;
2111 
2112         BUG_ON(o2net_wq == NULL);
2113         BUG_ON(o2net_listen_sock == NULL);
2114 
2115         /* stop the listening socket from generating work */
2116         write_lock_bh(&sock->sk->sk_callback_lock);
2117         sock->sk->sk_data_ready = sock->sk->sk_user_data;
2118         sock->sk->sk_user_data = NULL;
2119         write_unlock_bh(&sock->sk->sk_callback_lock);
2120 
2121         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2122                 struct o2nm_node *node = o2nm_get_node_by_num(i);
2123                 if (node) {
2124                         o2net_disconnect_node(node);
2125                         o2nm_node_put(node);
2126                 }
2127         }
2128 
2129         /* finish all work and tear down the work queue */
2130         mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2131         destroy_workqueue(o2net_wq);
2132         o2net_wq = NULL;
2133 
2134         sock_release(o2net_listen_sock);
2135         o2net_listen_sock = NULL;
2136 
2137         o2quo_conn_err(node->nd_num);
2138 }
2139 
2140 /* ------------------------------------------------------------ */
2141 
2142 int o2net_init(void)
2143 {
2144         unsigned long i;
2145 
2146         o2quo_init();
2147 
2148         if (o2net_debugfs_init())
2149                 goto out;
2150 
2151         o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2152         o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2153         o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2154         if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp)
2155                 goto out;
2156 
2157         o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2158         o2net_hand->connector_id = cpu_to_be64(1);
2159 
2160         o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2161         o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2162 
2163         for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2164                 struct o2net_node *nn = o2net_nn_from_num(i);
2165 
2166                 atomic_set(&nn->nn_timeout, 0);
2167                 spin_lock_init(&nn->nn_lock);
2168                 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2169                 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2170                                   o2net_connect_expired);
2171                 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2172                 /* until we see hb from a node we'll return einval */
2173                 nn->nn_persistent_error = -ENOTCONN;
2174                 init_waitqueue_head(&nn->nn_sc_wq);
2175                 idr_init(&nn->nn_status_idr);
2176                 INIT_LIST_HEAD(&nn->nn_status_list);
2177         }
2178 
2179         return 0;
2180 
2181 out:
2182         kfree(o2net_hand);
2183         kfree(o2net_keep_req);
2184         kfree(o2net_keep_resp);
2185         o2net_debugfs_exit();
2186         o2quo_exit();
2187         return -ENOMEM;
2188 }
2189 
2190 void o2net_exit(void)
2191 {
2192         o2quo_exit();
2193         kfree(o2net_hand);
2194         kfree(o2net_keep_req);
2195         kfree(o2net_keep_resp);
2196         o2net_debugfs_exit();
2197 }
2198 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp