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(unsigned long data);
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 setup_timer(&sc->sc_idle_timeout, o2net_idle_timer,
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_valid && !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) {
547 printk(KERN_NOTICE "o2net: No longer connected to "
548 SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
549 o2net_complete_nodes_nsw(nn);
552 if (!was_valid && valid) {
553 o2quo_conn_up(o2net_num_from_nn(nn));
554 cancel_delayed_work(&nn->nn_connect_expired);
555 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
556 o2nm_this_node() > sc->sc_node->nd_num ?
557 "Connected to" : "Accepted connection from",
561 /* trigger the connecting worker func as long as we're not valid,
562 * it will back off if it shouldn't connect. This can be called
563 * from node config teardown and so needs to be careful about
564 * the work queue actually being up. */
565 if (!valid && o2net_wq) {
567 /* delay if we're within a RECONNECT_DELAY of the
569 delay = (nn->nn_last_connect_attempt +
570 msecs_to_jiffies(o2net_reconnect_delay()))
572 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
574 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
575 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
578 * Delay the expired work after idle timeout.
580 * We might have lots of failed connection attempts that run
581 * through here but we only cancel the connect_expired work when
582 * a connection attempt succeeds. So only the first enqueue of
583 * the connect_expired work will do anything. The rest will see
584 * that it's already queued and do nothing.
586 delay += msecs_to_jiffies(o2net_idle_timeout());
587 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
590 /* keep track of the nn's sc ref for the caller */
591 if ((old_sc == NULL) && sc)
593 if (old_sc && (old_sc != sc)) {
594 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
599 /* see o2net_register_callbacks() */
600 static void o2net_data_ready(struct sock *sk)
602 void (*ready)(struct sock *sk);
603 struct o2net_sock_container *sc;
605 read_lock_bh(&sk->sk_callback_lock);
606 sc = sk->sk_user_data;
608 sclog(sc, "data_ready hit\n");
609 o2net_set_data_ready_time(sc);
610 o2net_sc_queue_work(sc, &sc->sc_rx_work);
611 ready = sc->sc_data_ready;
613 ready = sk->sk_data_ready;
615 read_unlock_bh(&sk->sk_callback_lock);
620 /* see o2net_register_callbacks() */
621 static void o2net_state_change(struct sock *sk)
623 void (*state_change)(struct sock *sk);
624 struct o2net_sock_container *sc;
626 read_lock_bh(&sk->sk_callback_lock);
627 sc = sk->sk_user_data;
629 state_change = sk->sk_state_change;
633 sclog(sc, "state_change to %d\n", sk->sk_state);
635 state_change = sc->sc_state_change;
637 switch(sk->sk_state) {
638 /* ignore connecting sockets as they make progress */
642 case TCP_ESTABLISHED:
643 o2net_sc_queue_work(sc, &sc->sc_connect_work);
646 printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
647 " shutdown, state %d\n",
648 SC_NODEF_ARGS(sc), sk->sk_state);
649 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
653 read_unlock_bh(&sk->sk_callback_lock);
658 * we register callbacks so we can queue work on events before calling
659 * the original callbacks. our callbacks our careful to test user_data
660 * to discover when they've reaced with o2net_unregister_callbacks().
662 static void o2net_register_callbacks(struct sock *sk,
663 struct o2net_sock_container *sc)
665 write_lock_bh(&sk->sk_callback_lock);
667 /* accepted sockets inherit the old listen socket data ready */
668 if (sk->sk_data_ready == o2net_listen_data_ready) {
669 sk->sk_data_ready = sk->sk_user_data;
670 sk->sk_user_data = NULL;
673 BUG_ON(sk->sk_user_data != NULL);
674 sk->sk_user_data = sc;
677 sc->sc_data_ready = sk->sk_data_ready;
678 sc->sc_state_change = sk->sk_state_change;
679 sk->sk_data_ready = o2net_data_ready;
680 sk->sk_state_change = o2net_state_change;
682 mutex_init(&sc->sc_send_lock);
684 write_unlock_bh(&sk->sk_callback_lock);
687 static int o2net_unregister_callbacks(struct sock *sk,
688 struct o2net_sock_container *sc)
692 write_lock_bh(&sk->sk_callback_lock);
693 if (sk->sk_user_data == sc) {
695 sk->sk_user_data = NULL;
696 sk->sk_data_ready = sc->sc_data_ready;
697 sk->sk_state_change = sc->sc_state_change;
699 write_unlock_bh(&sk->sk_callback_lock);
705 * this is a little helper that is called by callers who have seen a problem
706 * with an sc and want to detach it from the nn if someone already hasn't beat
707 * them to it. if an error is given then the shutdown will be persistent
708 * and pending transmits will be canceled.
710 static void o2net_ensure_shutdown(struct o2net_node *nn,
711 struct o2net_sock_container *sc,
714 spin_lock(&nn->nn_lock);
716 o2net_set_nn_state(nn, NULL, 0, err);
717 spin_unlock(&nn->nn_lock);
721 * This work queue function performs the blocking parts of socket shutdown. A
722 * few paths lead here. set_nn_state will trigger this callback if it sees an
723 * sc detached from the nn. state_change will also trigger this callback
724 * directly when it sees errors. In that case we need to call set_nn_state
725 * ourselves as state_change couldn't get the nn_lock and call set_nn_state
728 static void o2net_shutdown_sc(struct work_struct *work)
730 struct o2net_sock_container *sc =
731 container_of(work, struct o2net_sock_container,
733 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
735 sclog(sc, "shutting down\n");
737 /* drop the callbacks ref and call shutdown only once */
738 if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
739 /* we shouldn't flush as we're in the thread, the
740 * races with pending sc work structs are harmless */
741 del_timer_sync(&sc->sc_idle_timeout);
742 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
744 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
747 /* not fatal so failed connects before the other guy has our
748 * heartbeat can be retried */
749 o2net_ensure_shutdown(nn, sc, 0);
753 /* ------------------------------------------------------------ */
755 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
758 int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
761 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
766 static struct o2net_msg_handler *
767 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
768 struct rb_node **ret_parent)
770 struct rb_node **p = &o2net_handler_tree.rb_node;
771 struct rb_node *parent = NULL;
772 struct o2net_msg_handler *nmh, *ret = NULL;
777 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
778 cmp = o2net_handler_cmp(nmh, msg_type, key);
792 if (ret_parent != NULL)
793 *ret_parent = parent;
798 static void o2net_handler_kref_release(struct kref *kref)
800 struct o2net_msg_handler *nmh;
801 nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
806 static void o2net_handler_put(struct o2net_msg_handler *nmh)
808 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
811 /* max_len is protection for the handler func. incoming messages won't
812 * be given to the handler if their payload is longer than the max. */
813 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
814 o2net_msg_handler_func *func, void *data,
815 o2net_post_msg_handler_func *post_func,
816 struct list_head *unreg_list)
818 struct o2net_msg_handler *nmh = NULL;
819 struct rb_node **p, *parent;
822 if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
823 mlog(0, "max_len for message handler out of range: %u\n",
830 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
836 mlog(0, "no message handler provided: %u, %p\n",
842 nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
849 nmh->nh_func_data = data;
850 nmh->nh_post_func = post_func;
851 nmh->nh_msg_type = msg_type;
852 nmh->nh_max_len = max_len;
854 /* the tree and list get this ref.. they're both removed in
855 * unregister when this ref is dropped */
856 kref_init(&nmh->nh_kref);
857 INIT_LIST_HEAD(&nmh->nh_unregister_item);
859 write_lock(&o2net_handler_lock);
860 if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
863 rb_link_node(&nmh->nh_node, parent, p);
864 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
865 list_add_tail(&nmh->nh_unregister_item, unreg_list);
867 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
868 func, msg_type, key);
869 /* we've had some trouble with handlers seemingly vanishing. */
870 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
872 "couldn't find handler we *just* registered "
873 "for type %u key %08x\n", msg_type, key);
875 write_unlock(&o2net_handler_lock);
885 EXPORT_SYMBOL_GPL(o2net_register_handler);
887 void o2net_unregister_handler_list(struct list_head *list)
889 struct o2net_msg_handler *nmh, *n;
891 write_lock(&o2net_handler_lock);
892 list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
893 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
894 nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
895 rb_erase(&nmh->nh_node, &o2net_handler_tree);
896 list_del_init(&nmh->nh_unregister_item);
897 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
899 write_unlock(&o2net_handler_lock);
901 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
903 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
905 struct o2net_msg_handler *nmh;
907 read_lock(&o2net_handler_lock);
908 nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
910 kref_get(&nmh->nh_kref);
911 read_unlock(&o2net_handler_lock);
916 /* ------------------------------------------------------------ */
918 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
920 struct kvec vec = { .iov_len = len, .iov_base = data, };
921 struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
922 return kernel_recvmsg(sock, &msg, &vec, 1, len, msg.msg_flags);
925 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
926 size_t veclen, size_t total)
929 struct msghdr msg = {.msg_flags = 0,};
936 ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
937 if (likely(ret == total))
939 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
941 ret = -EPIPE; /* should be smarter, I bet */
943 mlog(0, "returning error: %d\n", ret);
947 static void o2net_sendpage(struct o2net_sock_container *sc,
948 void *kmalloced_virt,
951 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
955 mutex_lock(&sc->sc_send_lock);
956 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
957 virt_to_page(kmalloced_virt),
958 offset_in_page(kmalloced_virt),
960 mutex_unlock(&sc->sc_send_lock);
963 if (ret == (ssize_t)-EAGAIN) {
964 mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
965 " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
969 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
970 " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
971 o2net_ensure_shutdown(nn, sc, 0);
976 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
978 memset(msg, 0, sizeof(struct o2net_msg));
979 msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
980 msg->data_len = cpu_to_be16(data_len);
981 msg->msg_type = cpu_to_be16(msg_type);
982 msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
984 msg->key = cpu_to_be32(key);
987 static int o2net_tx_can_proceed(struct o2net_node *nn,
988 struct o2net_sock_container **sc_ret,
993 spin_lock(&nn->nn_lock);
994 if (nn->nn_persistent_error) {
997 *error = nn->nn_persistent_error;
998 } else if (nn->nn_sc_valid) {
999 kref_get(&nn->nn_sc->sc_kref);
1002 *sc_ret = nn->nn_sc;
1005 spin_unlock(&nn->nn_lock);
1010 /* Get a map of all nodes to which this node is currently connected to */
1011 void o2net_fill_node_map(unsigned long *map, unsigned bytes)
1013 struct o2net_sock_container *sc;
1016 BUG_ON(bytes < (BITS_TO_LONGS(O2NM_MAX_NODES) * sizeof(unsigned long)));
1018 memset(map, 0, bytes);
1019 for (node = 0; node < O2NM_MAX_NODES; ++node) {
1020 if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1028 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1030 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1031 size_t caller_veclen, u8 target_node, int *status)
1034 struct o2net_msg *msg = NULL;
1035 size_t veclen, caller_bytes = 0;
1036 struct kvec *vec = NULL;
1037 struct o2net_sock_container *sc = NULL;
1038 struct o2net_node *nn = o2net_nn_from_num(target_node);
1039 struct o2net_status_wait nsw = {
1040 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1042 struct o2net_send_tracking nst;
1044 o2net_init_nst(&nst, msg_type, key, current, target_node);
1046 if (o2net_wq == NULL) {
1047 mlog(0, "attempt to tx without o2netd running\n");
1052 if (caller_veclen == 0) {
1053 mlog(0, "bad kvec array length\n");
1058 caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1059 if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1060 mlog(0, "total payload len %zu too large\n", caller_bytes);
1065 if (target_node == o2nm_this_node()) {
1070 o2net_debug_add_nst(&nst);
1072 o2net_set_nst_sock_time(&nst);
1074 wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1078 o2net_set_nst_sock_container(&nst, sc);
1080 veclen = caller_veclen + 1;
1081 vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1083 mlog(0, "failed to %zu element kvec!\n", veclen);
1088 msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1090 mlog(0, "failed to allocate a o2net_msg!\n");
1095 o2net_init_msg(msg, caller_bytes, msg_type, key);
1097 vec[0].iov_len = sizeof(struct o2net_msg);
1098 vec[0].iov_base = msg;
1099 memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1101 ret = o2net_prep_nsw(nn, &nsw);
1105 msg->msg_num = cpu_to_be32(nsw.ns_id);
1106 o2net_set_nst_msg_id(&nst, nsw.ns_id);
1108 o2net_set_nst_send_time(&nst);
1110 /* finally, convert the message header to network byte-order
1112 mutex_lock(&sc->sc_send_lock);
1113 ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1114 sizeof(struct o2net_msg) + caller_bytes);
1115 mutex_unlock(&sc->sc_send_lock);
1116 msglog(msg, "sending returned %d\n", ret);
1118 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1122 /* wait on other node's handler */
1123 o2net_set_nst_status_time(&nst);
1124 wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1126 o2net_update_send_stats(&nst, sc);
1128 /* Note that we avoid overwriting the callers status return
1129 * variable if a system error was reported on the other
1130 * side. Callers beware. */
1131 ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1133 *status = nsw.ns_status;
1135 mlog(0, "woken, returning system status %d, user status %d\n",
1136 ret, nsw.ns_status);
1138 o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1143 o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1146 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1148 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1149 u8 target_node, int *status)
1155 return o2net_send_message_vec(msg_type, key, &vec, 1,
1156 target_node, status);
1158 EXPORT_SYMBOL_GPL(o2net_send_message);
1160 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1161 enum o2net_system_error syserr, int err)
1165 .iov_len = sizeof(struct o2net_msg),
1168 BUG_ON(syserr >= O2NET_ERR_MAX);
1170 /* leave other fields intact from the incoming message, msg_num
1172 hdr->sys_status = cpu_to_be32(syserr);
1173 hdr->status = cpu_to_be32(err);
1174 hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC); // twiddle the magic
1177 msglog(hdr, "about to send status magic %d\n", err);
1178 /* hdr has been in host byteorder this whole time */
1179 return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1182 /* this returns -errno if the header was unknown or too large, etc.
1183 * after this is called the buffer us reused for the next message */
1184 static int o2net_process_message(struct o2net_sock_container *sc,
1185 struct o2net_msg *hdr)
1187 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1188 int ret = 0, handler_status;
1189 enum o2net_system_error syserr;
1190 struct o2net_msg_handler *nmh = NULL;
1191 void *ret_data = NULL;
1193 msglog(hdr, "processing message\n");
1195 o2net_sc_postpone_idle(sc);
1197 switch(be16_to_cpu(hdr->magic)) {
1198 case O2NET_MSG_STATUS_MAGIC:
1199 /* special type for returning message status */
1200 o2net_complete_nsw(nn, NULL,
1201 be32_to_cpu(hdr->msg_num),
1202 be32_to_cpu(hdr->sys_status),
1203 be32_to_cpu(hdr->status));
1205 case O2NET_MSG_KEEP_REQ_MAGIC:
1206 o2net_sendpage(sc, o2net_keep_resp,
1207 sizeof(*o2net_keep_resp));
1209 case O2NET_MSG_KEEP_RESP_MAGIC:
1211 case O2NET_MSG_MAGIC:
1214 msglog(hdr, "bad magic\n");
1220 /* find a handler for it */
1222 nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1223 be32_to_cpu(hdr->key));
1225 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1226 be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1227 syserr = O2NET_ERR_NO_HNDLR;
1231 syserr = O2NET_ERR_NONE;
1233 if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1234 syserr = O2NET_ERR_OVERFLOW;
1236 if (syserr != O2NET_ERR_NONE)
1239 o2net_set_func_start_time(sc);
1240 sc->sc_msg_key = be32_to_cpu(hdr->key);
1241 sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1242 handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1243 be16_to_cpu(hdr->data_len),
1244 nmh->nh_func_data, &ret_data);
1245 o2net_set_func_stop_time(sc);
1247 o2net_update_recv_stats(sc);
1250 /* this destroys the hdr, so don't use it after this */
1251 mutex_lock(&sc->sc_send_lock);
1252 ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1254 mutex_unlock(&sc->sc_send_lock);
1256 mlog(0, "sending handler status %d, syserr %d returned %d\n",
1257 handler_status, syserr, ret);
1260 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1261 if (nmh->nh_post_func)
1262 (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1268 o2net_handler_put(nmh);
1272 static int o2net_check_handshake(struct o2net_sock_container *sc)
1274 struct o2net_handshake *hand = page_address(sc->sc_page);
1275 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1277 if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1278 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1279 "protocol version %llu but %llu is required. "
1280 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1281 (unsigned long long)be64_to_cpu(hand->protocol_version),
1282 O2NET_PROTOCOL_VERSION);
1284 /* don't bother reconnecting if its the wrong version. */
1285 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1290 * Ensure timeouts are consistent with other nodes, otherwise
1291 * we can end up with one node thinking that the other must be down,
1292 * but isn't. This can ultimately cause corruption.
1294 if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1295 o2net_idle_timeout()) {
1296 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1297 "idle timeout of %u ms, but we use %u ms locally. "
1298 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1299 be32_to_cpu(hand->o2net_idle_timeout_ms),
1300 o2net_idle_timeout());
1301 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1305 if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1306 o2net_keepalive_delay()) {
1307 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1308 "delay of %u ms, but we use %u ms locally. "
1309 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1310 be32_to_cpu(hand->o2net_keepalive_delay_ms),
1311 o2net_keepalive_delay());
1312 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1316 if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1317 O2HB_MAX_WRITE_TIMEOUT_MS) {
1318 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1319 "timeout of %u ms, but we use %u ms locally. "
1320 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1321 be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1322 O2HB_MAX_WRITE_TIMEOUT_MS);
1323 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1327 sc->sc_handshake_ok = 1;
1329 spin_lock(&nn->nn_lock);
1330 /* set valid and queue the idle timers only if it hasn't been
1331 * shut down already */
1332 if (nn->nn_sc == sc) {
1333 o2net_sc_reset_idle_timer(sc);
1334 atomic_set(&nn->nn_timeout, 0);
1335 o2net_set_nn_state(nn, sc, 1, 0);
1337 spin_unlock(&nn->nn_lock);
1339 /* shift everything up as though it wasn't there */
1340 sc->sc_page_off -= sizeof(struct o2net_handshake);
1341 if (sc->sc_page_off)
1342 memmove(hand, hand + 1, sc->sc_page_off);
1347 /* this demuxes the queued rx bytes into header or payload bits and calls
1348 * handlers as each full message is read off the socket. it returns -error,
1349 * == 0 eof, or > 0 for progress made.*/
1350 static int o2net_advance_rx(struct o2net_sock_container *sc)
1352 struct o2net_msg *hdr;
1357 sclog(sc, "receiving\n");
1358 o2net_set_advance_start_time(sc);
1360 if (unlikely(sc->sc_handshake_ok == 0)) {
1361 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1362 data = page_address(sc->sc_page) + sc->sc_page_off;
1363 datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1364 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1366 sc->sc_page_off += ret;
1369 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1370 o2net_check_handshake(sc);
1371 if (unlikely(sc->sc_handshake_ok == 0))
1377 /* do we need more header? */
1378 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1379 data = page_address(sc->sc_page) + sc->sc_page_off;
1380 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1381 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1383 sc->sc_page_off += ret;
1384 /* only swab incoming here.. we can
1385 * only get here once as we cross from
1386 * being under to over */
1387 if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1388 hdr = page_address(sc->sc_page);
1389 if (be16_to_cpu(hdr->data_len) >
1390 O2NET_MAX_PAYLOAD_BYTES)
1398 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1399 /* oof, still don't have a header */
1403 /* this was swabbed above when we first read it */
1404 hdr = page_address(sc->sc_page);
1406 msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1408 /* do we need more payload? */
1409 if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1410 /* need more payload */
1411 data = page_address(sc->sc_page) + sc->sc_page_off;
1412 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1414 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1416 sc->sc_page_off += ret;
1421 if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1422 /* we can only get here once, the first time we read
1423 * the payload.. so set ret to progress if the handler
1424 * works out. after calling this the message is toast */
1425 ret = o2net_process_message(sc, hdr);
1428 sc->sc_page_off = 0;
1432 sclog(sc, "ret = %d\n", ret);
1433 o2net_set_advance_stop_time(sc);
1437 /* this work func is triggerd by data ready. it reads until it can read no
1438 * more. it interprets 0, eof, as fatal. if data_ready hits while we're doing
1439 * our work the work struct will be marked and we'll be called again. */
1440 static void o2net_rx_until_empty(struct work_struct *work)
1442 struct o2net_sock_container *sc =
1443 container_of(work, struct o2net_sock_container, sc_rx_work);
1447 ret = o2net_advance_rx(sc);
1450 if (ret <= 0 && ret != -EAGAIN) {
1451 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1452 sclog(sc, "saw error %d, closing\n", ret);
1453 /* not permanent so read failed handshake can retry */
1454 o2net_ensure_shutdown(nn, sc, 0);
1460 static int o2net_set_nodelay(struct socket *sock)
1464 return kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
1465 (void *)&val, sizeof(val));
1468 static int o2net_set_usertimeout(struct socket *sock)
1470 int user_timeout = O2NET_TCP_USER_TIMEOUT;
1472 return kernel_setsockopt(sock, SOL_TCP, TCP_USER_TIMEOUT,
1473 (void *)&user_timeout, sizeof(user_timeout));
1476 static void o2net_initialize_handshake(void)
1478 o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1479 O2HB_MAX_WRITE_TIMEOUT_MS);
1480 o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1481 o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1482 o2net_keepalive_delay());
1483 o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1484 o2net_reconnect_delay());
1487 /* ------------------------------------------------------------ */
1489 /* called when a connect completes and after a sock is accepted. the
1490 * rx path will see the response and mark the sc valid */
1491 static void o2net_sc_connect_completed(struct work_struct *work)
1493 struct o2net_sock_container *sc =
1494 container_of(work, struct o2net_sock_container,
1497 mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1498 (unsigned long long)O2NET_PROTOCOL_VERSION,
1499 (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1501 o2net_initialize_handshake();
1502 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1506 /* this is called as a work_struct func. */
1507 static void o2net_sc_send_keep_req(struct work_struct *work)
1509 struct o2net_sock_container *sc =
1510 container_of(work, struct o2net_sock_container,
1511 sc_keepalive_work.work);
1513 o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1517 /* socket shutdown does a del_timer_sync against this as it tears down.
1518 * we can't start this timer until we've got to the point in sc buildup
1519 * where shutdown is going to be involved */
1520 static void o2net_idle_timer(unsigned long data)
1522 struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1523 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1524 #ifdef CONFIG_DEBUG_FS
1525 unsigned long msecs = ktime_to_ms(ktime_get()) -
1526 ktime_to_ms(sc->sc_tv_timer);
1528 unsigned long msecs = o2net_idle_timeout();
1531 printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1532 "idle for %lu.%lu secs.\n",
1533 SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1535 /* idle timerout happen, don't shutdown the connection, but
1536 * make fence decision. Maybe the connection can recover before
1537 * the decision is made.
1539 atomic_set(&nn->nn_timeout, 1);
1540 o2quo_conn_err(o2net_num_from_nn(nn));
1541 queue_delayed_work(o2net_wq, &nn->nn_still_up,
1542 msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1544 o2net_sc_reset_idle_timer(sc);
1548 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1550 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1551 o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1552 msecs_to_jiffies(o2net_keepalive_delay()));
1553 o2net_set_sock_timer(sc);
1554 mod_timer(&sc->sc_idle_timeout,
1555 jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1558 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1560 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1562 /* clear fence decision since the connection recover from timeout*/
1563 if (atomic_read(&nn->nn_timeout)) {
1564 o2quo_conn_up(o2net_num_from_nn(nn));
1565 cancel_delayed_work(&nn->nn_still_up);
1566 atomic_set(&nn->nn_timeout, 0);
1569 /* Only push out an existing timer */
1570 if (timer_pending(&sc->sc_idle_timeout))
1571 o2net_sc_reset_idle_timer(sc);
1574 /* this work func is kicked whenever a path sets the nn state which doesn't
1575 * have valid set. This includes seeing hb come up, losing a connection,
1576 * having a connect attempt fail, etc. This centralizes the logic which decides
1577 * if a connect attempt should be made or if we should give up and all future
1578 * transmit attempts should fail */
1579 static void o2net_start_connect(struct work_struct *work)
1581 struct o2net_node *nn =
1582 container_of(work, struct o2net_node, nn_connect_work.work);
1583 struct o2net_sock_container *sc = NULL;
1584 struct o2nm_node *node = NULL, *mynode = NULL;
1585 struct socket *sock = NULL;
1586 struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1588 unsigned int timeout;
1589 unsigned int noio_flag;
1592 * sock_create allocates the sock with GFP_KERNEL. We must set
1593 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1594 * by this process are done as if GFP_NOIO was specified. So we
1595 * are not reentering filesystem while doing memory reclaim.
1597 noio_flag = memalloc_noio_save();
1598 /* if we're greater we initiate tx, otherwise we accept */
1599 if (o2nm_this_node() <= o2net_num_from_nn(nn))
1602 /* watch for racing with tearing a node down */
1603 node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1607 mynode = o2nm_get_node_by_num(o2nm_this_node());
1611 spin_lock(&nn->nn_lock);
1613 * see if we already have one pending or have given up.
1614 * For nn_timeout, it is set when we close the connection
1615 * because of the idle time out. So it means that we have
1616 * at least connected to that node successfully once,
1617 * now try to connect to it again.
1619 timeout = atomic_read(&nn->nn_timeout);
1620 stop = (nn->nn_sc ||
1621 (nn->nn_persistent_error &&
1622 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1623 spin_unlock(&nn->nn_lock);
1627 nn->nn_last_connect_attempt = jiffies;
1629 sc = sc_alloc(node);
1631 mlog(0, "couldn't allocate sc\n");
1636 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1638 mlog(0, "can't create socket: %d\n", ret);
1641 sc->sc_sock = sock; /* freed by sc_kref_release */
1643 sock->sk->sk_allocation = GFP_ATOMIC;
1645 myaddr.sin_family = AF_INET;
1646 myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1647 myaddr.sin_port = htons(0); /* any port */
1649 ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1652 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1653 ret, &mynode->nd_ipv4_address);
1657 ret = o2net_set_nodelay(sc->sc_sock);
1659 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1663 ret = o2net_set_usertimeout(sock);
1665 mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1669 o2net_register_callbacks(sc->sc_sock->sk, sc);
1671 spin_lock(&nn->nn_lock);
1672 /* handshake completion will set nn->nn_sc_valid */
1673 o2net_set_nn_state(nn, sc, 0, 0);
1674 spin_unlock(&nn->nn_lock);
1676 remoteaddr.sin_family = AF_INET;
1677 remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1678 remoteaddr.sin_port = node->nd_ipv4_port;
1680 ret = sc->sc_sock->ops->connect(sc->sc_sock,
1681 (struct sockaddr *)&remoteaddr,
1684 if (ret == -EINPROGRESS)
1689 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1690 " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1691 /* 0 err so that another will be queued and attempted
1692 * from set_nn_state */
1693 o2net_ensure_shutdown(nn, sc, 0);
1698 o2nm_node_put(node);
1700 o2nm_node_put(mynode);
1702 memalloc_noio_restore(noio_flag);
1706 static void o2net_connect_expired(struct work_struct *work)
1708 struct o2net_node *nn =
1709 container_of(work, struct o2net_node, nn_connect_expired.work);
1711 spin_lock(&nn->nn_lock);
1712 if (!nn->nn_sc_valid) {
1713 printk(KERN_NOTICE "o2net: No connection established with "
1714 "node %u after %u.%u seconds, check network and"
1715 " cluster configuration.\n",
1716 o2net_num_from_nn(nn),
1717 o2net_idle_timeout() / 1000,
1718 o2net_idle_timeout() % 1000);
1720 o2net_set_nn_state(nn, NULL, 0, 0);
1722 spin_unlock(&nn->nn_lock);
1725 static void o2net_still_up(struct work_struct *work)
1727 struct o2net_node *nn =
1728 container_of(work, struct o2net_node, nn_still_up.work);
1730 o2quo_hb_still_up(o2net_num_from_nn(nn));
1733 /* ------------------------------------------------------------ */
1735 void o2net_disconnect_node(struct o2nm_node *node)
1737 struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1739 /* don't reconnect until it's heartbeating again */
1740 spin_lock(&nn->nn_lock);
1741 atomic_set(&nn->nn_timeout, 0);
1742 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1743 spin_unlock(&nn->nn_lock);
1746 cancel_delayed_work(&nn->nn_connect_expired);
1747 cancel_delayed_work(&nn->nn_connect_work);
1748 cancel_delayed_work(&nn->nn_still_up);
1749 flush_workqueue(o2net_wq);
1753 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1756 o2quo_hb_down(node_num);
1761 if (node_num != o2nm_this_node())
1762 o2net_disconnect_node(node);
1764 BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1767 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1770 struct o2net_node *nn = o2net_nn_from_num(node_num);
1772 o2quo_hb_up(node_num);
1776 /* ensure an immediate connect attempt */
1777 nn->nn_last_connect_attempt = jiffies -
1778 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1780 if (node_num != o2nm_this_node()) {
1781 /* believe it or not, accept and node hearbeating testing
1782 * can succeed for this node before we got here.. so
1783 * only use set_nn_state to clear the persistent error
1784 * if that hasn't already happened */
1785 spin_lock(&nn->nn_lock);
1786 atomic_set(&nn->nn_timeout, 0);
1787 if (nn->nn_persistent_error)
1788 o2net_set_nn_state(nn, NULL, 0, 0);
1789 spin_unlock(&nn->nn_lock);
1793 void o2net_unregister_hb_callbacks(void)
1795 o2hb_unregister_callback(NULL, &o2net_hb_up);
1796 o2hb_unregister_callback(NULL, &o2net_hb_down);
1799 int o2net_register_hb_callbacks(void)
1803 o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1804 o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1805 o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1806 o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1808 ret = o2hb_register_callback(NULL, &o2net_hb_up);
1810 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1813 o2net_unregister_hb_callbacks();
1818 /* ------------------------------------------------------------ */
1820 static int o2net_accept_one(struct socket *sock, int *more)
1823 struct sockaddr_in sin;
1824 struct socket *new_sock = NULL;
1825 struct o2nm_node *node = NULL;
1826 struct o2nm_node *local_node = NULL;
1827 struct o2net_sock_container *sc = NULL;
1828 struct o2net_node *nn;
1829 unsigned int noio_flag;
1832 * sock_create_lite allocates the sock with GFP_KERNEL. We must set
1833 * per-process flag PF_MEMALLOC_NOIO so that all allocations done
1834 * by this process are done as if GFP_NOIO was specified. So we
1835 * are not reentering filesystem while doing memory reclaim.
1837 noio_flag = memalloc_noio_save();
1839 BUG_ON(sock == NULL);
1841 ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1842 sock->sk->sk_protocol, &new_sock);
1846 new_sock->type = sock->type;
1847 new_sock->ops = sock->ops;
1848 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, false);
1853 new_sock->sk->sk_allocation = GFP_ATOMIC;
1855 ret = o2net_set_nodelay(new_sock);
1857 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1861 ret = o2net_set_usertimeout(new_sock);
1863 mlog(ML_ERROR, "set TCP_USER_TIMEOUT failed with %d\n", ret);
1868 ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1873 node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1875 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1876 "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1877 ntohs(sin.sin_port));
1882 if (o2nm_this_node() >= node->nd_num) {
1883 local_node = o2nm_get_node_by_num(o2nm_this_node());
1885 printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1886 "seen at node '%s' (%u, %pI4:%d) from "
1887 "node '%s' (%u, %pI4:%d)\n",
1888 local_node->nd_name, local_node->nd_num,
1889 &(local_node->nd_ipv4_address),
1890 ntohs(local_node->nd_ipv4_port),
1892 node->nd_num, &sin.sin_addr.s_addr,
1893 ntohs(sin.sin_port));
1898 /* this happens all the time when the other node sees our heartbeat
1899 * and tries to connect before we see their heartbeat */
1900 if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1901 mlog(ML_CONN, "attempt to connect from node '%s' at "
1902 "%pI4:%d but it isn't heartbeating\n",
1903 node->nd_name, &sin.sin_addr.s_addr,
1904 ntohs(sin.sin_port));
1909 nn = o2net_nn_from_num(node->nd_num);
1911 spin_lock(&nn->nn_lock);
1916 spin_unlock(&nn->nn_lock);
1918 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1919 "at %pI4:%d but it already has an open connection\n",
1920 node->nd_name, &sin.sin_addr.s_addr,
1921 ntohs(sin.sin_port));
1925 sc = sc_alloc(node);
1931 sc->sc_sock = new_sock;
1934 spin_lock(&nn->nn_lock);
1935 atomic_set(&nn->nn_timeout, 0);
1936 o2net_set_nn_state(nn, sc, 0, 0);
1937 spin_unlock(&nn->nn_lock);
1939 o2net_register_callbacks(sc->sc_sock->sk, sc);
1940 o2net_sc_queue_work(sc, &sc->sc_rx_work);
1942 o2net_initialize_handshake();
1943 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1947 sock_release(new_sock);
1949 o2nm_node_put(node);
1951 o2nm_node_put(local_node);
1955 memalloc_noio_restore(noio_flag);
1960 * This function is invoked in response to one or more
1961 * pending accepts at softIRQ level. We must drain the
1962 * entire que before returning.
1965 static void o2net_accept_many(struct work_struct *work)
1967 struct socket *sock = o2net_listen_sock;
1972 * It is critical to note that due to interrupt moderation
1973 * at the network driver level, we can't assume to get a
1974 * softIRQ for every single conn since tcp SYN packets
1975 * can arrive back-to-back, and therefore many pending
1976 * accepts may result in just 1 softIRQ. If we terminate
1977 * the o2net_accept_one() loop upon seeing an err, what happens
1978 * to the rest of the conns in the queue? If no new SYN
1979 * arrives for hours, no softIRQ will be delivered,
1980 * and the connections will just sit in the queue.
1984 err = o2net_accept_one(sock, &more);
1991 static void o2net_listen_data_ready(struct sock *sk)
1993 void (*ready)(struct sock *sk);
1995 read_lock_bh(&sk->sk_callback_lock);
1996 ready = sk->sk_user_data;
1997 if (ready == NULL) { /* check for teardown race */
1998 ready = sk->sk_data_ready;
2002 /* This callback may called twice when a new connection
2003 * is being established as a child socket inherits everything
2004 * from a parent LISTEN socket, including the data_ready cb of
2005 * the parent. This leads to a hazard. In o2net_accept_one()
2006 * we are still initializing the child socket but have not
2007 * changed the inherited data_ready callback yet when
2008 * data starts arriving.
2009 * We avoid this hazard by checking the state.
2010 * For the listening socket, the state will be TCP_LISTEN; for the new
2011 * socket, will be TCP_ESTABLISHED. Also, in this case,
2012 * sk->sk_user_data is not a valid function pointer.
2015 if (sk->sk_state == TCP_LISTEN) {
2016 queue_work(o2net_wq, &o2net_listen_work);
2022 read_unlock_bh(&sk->sk_callback_lock);
2027 static int o2net_open_listening_sock(__be32 addr, __be16 port)
2029 struct socket *sock = NULL;
2031 struct sockaddr_in sin = {
2032 .sin_family = PF_INET,
2033 .sin_addr = { .s_addr = addr },
2037 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2039 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
2043 sock->sk->sk_allocation = GFP_ATOMIC;
2045 write_lock_bh(&sock->sk->sk_callback_lock);
2046 sock->sk->sk_user_data = sock->sk->sk_data_ready;
2047 sock->sk->sk_data_ready = o2net_listen_data_ready;
2048 write_unlock_bh(&sock->sk->sk_callback_lock);
2050 o2net_listen_sock = sock;
2051 INIT_WORK(&o2net_listen_work, o2net_accept_many);
2053 sock->sk->sk_reuse = SK_CAN_REUSE;
2054 ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2056 printk(KERN_ERR "o2net: Error %d while binding socket at "
2057 "%pI4:%u\n", ret, &addr, ntohs(port));
2061 ret = sock->ops->listen(sock, 64);
2063 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2064 ret, &addr, ntohs(port));
2068 o2net_listen_sock = NULL;
2076 * called from node manager when we should bring up our network listening
2077 * socket. node manager handles all the serialization to only call this
2078 * once and to match it with o2net_stop_listening(). note,
2079 * o2nm_this_node() doesn't work yet as we're being called while it
2082 int o2net_start_listening(struct o2nm_node *node)
2086 BUG_ON(o2net_wq != NULL);
2087 BUG_ON(o2net_listen_sock != NULL);
2089 mlog(ML_KTHREAD, "starting o2net thread...\n");
2090 o2net_wq = alloc_ordered_workqueue("o2net", WQ_MEM_RECLAIM);
2091 if (o2net_wq == NULL) {
2092 mlog(ML_ERROR, "unable to launch o2net thread\n");
2093 return -ENOMEM; /* ? */
2096 ret = o2net_open_listening_sock(node->nd_ipv4_address,
2097 node->nd_ipv4_port);
2099 destroy_workqueue(o2net_wq);
2102 o2quo_conn_up(node->nd_num);
2107 /* again, o2nm_this_node() doesn't work here as we're involved in
2108 * tearing it down */
2109 void o2net_stop_listening(struct o2nm_node *node)
2111 struct socket *sock = o2net_listen_sock;
2114 BUG_ON(o2net_wq == NULL);
2115 BUG_ON(o2net_listen_sock == NULL);
2117 /* stop the listening socket from generating work */
2118 write_lock_bh(&sock->sk->sk_callback_lock);
2119 sock->sk->sk_data_ready = sock->sk->sk_user_data;
2120 sock->sk->sk_user_data = NULL;
2121 write_unlock_bh(&sock->sk->sk_callback_lock);
2123 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2124 struct o2nm_node *node = o2nm_get_node_by_num(i);
2126 o2net_disconnect_node(node);
2127 o2nm_node_put(node);
2131 /* finish all work and tear down the work queue */
2132 mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2133 destroy_workqueue(o2net_wq);
2136 sock_release(o2net_listen_sock);
2137 o2net_listen_sock = NULL;
2139 o2quo_conn_err(node->nd_num);
2142 /* ------------------------------------------------------------ */
2144 int o2net_init(void)
2150 if (o2net_debugfs_init())
2153 o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2154 o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2155 o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2156 if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp)
2159 o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2160 o2net_hand->connector_id = cpu_to_be64(1);
2162 o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2163 o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2165 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2166 struct o2net_node *nn = o2net_nn_from_num(i);
2168 atomic_set(&nn->nn_timeout, 0);
2169 spin_lock_init(&nn->nn_lock);
2170 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2171 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2172 o2net_connect_expired);
2173 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2174 /* until we see hb from a node we'll return einval */
2175 nn->nn_persistent_error = -ENOTCONN;
2176 init_waitqueue_head(&nn->nn_sc_wq);
2177 idr_init(&nn->nn_status_idr);
2178 INIT_LIST_HEAD(&nn->nn_status_list);
2185 kfree(o2net_keep_req);
2186 kfree(o2net_keep_resp);
2187 o2net_debugfs_exit();
2192 void o2net_exit(void)
2196 kfree(o2net_keep_req);
2197 kfree(o2net_keep_resp);
2198 o2net_debugfs_exit();