1 /* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
5 * Copyright (C) 2004 Oracle. All rights reserved.
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public
9 * License as published by the Free Software Foundation; either
10 * version 2 of the License, or (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License for more details.
17 * You should have received a copy of the GNU General Public
18 * License along with this program; if not, write to the
19 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
20 * Boston, MA 021110-1307, USA.
24 * Callers for this were originally written against a very simple synchronus
25 * API. This implementation reflects those simple callers. Some day I'm sure
26 * we'll need to move to a more robust posting/callback mechanism.
28 * Transmit calls pass in kernel virtual addresses and block copying this into
29 * the socket's tx buffers via a usual blocking sendmsg. They'll block waiting
30 * for a failed socket to timeout. TX callers can also pass in a poniter to an
31 * 'int' which gets filled with an errno off the wire in response to the
34 * Handlers for unsolicited messages are registered. Each socket has a page
35 * that incoming data is copied into. First the header, then the data.
36 * Handlers are called from only one thread with a reference to this per-socket
37 * page. This page is destroyed after the handler call, so it can't be
38 * referenced beyond the call. Handlers may block but are discouraged from
41 * Any framing errors (bad magic, large payload lengths) close a connection.
43 * Our sock_container holds the state we associate with a socket. It's current
44 * framing state is held there as well as the refcounting we do around when it
45 * is safe to tear down the socket. The socket is only finally torn down from
46 * the container when the container loses all of its references -- so as long
47 * as you hold a ref on the container you can trust that the socket is valid
48 * for use with kernel socket APIs.
50 * Connections are initiated between a pair of nodes when the node with the
51 * higher node number gets a heartbeat callback which indicates that the lower
52 * numbered node has started heartbeating. The lower numbered node is passive
53 * and only accepts the connection if the higher numbered node is heartbeating.
56 #include <linux/kernel.h>
57 #include <linux/jiffies.h>
58 #include <linux/slab.h>
59 #include <linux/idr.h>
60 #include <linux/kref.h>
61 #include <linux/net.h>
64 #include <asm/uaccess.h>
66 #include "heartbeat.h"
68 #include "nodemanager.h"
69 #define MLOG_MASK_PREFIX ML_TCP
73 #include "tcp_internal.h"
75 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
76 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num, \
77 &sc->sc_node->nd_ipv4_address, \
78 ntohs(sc->sc_node->nd_ipv4_port)
81 * In the following two log macros, the whitespace after the ',' just
82 * before ##args is intentional. Otherwise, gcc 2.95 will eat the
83 * previous token if args expands to nothing.
85 #define msglog(hdr, fmt, args...) do { \
86 typeof(hdr) __hdr = (hdr); \
87 mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d " \
88 "key %08x num %u] " fmt, \
89 be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len), \
90 be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status), \
91 be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key), \
92 be32_to_cpu(__hdr->msg_num) , ##args); \
95 #define sclog(sc, fmt, args...) do { \
96 typeof(sc) __sc = (sc); \
97 mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p " \
98 "pg_off %zu] " fmt, __sc, \
99 atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock, \
100 __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off , \
104 static DEFINE_RWLOCK(o2net_handler_lock);
105 static struct rb_root o2net_handler_tree = RB_ROOT;
107 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
109 /* XXX someday we'll need better accounting */
110 static struct socket *o2net_listen_sock = NULL;
113 * listen work is only queued by the listening socket callbacks on the
114 * o2net_wq. teardown detaches the callbacks before destroying the workqueue.
115 * quorum work is queued as sock containers are shutdown.. stop_listening
116 * tears down all the node's sock containers, preventing future shutdowns
117 * and queued quroum work, before canceling delayed quorum work and
118 * destroying the work queue.
120 static struct workqueue_struct *o2net_wq;
121 static struct work_struct o2net_listen_work;
123 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
124 #define O2NET_HB_PRI 0x1
126 static struct o2net_handshake *o2net_hand;
127 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
129 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
130 {[O2NET_ERR_NONE] = 0,
131 [O2NET_ERR_NO_HNDLR] = -ENOPROTOOPT,
132 [O2NET_ERR_OVERFLOW] = -EOVERFLOW,
133 [O2NET_ERR_DIED] = -EHOSTDOWN,};
135 /* can't quite avoid *all* internal declarations :/ */
136 static void o2net_sc_connect_completed(struct work_struct *work);
137 static void o2net_rx_until_empty(struct work_struct *work);
138 static void o2net_shutdown_sc(struct work_struct *work);
139 static void o2net_listen_data_ready(struct sock *sk, int bytes);
140 static void o2net_sc_send_keep_req(struct work_struct *work);
141 static void o2net_idle_timer(unsigned long data);
142 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
143 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
145 #ifdef CONFIG_DEBUG_FS
146 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
147 u32 msgkey, struct task_struct *task, u8 node)
149 INIT_LIST_HEAD(&nst->st_net_debug_item);
151 nst->st_msg_type = msgtype;
152 nst->st_msg_key = msgkey;
156 static void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
158 do_gettimeofday(&nst->st_sock_time);
161 static void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
163 do_gettimeofday(&nst->st_send_time);
166 static void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
168 do_gettimeofday(&nst->st_status_time);
171 static void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
172 struct o2net_sock_container *sc)
177 static void o2net_set_nst_msg_id(struct o2net_send_tracking *nst, u32 msg_id)
182 #else /* CONFIG_DEBUG_FS */
184 static inline void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
185 u32 msgkey, struct task_struct *task, u8 node)
189 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
193 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
197 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
201 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
202 struct o2net_sock_container *sc)
206 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
211 #endif /* CONFIG_DEBUG_FS */
213 static inline int o2net_reconnect_delay(void)
215 return o2nm_single_cluster->cl_reconnect_delay_ms;
218 static inline int o2net_keepalive_delay(void)
220 return o2nm_single_cluster->cl_keepalive_delay_ms;
223 static inline int o2net_idle_timeout(void)
225 return o2nm_single_cluster->cl_idle_timeout_ms;
228 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
231 BUG_ON(err >= O2NET_ERR_MAX);
232 trans = o2net_sys_err_translations[err];
234 /* Just in case we mess up the translation table above */
235 BUG_ON(err != O2NET_ERR_NONE && trans == 0);
239 static struct o2net_node * o2net_nn_from_num(u8 node_num)
241 BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
242 return &o2net_nodes[node_num];
245 static u8 o2net_num_from_nn(struct o2net_node *nn)
248 return nn - o2net_nodes;
251 /* ------------------------------------------------------------ */
253 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
258 if (!idr_pre_get(&nn->nn_status_idr, GFP_ATOMIC)) {
262 spin_lock(&nn->nn_lock);
263 ret = idr_get_new(&nn->nn_status_idr, nsw, &nsw->ns_id);
265 list_add_tail(&nsw->ns_node_item,
266 &nn->nn_status_list);
267 spin_unlock(&nn->nn_lock);
268 } while (ret == -EAGAIN);
271 init_waitqueue_head(&nsw->ns_wq);
272 nsw->ns_sys_status = O2NET_ERR_NONE;
279 static void o2net_complete_nsw_locked(struct o2net_node *nn,
280 struct o2net_status_wait *nsw,
281 enum o2net_system_error sys_status,
284 assert_spin_locked(&nn->nn_lock);
286 if (!list_empty(&nsw->ns_node_item)) {
287 list_del_init(&nsw->ns_node_item);
288 nsw->ns_sys_status = sys_status;
289 nsw->ns_status = status;
290 idr_remove(&nn->nn_status_idr, nsw->ns_id);
291 wake_up(&nsw->ns_wq);
295 static void o2net_complete_nsw(struct o2net_node *nn,
296 struct o2net_status_wait *nsw,
297 u64 id, enum o2net_system_error sys_status,
300 spin_lock(&nn->nn_lock);
305 nsw = idr_find(&nn->nn_status_idr, id);
310 o2net_complete_nsw_locked(nn, nsw, sys_status, status);
313 spin_unlock(&nn->nn_lock);
317 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
319 struct o2net_status_wait *nsw, *tmp;
320 unsigned int num_kills = 0;
322 assert_spin_locked(&nn->nn_lock);
324 list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
325 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
329 mlog(0, "completed %d messages for node %u\n", num_kills,
330 o2net_num_from_nn(nn));
333 static int o2net_nsw_completed(struct o2net_node *nn,
334 struct o2net_status_wait *nsw)
337 spin_lock(&nn->nn_lock);
338 completed = list_empty(&nsw->ns_node_item);
339 spin_unlock(&nn->nn_lock);
343 /* ------------------------------------------------------------ */
345 static void sc_kref_release(struct kref *kref)
347 struct o2net_sock_container *sc = container_of(kref,
348 struct o2net_sock_container, sc_kref);
349 BUG_ON(timer_pending(&sc->sc_idle_timeout));
351 sclog(sc, "releasing\n");
354 sock_release(sc->sc_sock);
358 o2nm_undepend_item(&sc->sc_node->nd_item);
359 o2nm_node_put(sc->sc_node);
362 o2net_debug_del_sc(sc);
366 static void sc_put(struct o2net_sock_container *sc)
369 kref_put(&sc->sc_kref, sc_kref_release);
371 static void sc_get(struct o2net_sock_container *sc)
374 kref_get(&sc->sc_kref);
376 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
378 struct o2net_sock_container *sc, *ret = NULL;
379 struct page *page = NULL;
382 page = alloc_page(GFP_NOFS);
383 sc = kzalloc(sizeof(*sc), GFP_NOFS);
384 if (sc == NULL || page == NULL)
387 kref_init(&sc->sc_kref);
391 /* pin the node item of the remote node */
392 status = o2nm_depend_item(&node->nd_item);
398 INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
399 INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
400 INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
401 INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
403 init_timer(&sc->sc_idle_timeout);
404 sc->sc_idle_timeout.function = o2net_idle_timer;
405 sc->sc_idle_timeout.data = (unsigned long)sc;
407 sclog(sc, "alloced\n");
411 o2net_debug_add_sc(sc);
423 /* ------------------------------------------------------------ */
425 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
426 struct work_struct *work)
429 if (!queue_work(o2net_wq, work))
432 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
433 struct delayed_work *work,
437 if (!queue_delayed_work(o2net_wq, work, delay))
440 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
441 struct delayed_work *work)
443 if (cancel_delayed_work(work))
447 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
449 int o2net_num_connected_peers(void)
451 return atomic_read(&o2net_connected_peers);
454 static void o2net_set_nn_state(struct o2net_node *nn,
455 struct o2net_sock_container *sc,
456 unsigned valid, int err)
458 int was_valid = nn->nn_sc_valid;
459 int was_err = nn->nn_persistent_error;
460 struct o2net_sock_container *old_sc = nn->nn_sc;
462 assert_spin_locked(&nn->nn_lock);
465 atomic_dec(&o2net_connected_peers);
466 else if (!old_sc && sc)
467 atomic_inc(&o2net_connected_peers);
469 /* the node num comparison and single connect/accept path should stop
470 * an non-null sc from being overwritten with another */
471 BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
472 mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
473 mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
475 if (was_valid && !valid && err == 0)
478 mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
479 o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
480 nn->nn_persistent_error, err);
483 nn->nn_sc_valid = valid ? 1 : 0;
484 nn->nn_persistent_error = err;
486 /* mirrors o2net_tx_can_proceed() */
487 if (nn->nn_persistent_error || nn->nn_sc_valid)
488 wake_up(&nn->nn_sc_wq);
490 if (!was_err && nn->nn_persistent_error) {
491 o2quo_conn_err(o2net_num_from_nn(nn));
492 queue_delayed_work(o2net_wq, &nn->nn_still_up,
493 msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
496 if (was_valid && !valid) {
497 printk(KERN_NOTICE "o2net: no longer connected to "
498 SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
499 o2net_complete_nodes_nsw(nn);
502 if (!was_valid && valid) {
503 o2quo_conn_up(o2net_num_from_nn(nn));
504 cancel_delayed_work(&nn->nn_connect_expired);
505 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
506 o2nm_this_node() > sc->sc_node->nd_num ?
507 "connected to" : "accepted connection from",
511 /* trigger the connecting worker func as long as we're not valid,
512 * it will back off if it shouldn't connect. This can be called
513 * from node config teardown and so needs to be careful about
514 * the work queue actually being up. */
515 if (!valid && o2net_wq) {
517 /* delay if we're withing a RECONNECT_DELAY of the
519 delay = (nn->nn_last_connect_attempt +
520 msecs_to_jiffies(o2net_reconnect_delay()))
522 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
524 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
525 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
528 * Delay the expired work after idle timeout.
530 * We might have lots of failed connection attempts that run
531 * through here but we only cancel the connect_expired work when
532 * a connection attempt succeeds. So only the first enqueue of
533 * the connect_expired work will do anything. The rest will see
534 * that it's already queued and do nothing.
536 delay += msecs_to_jiffies(o2net_idle_timeout());
537 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
540 /* keep track of the nn's sc ref for the caller */
541 if ((old_sc == NULL) && sc)
543 if (old_sc && (old_sc != sc)) {
544 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
549 /* see o2net_register_callbacks() */
550 static void o2net_data_ready(struct sock *sk, int bytes)
552 void (*ready)(struct sock *sk, int bytes);
554 read_lock(&sk->sk_callback_lock);
555 if (sk->sk_user_data) {
556 struct o2net_sock_container *sc = sk->sk_user_data;
557 sclog(sc, "data_ready hit\n");
558 do_gettimeofday(&sc->sc_tv_data_ready);
559 o2net_sc_queue_work(sc, &sc->sc_rx_work);
560 ready = sc->sc_data_ready;
562 ready = sk->sk_data_ready;
564 read_unlock(&sk->sk_callback_lock);
569 /* see o2net_register_callbacks() */
570 static void o2net_state_change(struct sock *sk)
572 void (*state_change)(struct sock *sk);
573 struct o2net_sock_container *sc;
575 read_lock(&sk->sk_callback_lock);
576 sc = sk->sk_user_data;
578 state_change = sk->sk_state_change;
582 sclog(sc, "state_change to %d\n", sk->sk_state);
584 state_change = sc->sc_state_change;
586 switch(sk->sk_state) {
587 /* ignore connecting sockets as they make progress */
591 case TCP_ESTABLISHED:
592 o2net_sc_queue_work(sc, &sc->sc_connect_work);
595 printk(KERN_INFO "o2net: connection to " SC_NODEF_FMT
596 " shutdown, state %d\n",
597 SC_NODEF_ARGS(sc), sk->sk_state);
598 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
602 read_unlock(&sk->sk_callback_lock);
607 * we register callbacks so we can queue work on events before calling
608 * the original callbacks. our callbacks our careful to test user_data
609 * to discover when they've reaced with o2net_unregister_callbacks().
611 static void o2net_register_callbacks(struct sock *sk,
612 struct o2net_sock_container *sc)
614 write_lock_bh(&sk->sk_callback_lock);
616 /* accepted sockets inherit the old listen socket data ready */
617 if (sk->sk_data_ready == o2net_listen_data_ready) {
618 sk->sk_data_ready = sk->sk_user_data;
619 sk->sk_user_data = NULL;
622 BUG_ON(sk->sk_user_data != NULL);
623 sk->sk_user_data = sc;
626 sc->sc_data_ready = sk->sk_data_ready;
627 sc->sc_state_change = sk->sk_state_change;
628 sk->sk_data_ready = o2net_data_ready;
629 sk->sk_state_change = o2net_state_change;
631 mutex_init(&sc->sc_send_lock);
633 write_unlock_bh(&sk->sk_callback_lock);
636 static int o2net_unregister_callbacks(struct sock *sk,
637 struct o2net_sock_container *sc)
641 write_lock_bh(&sk->sk_callback_lock);
642 if (sk->sk_user_data == sc) {
644 sk->sk_user_data = NULL;
645 sk->sk_data_ready = sc->sc_data_ready;
646 sk->sk_state_change = sc->sc_state_change;
648 write_unlock_bh(&sk->sk_callback_lock);
654 * this is a little helper that is called by callers who have seen a problem
655 * with an sc and want to detach it from the nn if someone already hasn't beat
656 * them to it. if an error is given then the shutdown will be persistent
657 * and pending transmits will be canceled.
659 static void o2net_ensure_shutdown(struct o2net_node *nn,
660 struct o2net_sock_container *sc,
663 spin_lock(&nn->nn_lock);
665 o2net_set_nn_state(nn, NULL, 0, err);
666 spin_unlock(&nn->nn_lock);
670 * This work queue function performs the blocking parts of socket shutdown. A
671 * few paths lead here. set_nn_state will trigger this callback if it sees an
672 * sc detached from the nn. state_change will also trigger this callback
673 * directly when it sees errors. In that case we need to call set_nn_state
674 * ourselves as state_change couldn't get the nn_lock and call set_nn_state
677 static void o2net_shutdown_sc(struct work_struct *work)
679 struct o2net_sock_container *sc =
680 container_of(work, struct o2net_sock_container,
682 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
684 sclog(sc, "shutting down\n");
686 /* drop the callbacks ref and call shutdown only once */
687 if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
688 /* we shouldn't flush as we're in the thread, the
689 * races with pending sc work structs are harmless */
690 del_timer_sync(&sc->sc_idle_timeout);
691 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
693 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
696 /* not fatal so failed connects before the other guy has our
697 * heartbeat can be retried */
698 o2net_ensure_shutdown(nn, sc, 0);
702 /* ------------------------------------------------------------ */
704 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
707 int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
710 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
715 static struct o2net_msg_handler *
716 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
717 struct rb_node **ret_parent)
719 struct rb_node **p = &o2net_handler_tree.rb_node;
720 struct rb_node *parent = NULL;
721 struct o2net_msg_handler *nmh, *ret = NULL;
726 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
727 cmp = o2net_handler_cmp(nmh, msg_type, key);
741 if (ret_parent != NULL)
742 *ret_parent = parent;
747 static void o2net_handler_kref_release(struct kref *kref)
749 struct o2net_msg_handler *nmh;
750 nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
755 static void o2net_handler_put(struct o2net_msg_handler *nmh)
757 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
760 /* max_len is protection for the handler func. incoming messages won't
761 * be given to the handler if their payload is longer than the max. */
762 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
763 o2net_msg_handler_func *func, void *data,
764 o2net_post_msg_handler_func *post_func,
765 struct list_head *unreg_list)
767 struct o2net_msg_handler *nmh = NULL;
768 struct rb_node **p, *parent;
771 if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
772 mlog(0, "max_len for message handler out of range: %u\n",
779 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
785 mlog(0, "no message handler provided: %u, %p\n",
791 nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
798 nmh->nh_func_data = data;
799 nmh->nh_post_func = post_func;
800 nmh->nh_msg_type = msg_type;
801 nmh->nh_max_len = max_len;
803 /* the tree and list get this ref.. they're both removed in
804 * unregister when this ref is dropped */
805 kref_init(&nmh->nh_kref);
806 INIT_LIST_HEAD(&nmh->nh_unregister_item);
808 write_lock(&o2net_handler_lock);
809 if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
812 rb_link_node(&nmh->nh_node, parent, p);
813 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
814 list_add_tail(&nmh->nh_unregister_item, unreg_list);
816 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
817 func, msg_type, key);
818 /* we've had some trouble with handlers seemingly vanishing. */
819 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
821 "couldn't find handler we *just* registerd "
822 "for type %u key %08x\n", msg_type, key);
824 write_unlock(&o2net_handler_lock);
834 EXPORT_SYMBOL_GPL(o2net_register_handler);
836 void o2net_unregister_handler_list(struct list_head *list)
838 struct o2net_msg_handler *nmh, *n;
840 write_lock(&o2net_handler_lock);
841 list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
842 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
843 nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
844 rb_erase(&nmh->nh_node, &o2net_handler_tree);
845 list_del_init(&nmh->nh_unregister_item);
846 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
848 write_unlock(&o2net_handler_lock);
850 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
852 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
854 struct o2net_msg_handler *nmh;
856 read_lock(&o2net_handler_lock);
857 nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
859 kref_get(&nmh->nh_kref);
860 read_unlock(&o2net_handler_lock);
865 /* ------------------------------------------------------------ */
867 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
875 struct msghdr msg = {
877 .msg_iov = (struct iovec *)&vec,
878 .msg_flags = MSG_DONTWAIT,
883 ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
889 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
890 size_t veclen, size_t total)
894 struct msghdr msg = {
895 .msg_iov = (struct iovec *)vec,
896 .msg_iovlen = veclen,
906 ret = sock_sendmsg(sock, &msg, total);
909 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
912 ret = -EPIPE; /* should be smarter, I bet */
919 mlog(0, "returning error: %d\n", ret);
923 static void o2net_sendpage(struct o2net_sock_container *sc,
924 void *kmalloced_virt,
927 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
931 mutex_lock(&sc->sc_send_lock);
932 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
933 virt_to_page(kmalloced_virt),
934 (long)kmalloced_virt & ~PAGE_MASK,
936 mutex_unlock(&sc->sc_send_lock);
939 if (ret == (ssize_t)-EAGAIN) {
940 mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
941 " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
945 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
946 " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
947 o2net_ensure_shutdown(nn, sc, 0);
952 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
954 memset(msg, 0, sizeof(struct o2net_msg));
955 msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
956 msg->data_len = cpu_to_be16(data_len);
957 msg->msg_type = cpu_to_be16(msg_type);
958 msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
960 msg->key = cpu_to_be32(key);
963 static int o2net_tx_can_proceed(struct o2net_node *nn,
964 struct o2net_sock_container **sc_ret,
969 spin_lock(&nn->nn_lock);
970 if (nn->nn_persistent_error) {
973 *error = nn->nn_persistent_error;
974 } else if (nn->nn_sc_valid) {
975 kref_get(&nn->nn_sc->sc_kref);
981 spin_unlock(&nn->nn_lock);
986 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
987 size_t caller_veclen, u8 target_node, int *status)
990 struct o2net_msg *msg = NULL;
991 size_t veclen, caller_bytes = 0;
992 struct kvec *vec = NULL;
993 struct o2net_sock_container *sc = NULL;
994 struct o2net_node *nn = o2net_nn_from_num(target_node);
995 struct o2net_status_wait nsw = {
996 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
998 struct o2net_send_tracking nst;
1000 o2net_init_nst(&nst, msg_type, key, current, target_node);
1002 if (o2net_wq == NULL) {
1003 mlog(0, "attempt to tx without o2netd running\n");
1008 if (caller_veclen == 0) {
1009 mlog(0, "bad kvec array length\n");
1014 caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1015 if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1016 mlog(0, "total payload len %zu too large\n", caller_bytes);
1021 if (target_node == o2nm_this_node()) {
1026 o2net_debug_add_nst(&nst);
1028 o2net_set_nst_sock_time(&nst);
1030 wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1034 o2net_set_nst_sock_container(&nst, sc);
1036 veclen = caller_veclen + 1;
1037 vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1039 mlog(0, "failed to %zu element kvec!\n", veclen);
1044 msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1046 mlog(0, "failed to allocate a o2net_msg!\n");
1051 o2net_init_msg(msg, caller_bytes, msg_type, key);
1053 vec[0].iov_len = sizeof(struct o2net_msg);
1054 vec[0].iov_base = msg;
1055 memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1057 ret = o2net_prep_nsw(nn, &nsw);
1061 msg->msg_num = cpu_to_be32(nsw.ns_id);
1062 o2net_set_nst_msg_id(&nst, nsw.ns_id);
1064 o2net_set_nst_send_time(&nst);
1066 /* finally, convert the message header to network byte-order
1068 mutex_lock(&sc->sc_send_lock);
1069 ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1070 sizeof(struct o2net_msg) + caller_bytes);
1071 mutex_unlock(&sc->sc_send_lock);
1072 msglog(msg, "sending returned %d\n", ret);
1074 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1078 /* wait on other node's handler */
1079 o2net_set_nst_status_time(&nst);
1080 wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1082 /* Note that we avoid overwriting the callers status return
1083 * variable if a system error was reported on the other
1084 * side. Callers beware. */
1085 ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1087 *status = nsw.ns_status;
1089 mlog(0, "woken, returning system status %d, user status %d\n",
1090 ret, nsw.ns_status);
1092 o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1099 o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1102 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1104 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1105 u8 target_node, int *status)
1111 return o2net_send_message_vec(msg_type, key, &vec, 1,
1112 target_node, status);
1114 EXPORT_SYMBOL_GPL(o2net_send_message);
1116 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1117 enum o2net_system_error syserr, int err)
1121 .iov_len = sizeof(struct o2net_msg),
1124 BUG_ON(syserr >= O2NET_ERR_MAX);
1126 /* leave other fields intact from the incoming message, msg_num
1128 hdr->sys_status = cpu_to_be32(syserr);
1129 hdr->status = cpu_to_be32(err);
1130 hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC); // twiddle the magic
1133 msglog(hdr, "about to send status magic %d\n", err);
1134 /* hdr has been in host byteorder this whole time */
1135 return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1138 /* this returns -errno if the header was unknown or too large, etc.
1139 * after this is called the buffer us reused for the next message */
1140 static int o2net_process_message(struct o2net_sock_container *sc,
1141 struct o2net_msg *hdr)
1143 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1144 int ret = 0, handler_status;
1145 enum o2net_system_error syserr;
1146 struct o2net_msg_handler *nmh = NULL;
1147 void *ret_data = NULL;
1149 msglog(hdr, "processing message\n");
1151 o2net_sc_postpone_idle(sc);
1153 switch(be16_to_cpu(hdr->magic)) {
1154 case O2NET_MSG_STATUS_MAGIC:
1155 /* special type for returning message status */
1156 o2net_complete_nsw(nn, NULL,
1157 be32_to_cpu(hdr->msg_num),
1158 be32_to_cpu(hdr->sys_status),
1159 be32_to_cpu(hdr->status));
1161 case O2NET_MSG_KEEP_REQ_MAGIC:
1162 o2net_sendpage(sc, o2net_keep_resp,
1163 sizeof(*o2net_keep_resp));
1165 case O2NET_MSG_KEEP_RESP_MAGIC:
1167 case O2NET_MSG_MAGIC:
1170 msglog(hdr, "bad magic\n");
1176 /* find a handler for it */
1178 nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1179 be32_to_cpu(hdr->key));
1181 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1182 be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1183 syserr = O2NET_ERR_NO_HNDLR;
1187 syserr = O2NET_ERR_NONE;
1189 if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1190 syserr = O2NET_ERR_OVERFLOW;
1192 if (syserr != O2NET_ERR_NONE)
1195 do_gettimeofday(&sc->sc_tv_func_start);
1196 sc->sc_msg_key = be32_to_cpu(hdr->key);
1197 sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1198 handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1199 be16_to_cpu(hdr->data_len),
1200 nmh->nh_func_data, &ret_data);
1201 do_gettimeofday(&sc->sc_tv_func_stop);
1204 /* this destroys the hdr, so don't use it after this */
1205 mutex_lock(&sc->sc_send_lock);
1206 ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1208 mutex_unlock(&sc->sc_send_lock);
1210 mlog(0, "sending handler status %d, syserr %d returned %d\n",
1211 handler_status, syserr, ret);
1214 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1215 if (nmh->nh_post_func)
1216 (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1222 o2net_handler_put(nmh);
1226 static int o2net_check_handshake(struct o2net_sock_container *sc)
1228 struct o2net_handshake *hand = page_address(sc->sc_page);
1229 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1231 if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1232 mlog(ML_NOTICE, SC_NODEF_FMT " advertised net protocol "
1233 "version %llu but %llu is required, disconnecting\n",
1235 (unsigned long long)be64_to_cpu(hand->protocol_version),
1236 O2NET_PROTOCOL_VERSION);
1238 /* don't bother reconnecting if its the wrong version. */
1239 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1244 * Ensure timeouts are consistent with other nodes, otherwise
1245 * we can end up with one node thinking that the other must be down,
1246 * but isn't. This can ultimately cause corruption.
1248 if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1249 o2net_idle_timeout()) {
1250 mlog(ML_NOTICE, SC_NODEF_FMT " uses a network idle timeout of "
1251 "%u ms, but we use %u ms locally. disconnecting\n",
1253 be32_to_cpu(hand->o2net_idle_timeout_ms),
1254 o2net_idle_timeout());
1255 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1259 if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1260 o2net_keepalive_delay()) {
1261 mlog(ML_NOTICE, SC_NODEF_FMT " uses a keepalive delay of "
1262 "%u ms, but we use %u ms locally. disconnecting\n",
1264 be32_to_cpu(hand->o2net_keepalive_delay_ms),
1265 o2net_keepalive_delay());
1266 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1270 if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1271 O2HB_MAX_WRITE_TIMEOUT_MS) {
1272 mlog(ML_NOTICE, SC_NODEF_FMT " uses a heartbeat timeout of "
1273 "%u ms, but we use %u ms locally. disconnecting\n",
1275 be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1276 O2HB_MAX_WRITE_TIMEOUT_MS);
1277 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1281 sc->sc_handshake_ok = 1;
1283 spin_lock(&nn->nn_lock);
1284 /* set valid and queue the idle timers only if it hasn't been
1285 * shut down already */
1286 if (nn->nn_sc == sc) {
1287 o2net_sc_reset_idle_timer(sc);
1288 atomic_set(&nn->nn_timeout, 0);
1289 o2net_set_nn_state(nn, sc, 1, 0);
1291 spin_unlock(&nn->nn_lock);
1293 /* shift everything up as though it wasn't there */
1294 sc->sc_page_off -= sizeof(struct o2net_handshake);
1295 if (sc->sc_page_off)
1296 memmove(hand, hand + 1, sc->sc_page_off);
1301 /* this demuxes the queued rx bytes into header or payload bits and calls
1302 * handlers as each full message is read off the socket. it returns -error,
1303 * == 0 eof, or > 0 for progress made.*/
1304 static int o2net_advance_rx(struct o2net_sock_container *sc)
1306 struct o2net_msg *hdr;
1311 sclog(sc, "receiving\n");
1312 do_gettimeofday(&sc->sc_tv_advance_start);
1314 if (unlikely(sc->sc_handshake_ok == 0)) {
1315 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1316 data = page_address(sc->sc_page) + sc->sc_page_off;
1317 datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1318 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1320 sc->sc_page_off += ret;
1323 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1324 o2net_check_handshake(sc);
1325 if (unlikely(sc->sc_handshake_ok == 0))
1331 /* do we need more header? */
1332 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1333 data = page_address(sc->sc_page) + sc->sc_page_off;
1334 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1335 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1337 sc->sc_page_off += ret;
1338 /* only swab incoming here.. we can
1339 * only get here once as we cross from
1340 * being under to over */
1341 if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1342 hdr = page_address(sc->sc_page);
1343 if (be16_to_cpu(hdr->data_len) >
1344 O2NET_MAX_PAYLOAD_BYTES)
1352 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1353 /* oof, still don't have a header */
1357 /* this was swabbed above when we first read it */
1358 hdr = page_address(sc->sc_page);
1360 msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1362 /* do we need more payload? */
1363 if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1364 /* need more payload */
1365 data = page_address(sc->sc_page) + sc->sc_page_off;
1366 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1368 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1370 sc->sc_page_off += ret;
1375 if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1376 /* we can only get here once, the first time we read
1377 * the payload.. so set ret to progress if the handler
1378 * works out. after calling this the message is toast */
1379 ret = o2net_process_message(sc, hdr);
1382 sc->sc_page_off = 0;
1386 sclog(sc, "ret = %d\n", ret);
1387 do_gettimeofday(&sc->sc_tv_advance_stop);
1391 /* this work func is triggerd by data ready. it reads until it can read no
1392 * more. it interprets 0, eof, as fatal. if data_ready hits while we're doing
1393 * our work the work struct will be marked and we'll be called again. */
1394 static void o2net_rx_until_empty(struct work_struct *work)
1396 struct o2net_sock_container *sc =
1397 container_of(work, struct o2net_sock_container, sc_rx_work);
1401 ret = o2net_advance_rx(sc);
1404 if (ret <= 0 && ret != -EAGAIN) {
1405 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1406 sclog(sc, "saw error %d, closing\n", ret);
1407 /* not permanent so read failed handshake can retry */
1408 o2net_ensure_shutdown(nn, sc, 0);
1414 static int o2net_set_nodelay(struct socket *sock)
1423 * Dear unsuspecting programmer,
1425 * Don't use sock_setsockopt() for SOL_TCP. It doesn't check its level
1426 * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1427 * silently turn into SO_DEBUG.
1430 * Keeper of hilariously fragile interfaces.
1432 ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1433 (char __user *)&val, sizeof(val));
1439 static void o2net_initialize_handshake(void)
1441 o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1442 O2HB_MAX_WRITE_TIMEOUT_MS);
1443 o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1444 o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1445 o2net_keepalive_delay());
1446 o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1447 o2net_reconnect_delay());
1450 /* ------------------------------------------------------------ */
1452 /* called when a connect completes and after a sock is accepted. the
1453 * rx path will see the response and mark the sc valid */
1454 static void o2net_sc_connect_completed(struct work_struct *work)
1456 struct o2net_sock_container *sc =
1457 container_of(work, struct o2net_sock_container,
1460 mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1461 (unsigned long long)O2NET_PROTOCOL_VERSION,
1462 (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1464 o2net_initialize_handshake();
1465 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1469 /* this is called as a work_struct func. */
1470 static void o2net_sc_send_keep_req(struct work_struct *work)
1472 struct o2net_sock_container *sc =
1473 container_of(work, struct o2net_sock_container,
1474 sc_keepalive_work.work);
1476 o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1480 /* socket shutdown does a del_timer_sync against this as it tears down.
1481 * we can't start this timer until we've got to the point in sc buildup
1482 * where shutdown is going to be involved */
1483 static void o2net_idle_timer(unsigned long data)
1485 struct o2net_sock_container *sc = (struct o2net_sock_container *)data;
1486 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1489 do_gettimeofday(&now);
1491 printk(KERN_NOTICE "o2net: connection to " SC_NODEF_FMT " has been idle for %u.%u "
1492 "seconds, shutting it down.\n", SC_NODEF_ARGS(sc),
1493 o2net_idle_timeout() / 1000,
1494 o2net_idle_timeout() % 1000);
1495 mlog(ML_NOTICE, "here are some times that might help debug the "
1496 "situation: (tmr %ld.%ld now %ld.%ld dr %ld.%ld adv "
1497 "%ld.%ld:%ld.%ld func (%08x:%u) %ld.%ld:%ld.%ld)\n",
1498 sc->sc_tv_timer.tv_sec, (long) sc->sc_tv_timer.tv_usec,
1499 now.tv_sec, (long) now.tv_usec,
1500 sc->sc_tv_data_ready.tv_sec, (long) sc->sc_tv_data_ready.tv_usec,
1501 sc->sc_tv_advance_start.tv_sec,
1502 (long) sc->sc_tv_advance_start.tv_usec,
1503 sc->sc_tv_advance_stop.tv_sec,
1504 (long) sc->sc_tv_advance_stop.tv_usec,
1505 sc->sc_msg_key, sc->sc_msg_type,
1506 sc->sc_tv_func_start.tv_sec, (long) sc->sc_tv_func_start.tv_usec,
1507 sc->sc_tv_func_stop.tv_sec, (long) sc->sc_tv_func_stop.tv_usec);
1510 * Initialize the nn_timeout so that the next connection attempt
1511 * will continue in o2net_start_connect.
1513 atomic_set(&nn->nn_timeout, 1);
1515 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1518 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1520 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1521 o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1522 msecs_to_jiffies(o2net_keepalive_delay()));
1523 do_gettimeofday(&sc->sc_tv_timer);
1524 mod_timer(&sc->sc_idle_timeout,
1525 jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1528 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1530 /* Only push out an existing timer */
1531 if (timer_pending(&sc->sc_idle_timeout))
1532 o2net_sc_reset_idle_timer(sc);
1535 /* this work func is kicked whenever a path sets the nn state which doesn't
1536 * have valid set. This includes seeing hb come up, losing a connection,
1537 * having a connect attempt fail, etc. This centralizes the logic which decides
1538 * if a connect attempt should be made or if we should give up and all future
1539 * transmit attempts should fail */
1540 static void o2net_start_connect(struct work_struct *work)
1542 struct o2net_node *nn =
1543 container_of(work, struct o2net_node, nn_connect_work.work);
1544 struct o2net_sock_container *sc = NULL;
1545 struct o2nm_node *node = NULL, *mynode = NULL;
1546 struct socket *sock = NULL;
1547 struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1549 unsigned int timeout;
1551 /* if we're greater we initiate tx, otherwise we accept */
1552 if (o2nm_this_node() <= o2net_num_from_nn(nn))
1555 /* watch for racing with tearing a node down */
1556 node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1562 mynode = o2nm_get_node_by_num(o2nm_this_node());
1563 if (mynode == NULL) {
1568 spin_lock(&nn->nn_lock);
1570 * see if we already have one pending or have given up.
1571 * For nn_timeout, it is set when we close the connection
1572 * because of the idle time out. So it means that we have
1573 * at least connected to that node successfully once,
1574 * now try to connect to it again.
1576 timeout = atomic_read(&nn->nn_timeout);
1577 stop = (nn->nn_sc ||
1578 (nn->nn_persistent_error &&
1579 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1580 spin_unlock(&nn->nn_lock);
1584 nn->nn_last_connect_attempt = jiffies;
1586 sc = sc_alloc(node);
1588 mlog(0, "couldn't allocate sc\n");
1593 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1595 mlog(0, "can't create socket: %d\n", ret);
1598 sc->sc_sock = sock; /* freed by sc_kref_release */
1600 sock->sk->sk_allocation = GFP_ATOMIC;
1602 myaddr.sin_family = AF_INET;
1603 myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1604 myaddr.sin_port = htons(0); /* any port */
1606 ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1609 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1610 ret, &mynode->nd_ipv4_address);
1614 ret = o2net_set_nodelay(sc->sc_sock);
1616 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1620 o2net_register_callbacks(sc->sc_sock->sk, sc);
1622 spin_lock(&nn->nn_lock);
1623 /* handshake completion will set nn->nn_sc_valid */
1624 o2net_set_nn_state(nn, sc, 0, 0);
1625 spin_unlock(&nn->nn_lock);
1627 remoteaddr.sin_family = AF_INET;
1628 remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1629 remoteaddr.sin_port = node->nd_ipv4_port;
1631 ret = sc->sc_sock->ops->connect(sc->sc_sock,
1632 (struct sockaddr *)&remoteaddr,
1635 if (ret == -EINPROGRESS)
1640 mlog(ML_NOTICE, "connect attempt to " SC_NODEF_FMT " failed "
1641 "with errno %d\n", SC_NODEF_ARGS(sc), ret);
1642 /* 0 err so that another will be queued and attempted
1643 * from set_nn_state */
1645 o2net_ensure_shutdown(nn, sc, 0);
1650 o2nm_node_put(node);
1652 o2nm_node_put(mynode);
1657 static void o2net_connect_expired(struct work_struct *work)
1659 struct o2net_node *nn =
1660 container_of(work, struct o2net_node, nn_connect_expired.work);
1662 spin_lock(&nn->nn_lock);
1663 if (!nn->nn_sc_valid) {
1664 mlog(ML_ERROR, "no connection established with node %u after "
1665 "%u.%u seconds, giving up and returning errors.\n",
1666 o2net_num_from_nn(nn),
1667 o2net_idle_timeout() / 1000,
1668 o2net_idle_timeout() % 1000);
1670 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1672 spin_unlock(&nn->nn_lock);
1675 static void o2net_still_up(struct work_struct *work)
1677 struct o2net_node *nn =
1678 container_of(work, struct o2net_node, nn_still_up.work);
1680 o2quo_hb_still_up(o2net_num_from_nn(nn));
1683 /* ------------------------------------------------------------ */
1685 void o2net_disconnect_node(struct o2nm_node *node)
1687 struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1689 /* don't reconnect until it's heartbeating again */
1690 spin_lock(&nn->nn_lock);
1691 atomic_set(&nn->nn_timeout, 0);
1692 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1693 spin_unlock(&nn->nn_lock);
1696 cancel_delayed_work(&nn->nn_connect_expired);
1697 cancel_delayed_work(&nn->nn_connect_work);
1698 cancel_delayed_work(&nn->nn_still_up);
1699 flush_workqueue(o2net_wq);
1703 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1706 o2quo_hb_down(node_num);
1711 if (node_num != o2nm_this_node())
1712 o2net_disconnect_node(node);
1714 BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1717 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1720 struct o2net_node *nn = o2net_nn_from_num(node_num);
1722 o2quo_hb_up(node_num);
1726 /* ensure an immediate connect attempt */
1727 nn->nn_last_connect_attempt = jiffies -
1728 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1730 if (node_num != o2nm_this_node()) {
1731 /* believe it or not, accept and node hearbeating testing
1732 * can succeed for this node before we got here.. so
1733 * only use set_nn_state to clear the persistent error
1734 * if that hasn't already happened */
1735 spin_lock(&nn->nn_lock);
1736 atomic_set(&nn->nn_timeout, 0);
1737 if (nn->nn_persistent_error)
1738 o2net_set_nn_state(nn, NULL, 0, 0);
1739 spin_unlock(&nn->nn_lock);
1743 void o2net_unregister_hb_callbacks(void)
1745 o2hb_unregister_callback(NULL, &o2net_hb_up);
1746 o2hb_unregister_callback(NULL, &o2net_hb_down);
1749 int o2net_register_hb_callbacks(void)
1753 o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1754 o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1755 o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1756 o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1758 ret = o2hb_register_callback(NULL, &o2net_hb_up);
1760 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1763 o2net_unregister_hb_callbacks();
1768 /* ------------------------------------------------------------ */
1770 static int o2net_accept_one(struct socket *sock)
1773 struct sockaddr_in sin;
1774 struct socket *new_sock = NULL;
1775 struct o2nm_node *node = NULL;
1776 struct o2nm_node *local_node = NULL;
1777 struct o2net_sock_container *sc = NULL;
1778 struct o2net_node *nn;
1780 BUG_ON(sock == NULL);
1781 ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1782 sock->sk->sk_protocol, &new_sock);
1786 new_sock->type = sock->type;
1787 new_sock->ops = sock->ops;
1788 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1792 new_sock->sk->sk_allocation = GFP_ATOMIC;
1794 ret = o2net_set_nodelay(new_sock);
1796 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1801 ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1806 node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1808 mlog(ML_NOTICE, "attempt to connect from unknown node at %pI4:%d\n",
1809 &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1814 if (o2nm_this_node() >= node->nd_num) {
1815 local_node = o2nm_get_node_by_num(o2nm_this_node());
1816 mlog(ML_NOTICE, "unexpected connect attempt seen at node '%s' ("
1817 "%u, %pI4:%d) from node '%s' (%u, %pI4:%d)\n",
1818 local_node->nd_name, local_node->nd_num,
1819 &(local_node->nd_ipv4_address),
1820 ntohs(local_node->nd_ipv4_port),
1821 node->nd_name, node->nd_num, &sin.sin_addr.s_addr,
1822 ntohs(sin.sin_port));
1827 /* this happens all the time when the other node sees our heartbeat
1828 * and tries to connect before we see their heartbeat */
1829 if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1830 mlog(ML_CONN, "attempt to connect from node '%s' at "
1831 "%pI4:%d but it isn't heartbeating\n",
1832 node->nd_name, &sin.sin_addr.s_addr,
1833 ntohs(sin.sin_port));
1838 nn = o2net_nn_from_num(node->nd_num);
1840 spin_lock(&nn->nn_lock);
1845 spin_unlock(&nn->nn_lock);
1847 mlog(ML_NOTICE, "attempt to connect from node '%s' at "
1848 "%pI4:%d but it already has an open connection\n",
1849 node->nd_name, &sin.sin_addr.s_addr,
1850 ntohs(sin.sin_port));
1854 sc = sc_alloc(node);
1860 sc->sc_sock = new_sock;
1863 spin_lock(&nn->nn_lock);
1864 atomic_set(&nn->nn_timeout, 0);
1865 o2net_set_nn_state(nn, sc, 0, 0);
1866 spin_unlock(&nn->nn_lock);
1868 o2net_register_callbacks(sc->sc_sock->sk, sc);
1869 o2net_sc_queue_work(sc, &sc->sc_rx_work);
1871 o2net_initialize_handshake();
1872 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1876 sock_release(new_sock);
1878 o2nm_node_put(node);
1880 o2nm_node_put(local_node);
1886 static void o2net_accept_many(struct work_struct *work)
1888 struct socket *sock = o2net_listen_sock;
1889 while (o2net_accept_one(sock) == 0)
1893 static void o2net_listen_data_ready(struct sock *sk, int bytes)
1895 void (*ready)(struct sock *sk, int bytes);
1897 read_lock(&sk->sk_callback_lock);
1898 ready = sk->sk_user_data;
1899 if (ready == NULL) { /* check for teardown race */
1900 ready = sk->sk_data_ready;
1904 /* ->sk_data_ready is also called for a newly established child socket
1905 * before it has been accepted and the acceptor has set up their
1906 * data_ready.. we only want to queue listen work for our listening
1908 if (sk->sk_state == TCP_LISTEN) {
1909 mlog(ML_TCP, "bytes: %d\n", bytes);
1910 queue_work(o2net_wq, &o2net_listen_work);
1914 read_unlock(&sk->sk_callback_lock);
1918 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1920 struct socket *sock = NULL;
1922 struct sockaddr_in sin = {
1923 .sin_family = PF_INET,
1924 .sin_addr = { .s_addr = addr },
1928 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1930 mlog(ML_ERROR, "unable to create socket, ret=%d\n", ret);
1934 sock->sk->sk_allocation = GFP_ATOMIC;
1936 write_lock_bh(&sock->sk->sk_callback_lock);
1937 sock->sk->sk_user_data = sock->sk->sk_data_ready;
1938 sock->sk->sk_data_ready = o2net_listen_data_ready;
1939 write_unlock_bh(&sock->sk->sk_callback_lock);
1941 o2net_listen_sock = sock;
1942 INIT_WORK(&o2net_listen_work, o2net_accept_many);
1944 sock->sk->sk_reuse = 1;
1945 ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
1947 mlog(ML_ERROR, "unable to bind socket at %pI4:%u, "
1948 "ret=%d\n", &addr, ntohs(port), ret);
1952 ret = sock->ops->listen(sock, 64);
1954 mlog(ML_ERROR, "unable to listen on %pI4:%u, ret=%d\n",
1955 &addr, ntohs(port), ret);
1960 o2net_listen_sock = NULL;
1968 * called from node manager when we should bring up our network listening
1969 * socket. node manager handles all the serialization to only call this
1970 * once and to match it with o2net_stop_listening(). note,
1971 * o2nm_this_node() doesn't work yet as we're being called while it
1974 int o2net_start_listening(struct o2nm_node *node)
1978 BUG_ON(o2net_wq != NULL);
1979 BUG_ON(o2net_listen_sock != NULL);
1981 mlog(ML_KTHREAD, "starting o2net thread...\n");
1982 o2net_wq = create_singlethread_workqueue("o2net");
1983 if (o2net_wq == NULL) {
1984 mlog(ML_ERROR, "unable to launch o2net thread\n");
1985 return -ENOMEM; /* ? */
1988 ret = o2net_open_listening_sock(node->nd_ipv4_address,
1989 node->nd_ipv4_port);
1991 destroy_workqueue(o2net_wq);
1994 o2quo_conn_up(node->nd_num);
1999 /* again, o2nm_this_node() doesn't work here as we're involved in
2000 * tearing it down */
2001 void o2net_stop_listening(struct o2nm_node *node)
2003 struct socket *sock = o2net_listen_sock;
2006 BUG_ON(o2net_wq == NULL);
2007 BUG_ON(o2net_listen_sock == NULL);
2009 /* stop the listening socket from generating work */
2010 write_lock_bh(&sock->sk->sk_callback_lock);
2011 sock->sk->sk_data_ready = sock->sk->sk_user_data;
2012 sock->sk->sk_user_data = NULL;
2013 write_unlock_bh(&sock->sk->sk_callback_lock);
2015 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2016 struct o2nm_node *node = o2nm_get_node_by_num(i);
2018 o2net_disconnect_node(node);
2019 o2nm_node_put(node);
2023 /* finish all work and tear down the work queue */
2024 mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2025 destroy_workqueue(o2net_wq);
2028 sock_release(o2net_listen_sock);
2029 o2net_listen_sock = NULL;
2031 o2quo_conn_err(node->nd_num);
2034 /* ------------------------------------------------------------ */
2036 int o2net_init(void)
2042 if (o2net_debugfs_init())
2045 o2net_hand = kzalloc(sizeof(struct o2net_handshake), GFP_KERNEL);
2046 o2net_keep_req = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2047 o2net_keep_resp = kzalloc(sizeof(struct o2net_msg), GFP_KERNEL);
2048 if (!o2net_hand || !o2net_keep_req || !o2net_keep_resp) {
2050 kfree(o2net_keep_req);
2051 kfree(o2net_keep_resp);
2055 o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2056 o2net_hand->connector_id = cpu_to_be64(1);
2058 o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2059 o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2061 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2062 struct o2net_node *nn = o2net_nn_from_num(i);
2064 atomic_set(&nn->nn_timeout, 0);
2065 spin_lock_init(&nn->nn_lock);
2066 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2067 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2068 o2net_connect_expired);
2069 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2070 /* until we see hb from a node we'll return einval */
2071 nn->nn_persistent_error = -ENOTCONN;
2072 init_waitqueue_head(&nn->nn_sc_wq);
2073 idr_init(&nn->nn_status_idr);
2074 INIT_LIST_HEAD(&nn->nn_status_list);
2080 void o2net_exit(void)
2084 kfree(o2net_keep_req);
2085 kfree(o2net_keep_resp);
2086 o2net_debugfs_exit();