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/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>
66 #include <linux/uaccess.h>
68 #include "heartbeat.h"
70 #include "nodemanager.h"
71 #define MLOG_MASK_PREFIX ML_TCP
75 #include "tcp_internal.h"
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)
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.
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); \
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 , \
106 static DEFINE_RWLOCK(o2net_handler_lock);
107 static struct rb_root o2net_handler_tree = RB_ROOT;
109 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
111 /* XXX someday we'll need better accounting */
112 static struct socket *o2net_listen_sock;
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.
122 static struct workqueue_struct *o2net_wq;
123 static struct work_struct o2net_listen_work;
125 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
126 #define O2NET_HB_PRI 0x1
128 static struct o2net_handshake *o2net_hand;
129 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
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,};
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);
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)
151 INIT_LIST_HEAD(&nst->st_net_debug_item);
153 nst->st_msg_type = msgtype;
154 nst->st_msg_key = msgkey;
158 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
160 nst->st_sock_time = ktime_get();
163 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
165 nst->st_send_time = ktime_get();
168 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
170 nst->st_status_time = ktime_get();
173 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
174 struct o2net_sock_container *sc)
179 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
185 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
187 sc->sc_tv_timer = ktime_get();
190 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
192 sc->sc_tv_data_ready = ktime_get();
195 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
197 sc->sc_tv_advance_start = ktime_get();
200 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
202 sc->sc_tv_advance_stop = ktime_get();
205 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
207 sc->sc_tv_func_start = ktime_get();
210 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
212 sc->sc_tv_func_stop = ktime_get();
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 */
230 #ifdef CONFIG_OCFS2_FS_STATS
231 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
233 return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
236 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
237 struct o2net_sock_container *sc)
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,
245 sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
246 ktime_sub(nst->st_send_time,
251 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
253 sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
254 o2net_get_func_run_time(sc));
260 # define o2net_update_send_stats(a, b)
262 # define o2net_update_recv_stats(sc)
264 #endif /* CONFIG_OCFS2_FS_STATS */
266 static inline unsigned int o2net_reconnect_delay(void)
268 return o2nm_single_cluster->cl_reconnect_delay_ms;
271 static inline unsigned int o2net_keepalive_delay(void)
273 return o2nm_single_cluster->cl_keepalive_delay_ms;
276 static inline unsigned int o2net_idle_timeout(void)
278 return o2nm_single_cluster->cl_idle_timeout_ms;
281 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
284 BUG_ON(err >= O2NET_ERR_MAX);
285 trans = o2net_sys_err_translations[err];
287 /* Just in case we mess up the translation table above */
288 BUG_ON(err != O2NET_ERR_NONE && trans == 0);
292 static struct o2net_node * o2net_nn_from_num(u8 node_num)
294 BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
295 return &o2net_nodes[node_num];
298 static u8 o2net_num_from_nn(struct o2net_node *nn)
301 return nn - o2net_nodes;
304 /* ------------------------------------------------------------ */
306 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
310 spin_lock(&nn->nn_lock);
311 ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
314 list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
316 spin_unlock(&nn->nn_lock);
320 init_waitqueue_head(&nsw->ns_wq);
321 nsw->ns_sys_status = O2NET_ERR_NONE;
326 static void o2net_complete_nsw_locked(struct o2net_node *nn,
327 struct o2net_status_wait *nsw,
328 enum o2net_system_error sys_status,
331 assert_spin_locked(&nn->nn_lock);
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);
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,
347 spin_lock(&nn->nn_lock);
352 nsw = idr_find(&nn->nn_status_idr, id);
357 o2net_complete_nsw_locked(nn, nsw, sys_status, status);
360 spin_unlock(&nn->nn_lock);
364 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
366 struct o2net_status_wait *nsw, *tmp;
367 unsigned int num_kills = 0;
369 assert_spin_locked(&nn->nn_lock);
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);
376 mlog(0, "completed %d messages for node %u\n", num_kills,
377 o2net_num_from_nn(nn));
380 static int o2net_nsw_completed(struct o2net_node *nn,
381 struct o2net_status_wait *nsw)
384 spin_lock(&nn->nn_lock);
385 completed = list_empty(&nsw->ns_node_item);
386 spin_unlock(&nn->nn_lock);
390 /* ------------------------------------------------------------ */
392 static void sc_kref_release(struct kref *kref)
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));
398 sclog(sc, "releasing\n");
401 sock_release(sc->sc_sock);
405 o2nm_undepend_item(&sc->sc_node->nd_item);
406 o2nm_node_put(sc->sc_node);
409 o2net_debug_del_sc(sc);
412 __free_page(sc->sc_page);
416 static void sc_put(struct o2net_sock_container *sc)
419 kref_put(&sc->sc_kref, sc_kref_release);
421 static void sc_get(struct o2net_sock_container *sc)
424 kref_get(&sc->sc_kref);
426 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
428 struct o2net_sock_container *sc, *ret = NULL;
429 struct page *page = NULL;
432 page = alloc_page(GFP_NOFS);
433 sc = kzalloc(sizeof(*sc), GFP_NOFS);
434 if (sc == NULL || page == NULL)
437 kref_init(&sc->sc_kref);
441 /* pin the node item of the remote node */
442 status = o2nm_depend_item(&node->nd_item);
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);
453 timer_setup(&sc->sc_idle_timeout, o2net_idle_timer, 0);
455 sclog(sc, "alloced\n");
459 o2net_debug_add_sc(sc);
471 /* ------------------------------------------------------------ */
473 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
474 struct work_struct *work)
477 if (!queue_work(o2net_wq, work))
480 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
481 struct delayed_work *work,
485 if (!queue_delayed_work(o2net_wq, work, delay))
488 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
489 struct delayed_work *work)
491 if (cancel_delayed_work(work))
495 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
497 int o2net_num_connected_peers(void)
499 return atomic_read(&o2net_connected_peers);
502 static void o2net_set_nn_state(struct o2net_node *nn,
503 struct o2net_sock_container *sc,
504 unsigned valid, int err)
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;
510 assert_spin_locked(&nn->nn_lock);
513 atomic_dec(&o2net_connected_peers);
514 else if (!old_sc && sc)
515 atomic_inc(&o2net_connected_peers);
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);
523 if (was_valid && !valid && err == 0)
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);
531 nn->nn_sc_valid = valid ? 1 : 0;
532 nn->nn_persistent_error = err;
534 /* mirrors o2net_tx_can_proceed() */
535 if (nn->nn_persistent_error || nn->nn_sc_valid)
536 wake_up(&nn->nn_sc_wq);
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));
544 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 within 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)
601 void (*ready)(struct sock *sk);
602 struct o2net_sock_container *sc;
604 read_lock_bh(&sk->sk_callback_lock);
605 sc = sk->sk_user_data;
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;
612 ready = sk->sk_data_ready;
614 read_unlock_bh(&sk->sk_callback_lock);
619 /* see o2net_register_callbacks() */
620 static void o2net_state_change(struct sock *sk)
622 void (*state_change)(struct sock *sk);
623 struct o2net_sock_container *sc;
625 read_lock_bh(&sk->sk_callback_lock);
626 sc = sk->sk_user_data;
628 state_change = sk->sk_state_change;
632 sclog(sc, "state_change to %d\n", sk->sk_state);
634 state_change = sc->sc_state_change;
636 switch(sk->sk_state) {
637 /* ignore connecting sockets as they make progress */
641 case TCP_ESTABLISHED:
642 o2net_sc_queue_work(sc, &sc->sc_connect_work);
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);
652 read_unlock_bh(&sk->sk_callback_lock);
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().
661 static void o2net_register_callbacks(struct sock *sk,
662 struct o2net_sock_container *sc)
664 write_lock_bh(&sk->sk_callback_lock);
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;
672 BUG_ON(sk->sk_user_data != NULL);
673 sk->sk_user_data = sc;
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;
681 mutex_init(&sc->sc_send_lock);
683 write_unlock_bh(&sk->sk_callback_lock);
686 static int o2net_unregister_callbacks(struct sock *sk,
687 struct o2net_sock_container *sc)
691 write_lock_bh(&sk->sk_callback_lock);
692 if (sk->sk_user_data == sc) {
694 sk->sk_user_data = NULL;
695 sk->sk_data_ready = sc->sc_data_ready;
696 sk->sk_state_change = sc->sc_state_change;
698 write_unlock_bh(&sk->sk_callback_lock);
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.
709 static void o2net_ensure_shutdown(struct o2net_node *nn,
710 struct o2net_sock_container *sc,
713 spin_lock(&nn->nn_lock);
715 o2net_set_nn_state(nn, NULL, 0, err);
716 spin_unlock(&nn->nn_lock);
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
727 static void o2net_shutdown_sc(struct work_struct *work)
729 struct o2net_sock_container *sc =
730 container_of(work, struct o2net_sock_container,
732 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
734 sclog(sc, "shutting down\n");
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);
743 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
746 /* not fatal so failed connects before the other guy has our
747 * heartbeat can be retried */
748 o2net_ensure_shutdown(nn, sc, 0);
752 /* ------------------------------------------------------------ */
754 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
757 int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
760 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
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)
769 struct rb_node **p = &o2net_handler_tree.rb_node;
770 struct rb_node *parent = NULL;
771 struct o2net_msg_handler *nmh, *ret = NULL;
776 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
777 cmp = o2net_handler_cmp(nmh, msg_type, key);
791 if (ret_parent != NULL)
792 *ret_parent = parent;
797 static void o2net_handler_kref_release(struct kref *kref)
799 struct o2net_msg_handler *nmh;
800 nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
805 static void o2net_handler_put(struct o2net_msg_handler *nmh)
807 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
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)
817 struct o2net_msg_handler *nmh = NULL;
818 struct rb_node **p, *parent;
821 if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
822 mlog(0, "max_len for message handler out of range: %u\n",
829 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
835 mlog(0, "no message handler provided: %u, %p\n",
841 nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
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;
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);
858 write_lock(&o2net_handler_lock);
859 if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
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);
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,
871 "couldn't find handler we *just* registered "
872 "for type %u key %08x\n", msg_type, key);
874 write_unlock(&o2net_handler_lock);
882 EXPORT_SYMBOL_GPL(o2net_register_handler);
884 void o2net_unregister_handler_list(struct list_head *list)
886 struct o2net_msg_handler *nmh, *n;
888 write_lock(&o2net_handler_lock);
889 list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
890 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
891 nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
892 rb_erase(&nmh->nh_node, &o2net_handler_tree);
893 list_del_init(&nmh->nh_unregister_item);
894 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
896 write_unlock(&o2net_handler_lock);
898 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
900 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
902 struct o2net_msg_handler *nmh;
904 read_lock(&o2net_handler_lock);
905 nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
907 kref_get(&nmh->nh_kref);
908 read_unlock(&o2net_handler_lock);
913 /* ------------------------------------------------------------ */
915 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
917 struct kvec vec = { .iov_len = len, .iov_base = data, };
918 struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
919 iov_iter_kvec(&msg.msg_iter, READ | ITER_KVEC, &vec, 1, len);
920 return sock_recvmsg(sock, &msg, MSG_DONTWAIT);
923 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
924 size_t veclen, size_t total)
927 struct msghdr msg = {.msg_flags = 0,};
934 ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
935 if (likely(ret == total))
937 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
939 ret = -EPIPE; /* should be smarter, I bet */
941 mlog(0, "returning error: %d\n", ret);
945 static void o2net_sendpage(struct o2net_sock_container *sc,
946 void *kmalloced_virt,
949 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
953 mutex_lock(&sc->sc_send_lock);
954 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
955 virt_to_page(kmalloced_virt),
956 offset_in_page(kmalloced_virt),
958 mutex_unlock(&sc->sc_send_lock);
961 if (ret == (ssize_t)-EAGAIN) {
962 mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
963 " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
967 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
968 " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
969 o2net_ensure_shutdown(nn, sc, 0);
974 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
976 memset(msg, 0, sizeof(struct o2net_msg));
977 msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
978 msg->data_len = cpu_to_be16(data_len);
979 msg->msg_type = cpu_to_be16(msg_type);
980 msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
982 msg->key = cpu_to_be32(key);
985 static int o2net_tx_can_proceed(struct o2net_node *nn,
986 struct o2net_sock_container **sc_ret,
991 spin_lock(&nn->nn_lock);
992 if (nn->nn_persistent_error) {
995 *error = nn->nn_persistent_error;
996 } else if (nn->nn_sc_valid) {
997 kref_get(&nn->nn_sc->sc_kref);
1000 *sc_ret = nn->nn_sc;
1003 spin_unlock(&nn->nn_lock);
1008 /* Get a map of all nodes to which this node is currently connected to */
1009 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1011 struct o2net_sock_container *sc;
1014 BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1016 memset(map, 0, bytes);
1017 for (node = 0; node < O2NM_MAX_NODES; ++node) {
1018 if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1026 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1028 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1029 size_t caller_veclen, u8 target_node, int *status)
1032 struct o2net_msg *msg = NULL;
1033 size_t veclen, caller_bytes = 0;
1034 struct kvec *vec = NULL;
1035 struct o2net_sock_container *sc = NULL;
1036 struct o2net_node *nn = o2net_nn_from_num(target_node);
1037 struct o2net_status_wait nsw = {
1038 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1040 struct o2net_send_tracking nst;
1042 o2net_init_nst(&nst, msg_type, key, current, target_node);
1044 if (o2net_wq == NULL) {
1045 mlog(0, "attempt to tx without o2netd running\n");
1050 if (caller_veclen == 0) {
1051 mlog(0, "bad kvec array length\n");
1056 caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1057 if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1058 mlog(0, "total payload len %zu too large\n", caller_bytes);
1063 if (target_node == o2nm_this_node()) {
1068 o2net_debug_add_nst(&nst);
1070 o2net_set_nst_sock_time(&nst);
1072 wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1076 o2net_set_nst_sock_container(&nst, sc);
1078 veclen = caller_veclen + 1;
1079 vec = kmalloc_array(veclen, sizeof(struct kvec), GFP_ATOMIC);
1081 mlog(0, "failed to %zu element kvec!\n", veclen);
1086 msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1088 mlog(0, "failed to allocate a o2net_msg!\n");
1093 o2net_init_msg(msg, caller_bytes, msg_type, key);
1095 vec[0].iov_len = sizeof(struct o2net_msg);
1096 vec[0].iov_base = msg;
1097 memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1099 ret = o2net_prep_nsw(nn, &nsw);
1103 msg->msg_num = cpu_to_be32(nsw.ns_id);
1104 o2net_set_nst_msg_id(&nst, nsw.ns_id);
1106 o2net_set_nst_send_time(&nst);
1108 /* finally, convert the message header to network byte-order
1110 mutex_lock(&sc->sc_send_lock);
1111 ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1112 sizeof(struct o2net_msg) + caller_bytes);
1113 mutex_unlock(&sc->sc_send_lock);
1114 msglog(msg, "sending returned %d\n", ret);
1116 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1120 /* wait on other node's handler */
1121 o2net_set_nst_status_time(&nst);
1122 wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1124 o2net_update_send_stats(&nst, sc);
1126 /* Note that we avoid overwriting the callers status return
1127 * variable if a system error was reported on the other
1128 * side. Callers beware. */
1129 ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1131 *status = nsw.ns_status;
1133 mlog(0, "woken, returning system status %d, user status %d\n",
1134 ret, nsw.ns_status);
1136 o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1141 o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1144 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1146 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1147 u8 target_node, int *status)
1153 return o2net_send_message_vec(msg_type, key, &vec, 1,
1154 target_node, status);
1156 EXPORT_SYMBOL_GPL(o2net_send_message);
1158 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1159 enum o2net_system_error syserr, int err)
1163 .iov_len = sizeof(struct o2net_msg),
1166 BUG_ON(syserr >= O2NET_ERR_MAX);
1168 /* leave other fields intact from the incoming message, msg_num
1170 hdr->sys_status = cpu_to_be32(syserr);
1171 hdr->status = cpu_to_be32(err);
1172 hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC); // twiddle the magic
1175 msglog(hdr, "about to send status magic %d\n", err);
1176 /* hdr has been in host byteorder this whole time */
1177 return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1180 /* this returns -errno if the header was unknown or too large, etc.
1181 * after this is called the buffer us reused for the next message */
1182 static int o2net_process_message(struct o2net_sock_container *sc,
1183 struct o2net_msg *hdr)
1185 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1186 int ret = 0, handler_status;
1187 enum o2net_system_error syserr;
1188 struct o2net_msg_handler *nmh = NULL;
1189 void *ret_data = NULL;
1191 msglog(hdr, "processing message\n");
1193 o2net_sc_postpone_idle(sc);
1195 switch(be16_to_cpu(hdr->magic)) {
1196 case O2NET_MSG_STATUS_MAGIC:
1197 /* special type for returning message status */
1198 o2net_complete_nsw(nn, NULL,
1199 be32_to_cpu(hdr->msg_num),
1200 be32_to_cpu(hdr->sys_status),
1201 be32_to_cpu(hdr->status));
1203 case O2NET_MSG_KEEP_REQ_MAGIC:
1204 o2net_sendpage(sc, o2net_keep_resp,
1205 sizeof(*o2net_keep_resp));
1207 case O2NET_MSG_KEEP_RESP_MAGIC:
1209 case O2NET_MSG_MAGIC:
1212 msglog(hdr, "bad magic\n");
1218 /* find a handler for it */
1220 nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1221 be32_to_cpu(hdr->key));
1223 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1224 be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1225 syserr = O2NET_ERR_NO_HNDLR;
1229 syserr = O2NET_ERR_NONE;
1231 if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1232 syserr = O2NET_ERR_OVERFLOW;
1234 if (syserr != O2NET_ERR_NONE)
1237 o2net_set_func_start_time(sc);
1238 sc->sc_msg_key = be32_to_cpu(hdr->key);
1239 sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1240 handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1241 be16_to_cpu(hdr->data_len),
1242 nmh->nh_func_data, &ret_data);
1243 o2net_set_func_stop_time(sc);
1245 o2net_update_recv_stats(sc);
1248 /* this destroys the hdr, so don't use it after this */
1249 mutex_lock(&sc->sc_send_lock);
1250 ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1252 mutex_unlock(&sc->sc_send_lock);
1254 mlog(0, "sending handler status %d, syserr %d returned %d\n",
1255 handler_status, syserr, ret);
1258 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1259 if (nmh->nh_post_func)
1260 (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1266 o2net_handler_put(nmh);
1270 static int o2net_check_handshake(struct o2net_sock_container *sc)
1272 struct o2net_handshake *hand = page_address(sc->sc_page);
1273 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1275 if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1276 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1277 "protocol version %llu but %llu is required. "
1278 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1279 (unsigned long long)be64_to_cpu(hand->protocol_version),
1280 O2NET_PROTOCOL_VERSION);
1282 /* don't bother reconnecting if its the wrong version. */
1283 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1288 * Ensure timeouts are consistent with other nodes, otherwise
1289 * we can end up with one node thinking that the other must be down,
1290 * but isn't. This can ultimately cause corruption.
1292 if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1293 o2net_idle_timeout()) {
1294 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1295 "idle timeout of %u ms, but we use %u ms locally. "
1296 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1297 be32_to_cpu(hand->o2net_idle_timeout_ms),
1298 o2net_idle_timeout());
1299 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1303 if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1304 o2net_keepalive_delay()) {
1305 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1306 "delay of %u ms, but we use %u ms locally. "
1307 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1308 be32_to_cpu(hand->o2net_keepalive_delay_ms),
1309 o2net_keepalive_delay());
1310 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1314 if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1315 O2HB_MAX_WRITE_TIMEOUT_MS) {
1316 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1317 "timeout of %u ms, but we use %u ms locally. "
1318 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1319 be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1320 O2HB_MAX_WRITE_TIMEOUT_MS);
1321 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1325 sc->sc_handshake_ok = 1;
1327 spin_lock(&nn->nn_lock);
1328 /* set valid and queue the idle timers only if it hasn't been
1329 * shut down already */
1330 if (nn->nn_sc == sc) {
1331 o2net_sc_reset_idle_timer(sc);
1332 atomic_set(&nn->nn_timeout, 0);
1333 o2net_set_nn_state(nn, sc, 1, 0);
1335 spin_unlock(&nn->nn_lock);
1337 /* shift everything up as though it wasn't there */
1338 sc->sc_page_off -= sizeof(struct o2net_handshake);
1339 if (sc->sc_page_off)
1340 memmove(hand, hand + 1, sc->sc_page_off);
1345 /* this demuxes the queued rx bytes into header or payload bits and calls
1346 * handlers as each full message is read off the socket. it returns -error,
1347 * == 0 eof, or > 0 for progress made.*/
1348 static int o2net_advance_rx(struct o2net_sock_container *sc)
1350 struct o2net_msg *hdr;
1355 sclog(sc, "receiving\n");
1356 o2net_set_advance_start_time(sc);
1358 if (unlikely(sc->sc_handshake_ok == 0)) {
1359 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1360 data = page_address(sc->sc_page) + sc->sc_page_off;
1361 datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1362 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1364 sc->sc_page_off += ret;
1367 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1368 o2net_check_handshake(sc);
1369 if (unlikely(sc->sc_handshake_ok == 0))
1375 /* do we need more header? */
1376 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1377 data = page_address(sc->sc_page) + sc->sc_page_off;
1378 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1379 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1381 sc->sc_page_off += ret;
1382 /* only swab incoming here.. we can
1383 * only get here once as we cross from
1384 * being under to over */
1385 if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1386 hdr = page_address(sc->sc_page);
1387 if (be16_to_cpu(hdr->data_len) >
1388 O2NET_MAX_PAYLOAD_BYTES)
1396 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1397 /* oof, still don't have a header */
1401 /* this was swabbed above when we first read it */
1402 hdr = page_address(sc->sc_page);
1404 msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1406 /* do we need more payload? */
1407 if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1408 /* need more payload */
1409 data = page_address(sc->sc_page) + sc->sc_page_off;
1410 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1412 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1414 sc->sc_page_off += ret;
1419 if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1420 /* we can only get here once, the first time we read
1421 * the payload.. so set ret to progress if the handler
1422 * works out. after calling this the message is toast */
1423 ret = o2net_process_message(sc, hdr);
1426 sc->sc_page_off = 0;
1430 sclog(sc, "ret = %d\n", ret);
1431 o2net_set_advance_stop_time(sc);
1435 /* this work func is triggerd by data ready. it reads until it can read no
1436 * more. it interprets 0, eof, as fatal. if data_ready hits while we're doing
1437 * our work the work struct will be marked and we'll be called again. */
1438 static void o2net_rx_until_empty(struct work_struct *work)
1440 struct o2net_sock_container *sc =
1441 container_of(work, struct o2net_sock_container, sc_rx_work);
1445 ret = o2net_advance_rx(sc);
1448 if (ret <= 0 && ret != -EAGAIN) {
1449 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1450 sclog(sc, "saw error %d, closing\n", ret);
1451 /* not permanent so read failed handshake can retry */
1452 o2net_ensure_shutdown(nn, sc, 0);
1458 static int o2net_set_nodelay(struct socket *sock)
1462 return kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
1463 (void *)&val, sizeof(val));
1466 static int o2net_set_usertimeout(struct socket *sock)
1468 int user_timeout = O2NET_TCP_USER_TIMEOUT;
1470 return kernel_setsockopt(sock, SOL_TCP, TCP_USER_TIMEOUT,
1471 (void *)&user_timeout, sizeof(user_timeout));
1474 static void o2net_initialize_handshake(void)
1476 o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1477 O2HB_MAX_WRITE_TIMEOUT_MS);
1478 o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1479 o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1480 o2net_keepalive_delay());
1481 o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1482 o2net_reconnect_delay());
1485 /* ------------------------------------------------------------ */
1487 /* called when a connect completes and after a sock is accepted. the
1488 * rx path will see the response and mark the sc valid */
1489 static void o2net_sc_connect_completed(struct work_struct *work)
1491 struct o2net_sock_container *sc =
1492 container_of(work, struct o2net_sock_container,
1495 mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1496 (unsigned long long)O2NET_PROTOCOL_VERSION,
1497 (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1499 o2net_initialize_handshake();
1500 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1504 /* this is called as a work_struct func. */
1505 static void o2net_sc_send_keep_req(struct work_struct *work)
1507 struct o2net_sock_container *sc =
1508 container_of(work, struct o2net_sock_container,
1509 sc_keepalive_work.work);
1511 o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1515 /* socket shutdown does a del_timer_sync against this as it tears down.
1516 * we can't start this timer until we've got to the point in sc buildup
1517 * where shutdown is going to be involved */
1518 static void o2net_idle_timer(struct timer_list *t)
1520 struct o2net_sock_container *sc = from_timer(sc, t, sc_idle_timeout);
1521 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1522 #ifdef CONFIG_DEBUG_FS
1523 unsigned long msecs = ktime_to_ms(ktime_get()) -
1524 ktime_to_ms(sc->sc_tv_timer);
1526 unsigned long msecs = o2net_idle_timeout();
1529 printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1530 "idle for %lu.%lu secs.\n",
1531 SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1533 /* idle timerout happen, don't shutdown the connection, but
1534 * make fence decision. Maybe the connection can recover before
1535 * the decision is made.
1537 atomic_set(&nn->nn_timeout, 1);
1538 o2quo_conn_err(o2net_num_from_nn(nn));
1539 queue_delayed_work(o2net_wq, &nn->nn_still_up,
1540 msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1542 o2net_sc_reset_idle_timer(sc);
1546 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1548 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1549 o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1550 msecs_to_jiffies(o2net_keepalive_delay()));
1551 o2net_set_sock_timer(sc);
1552 mod_timer(&sc->sc_idle_timeout,
1553 jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1556 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1558 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1560 /* clear fence decision since the connection recover from timeout*/
1561 if (atomic_read(&nn->nn_timeout)) {
1562 o2quo_conn_up(o2net_num_from_nn(nn));
1563 cancel_delayed_work(&nn->nn_still_up);
1564 atomic_set(&nn->nn_timeout, 0);
1567 /* Only push out an existing timer */
1568 if (timer_pending(&sc->sc_idle_timeout))
1569 o2net_sc_reset_idle_timer(sc);
1572 /* this work func is kicked whenever a path sets the nn state which doesn't
1573 * have valid set. This includes seeing hb come up, losing a connection,
1574 * having a connect attempt fail, etc. This centralizes the logic which decides
1575 * if a connect attempt should be made or if we should give up and all future
1576 * transmit attempts should fail */
1577 static void o2net_start_connect(struct work_struct *work)
1579 struct o2net_node *nn =
1580 container_of(work, struct o2net_node, nn_connect_work.work);
1581 struct o2net_sock_container *sc = NULL;
1582 struct o2nm_node *node = NULL, *mynode = NULL;
1583 struct socket *sock = NULL;
1584 struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1586 unsigned int timeout;
1587 unsigned int noio_flag;
1590 * sock_create allocates the sock with GFP_KERNEL. We must set
1591 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1592 * by this process are done as if GFP_NOIO was specified. So we
1593 * are not reentering filesystem while doing memory reclaim.
1595 noio_flag = memalloc_noio_save();
1596 /* if we're greater we initiate tx, otherwise we accept */
1597 if (o2nm_this_node() <= o2net_num_from_nn(nn))
1600 /* watch for racing with tearing a node down */
1601 node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1605 mynode = o2nm_get_node_by_num(o2nm_this_node());
1609 spin_lock(&nn->nn_lock);
1611 * see if we already have one pending or have given up.
1612 * For nn_timeout, it is set when we close the connection
1613 * because of the idle time out. So it means that we have
1614 * at least connected to that node successfully once,
1615 * now try to connect to it again.
1617 timeout = atomic_read(&nn->nn_timeout);
1618 stop = (nn->nn_sc ||
1619 (nn->nn_persistent_error &&
1620 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1621 spin_unlock(&nn->nn_lock);
1625 nn->nn_last_connect_attempt = jiffies;
1627 sc = sc_alloc(node);
1629 mlog(0, "couldn't allocate sc\n");
1634 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1636 mlog(0, "can't create socket: %d\n", ret);
1639 sc->sc_sock = sock; /* freed by sc_kref_release */
1641 sock->sk->sk_allocation = GFP_ATOMIC;
1643 myaddr.sin_family = AF_INET;
1644 myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1645 myaddr.sin_port = htons(0); /* any port */
1647 ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1650 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1651 ret, &mynode->nd_ipv4_address);
1655 ret = o2net_set_nodelay(sc->sc_sock);
1657 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1661 ret = o2net_set_usertimeout(sock);
1663 mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1667 o2net_register_callbacks(sc->sc_sock->sk, sc);
1669 spin_lock(&nn->nn_lock);
1670 /* handshake completion will set nn->nn_sc_valid */
1671 o2net_set_nn_state(nn, sc, 0, 0);
1672 spin_unlock(&nn->nn_lock);
1674 remoteaddr.sin_family = AF_INET;
1675 remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1676 remoteaddr.sin_port = node->nd_ipv4_port;
1678 ret = sc->sc_sock->ops->connect(sc->sc_sock,
1679 (struct sockaddr *)&remoteaddr,
1682 if (ret == -EINPROGRESS)
1687 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1688 " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1689 /* 0 err so that another will be queued and attempted
1690 * from set_nn_state */
1691 o2net_ensure_shutdown(nn, sc, 0);
1696 o2nm_node_put(node);
1698 o2nm_node_put(mynode);
1700 memalloc_noio_restore(noio_flag);
1704 static void o2net_connect_expired(struct work_struct *work)
1706 struct o2net_node *nn =
1707 container_of(work, struct o2net_node, nn_connect_expired.work);
1709 spin_lock(&nn->nn_lock);
1710 if (!nn->nn_sc_valid) {
1711 printk(KERN_NOTICE "o2net: No connection established with "
1712 "node %u after %u.%u seconds, check network and"
1713 " cluster configuration.\n",
1714 o2net_num_from_nn(nn),
1715 o2net_idle_timeout() / 1000,
1716 o2net_idle_timeout() % 1000);
1718 o2net_set_nn_state(nn, NULL, 0, 0);
1720 spin_unlock(&nn->nn_lock);
1723 static void o2net_still_up(struct work_struct *work)
1725 struct o2net_node *nn =
1726 container_of(work, struct o2net_node, nn_still_up.work);
1728 o2quo_hb_still_up(o2net_num_from_nn(nn));
1731 /* ------------------------------------------------------------ */
1733 void o2net_disconnect_node(struct o2nm_node *node)
1735 struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1737 /* don't reconnect until it's heartbeating again */
1738 spin_lock(&nn->nn_lock);
1739 atomic_set(&nn->nn_timeout, 0);
1740 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1741 spin_unlock(&nn->nn_lock);
1744 cancel_delayed_work(&nn->nn_connect_expired);
1745 cancel_delayed_work(&nn->nn_connect_work);
1746 cancel_delayed_work(&nn->nn_still_up);
1747 flush_workqueue(o2net_wq);
1751 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1754 o2quo_hb_down(node_num);
1759 if (node_num != o2nm_this_node())
1760 o2net_disconnect_node(node);
1762 BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1765 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1768 struct o2net_node *nn = o2net_nn_from_num(node_num);
1770 o2quo_hb_up(node_num);
1774 /* ensure an immediate connect attempt */
1775 nn->nn_last_connect_attempt = jiffies -
1776 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1778 if (node_num != o2nm_this_node()) {
1779 /* believe it or not, accept and node hearbeating testing
1780 * can succeed for this node before we got here.. so
1781 * only use set_nn_state to clear the persistent error
1782 * if that hasn't already happened */
1783 spin_lock(&nn->nn_lock);
1784 atomic_set(&nn->nn_timeout, 0);
1785 if (nn->nn_persistent_error)
1786 o2net_set_nn_state(nn, NULL, 0, 0);
1787 spin_unlock(&nn->nn_lock);
1791 void o2net_unregister_hb_callbacks(void)
1793 o2hb_unregister_callback(NULL, &o2net_hb_up);
1794 o2hb_unregister_callback(NULL, &o2net_hb_down);
1797 int o2net_register_hb_callbacks(void)
1801 o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1802 o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1803 o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1804 o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1806 ret = o2hb_register_callback(NULL, &o2net_hb_up);
1808 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1811 o2net_unregister_hb_callbacks();
1816 /* ------------------------------------------------------------ */
1818 static int o2net_accept_one(struct socket *sock, int *more)
1821 struct sockaddr_in sin;
1822 struct socket *new_sock = NULL;
1823 struct o2nm_node *node = NULL;
1824 struct o2nm_node *local_node = NULL;
1825 struct o2net_sock_container *sc = NULL;
1826 struct o2net_node *nn;
1827 unsigned int noio_flag;
1830 * sock_create_lite allocates the sock with GFP_KERNEL. We must set
1831 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1832 * by this process are done as if GFP_NOIO was specified. So we
1833 * are not reentering filesystem while doing memory reclaim.
1835 noio_flag = memalloc_noio_save();
1837 BUG_ON(sock == NULL);
1839 ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1840 sock->sk->sk_protocol, &new_sock);
1844 new_sock->type = sock->type;
1845 new_sock->ops = sock->ops;
1846 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, false);
1851 new_sock->sk->sk_allocation = GFP_ATOMIC;
1853 ret = o2net_set_nodelay(new_sock);
1855 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1859 ret = o2net_set_usertimeout(new_sock);
1861 mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1865 ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin, 1);
1869 node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1871 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1872 "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1873 ntohs(sin.sin_port));
1878 if (o2nm_this_node() >= node->nd_num) {
1879 local_node = o2nm_get_node_by_num(o2nm_this_node());
1881 printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1882 "seen at node '%s' (%u, %pI4:%d) from "
1883 "node '%s' (%u, %pI4:%d)\n",
1884 local_node->nd_name, local_node->nd_num,
1885 &(local_node->nd_ipv4_address),
1886 ntohs(local_node->nd_ipv4_port),
1888 node->nd_num, &sin.sin_addr.s_addr,
1889 ntohs(sin.sin_port));
1894 /* this happens all the time when the other node sees our heartbeat
1895 * and tries to connect before we see their heartbeat */
1896 if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1897 mlog(ML_CONN, "attempt to connect from node '%s' at "
1898 "%pI4:%d but it isn't heartbeating\n",
1899 node->nd_name, &sin.sin_addr.s_addr,
1900 ntohs(sin.sin_port));
1905 nn = o2net_nn_from_num(node->nd_num);
1907 spin_lock(&nn->nn_lock);
1912 spin_unlock(&nn->nn_lock);
1914 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1915 "at %pI4:%d but it already has an open connection\n",
1916 node->nd_name, &sin.sin_addr.s_addr,
1917 ntohs(sin.sin_port));
1921 sc = sc_alloc(node);
1927 sc->sc_sock = new_sock;
1930 spin_lock(&nn->nn_lock);
1931 atomic_set(&nn->nn_timeout, 0);
1932 o2net_set_nn_state(nn, sc, 0, 0);
1933 spin_unlock(&nn->nn_lock);
1935 o2net_register_callbacks(sc->sc_sock->sk, sc);
1936 o2net_sc_queue_work(sc, &sc->sc_rx_work);
1938 o2net_initialize_handshake();
1939 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1943 sock_release(new_sock);
1945 o2nm_node_put(node);
1947 o2nm_node_put(local_node);
1951 memalloc_noio_restore(noio_flag);
1956 * This function is invoked in response to one or more
1957 * pending accepts at softIRQ level. We must drain the
1958 * entire que before returning.
1961 static void o2net_accept_many(struct work_struct *work)
1963 struct socket *sock = o2net_listen_sock;
1968 * It is critical to note that due to interrupt moderation
1969 * at the network driver level, we can't assume to get a
1970 * softIRQ for every single conn since tcp SYN packets
1971 * can arrive back-to-back, and therefore many pending
1972 * accepts may result in just 1 softIRQ. If we terminate
1973 * the o2net_accept_one() loop upon seeing an err, what happens
1974 * to the rest of the conns in the queue? If no new SYN
1975 * arrives for hours, no softIRQ will be delivered,
1976 * and the connections will just sit in the queue.
1980 err = o2net_accept_one(sock, &more);
1987 static void o2net_listen_data_ready(struct sock *sk)
1989 void (*ready)(struct sock *sk);
1991 read_lock_bh(&sk->sk_callback_lock);
1992 ready = sk->sk_user_data;
1993 if (ready == NULL) { /* check for teardown race */
1994 ready = sk->sk_data_ready;
1998 /* This callback may called twice when a new connection
1999 * is being established as a child socket inherits everything
2000 * from a parent LISTEN socket, including the data_ready cb of
2001 * the parent. This leads to a hazard. In o2net_accept_one()
2002 * we are still initializing the child socket but have not
2003 * changed the inherited data_ready callback yet when
2004 * data starts arriving.
2005 * We avoid this hazard by checking the state.
2006 * For the listening socket, the state will be TCP_LISTEN; for the new
2007 * socket, will be TCP_ESTABLISHED. Also, in this case,
2008 * sk->sk_user_data is not a valid function pointer.
2011 if (sk->sk_state == TCP_LISTEN) {
2012 queue_work(o2net_wq, &o2net_listen_work);
2018 read_unlock_bh(&sk->sk_callback_lock);
2023 static int o2net_open_listening_sock(__be32 addr, __be16 port)
2025 struct socket *sock = NULL;
2027 struct sockaddr_in sin = {
2028 .sin_family = PF_INET,
2029 .sin_addr = { .s_addr = addr },
2033 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2035 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
2039 sock->sk->sk_allocation = GFP_ATOMIC;
2041 write_lock_bh(&sock->sk->sk_callback_lock);
2042 sock->sk->sk_user_data = sock->sk->sk_data_ready;
2043 sock->sk->sk_data_ready = o2net_listen_data_ready;
2044 write_unlock_bh(&sock->sk->sk_callback_lock);
2046 o2net_listen_sock = sock;
2047 INIT_WORK(&o2net_listen_work, o2net_accept_many);
2049 sock->sk->sk_reuse = SK_CAN_REUSE;
2050 ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2052 printk(KERN_ERR "o2net: Error %d while binding socket at "
2053 "%pI4:%u\n", ret, &addr, ntohs(port));
2057 ret = sock->ops->listen(sock, 64);
2059 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2060 ret, &addr, ntohs(port));
2064 o2net_listen_sock = NULL;
2072 * called from node manager when we should bring up our network listening
2073 * socket. node manager handles all the serialization to only call this
2074 * once and to match it with o2net_stop_listening(). note,
2075 * o2nm_this_node() doesn't work yet as we're being called while it
2078 int o2net_start_listening(struct o2nm_node *node)
2082 BUG_ON(o2net_wq != NULL);
2083 BUG_ON(o2net_listen_sock != NULL);
2085 mlog(ML_KTHREAD, "starting o2net thread...\n");
2086 o2net_wq = alloc_ordered_workqueue("o2net", WQ_MEM_RECLAIM);
2087 if (o2net_wq == NULL) {
2088 mlog(ML_ERROR, "unable to launch o2net thread\n");
2089 return -ENOMEM; /* ? */
2092 ret = o2net_open_listening_sock(node->nd_ipv4_address,
2093 node->nd_ipv4_port);
2095 destroy_workqueue(o2net_wq);
2098 o2quo_conn_up(node->nd_num);
2103 /* again, o2nm_this_node() doesn't work here as we're involved in
2104 * tearing it down */
2105 void o2net_stop_listening(struct o2nm_node *node)
2107 struct socket *sock = o2net_listen_sock;
2110 BUG_ON(o2net_wq == NULL);
2111 BUG_ON(o2net_listen_sock == NULL);
2113 /* stop the listening socket from generating work */
2114 write_lock_bh(&sock->sk->sk_callback_lock);
2115 sock->sk->sk_data_ready = sock->sk->sk_user_data;
2116 sock->sk->sk_user_data = NULL;
2117 write_unlock_bh(&sock->sk->sk_callback_lock);
2119 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2120 struct o2nm_node *node = o2nm_get_node_by_num(i);
2122 o2net_disconnect_node(node);
2123 o2nm_node_put(node);
2127 /* finish all work and tear down the work queue */
2128 mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2129 destroy_workqueue(o2net_wq);
2132 sock_release(o2net_listen_sock);
2133 o2net_listen_sock = NULL;
2135 o2quo_conn_err(node->nd_num);
2138 /* ------------------------------------------------------------ */
2140 int o2net_init(void)
2146 if (o2net_debugfs_init())
2149 o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2150 o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2151 o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2152 if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp)
2155 o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2156 o2net_hand->connector_id = cpu_to_be64(1);
2158 o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2159 o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2161 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2162 struct o2net_node *nn = o2net_nn_from_num(i);
2164 atomic_set(&nn->nn_timeout, 0);
2165 spin_lock_init(&nn->nn_lock);
2166 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2167 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2168 o2net_connect_expired);
2169 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2170 /* until we see hb from a node we'll return einval */
2171 nn->nn_persistent_error = -ENOTCONN;
2172 init_waitqueue_head(&nn->nn_sc_wq);
2173 idr_init(&nn->nn_status_idr);
2174 INIT_LIST_HEAD(&nn->nn_status_list);
2181 kfree(o2net_keep_req);
2182 kfree(o2net_keep_resp);
2183 o2net_debugfs_exit();
2188 void o2net_exit(void)
2192 kfree(o2net_keep_req);
2193 kfree(o2net_keep_resp);
2194 o2net_debugfs_exit();