1 /* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
5 * Copyright (C) 2004 Oracle. All rights reserved.
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.
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.
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.
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.
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
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
41 * Any framing errors (bad magic, large payload lengths) close a connection.
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.
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.
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>
64 #include <asm/uaccess.h>
66 #include "heartbeat.h"
68 #include "nodemanager.h"
69 #define MLOG_MASK_PREFIX ML_TCP
73 #include "tcp_internal.h"
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)
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.
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); \
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 , \
104 static DEFINE_RWLOCK(o2net_handler_lock);
105 static struct rb_root o2net_handler_tree = RB_ROOT;
107 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
109 /* XXX someday we'll need better accounting */
110 static struct socket *o2net_listen_sock = NULL;
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.
120 static struct workqueue_struct *o2net_wq;
121 static struct work_struct o2net_listen_work;
123 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
124 #define O2NET_HB_PRI 0x1
126 static struct o2net_handshake *o2net_hand;
127 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
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,};
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);
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)
149 INIT_LIST_HEAD(&nst->st_net_debug_item);
151 nst->st_msg_type = msgtype;
152 nst->st_msg_key = msgkey;
156 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
158 nst->st_sock_time = ktime_get();
161 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
163 nst->st_send_time = ktime_get();
166 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
168 nst->st_status_time = ktime_get();
171 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
172 struct o2net_sock_container *sc)
177 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
183 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
185 sc->sc_tv_timer = ktime_get();
188 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
190 sc->sc_tv_data_ready = ktime_get();
193 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
195 sc->sc_tv_advance_start = ktime_get();
198 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
200 sc->sc_tv_advance_stop = ktime_get();
203 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
205 sc->sc_tv_func_start = ktime_get();
208 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
210 sc->sc_tv_func_stop = ktime_get();
213 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
215 return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
217 #else /* CONFIG_DEBUG_FS */
218 # define o2net_init_nst(a, b, c, d, e)
219 # define o2net_set_nst_sock_time(a)
220 # define o2net_set_nst_send_time(a)
221 # define o2net_set_nst_status_time(a)
222 # define o2net_set_nst_sock_container(a, b)
223 # define o2net_set_nst_msg_id(a, b)
224 # define o2net_set_sock_timer(a)
225 # define o2net_set_data_ready_time(a)
226 # define o2net_set_advance_start_time(a)
227 # define o2net_set_advance_stop_time(a)
228 # define o2net_set_func_start_time(a)
229 # define o2net_set_func_stop_time(a)
230 # define o2net_get_func_run_time(a) (ktime_t)0
231 #endif /* CONFIG_DEBUG_FS */
233 #ifdef CONFIG_OCFS2_FS_STATS
234 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
235 struct o2net_sock_container *sc)
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,
243 sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
244 ktime_sub(nst->st_send_time,
249 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
251 sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
252 o2net_get_func_run_time(sc));
258 # define o2net_update_send_stats(a, b)
260 #endif /* CONFIG_OCFS2_FS_STATS */
262 static inline int o2net_reconnect_delay(void)
264 return o2nm_single_cluster->cl_reconnect_delay_ms;
267 static inline int o2net_keepalive_delay(void)
269 return o2nm_single_cluster->cl_keepalive_delay_ms;
272 static inline int o2net_idle_timeout(void)
274 return o2nm_single_cluster->cl_idle_timeout_ms;
277 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
280 BUG_ON(err >= O2NET_ERR_MAX);
281 trans = o2net_sys_err_translations[err];
283 /* Just in case we mess up the translation table above */
284 BUG_ON(err != O2NET_ERR_NONE && trans == 0);
288 static struct o2net_node * o2net_nn_from_num(u8 node_num)
290 BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
291 return &o2net_nodes[node_num];
294 static u8 o2net_num_from_nn(struct o2net_node *nn)
297 return nn - o2net_nodes;
300 /* ------------------------------------------------------------ */
302 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
307 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
311 spin_lock(&nn->nn_lock);
312 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
314 list_add_tail(&nsw->ns_node_item,
315 &nn->nn_status_list);
316 spin_unlock(&nn->nn_lock);
317 } while (ret == -EAGAIN);
320 init_waitqueue_head(&nsw->ns_wq);
321 nsw->ns_sys_status = O2NET_ERR_NONE;
328 static void o2net_complete_nsw_locked(struct o2net_node *nn,
329 struct o2net_status_wait *nsw,
330 enum o2net_system_error sys_status,
333 assert_spin_locked(&nn->nn_lock);
335 if (!list_empty(&nsw->ns_node_item)) {
336 list_del_init(&nsw->ns_node_item);
337 nsw->ns_sys_status = sys_status;
338 nsw->ns_status = status;
339 idr_remove(&nn->nn_status_idr, nsw->ns_id);
340 wake_up(&nsw->ns_wq);
344 static void o2net_complete_nsw(struct o2net_node *nn,
345 struct o2net_status_wait *nsw,
346 u64 id, enum o2net_system_error sys_status,
349 spin_lock(&nn->nn_lock);
354 nsw = idr_find(&nn->nn_status_idr, id);
359 o2net_complete_nsw_locked(nn, nsw, sys_status, status);
362 spin_unlock(&nn->nn_lock);
366 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
368 struct o2net_status_wait *nsw, *tmp;
369 unsigned int num_kills = 0;
371 assert_spin_locked(&nn->nn_lock);
373 list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
374 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
378 mlog(0, "completed %d messages for node %u\n", num_kills,
379 o2net_num_from_nn(nn));
382 static int o2net_nsw_completed(struct o2net_node *nn,
383 struct o2net_status_wait *nsw)
386 spin_lock(&nn->nn_lock);
387 completed = list_empty(&nsw->ns_node_item);
388 spin_unlock(&nn->nn_lock);
392 /* ------------------------------------------------------------ */
394 static void sc_kref_release(struct kref *kref)
396 struct o2net_sock_container *sc = container_of(kref,
397 struct o2net_sock_container, sc_kref);
398 BUG_ON(timer_pending(&sc->sc_idle_timeout));
400 sclog(sc, "releasing\n");
403 sock_release(sc->sc_sock);
407 o2nm_undepend_item(&sc->sc_node->nd_item);
408 o2nm_node_put(sc->sc_node);
411 o2net_debug_del_sc(sc);
415 static void sc_put(struct o2net_sock_container *sc)
418 kref_put(&sc->sc_kref, sc_kref_release);
420 static void sc_get(struct o2net_sock_container *sc)
423 kref_get(&sc->sc_kref);
425 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
427 struct o2net_sock_container *sc, *ret = NULL;
428 struct page *page = NULL;
431 page = alloc_page(GFP_NOFS);
432 sc = kzalloc(sizeof(*sc), GFP_NOFS);
433 if (sc == NULL || page == NULL)
436 kref_init(&sc->sc_kref);
440 /* pin the node item of the remote node */
441 status = o2nm_depend_item(&node->nd_item);
447 INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
448 INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
449 INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
450 INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
452 init_timer(&sc->sc_idle_timeout);
453 sc->sc_idle_timeout.function = o2net_idle_timer;
454 sc->sc_idle_timeout.data = (unsigned long)sc;
456 sclog(sc, "alloced\n");
460 o2net_debug_add_sc(sc);
472 /* ------------------------------------------------------------ */
474 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
475 struct work_struct *work)
478 if (!queue_work(o2net_wq, work))
481 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
482 struct delayed_work *work,
486 if (!queue_delayed_work(o2net_wq, work, delay))
489 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
490 struct delayed_work *work)
492 if (cancel_delayed_work(work))
496 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
498 int o2net_num_connected_peers(void)
500 return atomic_read(&o2net_connected_peers);
503 static void o2net_set_nn_state(struct o2net_node *nn,
504 struct o2net_sock_container *sc,
505 unsigned valid, int err)
507 int was_valid = nn->nn_sc_valid;
508 int was_err = nn->nn_persistent_error;
509 struct o2net_sock_container *old_sc = nn->nn_sc;
511 assert_spin_locked(&nn->nn_lock);
514 atomic_dec(&o2net_connected_peers);
515 else if (!old_sc && sc)
516 atomic_inc(&o2net_connected_peers);
518 /* the node num comparison and single connect/accept path should stop
519 * an non-null sc from being overwritten with another */
520 BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
521 mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
522 mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
524 if (was_valid && !valid && err == 0)
527 mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
528 o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
529 nn->nn_persistent_error, err);
532 nn->nn_sc_valid = valid ? 1 : 0;
533 nn->nn_persistent_error = err;
535 /* mirrors o2net_tx_can_proceed() */
536 if (nn->nn_persistent_error || nn->nn_sc_valid)
537 wake_up(&nn->nn_sc_wq);
539 if (!was_err && nn->nn_persistent_error) {
540 o2quo_conn_err(o2net_num_from_nn(nn));
541 queue_delayed_work(o2net_wq, &nn->nn_still_up,
542 msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
545 if (was_valid && !valid) {
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);
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",
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) {
566 /* delay if we're withing a RECONNECT_DELAY of the
568 delay = (nn->nn_last_connect_attempt +
569 msecs_to_jiffies(o2net_reconnect_delay()))
571 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
573 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
574 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
577 * Delay the expired work after idle timeout.
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.
585 delay += msecs_to_jiffies(o2net_idle_timeout());
586 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
589 /* keep track of the nn's sc ref for the caller */
590 if ((old_sc == NULL) && sc)
592 if (old_sc && (old_sc != sc)) {
593 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
598 /* see o2net_register_callbacks() */
599 static void o2net_data_ready(struct sock *sk, int bytes)
601 void (*ready)(struct sock *sk, int bytes);
603 read_lock(&sk->sk_callback_lock);
604 if (sk->sk_user_data) {
605 struct o2net_sock_container *sc = sk->sk_user_data;
606 sclog(sc, "data_ready hit\n");
607 o2net_set_data_ready_time(sc);
608 o2net_sc_queue_work(sc, &sc->sc_rx_work);
609 ready = sc->sc_data_ready;
611 ready = sk->sk_data_ready;
613 read_unlock(&sk->sk_callback_lock);
618 /* see o2net_register_callbacks() */
619 static void o2net_state_change(struct sock *sk)
621 void (*state_change)(struct sock *sk);
622 struct o2net_sock_container *sc;
624 read_lock(&sk->sk_callback_lock);
625 sc = sk->sk_user_data;
627 state_change = sk->sk_state_change;
631 sclog(sc, "state_change to %d\n", sk->sk_state);
633 state_change = sc->sc_state_change;
635 switch(sk->sk_state) {
636 /* ignore connecting sockets as they make progress */
640 case TCP_ESTABLISHED:
641 o2net_sc_queue_work(sc, &sc->sc_connect_work);
644 printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT
645 " shutdown, state %d\n",
646 SC_NODEF_ARGS(sc), sk->sk_state);
647 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
651 read_unlock(&sk->sk_callback_lock);
656 * we register callbacks so we can queue work on events before calling
657 * the original callbacks. our callbacks our careful to test user_data
658 * to discover when they've reaced with o2net_unregister_callbacks().
660 static void o2net_register_callbacks(struct sock *sk,
661 struct o2net_sock_container *sc)
663 write_lock_bh(&sk->sk_callback_lock);
665 /* accepted sockets inherit the old listen socket data ready */
666 if (sk->sk_data_ready == o2net_listen_data_ready) {
667 sk->sk_data_ready = sk->sk_user_data;
668 sk->sk_user_data = NULL;
671 BUG_ON(sk->sk_user_data != NULL);
672 sk->sk_user_data = sc;
675 sc->sc_data_ready = sk->sk_data_ready;
676 sc->sc_state_change = sk->sk_state_change;
677 sk->sk_data_ready = o2net_data_ready;
678 sk->sk_state_change = o2net_state_change;
680 mutex_init(&sc->sc_send_lock);
682 write_unlock_bh(&sk->sk_callback_lock);
685 static int o2net_unregister_callbacks(struct sock *sk,
686 struct o2net_sock_container *sc)
690 write_lock_bh(&sk->sk_callback_lock);
691 if (sk->sk_user_data == sc) {
693 sk->sk_user_data = NULL;
694 sk->sk_data_ready = sc->sc_data_ready;
695 sk->sk_state_change = sc->sc_state_change;
697 write_unlock_bh(&sk->sk_callback_lock);
703 * this is a little helper that is called by callers who have seen a problem
704 * with an sc and want to detach it from the nn if someone already hasn't beat
705 * them to it. if an error is given then the shutdown will be persistent
706 * and pending transmits will be canceled.
708 static void o2net_ensure_shutdown(struct o2net_node *nn,
709 struct o2net_sock_container *sc,
712 spin_lock(&nn->nn_lock);
714 o2net_set_nn_state(nn, NULL, 0, err);
715 spin_unlock(&nn->nn_lock);
719 * This work queue function performs the blocking parts of socket shutdown. A
720 * few paths lead here. set_nn_state will trigger this callback if it sees an
721 * sc detached from the nn. state_change will also trigger this callback
722 * directly when it sees errors. In that case we need to call set_nn_state
723 * ourselves as state_change couldn't get the nn_lock and call set_nn_state
726 static void o2net_shutdown_sc(struct work_struct *work)
728 struct o2net_sock_container *sc =
729 container_of(work, struct o2net_sock_container,
731 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
733 sclog(sc, "shutting down\n");
735 /* drop the callbacks ref and call shutdown only once */
736 if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
737 /* we shouldn't flush as we're in the thread, the
738 * races with pending sc work structs are harmless */
739 del_timer_sync(&sc->sc_idle_timeout);
740 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
742 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
745 /* not fatal so failed connects before the other guy has our
746 * heartbeat can be retried */
747 o2net_ensure_shutdown(nn, sc, 0);
751 /* ------------------------------------------------------------ */
753 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
756 int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
759 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
764 static struct o2net_msg_handler *
765 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
766 struct rb_node **ret_parent)
768 struct rb_node **p = &o2net_handler_tree.rb_node;
769 struct rb_node *parent = NULL;
770 struct o2net_msg_handler *nmh, *ret = NULL;
775 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
776 cmp = o2net_handler_cmp(nmh, msg_type, key);
790 if (ret_parent != NULL)
791 *ret_parent = parent;
796 static void o2net_handler_kref_release(struct kref *kref)
798 struct o2net_msg_handler *nmh;
799 nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
804 static void o2net_handler_put(struct o2net_msg_handler *nmh)
806 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
809 /* max_len is protection for the handler func. incoming messages won't
810 * be given to the handler if their payload is longer than the max. */
811 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
812 o2net_msg_handler_func *func, void *data,
813 o2net_post_msg_handler_func *post_func,
814 struct list_head *unreg_list)
816 struct o2net_msg_handler *nmh = NULL;
817 struct rb_node **p, *parent;
820 if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
821 mlog(0, "max_len for message handler out of range: %u\n",
828 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
834 mlog(0, "no message handler provided: %u, %p\n",
840 nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
847 nmh->nh_func_data = data;
848 nmh->nh_post_func = post_func;
849 nmh->nh_msg_type = msg_type;
850 nmh->nh_max_len = max_len;
852 /* the tree and list get this ref.. they're both removed in
853 * unregister when this ref is dropped */
854 kref_init(&nmh->nh_kref);
855 INIT_LIST_HEAD(&nmh->nh_unregister_item);
857 write_lock(&o2net_handler_lock);
858 if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
861 rb_link_node(&nmh->nh_node, parent, p);
862 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
863 list_add_tail(&nmh->nh_unregister_item, unreg_list);
865 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
866 func, msg_type, key);
867 /* we've had some trouble with handlers seemingly vanishing. */
868 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
870 "couldn't find handler we *just* registerd "
871 "for type %u key %08x\n", msg_type, key);
873 write_unlock(&o2net_handler_lock);
883 EXPORT_SYMBOL_GPL(o2net_register_handler);
885 void o2net_unregister_handler_list(struct list_head *list)
887 struct o2net_msg_handler *nmh, *n;
889 write_lock(&o2net_handler_lock);
890 list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
891 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
892 nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
893 rb_erase(&nmh->nh_node, &o2net_handler_tree);
894 list_del_init(&nmh->nh_unregister_item);
895 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
897 write_unlock(&o2net_handler_lock);
899 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
901 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
903 struct o2net_msg_handler *nmh;
905 read_lock(&o2net_handler_lock);
906 nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
908 kref_get(&nmh->nh_kref);
909 read_unlock(&o2net_handler_lock);
914 /* ------------------------------------------------------------ */
916 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
924 struct msghdr msg = {
926 .msg_iov = (struct iovec *)&vec,
927 .msg_flags = MSG_DONTWAIT,
932 ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
938 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
939 size_t veclen, size_t total)
943 struct msghdr msg = {
944 .msg_iov = (struct iovec *)vec,
945 .msg_iovlen = veclen,
955 ret = sock_sendmsg(sock, &msg, total);
958 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
961 ret = -EPIPE; /* should be smarter, I bet */
968 mlog(0, "returning error: %d\n", ret);
972 static void o2net_sendpage(struct o2net_sock_container *sc,
973 void *kmalloced_virt,
976 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
980 mutex_lock(&sc->sc_send_lock);
981 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
982 virt_to_page(kmalloced_virt),
983 (long)kmalloced_virt & ~PAGE_MASK,
985 mutex_unlock(&sc->sc_send_lock);
988 if (ret == (ssize_t)-EAGAIN) {
989 mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
990 " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
994 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
995 " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
996 o2net_ensure_shutdown(nn, sc, 0);
1001 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
1003 memset(msg, 0, sizeof(struct o2net_msg));
1004 msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
1005 msg->data_len = cpu_to_be16(data_len);
1006 msg->msg_type = cpu_to_be16(msg_type);
1007 msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
1009 msg->key = cpu_to_be32(key);
1012 static int o2net_tx_can_proceed(struct o2net_node *nn,
1013 struct o2net_sock_container **sc_ret,
1018 spin_lock(&nn->nn_lock);
1019 if (nn->nn_persistent_error) {
1022 *error = nn->nn_persistent_error;
1023 } else if (nn->nn_sc_valid) {
1024 kref_get(&nn->nn_sc->sc_kref);
1027 *sc_ret = nn->nn_sc;
1030 spin_unlock(&nn->nn_lock);
1035 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1036 size_t caller_veclen, u8 target_node, int *status)
1039 struct o2net_msg *msg = NULL;
1040 size_t veclen, caller_bytes = 0;
1041 struct kvec *vec = NULL;
1042 struct o2net_sock_container *sc = NULL;
1043 struct o2net_node *nn = o2net_nn_from_num(target_node);
1044 struct o2net_status_wait nsw = {
1045 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1047 struct o2net_send_tracking nst;
1049 o2net_init_nst(&nst, msg_type, key, current, target_node);
1051 if (o2net_wq == NULL) {
1052 mlog(0, "attempt to tx without o2netd running\n");
1057 if (caller_veclen == 0) {
1058 mlog(0, "bad kvec array length\n");
1063 caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1064 if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1065 mlog(0, "total payload len %zu too large\n", caller_bytes);
1070 if (target_node == o2nm_this_node()) {
1075 o2net_debug_add_nst(&nst);
1077 o2net_set_nst_sock_time(&nst);
1079 wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1083 o2net_set_nst_sock_container(&nst, sc);
1085 veclen = caller_veclen + 1;
1086 vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1088 mlog(0, "failed to %zu element kvec!\n", veclen);
1093 msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1095 mlog(0, "failed to allocate a o2net_msg!\n");
1100 o2net_init_msg(msg, caller_bytes, msg_type, key);
1102 vec[0].iov_len = sizeof(struct o2net_msg);
1103 vec[0].iov_base = msg;
1104 memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1106 ret = o2net_prep_nsw(nn, &nsw);
1110 msg->msg_num = cpu_to_be32(nsw.ns_id);
1111 o2net_set_nst_msg_id(&nst, nsw.ns_id);
1113 o2net_set_nst_send_time(&nst);
1115 /* finally, convert the message header to network byte-order
1117 mutex_lock(&sc->sc_send_lock);
1118 ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1119 sizeof(struct o2net_msg) + caller_bytes);
1120 mutex_unlock(&sc->sc_send_lock);
1121 msglog(msg, "sending returned %d\n", ret);
1123 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1127 /* wait on other node's handler */
1128 o2net_set_nst_status_time(&nst);
1129 wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1131 o2net_update_send_stats(&nst, sc);
1133 /* Note that we avoid overwriting the callers status return
1134 * variable if a system error was reported on the other
1135 * side. Callers beware. */
1136 ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1138 *status = nsw.ns_status;
1140 mlog(0, "woken, returning system status %d, user status %d\n",
1141 ret, nsw.ns_status);
1143 o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1150 o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1153 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1155 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1156 u8 target_node, int *status)
1162 return o2net_send_message_vec(msg_type, key, &vec, 1,
1163 target_node, status);
1165 EXPORT_SYMBOL_GPL(o2net_send_message);
1167 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1168 enum o2net_system_error syserr, int err)
1172 .iov_len = sizeof(struct o2net_msg),
1175 BUG_ON(syserr >= O2NET_ERR_MAX);
1177 /* leave other fields intact from the incoming message, msg_num
1179 hdr->sys_status = cpu_to_be32(syserr);
1180 hdr->status = cpu_to_be32(err);
1181 hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC); // twiddle the magic
1184 msglog(hdr, "about to send status magic %d\n", err);
1185 /* hdr has been in host byteorder this whole time */
1186 return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1189 /* this returns -errno if the header was unknown or too large, etc.
1190 * after this is called the buffer us reused for the next message */
1191 static int o2net_process_message(struct o2net_sock_container *sc,
1192 struct o2net_msg *hdr)
1194 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1195 int ret = 0, handler_status;
1196 enum o2net_system_error syserr;
1197 struct o2net_msg_handler *nmh = NULL;
1198 void *ret_data = NULL;
1200 msglog(hdr, "processing message\n");
1202 o2net_sc_postpone_idle(sc);
1204 switch(be16_to_cpu(hdr->magic)) {
1205 case O2NET_MSG_STATUS_MAGIC:
1206 /* special type for returning message status */
1207 o2net_complete_nsw(nn, NULL,
1208 be32_to_cpu(hdr->msg_num),
1209 be32_to_cpu(hdr->sys_status),
1210 be32_to_cpu(hdr->status));
1212 case O2NET_MSG_KEEP_REQ_MAGIC:
1213 o2net_sendpage(sc, o2net_keep_resp,
1214 sizeof(*o2net_keep_resp));
1216 case O2NET_MSG_KEEP_RESP_MAGIC:
1218 case O2NET_MSG_MAGIC:
1221 msglog(hdr, "bad magic\n");
1227 /* find a handler for it */
1229 nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1230 be32_to_cpu(hdr->key));
1232 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1233 be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1234 syserr = O2NET_ERR_NO_HNDLR;
1238 syserr = O2NET_ERR_NONE;
1240 if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1241 syserr = O2NET_ERR_OVERFLOW;
1243 if (syserr != O2NET_ERR_NONE)
1246 o2net_set_func_start_time(sc);
1247 sc->sc_msg_key = be32_to_cpu(hdr->key);
1248 sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1249 handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1250 be16_to_cpu(hdr->data_len),
1251 nmh->nh_func_data, &ret_data);
1252 o2net_set_func_stop_time(sc);
1254 o2net_update_recv_stats(sc);
1257 /* this destroys the hdr, so don't use it after this */
1258 mutex_lock(&sc->sc_send_lock);
1259 ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1261 mutex_unlock(&sc->sc_send_lock);
1263 mlog(0, "sending handler status %d, syserr %d returned %d\n",
1264 handler_status, syserr, ret);
1267 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1268 if (nmh->nh_post_func)
1269 (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1275 o2net_handler_put(nmh);
1279 static int o2net_check_handshake(struct o2net_sock_container *sc)
1281 struct o2net_handshake *hand = page_address(sc->sc_page);
1282 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1284 if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1285 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1286 "version %llu but %llu is required, disconnecting\n",
1288 (unsigned long long)be64_to_cpu(hand->protocol_version),
1289 O2NET_PROTOCOL_VERSION);
1291 /* don't bother reconnecting if its the wrong version. */
1292 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1297 * Ensure timeouts are consistent with other nodes, otherwise
1298 * we can end up with one node thinking that the other must be down,
1299 * but isn't. This can ultimately cause corruption.
1301 if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1302 o2net_idle_timeout()) {
1303 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1304 "%u ms, but we use %u ms locally. disconnecting\n",
1306 be32_to_cpu(hand->o2net_idle_timeout_ms),
1307 o2net_idle_timeout());
1308 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1312 if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1313 o2net_keepalive_delay()) {
1314 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1315 "%u ms, but we use %u ms locally. disconnecting\n",
1317 be32_to_cpu(hand->o2net_keepalive_delay_ms),
1318 o2net_keepalive_delay());
1319 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1323 if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1324 O2HB_MAX_WRITE_TIMEOUT_MS) {
1325 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1326 "%u ms, but we use %u ms locally. disconnecting\n",
1328 be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1329 O2HB_MAX_WRITE_TIMEOUT_MS);
1330 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1334 sc->sc_handshake_ok = 1;
1336 spin_lock(&nn->nn_lock);
1337 /* set valid and queue the idle timers only if it hasn't been
1338 * shut down already */
1339 if (nn->nn_sc == sc) {
1340 o2net_sc_reset_idle_timer(sc);
1341 atomic_set(&nn->nn_timeout, 0);
1342 o2net_set_nn_state(nn, sc, 1, 0);
1344 spin_unlock(&nn->nn_lock);
1346 /* shift everything up as though it wasn't there */
1347 sc->sc_page_off -= sizeof(struct o2net_handshake);
1348 if (sc->sc_page_off)
1349 memmove(hand, hand + 1, sc->sc_page_off);
1354 /* this demuxes the queued rx bytes into header or payload bits and calls
1355 * handlers as each full message is read off the socket. it returns -error,
1356 * == 0 eof, or > 0 for progress made.*/
1357 static int o2net_advance_rx(struct o2net_sock_container *sc)
1359 struct o2net_msg *hdr;
1364 sclog(sc, "receiving\n");
1365 o2net_set_advance_start_time(sc);
1367 if (unlikely(sc->sc_handshake_ok == 0)) {
1368 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1369 data = page_address(sc->sc_page) + sc->sc_page_off;
1370 datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1371 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1373 sc->sc_page_off += ret;
1376 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1377 o2net_check_handshake(sc);
1378 if (unlikely(sc->sc_handshake_ok == 0))
1384 /* do we need more header? */
1385 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1386 data = page_address(sc->sc_page) + sc->sc_page_off;
1387 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1388 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1390 sc->sc_page_off += ret;
1391 /* only swab incoming here.. we can
1392 * only get here once as we cross from
1393 * being under to over */
1394 if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1395 hdr = page_address(sc->sc_page);
1396 if (be16_to_cpu(hdr->data_len) >
1397 O2NET_MAX_PAYLOAD_BYTES)
1405 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1406 /* oof, still don't have a header */
1410 /* this was swabbed above when we first read it */
1411 hdr = page_address(sc->sc_page);
1413 msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1415 /* do we need more payload? */
1416 if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1417 /* need more payload */
1418 data = page_address(sc->sc_page) + sc->sc_page_off;
1419 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1421 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1423 sc->sc_page_off += ret;
1428 if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1429 /* we can only get here once, the first time we read
1430 * the payload.. so set ret to progress if the handler
1431 * works out. after calling this the message is toast */
1432 ret = o2net_process_message(sc, hdr);
1435 sc->sc_page_off = 0;
1439 sclog(sc, "ret = %d\n", ret);
1440 o2net_set_advance_stop_time(sc);
1444 /* this work func is triggerd by data ready. it reads until it can read no
1445 * more. it interprets 0, eof, as fatal. if data_ready hits while we're doing
1446 * our work the work struct will be marked and we'll be called again. */
1447 static void o2net_rx_until_empty(struct work_struct *work)
1449 struct o2net_sock_container *sc =
1450 container_of(work, struct o2net_sock_container, sc_rx_work);
1454 ret = o2net_advance_rx(sc);
1457 if (ret <= 0 && ret != -EAGAIN) {
1458 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1459 sclog(sc, "saw error %d, closing\n", ret);
1460 /* not permanent so read failed handshake can retry */
1461 o2net_ensure_shutdown(nn, sc, 0);
1467 static int o2net_set_nodelay(struct socket *sock)
1476 * Dear unsuspecting programmer,
1478 * Don't use sock_setsockopt() for SOL_TCP. It doesn't check its level
1479 * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1480 * silently turn into SO_DEBUG.
1483 * Keeper of hilariously fragile interfaces.
1485 ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1486 (char __user *)&val, sizeof(val));
1492 static void o2net_initialize_handshake(void)
1494 o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1495 O2HB_MAX_WRITE_TIMEOUT_MS);
1496 o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1497 o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1498 o2net_keepalive_delay());
1499 o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1500 o2net_reconnect_delay());
1503 /* ------------------------------------------------------------ */
1505 /* called when a connect completes and after a sock is accepted. the
1506 * rx path will see the response and mark the sc valid */
1507 static void o2net_sc_connect_completed(struct work_struct *work)
1509 struct o2net_sock_container *sc =
1510 container_of(work, struct o2net_sock_container,
1513 mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1514 (unsigned long long)O2NET_PROTOCOL_VERSION,
1515 (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1517 o2net_initialize_handshake();
1518 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1522 /* this is called as a work_struct func. */
1523 static void o2net_sc_send_keep_req(struct work_struct *work)
1525 struct o2net_sock_container *sc =
1526 container_of(work, struct o2net_sock_container,
1527 sc_keepalive_work.work);
1529 o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1533 /* socket shutdown does a del_timer_sync against this as it tears down.
1534 * we can't start this timer until we've got to the point in sc buildup
1535 * where shutdown is going to be involved */
1536 static void o2net_idle_timer(unsigned long data)
1538 struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1539 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1541 #ifdef CONFIG_DEBUG_FS
1542 ktime_t now = ktime_get();
1545 printk(KERN_NOTICE "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1546 "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1547 o2net_idle_timeout() / 1000,
1548 o2net_idle_timeout() % 1000);
1550 #ifdef CONFIG_DEBUG_FS
1551 mlog(ML_NOTICE, "Here are some times that might help debug the "
1552 "situation: (Timer: %lld, Now %lld, DataReady %lld, Advance %lld-%lld, "
1553 "Key 0x%08x, Func %u, FuncTime %lld-%lld)\n",
1554 (long long)ktime_to_us(sc->sc_tv_timer), (long long)ktime_to_us(now),
1555 (long long)ktime_to_us(sc->sc_tv_data_ready),
1556 (long long)ktime_to_us(sc->sc_tv_advance_start),
1557 (long long)ktime_to_us(sc->sc_tv_advance_stop),
1558 sc->sc_msg_key, sc->sc_msg_type,
1559 (long long)ktime_to_us(sc->sc_tv_func_start),
1560 (long long)ktime_to_us(sc->sc_tv_func_stop));
1564 * Initialize the nn_timeout so that the next connection attempt
1565 * will continue in o2net_start_connect.
1567 atomic_set(&nn->nn_timeout, 1);
1569 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1572 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1574 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1575 o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1576 msecs_to_jiffies(o2net_keepalive_delay()));
1577 o2net_set_sock_timer(sc);
1578 mod_timer(&sc->sc_idle_timeout,
1579 jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1582 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1584 /* Only push out an existing timer */
1585 if (timer_pending(&sc->sc_idle_timeout))
1586 o2net_sc_reset_idle_timer(sc);
1589 /* this work func is kicked whenever a path sets the nn state which doesn't
1590 * have valid set. This includes seeing hb come up, losing a connection,
1591 * having a connect attempt fail, etc. This centralizes the logic which decides
1592 * if a connect attempt should be made or if we should give up and all future
1593 * transmit attempts should fail */
1594 static void o2net_start_connect(struct work_struct *work)
1596 struct o2net_node *nn =
1597 container_of(work, struct o2net_node, nn_connect_work.work);
1598 struct o2net_sock_container *sc = NULL;
1599 struct o2nm_node *node = NULL, *mynode = NULL;
1600 struct socket *sock = NULL;
1601 struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1603 unsigned int timeout;
1605 /* if we're greater we initiate tx, otherwise we accept */
1606 if (o2nm_this_node() <= o2net_num_from_nn(nn))
1609 /* watch for racing with tearing a node down */
1610 node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1616 mynode = o2nm_get_node_by_num(o2nm_this_node());
1617 if (mynode == NULL) {
1622 spin_lock(&nn->nn_lock);
1624 * see if we already have one pending or have given up.
1625 * For nn_timeout, it is set when we close the connection
1626 * because of the idle time out. So it means that we have
1627 * at least connected to that node successfully once,
1628 * now try to connect to it again.
1630 timeout = atomic_read(&nn->nn_timeout);
1631 stop = (nn->nn_sc ||
1632 (nn->nn_persistent_error &&
1633 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1634 spin_unlock(&nn->nn_lock);
1638 nn->nn_last_connect_attempt = jiffies;
1640 sc = sc_alloc(node);
1642 mlog(0, "couldn't allocate sc\n");
1647 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1649 mlog(0, "can't create socket: %d\n", ret);
1652 sc->sc_sock = sock; /* freed by sc_kref_release */
1654 sock->sk->sk_allocation = GFP_ATOMIC;
1656 myaddr.sin_family = AF_INET;
1657 myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1658 myaddr.sin_port = htons(0); /* any port */
1660 ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1663 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1664 ret, &mynode->nd_ipv4_address);
1668 ret = o2net_set_nodelay(sc->sc_sock);
1670 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1674 o2net_register_callbacks(sc->sc_sock->sk, sc);
1676 spin_lock(&nn->nn_lock);
1677 /* handshake completion will set nn->nn_sc_valid */
1678 o2net_set_nn_state(nn, sc, 0, 0);
1679 spin_unlock(&nn->nn_lock);
1681 remoteaddr.sin_family = AF_INET;
1682 remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1683 remoteaddr.sin_port = node->nd_ipv4_port;
1685 ret = sc->sc_sock->ops->connect(sc->sc_sock,
1686 (struct sockaddr *)&remoteaddr,
1689 if (ret == -EINPROGRESS)
1694 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1695 "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1696 /* 0 err so that another will be queued and attempted
1697 * from set_nn_state */
1699 o2net_ensure_shutdown(nn, sc, 0);
1704 o2nm_node_put(node);
1706 o2nm_node_put(mynode);
1711 static void o2net_connect_expired(struct work_struct *work)
1713 struct o2net_node *nn =
1714 container_of(work, struct o2net_node, nn_connect_expired.work);
1716 spin_lock(&nn->nn_lock);
1717 if (!nn->nn_sc_valid) {
1718 mlog(ML_ERROR, "no connection established with node %u after "
1719 "%u.%u seconds, giving up and returning errors.\n",
1720 o2net_num_from_nn(nn),
1721 o2net_idle_timeout() / 1000,
1722 o2net_idle_timeout() % 1000);
1724 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1726 spin_unlock(&nn->nn_lock);
1729 static void o2net_still_up(struct work_struct *work)
1731 struct o2net_node *nn =
1732 container_of(work, struct o2net_node, nn_still_up.work);
1734 o2quo_hb_still_up(o2net_num_from_nn(nn));
1737 /* ------------------------------------------------------------ */
1739 void o2net_disconnect_node(struct o2nm_node *node)
1741 struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1743 /* don't reconnect until it's heartbeating again */
1744 spin_lock(&nn->nn_lock);
1745 atomic_set(&nn->nn_timeout, 0);
1746 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1747 spin_unlock(&nn->nn_lock);
1750 cancel_delayed_work(&nn->nn_connect_expired);
1751 cancel_delayed_work(&nn->nn_connect_work);
1752 cancel_delayed_work(&nn->nn_still_up);
1753 flush_workqueue(o2net_wq);
1757 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1760 o2quo_hb_down(node_num);
1765 if (node_num != o2nm_this_node())
1766 o2net_disconnect_node(node);
1768 BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1771 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1774 struct o2net_node *nn = o2net_nn_from_num(node_num);
1776 o2quo_hb_up(node_num);
1780 /* ensure an immediate connect attempt */
1781 nn->nn_last_connect_attempt = jiffies -
1782 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1784 if (node_num != o2nm_this_node()) {
1785 /* believe it or not, accept and node hearbeating testing
1786 * can succeed for this node before we got here.. so
1787 * only use set_nn_state to clear the persistent error
1788 * if that hasn't already happened */
1789 spin_lock(&nn->nn_lock);
1790 atomic_set(&nn->nn_timeout, 0);
1791 if (nn->nn_persistent_error)
1792 o2net_set_nn_state(nn, NULL, 0, 0);
1793 spin_unlock(&nn->nn_lock);
1797 void o2net_unregister_hb_callbacks(void)
1799 o2hb_unregister_callback(NULL, &o2net_hb_up);
1800 o2hb_unregister_callback(NULL, &o2net_hb_down);
1803 int o2net_register_hb_callbacks(void)
1807 o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1808 o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1809 o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1810 o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1812 ret = o2hb_register_callback(NULL, &o2net_hb_up);
1814 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1817 o2net_unregister_hb_callbacks();
1822 /* ------------------------------------------------------------ */
1824 static int o2net_accept_one(struct socket *sock)
1827 struct sockaddr_in sin;
1828 struct socket *new_sock = NULL;
1829 struct o2nm_node *node = NULL;
1830 struct o2nm_node *local_node = NULL;
1831 struct o2net_sock_container *sc = NULL;
1832 struct o2net_node *nn;
1834 BUG_ON(sock == NULL);
1835 ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1836 sock->sk->sk_protocol, &new_sock);
1840 new_sock->type = sock->type;
1841 new_sock->ops = sock->ops;
1842 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1846 new_sock->sk->sk_allocation = GFP_ATOMIC;
1848 ret = o2net_set_nodelay(new_sock);
1850 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1855 ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1860 node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1862 mlog(ML_NOTICE, "attempt to connect from unknown node at %pI4:%d\n",
1863 &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1868 if (o2nm_this_node() >= node->nd_num) {
1869 local_node = o2nm_get_node_by_num(o2nm_this_node());
1870 mlog(ML_NOTICE, "unexpected connect attempt seen at node '%s' ("
1871 "%u, %pI4:%d) from node '%s' (%u, %pI4:%d)\n",
1872 local_node->nd_name, local_node->nd_num,
1873 &(local_node->nd_ipv4_address),
1874 ntohs(local_node->nd_ipv4_port),
1875 node->nd_name, node->nd_num, &sin.sin_addr.s_addr,
1876 ntohs(sin.sin_port));
1881 /* this happens all the time when the other node sees our heartbeat
1882 * and tries to connect before we see their heartbeat */
1883 if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1884 mlog(ML_CONN, "attempt to connect from node '%s' at "
1885 "%pI4:%d but it isn't heartbeating\n",
1886 node->nd_name, &sin.sin_addr.s_addr,
1887 ntohs(sin.sin_port));
1892 nn = o2net_nn_from_num(node->nd_num);
1894 spin_lock(&nn->nn_lock);
1899 spin_unlock(&nn->nn_lock);
1901 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1902 "%pI4:%d but it already has an open connection\n",
1903 node->nd_name, &sin.sin_addr.s_addr,
1904 ntohs(sin.sin_port));
1908 sc = sc_alloc(node);
1914 sc->sc_sock = new_sock;
1917 spin_lock(&nn->nn_lock);
1918 atomic_set(&nn->nn_timeout, 0);
1919 o2net_set_nn_state(nn, sc, 0, 0);
1920 spin_unlock(&nn->nn_lock);
1922 o2net_register_callbacks(sc->sc_sock->sk, sc);
1923 o2net_sc_queue_work(sc, &sc->sc_rx_work);
1925 o2net_initialize_handshake();
1926 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1930 sock_release(new_sock);
1932 o2nm_node_put(node);
1934 o2nm_node_put(local_node);
1940 static void o2net_accept_many(struct work_struct *work)
1942 struct socket *sock = o2net_listen_sock;
1943 while (o2net_accept_one(sock) == 0)
1947 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1949 void (*ready)(struct sock *sk, int bytes);
1951 read_lock(&sk->sk_callback_lock);
1952 ready = sk->sk_user_data;
1953 if (ready == NULL) { /* check for teardown race */
1954 ready = sk->sk_data_ready;
1958 /* ->sk_data_ready is also called for a newly established child socket
1959 * before it has been accepted and the acceptor has set up their
1960 * data_ready.. we only want to queue listen work for our listening
1962 if (sk->sk_state == TCP_LISTEN) {
1963 mlog(ML_TCP, "bytes: %d\n", bytes);
1964 queue_work(o2net_wq, &o2net_listen_work);
1968 read_unlock(&sk->sk_callback_lock);
1972 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1974 struct socket *sock = NULL;
1976 struct sockaddr_in sin = {
1977 .sin_family = PF_INET,
1978 .sin_addr = { .s_addr = addr },
1982 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1984 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1988 sock->sk->sk_allocation = GFP_ATOMIC;
1990 write_lock_bh(&sock->sk->sk_callback_lock);
1991 sock->sk->sk_user_data = sock->sk->sk_data_ready;
1992 sock->sk->sk_data_ready = o2net_listen_data_ready;
1993 write_unlock_bh(&sock->sk->sk_callback_lock);
1995 o2net_listen_sock = sock;
1996 INIT_WORK(&o2net_listen_work, o2net_accept_many);
1998 sock->sk->sk_reuse = 1;
1999 ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2001 mlog(ML_ERROR, "unable to bind socket at %pI4:%u, "
2002 "ret=%d\n", &addr, ntohs(port), ret);
2006 ret = sock->ops->listen(sock, 64);
2008 mlog(ML_ERROR, "unable to listen on %pI4:%u, ret=%d\n",
2009 &addr, ntohs(port), ret);
2014 o2net_listen_sock = NULL;
2022 * called from node manager when we should bring up our network listening
2023 * socket. node manager handles all the serialization to only call this
2024 * once and to match it with o2net_stop_listening(). note,
2025 * o2nm_this_node() doesn't work yet as we're being called while it
2028 int o2net_start_listening(struct o2nm_node *node)
2032 BUG_ON(o2net_wq != NULL);
2033 BUG_ON(o2net_listen_sock != NULL);
2035 mlog(ML_KTHREAD, "starting o2net thread...\n");
2036 o2net_wq = create_singlethread_workqueue("o2net");
2037 if (o2net_wq == NULL) {
2038 mlog(ML_ERROR, "unable to launch o2net thread\n");
2039 return -ENOMEM; /* ? */
2042 ret = o2net_open_listening_sock(node->nd_ipv4_address,
2043 node->nd_ipv4_port);
2045 destroy_workqueue(o2net_wq);
2048 o2quo_conn_up(node->nd_num);
2053 /* again, o2nm_this_node() doesn't work here as we're involved in
2054 * tearing it down */
2055 void o2net_stop_listening(struct o2nm_node *node)
2057 struct socket *sock = o2net_listen_sock;
2060 BUG_ON(o2net_wq == NULL);
2061 BUG_ON(o2net_listen_sock == NULL);
2063 /* stop the listening socket from generating work */
2064 write_lock_bh(&sock->sk->sk_callback_lock);
2065 sock->sk->sk_data_ready = sock->sk->sk_user_data;
2066 sock->sk->sk_user_data = NULL;
2067 write_unlock_bh(&sock->sk->sk_callback_lock);
2069 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2070 struct o2nm_node *node = o2nm_get_node_by_num(i);
2072 o2net_disconnect_node(node);
2073 o2nm_node_put(node);
2077 /* finish all work and tear down the work queue */
2078 mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2079 destroy_workqueue(o2net_wq);
2082 sock_release(o2net_listen_sock);
2083 o2net_listen_sock = NULL;
2085 o2quo_conn_err(node->nd_num);
2088 /* ------------------------------------------------------------ */
2090 int o2net_init(void)
2096 if (o2net_debugfs_init())
2099 o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2100 o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2101 o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2102 if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2104 kfree(o2net_keep_req);
2105 kfree(o2net_keep_resp);
2109 o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2110 o2net_hand->connector_id = cpu_to_be64(1);
2112 o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2113 o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2115 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2116 struct o2net_node *nn = o2net_nn_from_num(i);
2118 atomic_set(&nn->nn_timeout, 0);
2119 spin_lock_init(&nn->nn_lock);
2120 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2121 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2122 o2net_connect_expired);
2123 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2124 /* until we see hb from a node we'll return einval */
2125 nn->nn_persistent_error = -ENOTCONN;
2126 init_waitqueue_head(&nn->nn_sc_wq);
2127 idr_init(&nn->nn_status_idr);
2128 INIT_LIST_HEAD(&nn->nn_status_list);
2134 void o2net_exit(void)
2138 kfree(o2net_keep_req);
2139 kfree(o2net_keep_resp);
2140 o2net_debugfs_exit();