1 // SPDX-License-Identifier: GPL-2.0-or-later
4 * Copyright (C) 2004 Oracle. All rights reserved.
8 * Callers for this were originally written against a very simple synchronus
9 * API. This implementation reflects those simple callers. Some day I'm sure
10 * we'll need to move to a more robust posting/callback mechanism.
12 * Transmit calls pass in kernel virtual addresses and block copying this into
13 * the socket's tx buffers via a usual blocking sendmsg. They'll block waiting
14 * for a failed socket to timeout. TX callers can also pass in a poniter to an
15 * 'int' which gets filled with an errno off the wire in response to the
18 * Handlers for unsolicited messages are registered. Each socket has a page
19 * that incoming data is copied into. First the header, then the data.
20 * Handlers are called from only one thread with a reference to this per-socket
21 * page. This page is destroyed after the handler call, so it can't be
22 * referenced beyond the call. Handlers may block but are discouraged from
25 * Any framing errors (bad magic, large payload lengths) close a connection.
27 * Our sock_container holds the state we associate with a socket. It's current
28 * framing state is held there as well as the refcounting we do around when it
29 * is safe to tear down the socket. The socket is only finally torn down from
30 * the container when the container loses all of its references -- so as long
31 * as you hold a ref on the container you can trust that the socket is valid
32 * for use with kernel socket APIs.
34 * Connections are initiated between a pair of nodes when the node with the
35 * higher node number gets a heartbeat callback which indicates that the lower
36 * numbered node has started heartbeating. The lower numbered node is passive
37 * and only accepts the connection if the higher numbered node is heartbeating.
40 #include <linux/kernel.h>
41 #include <linux/sched/mm.h>
42 #include <linux/jiffies.h>
43 #include <linux/slab.h>
44 #include <linux/idr.h>
45 #include <linux/kref.h>
46 #include <linux/net.h>
47 #include <linux/export.h>
49 #include <trace/events/sock.h>
51 #include <linux/uaccess.h>
53 #include "heartbeat.h"
55 #include "nodemanager.h"
56 #define MLOG_MASK_PREFIX ML_TCP
60 #include "tcp_internal.h"
62 #define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
63 #define SC_NODEF_ARGS(sc) sc->sc_node->nd_name, sc->sc_node->nd_num, \
64 &sc->sc_node->nd_ipv4_address, \
65 ntohs(sc->sc_node->nd_ipv4_port)
68 * In the following two log macros, the whitespace after the ',' just
69 * before ##args is intentional. Otherwise, gcc 2.95 will eat the
70 * previous token if args expands to nothing.
72 #define msglog(hdr, fmt, args...) do { \
73 typeof(hdr) __hdr = (hdr); \
74 mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d " \
75 "key %08x num %u] " fmt, \
76 be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len), \
77 be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status), \
78 be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key), \
79 be32_to_cpu(__hdr->msg_num) , ##args); \
82 #define sclog(sc, fmt, args...) do { \
83 typeof(sc) __sc = (sc); \
84 mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p " \
85 "pg_off %zu] " fmt, __sc, \
86 kref_read(&__sc->sc_kref), __sc->sc_sock, \
87 __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off , \
91 static DEFINE_RWLOCK(o2net_handler_lock);
92 static struct rb_root o2net_handler_tree = RB_ROOT;
94 static struct o2net_node o2net_nodes[O2NM_MAX_NODES];
96 /* XXX someday we'll need better accounting */
97 static struct socket *o2net_listen_sock;
100 * listen work is only queued by the listening socket callbacks on the
101 * o2net_wq. teardown detaches the callbacks before destroying the workqueue.
102 * quorum work is queued as sock containers are shutdown.. stop_listening
103 * tears down all the node's sock containers, preventing future shutdowns
104 * and queued quroum work, before canceling delayed quorum work and
105 * destroying the work queue.
107 static struct workqueue_struct *o2net_wq;
108 static struct work_struct o2net_listen_work;
110 static struct o2hb_callback_func o2net_hb_up, o2net_hb_down;
111 #define O2NET_HB_PRI 0x1
113 static struct o2net_handshake *o2net_hand;
114 static struct o2net_msg *o2net_keep_req, *o2net_keep_resp;
116 static int o2net_sys_err_translations[O2NET_ERR_MAX] =
117 {[O2NET_ERR_NONE] = 0,
118 [O2NET_ERR_NO_HNDLR] = -ENOPROTOOPT,
119 [O2NET_ERR_OVERFLOW] = -EOVERFLOW,
120 [O2NET_ERR_DIED] = -EHOSTDOWN,};
122 /* can't quite avoid *all* internal declarations :/ */
123 static void o2net_sc_connect_completed(struct work_struct *work);
124 static void o2net_rx_until_empty(struct work_struct *work);
125 static void o2net_shutdown_sc(struct work_struct *work);
126 static void o2net_listen_data_ready(struct sock *sk);
127 static void o2net_sc_send_keep_req(struct work_struct *work);
128 static void o2net_idle_timer(struct timer_list *t);
129 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc);
130 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc);
132 #ifdef CONFIG_DEBUG_FS
133 static void o2net_init_nst(struct o2net_send_tracking *nst, u32 msgtype,
134 u32 msgkey, struct task_struct *task, u8 node)
136 INIT_LIST_HEAD(&nst->st_net_debug_item);
138 nst->st_msg_type = msgtype;
139 nst->st_msg_key = msgkey;
143 static inline void o2net_set_nst_sock_time(struct o2net_send_tracking *nst)
145 nst->st_sock_time = ktime_get();
148 static inline void o2net_set_nst_send_time(struct o2net_send_tracking *nst)
150 nst->st_send_time = ktime_get();
153 static inline void o2net_set_nst_status_time(struct o2net_send_tracking *nst)
155 nst->st_status_time = ktime_get();
158 static inline void o2net_set_nst_sock_container(struct o2net_send_tracking *nst,
159 struct o2net_sock_container *sc)
164 static inline void o2net_set_nst_msg_id(struct o2net_send_tracking *nst,
170 static inline void o2net_set_sock_timer(struct o2net_sock_container *sc)
172 sc->sc_tv_timer = ktime_get();
175 static inline void o2net_set_data_ready_time(struct o2net_sock_container *sc)
177 sc->sc_tv_data_ready = ktime_get();
180 static inline void o2net_set_advance_start_time(struct o2net_sock_container *sc)
182 sc->sc_tv_advance_start = ktime_get();
185 static inline void o2net_set_advance_stop_time(struct o2net_sock_container *sc)
187 sc->sc_tv_advance_stop = ktime_get();
190 static inline void o2net_set_func_start_time(struct o2net_sock_container *sc)
192 sc->sc_tv_func_start = ktime_get();
195 static inline void o2net_set_func_stop_time(struct o2net_sock_container *sc)
197 sc->sc_tv_func_stop = ktime_get();
200 #else /* CONFIG_DEBUG_FS */
201 # define o2net_init_nst(a, b, c, d, e)
202 # define o2net_set_nst_sock_time(a)
203 # define o2net_set_nst_send_time(a)
204 # define o2net_set_nst_status_time(a)
205 # define o2net_set_nst_sock_container(a, b)
206 # define o2net_set_nst_msg_id(a, b)
207 # define o2net_set_sock_timer(a)
208 # define o2net_set_data_ready_time(a)
209 # define o2net_set_advance_start_time(a)
210 # define o2net_set_advance_stop_time(a)
211 # define o2net_set_func_start_time(a)
212 # define o2net_set_func_stop_time(a)
213 #endif /* CONFIG_DEBUG_FS */
215 #ifdef CONFIG_OCFS2_FS_STATS
216 static ktime_t o2net_get_func_run_time(struct o2net_sock_container *sc)
218 return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
221 static void o2net_update_send_stats(struct o2net_send_tracking *nst,
222 struct o2net_sock_container *sc)
224 sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
225 ktime_sub(ktime_get(),
226 nst->st_status_time));
227 sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
228 ktime_sub(nst->st_status_time,
230 sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
231 ktime_sub(nst->st_send_time,
236 static void o2net_update_recv_stats(struct o2net_sock_container *sc)
238 sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
239 o2net_get_func_run_time(sc));
245 # define o2net_update_send_stats(a, b)
247 # define o2net_update_recv_stats(sc)
249 #endif /* CONFIG_OCFS2_FS_STATS */
251 static inline unsigned int o2net_reconnect_delay(void)
253 return o2nm_single_cluster->cl_reconnect_delay_ms;
256 static inline unsigned int o2net_keepalive_delay(void)
258 return o2nm_single_cluster->cl_keepalive_delay_ms;
261 static inline unsigned int o2net_idle_timeout(void)
263 return o2nm_single_cluster->cl_idle_timeout_ms;
266 static inline int o2net_sys_err_to_errno(enum o2net_system_error err)
269 BUG_ON(err >= O2NET_ERR_MAX);
270 trans = o2net_sys_err_translations[err];
272 /* Just in case we mess up the translation table above */
273 BUG_ON(err != O2NET_ERR_NONE && trans == 0);
277 static struct o2net_node * o2net_nn_from_num(u8 node_num)
279 BUG_ON(node_num >= ARRAY_SIZE(o2net_nodes));
280 return &o2net_nodes[node_num];
283 static u8 o2net_num_from_nn(struct o2net_node *nn)
286 return nn - o2net_nodes;
289 /* ------------------------------------------------------------ */
291 static int o2net_prep_nsw(struct o2net_node *nn, struct o2net_status_wait *nsw)
295 spin_lock(&nn->nn_lock);
296 ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
299 list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
301 spin_unlock(&nn->nn_lock);
305 init_waitqueue_head(&nsw->ns_wq);
306 nsw->ns_sys_status = O2NET_ERR_NONE;
311 static void o2net_complete_nsw_locked(struct o2net_node *nn,
312 struct o2net_status_wait *nsw,
313 enum o2net_system_error sys_status,
316 assert_spin_locked(&nn->nn_lock);
318 if (!list_empty(&nsw->ns_node_item)) {
319 list_del_init(&nsw->ns_node_item);
320 nsw->ns_sys_status = sys_status;
321 nsw->ns_status = status;
322 idr_remove(&nn->nn_status_idr, nsw->ns_id);
323 wake_up(&nsw->ns_wq);
327 static void o2net_complete_nsw(struct o2net_node *nn,
328 struct o2net_status_wait *nsw,
329 u64 id, enum o2net_system_error sys_status,
332 spin_lock(&nn->nn_lock);
337 nsw = idr_find(&nn->nn_status_idr, id);
342 o2net_complete_nsw_locked(nn, nsw, sys_status, status);
345 spin_unlock(&nn->nn_lock);
349 static void o2net_complete_nodes_nsw(struct o2net_node *nn)
351 struct o2net_status_wait *nsw, *tmp;
352 unsigned int num_kills = 0;
354 assert_spin_locked(&nn->nn_lock);
356 list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
357 o2net_complete_nsw_locked(nn, nsw, O2NET_ERR_DIED, 0);
361 mlog(0, "completed %d messages for node %u\n", num_kills,
362 o2net_num_from_nn(nn));
365 static int o2net_nsw_completed(struct o2net_node *nn,
366 struct o2net_status_wait *nsw)
369 spin_lock(&nn->nn_lock);
370 completed = list_empty(&nsw->ns_node_item);
371 spin_unlock(&nn->nn_lock);
375 /* ------------------------------------------------------------ */
377 static void sc_kref_release(struct kref *kref)
379 struct o2net_sock_container *sc = container_of(kref,
380 struct o2net_sock_container, sc_kref);
381 BUG_ON(timer_pending(&sc->sc_idle_timeout));
383 sclog(sc, "releasing\n");
386 sock_release(sc->sc_sock);
390 o2nm_undepend_item(&sc->sc_node->nd_item);
391 o2nm_node_put(sc->sc_node);
394 o2net_debug_del_sc(sc);
397 __free_page(sc->sc_page);
401 static void sc_put(struct o2net_sock_container *sc)
404 kref_put(&sc->sc_kref, sc_kref_release);
406 static void sc_get(struct o2net_sock_container *sc)
409 kref_get(&sc->sc_kref);
411 static struct o2net_sock_container *sc_alloc(struct o2nm_node *node)
413 struct o2net_sock_container *sc, *ret = NULL;
414 struct page *page = NULL;
417 page = alloc_page(GFP_NOFS);
418 sc = kzalloc(sizeof(*sc), GFP_NOFS);
419 if (sc == NULL || page == NULL)
422 kref_init(&sc->sc_kref);
426 /* pin the node item of the remote node */
427 status = o2nm_depend_item(&node->nd_item);
433 INIT_WORK(&sc->sc_connect_work, o2net_sc_connect_completed);
434 INIT_WORK(&sc->sc_rx_work, o2net_rx_until_empty);
435 INIT_WORK(&sc->sc_shutdown_work, o2net_shutdown_sc);
436 INIT_DELAYED_WORK(&sc->sc_keepalive_work, o2net_sc_send_keep_req);
438 timer_setup(&sc->sc_idle_timeout, o2net_idle_timer, 0);
440 sclog(sc, "alloced\n");
444 o2net_debug_add_sc(sc);
456 /* ------------------------------------------------------------ */
458 static void o2net_sc_queue_work(struct o2net_sock_container *sc,
459 struct work_struct *work)
462 if (!queue_work(o2net_wq, work))
465 static void o2net_sc_queue_delayed_work(struct o2net_sock_container *sc,
466 struct delayed_work *work,
470 if (!queue_delayed_work(o2net_wq, work, delay))
473 static void o2net_sc_cancel_delayed_work(struct o2net_sock_container *sc,
474 struct delayed_work *work)
476 if (cancel_delayed_work(work))
480 static atomic_t o2net_connected_peers = ATOMIC_INIT(0);
482 int o2net_num_connected_peers(void)
484 return atomic_read(&o2net_connected_peers);
487 static void o2net_set_nn_state(struct o2net_node *nn,
488 struct o2net_sock_container *sc,
489 unsigned valid, int err)
491 int was_valid = nn->nn_sc_valid;
492 int was_err = nn->nn_persistent_error;
493 struct o2net_sock_container *old_sc = nn->nn_sc;
495 assert_spin_locked(&nn->nn_lock);
498 atomic_dec(&o2net_connected_peers);
499 else if (!old_sc && sc)
500 atomic_inc(&o2net_connected_peers);
502 /* the node num comparison and single connect/accept path should stop
503 * an non-null sc from being overwritten with another */
504 BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
505 mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
506 mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
508 if (was_valid && !valid && err == 0)
511 mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
512 o2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
513 nn->nn_persistent_error, err);
516 nn->nn_sc_valid = valid ? 1 : 0;
517 nn->nn_persistent_error = err;
519 /* mirrors o2net_tx_can_proceed() */
520 if (nn->nn_persistent_error || nn->nn_sc_valid)
521 wake_up(&nn->nn_sc_wq);
523 if (was_valid && !was_err && nn->nn_persistent_error) {
524 o2quo_conn_err(o2net_num_from_nn(nn));
525 queue_delayed_work(o2net_wq, &nn->nn_still_up,
526 msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
529 if (was_valid && !valid) {
531 printk(KERN_NOTICE "o2net: No longer connected to "
532 SC_NODEF_FMT "\n", SC_NODEF_ARGS(old_sc));
533 o2net_complete_nodes_nsw(nn);
536 if (!was_valid && valid) {
537 o2quo_conn_up(o2net_num_from_nn(nn));
538 cancel_delayed_work(&nn->nn_connect_expired);
539 printk(KERN_NOTICE "o2net: %s " SC_NODEF_FMT "\n",
540 o2nm_this_node() > sc->sc_node->nd_num ?
541 "Connected to" : "Accepted connection from",
545 /* trigger the connecting worker func as long as we're not valid,
546 * it will back off if it shouldn't connect. This can be called
547 * from node config teardown and so needs to be careful about
548 * the work queue actually being up. */
549 if (!valid && o2net_wq) {
551 /* delay if we're within a RECONNECT_DELAY of the
553 delay = (nn->nn_last_connect_attempt +
554 msecs_to_jiffies(o2net_reconnect_delay()))
556 if (delay > msecs_to_jiffies(o2net_reconnect_delay()))
558 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
559 queue_delayed_work(o2net_wq, &nn->nn_connect_work, delay);
562 * Delay the expired work after idle timeout.
564 * We might have lots of failed connection attempts that run
565 * through here but we only cancel the connect_expired work when
566 * a connection attempt succeeds. So only the first enqueue of
567 * the connect_expired work will do anything. The rest will see
568 * that it's already queued and do nothing.
570 delay += msecs_to_jiffies(o2net_idle_timeout());
571 queue_delayed_work(o2net_wq, &nn->nn_connect_expired, delay);
574 /* keep track of the nn's sc ref for the caller */
575 if ((old_sc == NULL) && sc)
577 if (old_sc && (old_sc != sc)) {
578 o2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
583 /* see o2net_register_callbacks() */
584 static void o2net_data_ready(struct sock *sk)
586 void (*ready)(struct sock *sk);
587 struct o2net_sock_container *sc;
589 trace_sk_data_ready(sk);
591 read_lock_bh(&sk->sk_callback_lock);
592 sc = sk->sk_user_data;
594 sclog(sc, "data_ready hit\n");
595 o2net_set_data_ready_time(sc);
596 o2net_sc_queue_work(sc, &sc->sc_rx_work);
597 ready = sc->sc_data_ready;
599 ready = sk->sk_data_ready;
601 read_unlock_bh(&sk->sk_callback_lock);
606 /* see o2net_register_callbacks() */
607 static void o2net_state_change(struct sock *sk)
609 void (*state_change)(struct sock *sk);
610 struct o2net_sock_container *sc;
612 read_lock_bh(&sk->sk_callback_lock);
613 sc = sk->sk_user_data;
615 state_change = sk->sk_state_change;
619 sclog(sc, "state_change to %d\n", sk->sk_state);
621 state_change = sc->sc_state_change;
623 switch(sk->sk_state) {
624 /* ignore connecting sockets as they make progress */
628 case TCP_ESTABLISHED:
629 o2net_sc_queue_work(sc, &sc->sc_connect_work);
632 printk(KERN_INFO "o2net: Connection to " SC_NODEF_FMT
633 " shutdown, state %d\n",
634 SC_NODEF_ARGS(sc), sk->sk_state);
635 o2net_sc_queue_work(sc, &sc->sc_shutdown_work);
639 read_unlock_bh(&sk->sk_callback_lock);
644 * we register callbacks so we can queue work on events before calling
645 * the original callbacks. our callbacks our careful to test user_data
646 * to discover when they've reaced with o2net_unregister_callbacks().
648 static void o2net_register_callbacks(struct sock *sk,
649 struct o2net_sock_container *sc)
651 write_lock_bh(&sk->sk_callback_lock);
653 /* accepted sockets inherit the old listen socket data ready */
654 if (sk->sk_data_ready == o2net_listen_data_ready) {
655 sk->sk_data_ready = sk->sk_user_data;
656 sk->sk_user_data = NULL;
659 BUG_ON(sk->sk_user_data != NULL);
660 sk->sk_user_data = sc;
663 sc->sc_data_ready = sk->sk_data_ready;
664 sc->sc_state_change = sk->sk_state_change;
665 sk->sk_data_ready = o2net_data_ready;
666 sk->sk_state_change = o2net_state_change;
668 mutex_init(&sc->sc_send_lock);
670 write_unlock_bh(&sk->sk_callback_lock);
673 static int o2net_unregister_callbacks(struct sock *sk,
674 struct o2net_sock_container *sc)
678 write_lock_bh(&sk->sk_callback_lock);
679 if (sk->sk_user_data == sc) {
681 sk->sk_user_data = NULL;
682 sk->sk_data_ready = sc->sc_data_ready;
683 sk->sk_state_change = sc->sc_state_change;
685 write_unlock_bh(&sk->sk_callback_lock);
691 * this is a little helper that is called by callers who have seen a problem
692 * with an sc and want to detach it from the nn if someone already hasn't beat
693 * them to it. if an error is given then the shutdown will be persistent
694 * and pending transmits will be canceled.
696 static void o2net_ensure_shutdown(struct o2net_node *nn,
697 struct o2net_sock_container *sc,
700 spin_lock(&nn->nn_lock);
702 o2net_set_nn_state(nn, NULL, 0, err);
703 spin_unlock(&nn->nn_lock);
707 * This work queue function performs the blocking parts of socket shutdown. A
708 * few paths lead here. set_nn_state will trigger this callback if it sees an
709 * sc detached from the nn. state_change will also trigger this callback
710 * directly when it sees errors. In that case we need to call set_nn_state
711 * ourselves as state_change couldn't get the nn_lock and call set_nn_state
714 static void o2net_shutdown_sc(struct work_struct *work)
716 struct o2net_sock_container *sc =
717 container_of(work, struct o2net_sock_container,
719 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
721 sclog(sc, "shutting down\n");
723 /* drop the callbacks ref and call shutdown only once */
724 if (o2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
725 /* we shouldn't flush as we're in the thread, the
726 * races with pending sc work structs are harmless */
727 del_timer_sync(&sc->sc_idle_timeout);
728 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
730 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
733 /* not fatal so failed connects before the other guy has our
734 * heartbeat can be retried */
735 o2net_ensure_shutdown(nn, sc, 0);
739 /* ------------------------------------------------------------ */
741 static int o2net_handler_cmp(struct o2net_msg_handler *nmh, u32 msg_type,
744 int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
747 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
752 static struct o2net_msg_handler *
753 o2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
754 struct rb_node **ret_parent)
756 struct rb_node **p = &o2net_handler_tree.rb_node;
757 struct rb_node *parent = NULL;
758 struct o2net_msg_handler *nmh, *ret = NULL;
763 nmh = rb_entry(parent, struct o2net_msg_handler, nh_node);
764 cmp = o2net_handler_cmp(nmh, msg_type, key);
778 if (ret_parent != NULL)
779 *ret_parent = parent;
784 static void o2net_handler_kref_release(struct kref *kref)
786 struct o2net_msg_handler *nmh;
787 nmh = container_of(kref, struct o2net_msg_handler, nh_kref);
792 static void o2net_handler_put(struct o2net_msg_handler *nmh)
794 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
797 /* max_len is protection for the handler func. incoming messages won't
798 * be given to the handler if their payload is longer than the max. */
799 int o2net_register_handler(u32 msg_type, u32 key, u32 max_len,
800 o2net_msg_handler_func *func, void *data,
801 o2net_post_msg_handler_func *post_func,
802 struct list_head *unreg_list)
804 struct o2net_msg_handler *nmh = NULL;
805 struct rb_node **p, *parent;
808 if (max_len > O2NET_MAX_PAYLOAD_BYTES) {
809 mlog(0, "max_len for message handler out of range: %u\n",
816 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
822 mlog(0, "no message handler provided: %u, %p\n",
828 nmh = kzalloc(sizeof(struct o2net_msg_handler), GFP_NOFS);
835 nmh->nh_func_data = data;
836 nmh->nh_post_func = post_func;
837 nmh->nh_msg_type = msg_type;
838 nmh->nh_max_len = max_len;
840 /* the tree and list get this ref.. they're both removed in
841 * unregister when this ref is dropped */
842 kref_init(&nmh->nh_kref);
843 INIT_LIST_HEAD(&nmh->nh_unregister_item);
845 write_lock(&o2net_handler_lock);
846 if (o2net_handler_tree_lookup(msg_type, key, &p, &parent))
849 rb_link_node(&nmh->nh_node, parent, p);
850 rb_insert_color(&nmh->nh_node, &o2net_handler_tree);
851 list_add_tail(&nmh->nh_unregister_item, unreg_list);
853 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
854 func, msg_type, key);
855 /* we've had some trouble with handlers seemingly vanishing. */
856 mlog_bug_on_msg(o2net_handler_tree_lookup(msg_type, key, &p,
858 "couldn't find handler we *just* registered "
859 "for type %u key %08x\n", msg_type, key);
861 write_unlock(&o2net_handler_lock);
869 EXPORT_SYMBOL_GPL(o2net_register_handler);
871 void o2net_unregister_handler_list(struct list_head *list)
873 struct o2net_msg_handler *nmh, *n;
875 write_lock(&o2net_handler_lock);
876 list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
877 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
878 nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
879 rb_erase(&nmh->nh_node, &o2net_handler_tree);
880 list_del_init(&nmh->nh_unregister_item);
881 kref_put(&nmh->nh_kref, o2net_handler_kref_release);
883 write_unlock(&o2net_handler_lock);
885 EXPORT_SYMBOL_GPL(o2net_unregister_handler_list);
887 static struct o2net_msg_handler *o2net_handler_get(u32 msg_type, u32 key)
889 struct o2net_msg_handler *nmh;
891 read_lock(&o2net_handler_lock);
892 nmh = o2net_handler_tree_lookup(msg_type, key, NULL, NULL);
894 kref_get(&nmh->nh_kref);
895 read_unlock(&o2net_handler_lock);
900 /* ------------------------------------------------------------ */
902 static int o2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
904 struct kvec vec = { .iov_len = len, .iov_base = data, };
905 struct msghdr msg = { .msg_flags = MSG_DONTWAIT, };
906 iov_iter_kvec(&msg.msg_iter, ITER_DEST, &vec, 1, len);
907 return sock_recvmsg(sock, &msg, MSG_DONTWAIT);
910 static int o2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
911 size_t veclen, size_t total)
914 struct msghdr msg = {.msg_flags = 0,};
921 ret = kernel_sendmsg(sock, &msg, vec, veclen, total);
922 if (likely(ret == total))
924 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret, total);
926 ret = -EPIPE; /* should be smarter, I bet */
928 mlog(0, "returning error: %d\n", ret);
932 static void o2net_sendpage(struct o2net_sock_container *sc,
933 void *virt, size_t size)
935 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
936 struct msghdr msg = {};
940 bvec_set_virt(&bv, virt, size);
941 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bv, 1, size);
944 msg.msg_flags = MSG_DONTWAIT | MSG_SPLICE_PAGES;
945 mutex_lock(&sc->sc_send_lock);
946 ret = sock_sendmsg(sc->sc_sock, &msg);
947 mutex_unlock(&sc->sc_send_lock);
951 if (ret == (ssize_t)-EAGAIN) {
952 mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
953 " returned EAGAIN\n", size, SC_NODEF_ARGS(sc));
957 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
958 " failed with %zd\n", size, SC_NODEF_ARGS(sc), ret);
959 o2net_ensure_shutdown(nn, sc, 0);
964 static void o2net_init_msg(struct o2net_msg *msg, u16 data_len, u16 msg_type, u32 key)
966 memset(msg, 0, sizeof(struct o2net_msg));
967 msg->magic = cpu_to_be16(O2NET_MSG_MAGIC);
968 msg->data_len = cpu_to_be16(data_len);
969 msg->msg_type = cpu_to_be16(msg_type);
970 msg->sys_status = cpu_to_be32(O2NET_ERR_NONE);
972 msg->key = cpu_to_be32(key);
975 static int o2net_tx_can_proceed(struct o2net_node *nn,
976 struct o2net_sock_container **sc_ret,
981 spin_lock(&nn->nn_lock);
982 if (nn->nn_persistent_error) {
985 *error = nn->nn_persistent_error;
986 } else if (nn->nn_sc_valid) {
987 kref_get(&nn->nn_sc->sc_kref);
993 spin_unlock(&nn->nn_lock);
998 /* Get a map of all nodes to which this node is currently connected to */
999 void o2net_fill_node_map(unsigned long *map, unsigned int bits)
1001 struct o2net_sock_container *sc;
1004 bitmap_zero(map, bits);
1005 for (node = 0; node < O2NM_MAX_NODES; ++node) {
1006 if (!o2net_tx_can_proceed(o2net_nn_from_num(node), &sc, &ret))
1014 EXPORT_SYMBOL_GPL(o2net_fill_node_map);
1016 int o2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1017 size_t caller_veclen, u8 target_node, int *status)
1020 struct o2net_msg *msg = NULL;
1021 size_t veclen, caller_bytes = 0;
1022 struct kvec *vec = NULL;
1023 struct o2net_sock_container *sc = NULL;
1024 struct o2net_node *nn = o2net_nn_from_num(target_node);
1025 struct o2net_status_wait nsw = {
1026 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1028 struct o2net_send_tracking nst;
1030 o2net_init_nst(&nst, msg_type, key, current, target_node);
1032 if (o2net_wq == NULL) {
1033 mlog(0, "attempt to tx without o2netd running\n");
1038 if (caller_veclen == 0) {
1039 mlog(0, "bad kvec array length\n");
1044 caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1045 if (caller_bytes > O2NET_MAX_PAYLOAD_BYTES) {
1046 mlog(0, "total payload len %zu too large\n", caller_bytes);
1051 if (target_node == o2nm_this_node()) {
1056 o2net_debug_add_nst(&nst);
1058 o2net_set_nst_sock_time(&nst);
1060 wait_event(nn->nn_sc_wq, o2net_tx_can_proceed(nn, &sc, &ret));
1064 o2net_set_nst_sock_container(&nst, sc);
1066 veclen = caller_veclen + 1;
1067 vec = kmalloc_array(veclen, sizeof(struct kvec), GFP_ATOMIC);
1069 mlog(0, "failed to %zu element kvec!\n", veclen);
1074 msg = kmalloc(sizeof(struct o2net_msg), GFP_ATOMIC);
1076 mlog(0, "failed to allocate a o2net_msg!\n");
1081 o2net_init_msg(msg, caller_bytes, msg_type, key);
1083 vec[0].iov_len = sizeof(struct o2net_msg);
1084 vec[0].iov_base = msg;
1085 memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1087 ret = o2net_prep_nsw(nn, &nsw);
1091 msg->msg_num = cpu_to_be32(nsw.ns_id);
1092 o2net_set_nst_msg_id(&nst, nsw.ns_id);
1094 o2net_set_nst_send_time(&nst);
1096 /* finally, convert the message header to network byte-order
1098 mutex_lock(&sc->sc_send_lock);
1099 ret = o2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1100 sizeof(struct o2net_msg) + caller_bytes);
1101 mutex_unlock(&sc->sc_send_lock);
1102 msglog(msg, "sending returned %d\n", ret);
1104 mlog(0, "error returned from o2net_send_tcp_msg=%d\n", ret);
1108 /* wait on other node's handler */
1109 o2net_set_nst_status_time(&nst);
1110 wait_event(nsw.ns_wq, o2net_nsw_completed(nn, &nsw));
1112 o2net_update_send_stats(&nst, sc);
1114 /* Note that we avoid overwriting the callers status return
1115 * variable if a system error was reported on the other
1116 * side. Callers beware. */
1117 ret = o2net_sys_err_to_errno(nsw.ns_sys_status);
1119 *status = nsw.ns_status;
1121 mlog(0, "woken, returning system status %d, user status %d\n",
1122 ret, nsw.ns_status);
1124 o2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1129 o2net_complete_nsw(nn, &nsw, 0, 0, 0);
1132 EXPORT_SYMBOL_GPL(o2net_send_message_vec);
1134 int o2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1135 u8 target_node, int *status)
1141 return o2net_send_message_vec(msg_type, key, &vec, 1,
1142 target_node, status);
1144 EXPORT_SYMBOL_GPL(o2net_send_message);
1146 static int o2net_send_status_magic(struct socket *sock, struct o2net_msg *hdr,
1147 enum o2net_system_error syserr, int err)
1151 .iov_len = sizeof(struct o2net_msg),
1154 BUG_ON(syserr >= O2NET_ERR_MAX);
1156 /* leave other fields intact from the incoming message, msg_num
1158 hdr->sys_status = cpu_to_be32(syserr);
1159 hdr->status = cpu_to_be32(err);
1160 hdr->magic = cpu_to_be16(O2NET_MSG_STATUS_MAGIC); // twiddle the magic
1163 msglog(hdr, "about to send status magic %d\n", err);
1164 /* hdr has been in host byteorder this whole time */
1165 return o2net_send_tcp_msg(sock, &vec, 1, sizeof(struct o2net_msg));
1168 /* this returns -errno if the header was unknown or too large, etc.
1169 * after this is called the buffer us reused for the next message */
1170 static int o2net_process_message(struct o2net_sock_container *sc,
1171 struct o2net_msg *hdr)
1173 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1174 int ret = 0, handler_status;
1175 enum o2net_system_error syserr;
1176 struct o2net_msg_handler *nmh = NULL;
1177 void *ret_data = NULL;
1179 msglog(hdr, "processing message\n");
1181 o2net_sc_postpone_idle(sc);
1183 switch(be16_to_cpu(hdr->magic)) {
1184 case O2NET_MSG_STATUS_MAGIC:
1185 /* special type for returning message status */
1186 o2net_complete_nsw(nn, NULL,
1187 be32_to_cpu(hdr->msg_num),
1188 be32_to_cpu(hdr->sys_status),
1189 be32_to_cpu(hdr->status));
1191 case O2NET_MSG_KEEP_REQ_MAGIC:
1192 o2net_sendpage(sc, o2net_keep_resp,
1193 sizeof(*o2net_keep_resp));
1195 case O2NET_MSG_KEEP_RESP_MAGIC:
1197 case O2NET_MSG_MAGIC:
1200 msglog(hdr, "bad magic\n");
1205 /* find a handler for it */
1207 nmh = o2net_handler_get(be16_to_cpu(hdr->msg_type),
1208 be32_to_cpu(hdr->key));
1210 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1211 be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1212 syserr = O2NET_ERR_NO_HNDLR;
1216 syserr = O2NET_ERR_NONE;
1218 if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1219 syserr = O2NET_ERR_OVERFLOW;
1221 if (syserr != O2NET_ERR_NONE)
1224 o2net_set_func_start_time(sc);
1225 sc->sc_msg_key = be32_to_cpu(hdr->key);
1226 sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1227 handler_status = (nmh->nh_func)(hdr, sizeof(struct o2net_msg) +
1228 be16_to_cpu(hdr->data_len),
1229 nmh->nh_func_data, &ret_data);
1230 o2net_set_func_stop_time(sc);
1232 o2net_update_recv_stats(sc);
1235 /* this destroys the hdr, so don't use it after this */
1236 mutex_lock(&sc->sc_send_lock);
1237 ret = o2net_send_status_magic(sc->sc_sock, hdr, syserr,
1239 mutex_unlock(&sc->sc_send_lock);
1241 mlog(0, "sending handler status %d, syserr %d returned %d\n",
1242 handler_status, syserr, ret);
1245 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1246 if (nmh->nh_post_func)
1247 (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1253 o2net_handler_put(nmh);
1257 static int o2net_check_handshake(struct o2net_sock_container *sc)
1259 struct o2net_handshake *hand = page_address(sc->sc_page);
1260 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1262 if (hand->protocol_version != cpu_to_be64(O2NET_PROTOCOL_VERSION)) {
1263 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " Advertised net "
1264 "protocol version %llu but %llu is required. "
1265 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1266 (unsigned long long)be64_to_cpu(hand->protocol_version),
1267 O2NET_PROTOCOL_VERSION);
1269 /* don't bother reconnecting if its the wrong version. */
1270 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1275 * Ensure timeouts are consistent with other nodes, otherwise
1276 * we can end up with one node thinking that the other must be down,
1277 * but isn't. This can ultimately cause corruption.
1279 if (be32_to_cpu(hand->o2net_idle_timeout_ms) !=
1280 o2net_idle_timeout()) {
1281 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a network "
1282 "idle timeout of %u ms, but we use %u ms locally. "
1283 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1284 be32_to_cpu(hand->o2net_idle_timeout_ms),
1285 o2net_idle_timeout());
1286 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1290 if (be32_to_cpu(hand->o2net_keepalive_delay_ms) !=
1291 o2net_keepalive_delay()) {
1292 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a keepalive "
1293 "delay of %u ms, but we use %u ms locally. "
1294 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1295 be32_to_cpu(hand->o2net_keepalive_delay_ms),
1296 o2net_keepalive_delay());
1297 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1301 if (be32_to_cpu(hand->o2hb_heartbeat_timeout_ms) !=
1302 O2HB_MAX_WRITE_TIMEOUT_MS) {
1303 printk(KERN_NOTICE "o2net: " SC_NODEF_FMT " uses a heartbeat "
1304 "timeout of %u ms, but we use %u ms locally. "
1305 "Disconnecting.\n", SC_NODEF_ARGS(sc),
1306 be32_to_cpu(hand->o2hb_heartbeat_timeout_ms),
1307 O2HB_MAX_WRITE_TIMEOUT_MS);
1308 o2net_ensure_shutdown(nn, sc, -ENOTCONN);
1312 sc->sc_handshake_ok = 1;
1314 spin_lock(&nn->nn_lock);
1315 /* set valid and queue the idle timers only if it hasn't been
1316 * shut down already */
1317 if (nn->nn_sc == sc) {
1318 o2net_sc_reset_idle_timer(sc);
1319 atomic_set(&nn->nn_timeout, 0);
1320 o2net_set_nn_state(nn, sc, 1, 0);
1322 spin_unlock(&nn->nn_lock);
1324 /* shift everything up as though it wasn't there */
1325 sc->sc_page_off -= sizeof(struct o2net_handshake);
1326 if (sc->sc_page_off)
1327 memmove(hand, hand + 1, sc->sc_page_off);
1332 /* this demuxes the queued rx bytes into header or payload bits and calls
1333 * handlers as each full message is read off the socket. it returns -error,
1334 * == 0 eof, or > 0 for progress made.*/
1335 static int o2net_advance_rx(struct o2net_sock_container *sc)
1337 struct o2net_msg *hdr;
1342 sclog(sc, "receiving\n");
1343 o2net_set_advance_start_time(sc);
1345 if (unlikely(sc->sc_handshake_ok == 0)) {
1346 if(sc->sc_page_off < sizeof(struct o2net_handshake)) {
1347 data = page_address(sc->sc_page) + sc->sc_page_off;
1348 datalen = sizeof(struct o2net_handshake) - sc->sc_page_off;
1349 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1351 sc->sc_page_off += ret;
1354 if (sc->sc_page_off == sizeof(struct o2net_handshake)) {
1355 o2net_check_handshake(sc);
1356 if (unlikely(sc->sc_handshake_ok == 0))
1362 /* do we need more header? */
1363 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1364 data = page_address(sc->sc_page) + sc->sc_page_off;
1365 datalen = sizeof(struct o2net_msg) - sc->sc_page_off;
1366 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1368 sc->sc_page_off += ret;
1369 /* only swab incoming here.. we can
1370 * only get here once as we cross from
1371 * being under to over */
1372 if (sc->sc_page_off == sizeof(struct o2net_msg)) {
1373 hdr = page_address(sc->sc_page);
1374 if (be16_to_cpu(hdr->data_len) >
1375 O2NET_MAX_PAYLOAD_BYTES)
1383 if (sc->sc_page_off < sizeof(struct o2net_msg)) {
1384 /* oof, still don't have a header */
1388 /* this was swabbed above when we first read it */
1389 hdr = page_address(sc->sc_page);
1391 msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1393 /* do we need more payload? */
1394 if (sc->sc_page_off - sizeof(struct o2net_msg) < be16_to_cpu(hdr->data_len)) {
1395 /* need more payload */
1396 data = page_address(sc->sc_page) + sc->sc_page_off;
1397 datalen = (sizeof(struct o2net_msg) + be16_to_cpu(hdr->data_len)) -
1399 ret = o2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1401 sc->sc_page_off += ret;
1406 if (sc->sc_page_off - sizeof(struct o2net_msg) == be16_to_cpu(hdr->data_len)) {
1407 /* we can only get here once, the first time we read
1408 * the payload.. so set ret to progress if the handler
1409 * works out. after calling this the message is toast */
1410 ret = o2net_process_message(sc, hdr);
1413 sc->sc_page_off = 0;
1417 sclog(sc, "ret = %d\n", ret);
1418 o2net_set_advance_stop_time(sc);
1422 /* this work func is triggerd by data ready. it reads until it can read no
1423 * more. it interprets 0, eof, as fatal. if data_ready hits while we're doing
1424 * our work the work struct will be marked and we'll be called again. */
1425 static void o2net_rx_until_empty(struct work_struct *work)
1427 struct o2net_sock_container *sc =
1428 container_of(work, struct o2net_sock_container, sc_rx_work);
1432 ret = o2net_advance_rx(sc);
1435 if (ret <= 0 && ret != -EAGAIN) {
1436 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1437 sclog(sc, "saw error %d, closing\n", ret);
1438 /* not permanent so read failed handshake can retry */
1439 o2net_ensure_shutdown(nn, sc, 0);
1445 static void o2net_initialize_handshake(void)
1447 o2net_hand->o2hb_heartbeat_timeout_ms = cpu_to_be32(
1448 O2HB_MAX_WRITE_TIMEOUT_MS);
1449 o2net_hand->o2net_idle_timeout_ms = cpu_to_be32(o2net_idle_timeout());
1450 o2net_hand->o2net_keepalive_delay_ms = cpu_to_be32(
1451 o2net_keepalive_delay());
1452 o2net_hand->o2net_reconnect_delay_ms = cpu_to_be32(
1453 o2net_reconnect_delay());
1456 /* ------------------------------------------------------------ */
1458 /* called when a connect completes and after a sock is accepted. the
1459 * rx path will see the response and mark the sc valid */
1460 static void o2net_sc_connect_completed(struct work_struct *work)
1462 struct o2net_sock_container *sc =
1463 container_of(work, struct o2net_sock_container,
1466 mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1467 (unsigned long long)O2NET_PROTOCOL_VERSION,
1468 (unsigned long long)be64_to_cpu(o2net_hand->connector_id));
1470 o2net_initialize_handshake();
1471 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1475 /* this is called as a work_struct func. */
1476 static void o2net_sc_send_keep_req(struct work_struct *work)
1478 struct o2net_sock_container *sc =
1479 container_of(work, struct o2net_sock_container,
1480 sc_keepalive_work.work);
1482 o2net_sendpage(sc, o2net_keep_req, sizeof(*o2net_keep_req));
1486 /* socket shutdown does a del_timer_sync against this as it tears down.
1487 * we can't start this timer until we've got to the point in sc buildup
1488 * where shutdown is going to be involved */
1489 static void o2net_idle_timer(struct timer_list *t)
1491 struct o2net_sock_container *sc = from_timer(sc, t, sc_idle_timeout);
1492 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1493 #ifdef CONFIG_DEBUG_FS
1494 unsigned long msecs = ktime_to_ms(ktime_get()) -
1495 ktime_to_ms(sc->sc_tv_timer);
1497 unsigned long msecs = o2net_idle_timeout();
1500 printk(KERN_NOTICE "o2net: Connection to " SC_NODEF_FMT " has been "
1501 "idle for %lu.%lu secs.\n",
1502 SC_NODEF_ARGS(sc), msecs / 1000, msecs % 1000);
1504 /* idle timerout happen, don't shutdown the connection, but
1505 * make fence decision. Maybe the connection can recover before
1506 * the decision is made.
1508 atomic_set(&nn->nn_timeout, 1);
1509 o2quo_conn_err(o2net_num_from_nn(nn));
1510 queue_delayed_work(o2net_wq, &nn->nn_still_up,
1511 msecs_to_jiffies(O2NET_QUORUM_DELAY_MS));
1513 o2net_sc_reset_idle_timer(sc);
1517 static void o2net_sc_reset_idle_timer(struct o2net_sock_container *sc)
1519 o2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1520 o2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1521 msecs_to_jiffies(o2net_keepalive_delay()));
1522 o2net_set_sock_timer(sc);
1523 mod_timer(&sc->sc_idle_timeout,
1524 jiffies + msecs_to_jiffies(o2net_idle_timeout()));
1527 static void o2net_sc_postpone_idle(struct o2net_sock_container *sc)
1529 struct o2net_node *nn = o2net_nn_from_num(sc->sc_node->nd_num);
1531 /* clear fence decision since the connection recover from timeout*/
1532 if (atomic_read(&nn->nn_timeout)) {
1533 o2quo_conn_up(o2net_num_from_nn(nn));
1534 cancel_delayed_work(&nn->nn_still_up);
1535 atomic_set(&nn->nn_timeout, 0);
1538 /* Only push out an existing timer */
1539 if (timer_pending(&sc->sc_idle_timeout))
1540 o2net_sc_reset_idle_timer(sc);
1543 /* this work func is kicked whenever a path sets the nn state which doesn't
1544 * have valid set. This includes seeing hb come up, losing a connection,
1545 * having a connect attempt fail, etc. This centralizes the logic which decides
1546 * if a connect attempt should be made or if we should give up and all future
1547 * transmit attempts should fail */
1548 static void o2net_start_connect(struct work_struct *work)
1550 struct o2net_node *nn =
1551 container_of(work, struct o2net_node, nn_connect_work.work);
1552 struct o2net_sock_container *sc = NULL;
1553 struct o2nm_node *node = NULL, *mynode = NULL;
1554 struct socket *sock = NULL;
1555 struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1557 unsigned int timeout;
1558 unsigned int nofs_flag;
1561 * sock_create allocates the sock with GFP_KERNEL. We must
1562 * prevent the filesystem from being reentered by memory reclaim.
1564 nofs_flag = memalloc_nofs_save();
1565 /* if we're greater we initiate tx, otherwise we accept */
1566 if (o2nm_this_node() <= o2net_num_from_nn(nn))
1569 /* watch for racing with tearing a node down */
1570 node = o2nm_get_node_by_num(o2net_num_from_nn(nn));
1574 mynode = o2nm_get_node_by_num(o2nm_this_node());
1578 spin_lock(&nn->nn_lock);
1580 * see if we already have one pending or have given up.
1581 * For nn_timeout, it is set when we close the connection
1582 * because of the idle time out. So it means that we have
1583 * at least connected to that node successfully once,
1584 * now try to connect to it again.
1586 timeout = atomic_read(&nn->nn_timeout);
1587 stop = (nn->nn_sc ||
1588 (nn->nn_persistent_error &&
1589 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1590 spin_unlock(&nn->nn_lock);
1594 nn->nn_last_connect_attempt = jiffies;
1596 sc = sc_alloc(node);
1598 mlog(0, "couldn't allocate sc\n");
1603 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1605 mlog(0, "can't create socket: %d\n", ret);
1608 sc->sc_sock = sock; /* freed by sc_kref_release */
1610 sock->sk->sk_allocation = GFP_ATOMIC;
1611 sock->sk->sk_use_task_frag = false;
1613 myaddr.sin_family = AF_INET;
1614 myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1615 myaddr.sin_port = htons(0); /* any port */
1617 ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1620 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1621 ret, &mynode->nd_ipv4_address);
1625 tcp_sock_set_nodelay(sc->sc_sock->sk);
1626 tcp_sock_set_user_timeout(sock->sk, O2NET_TCP_USER_TIMEOUT);
1628 o2net_register_callbacks(sc->sc_sock->sk, sc);
1630 spin_lock(&nn->nn_lock);
1631 /* handshake completion will set nn->nn_sc_valid */
1632 o2net_set_nn_state(nn, sc, 0, 0);
1633 spin_unlock(&nn->nn_lock);
1635 remoteaddr.sin_family = AF_INET;
1636 remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1637 remoteaddr.sin_port = node->nd_ipv4_port;
1639 ret = sc->sc_sock->ops->connect(sc->sc_sock,
1640 (struct sockaddr *)&remoteaddr,
1643 if (ret == -EINPROGRESS)
1648 printk(KERN_NOTICE "o2net: Connect attempt to " SC_NODEF_FMT
1649 " failed with errno %d\n", SC_NODEF_ARGS(sc), ret);
1650 /* 0 err so that another will be queued and attempted
1651 * from set_nn_state */
1652 o2net_ensure_shutdown(nn, sc, 0);
1657 o2nm_node_put(node);
1659 o2nm_node_put(mynode);
1661 memalloc_nofs_restore(nofs_flag);
1665 static void o2net_connect_expired(struct work_struct *work)
1667 struct o2net_node *nn =
1668 container_of(work, struct o2net_node, nn_connect_expired.work);
1670 spin_lock(&nn->nn_lock);
1671 if (!nn->nn_sc_valid) {
1672 printk(KERN_NOTICE "o2net: No connection established with "
1673 "node %u after %u.%u seconds, check network and"
1674 " cluster configuration.\n",
1675 o2net_num_from_nn(nn),
1676 o2net_idle_timeout() / 1000,
1677 o2net_idle_timeout() % 1000);
1679 o2net_set_nn_state(nn, NULL, 0, 0);
1681 spin_unlock(&nn->nn_lock);
1684 static void o2net_still_up(struct work_struct *work)
1686 struct o2net_node *nn =
1687 container_of(work, struct o2net_node, nn_still_up.work);
1689 o2quo_hb_still_up(o2net_num_from_nn(nn));
1692 /* ------------------------------------------------------------ */
1694 void o2net_disconnect_node(struct o2nm_node *node)
1696 struct o2net_node *nn = o2net_nn_from_num(node->nd_num);
1698 /* don't reconnect until it's heartbeating again */
1699 spin_lock(&nn->nn_lock);
1700 atomic_set(&nn->nn_timeout, 0);
1701 o2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1702 spin_unlock(&nn->nn_lock);
1705 cancel_delayed_work(&nn->nn_connect_expired);
1706 cancel_delayed_work(&nn->nn_connect_work);
1707 cancel_delayed_work(&nn->nn_still_up);
1708 flush_workqueue(o2net_wq);
1712 static void o2net_hb_node_down_cb(struct o2nm_node *node, int node_num,
1715 o2quo_hb_down(node_num);
1720 if (node_num != o2nm_this_node())
1721 o2net_disconnect_node(node);
1723 BUG_ON(atomic_read(&o2net_connected_peers) < 0);
1726 static void o2net_hb_node_up_cb(struct o2nm_node *node, int node_num,
1729 struct o2net_node *nn = o2net_nn_from_num(node_num);
1731 o2quo_hb_up(node_num);
1735 /* ensure an immediate connect attempt */
1736 nn->nn_last_connect_attempt = jiffies -
1737 (msecs_to_jiffies(o2net_reconnect_delay()) + 1);
1739 if (node_num != o2nm_this_node()) {
1740 /* believe it or not, accept and node heartbeating testing
1741 * can succeed for this node before we got here.. so
1742 * only use set_nn_state to clear the persistent error
1743 * if that hasn't already happened */
1744 spin_lock(&nn->nn_lock);
1745 atomic_set(&nn->nn_timeout, 0);
1746 if (nn->nn_persistent_error)
1747 o2net_set_nn_state(nn, NULL, 0, 0);
1748 spin_unlock(&nn->nn_lock);
1752 void o2net_unregister_hb_callbacks(void)
1754 o2hb_unregister_callback(NULL, &o2net_hb_up);
1755 o2hb_unregister_callback(NULL, &o2net_hb_down);
1758 int o2net_register_hb_callbacks(void)
1762 o2hb_setup_callback(&o2net_hb_down, O2HB_NODE_DOWN_CB,
1763 o2net_hb_node_down_cb, NULL, O2NET_HB_PRI);
1764 o2hb_setup_callback(&o2net_hb_up, O2HB_NODE_UP_CB,
1765 o2net_hb_node_up_cb, NULL, O2NET_HB_PRI);
1767 ret = o2hb_register_callback(NULL, &o2net_hb_up);
1769 ret = o2hb_register_callback(NULL, &o2net_hb_down);
1772 o2net_unregister_hb_callbacks();
1777 /* ------------------------------------------------------------ */
1779 static int o2net_accept_one(struct socket *sock, int *more)
1782 struct sockaddr_in sin;
1783 struct socket *new_sock = NULL;
1784 struct o2nm_node *node = NULL;
1785 struct o2nm_node *local_node = NULL;
1786 struct o2net_sock_container *sc = NULL;
1787 struct o2net_node *nn;
1788 unsigned int nofs_flag;
1791 * sock_create_lite allocates the sock with GFP_KERNEL. We must
1792 * prevent the filesystem from being reentered by memory reclaim.
1794 nofs_flag = memalloc_nofs_save();
1796 BUG_ON(sock == NULL);
1798 ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1799 sock->sk->sk_protocol, &new_sock);
1803 new_sock->type = sock->type;
1804 new_sock->ops = sock->ops;
1805 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, false);
1810 new_sock->sk->sk_allocation = GFP_ATOMIC;
1812 tcp_sock_set_nodelay(new_sock->sk);
1813 tcp_sock_set_user_timeout(new_sock->sk, O2NET_TCP_USER_TIMEOUT);
1815 ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin, 1);
1819 node = o2nm_get_node_by_ip(sin.sin_addr.s_addr);
1821 printk(KERN_NOTICE "o2net: Attempt to connect from unknown "
1822 "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1823 ntohs(sin.sin_port));
1828 if (o2nm_this_node() >= node->nd_num) {
1829 local_node = o2nm_get_node_by_num(o2nm_this_node());
1831 printk(KERN_NOTICE "o2net: Unexpected connect attempt "
1832 "seen at node '%s' (%u, %pI4:%d) from "
1833 "node '%s' (%u, %pI4:%d)\n",
1834 local_node->nd_name, local_node->nd_num,
1835 &(local_node->nd_ipv4_address),
1836 ntohs(local_node->nd_ipv4_port),
1838 node->nd_num, &sin.sin_addr.s_addr,
1839 ntohs(sin.sin_port));
1844 /* this happens all the time when the other node sees our heartbeat
1845 * and tries to connect before we see their heartbeat */
1846 if (!o2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1847 mlog(ML_CONN, "attempt to connect from node '%s' at "
1848 "%pI4:%d but it isn't heartbeating\n",
1849 node->nd_name, &sin.sin_addr.s_addr,
1850 ntohs(sin.sin_port));
1855 nn = o2net_nn_from_num(node->nd_num);
1857 spin_lock(&nn->nn_lock);
1862 spin_unlock(&nn->nn_lock);
1864 printk(KERN_NOTICE "o2net: Attempt to connect from node '%s' "
1865 "at %pI4:%d but it already has an open connection\n",
1866 node->nd_name, &sin.sin_addr.s_addr,
1867 ntohs(sin.sin_port));
1871 sc = sc_alloc(node);
1877 sc->sc_sock = new_sock;
1880 spin_lock(&nn->nn_lock);
1881 atomic_set(&nn->nn_timeout, 0);
1882 o2net_set_nn_state(nn, sc, 0, 0);
1883 spin_unlock(&nn->nn_lock);
1885 o2net_register_callbacks(sc->sc_sock->sk, sc);
1886 o2net_sc_queue_work(sc, &sc->sc_rx_work);
1888 o2net_initialize_handshake();
1889 o2net_sendpage(sc, o2net_hand, sizeof(*o2net_hand));
1893 sock_release(new_sock);
1895 o2nm_node_put(node);
1897 o2nm_node_put(local_node);
1901 memalloc_nofs_restore(nofs_flag);
1906 * This function is invoked in response to one or more
1907 * pending accepts at softIRQ level. We must drain the
1908 * entire que before returning.
1911 static void o2net_accept_many(struct work_struct *work)
1913 struct socket *sock = o2net_listen_sock;
1917 * It is critical to note that due to interrupt moderation
1918 * at the network driver level, we can't assume to get a
1919 * softIRQ for every single conn since tcp SYN packets
1920 * can arrive back-to-back, and therefore many pending
1921 * accepts may result in just 1 softIRQ. If we terminate
1922 * the o2net_accept_one() loop upon seeing an err, what happens
1923 * to the rest of the conns in the queue? If no new SYN
1924 * arrives for hours, no softIRQ will be delivered,
1925 * and the connections will just sit in the queue.
1929 o2net_accept_one(sock, &more);
1936 static void o2net_listen_data_ready(struct sock *sk)
1938 void (*ready)(struct sock *sk);
1940 trace_sk_data_ready(sk);
1942 read_lock_bh(&sk->sk_callback_lock);
1943 ready = sk->sk_user_data;
1944 if (ready == NULL) { /* check for teardown race */
1945 ready = sk->sk_data_ready;
1949 /* This callback may called twice when a new connection
1950 * is being established as a child socket inherits everything
1951 * from a parent LISTEN socket, including the data_ready cb of
1952 * the parent. This leads to a hazard. In o2net_accept_one()
1953 * we are still initializing the child socket but have not
1954 * changed the inherited data_ready callback yet when
1955 * data starts arriving.
1956 * We avoid this hazard by checking the state.
1957 * For the listening socket, the state will be TCP_LISTEN; for the new
1958 * socket, will be TCP_ESTABLISHED. Also, in this case,
1959 * sk->sk_user_data is not a valid function pointer.
1962 if (sk->sk_state == TCP_LISTEN) {
1963 queue_work(o2net_wq, &o2net_listen_work);
1969 read_unlock_bh(&sk->sk_callback_lock);
1974 static int o2net_open_listening_sock(__be32 addr, __be16 port)
1976 struct socket *sock = NULL;
1978 struct sockaddr_in sin = {
1979 .sin_family = PF_INET,
1980 .sin_addr = { .s_addr = addr },
1984 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1986 printk(KERN_ERR "o2net: Error %d while creating socket\n", ret);
1990 sock->sk->sk_allocation = GFP_ATOMIC;
1992 write_lock_bh(&sock->sk->sk_callback_lock);
1993 sock->sk->sk_user_data = sock->sk->sk_data_ready;
1994 sock->sk->sk_data_ready = o2net_listen_data_ready;
1995 write_unlock_bh(&sock->sk->sk_callback_lock);
1997 o2net_listen_sock = sock;
1998 INIT_WORK(&o2net_listen_work, o2net_accept_many);
2000 sock->sk->sk_reuse = SK_CAN_REUSE;
2001 ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2003 printk(KERN_ERR "o2net: Error %d while binding socket at "
2004 "%pI4:%u\n", ret, &addr, ntohs(port));
2008 ret = sock->ops->listen(sock, 64);
2010 printk(KERN_ERR "o2net: Error %d while listening on %pI4:%u\n",
2011 ret, &addr, ntohs(port));
2015 o2net_listen_sock = NULL;
2023 * called from node manager when we should bring up our network listening
2024 * socket. node manager handles all the serialization to only call this
2025 * once and to match it with o2net_stop_listening(). note,
2026 * o2nm_this_node() doesn't work yet as we're being called while it
2029 int o2net_start_listening(struct o2nm_node *node)
2033 BUG_ON(o2net_wq != NULL);
2034 BUG_ON(o2net_listen_sock != NULL);
2036 mlog(ML_KTHREAD, "starting o2net thread...\n");
2037 o2net_wq = alloc_ordered_workqueue("o2net", WQ_MEM_RECLAIM);
2038 if (o2net_wq == NULL) {
2039 mlog(ML_ERROR, "unable to launch o2net thread\n");
2040 return -ENOMEM; /* ? */
2043 ret = o2net_open_listening_sock(node->nd_ipv4_address,
2044 node->nd_ipv4_port);
2046 destroy_workqueue(o2net_wq);
2049 o2quo_conn_up(node->nd_num);
2054 /* again, o2nm_this_node() doesn't work here as we're involved in
2055 * tearing it down */
2056 void o2net_stop_listening(struct o2nm_node *node)
2058 struct socket *sock = o2net_listen_sock;
2061 BUG_ON(o2net_wq == NULL);
2062 BUG_ON(o2net_listen_sock == NULL);
2064 /* stop the listening socket from generating work */
2065 write_lock_bh(&sock->sk->sk_callback_lock);
2066 sock->sk->sk_data_ready = sock->sk->sk_user_data;
2067 sock->sk->sk_user_data = NULL;
2068 write_unlock_bh(&sock->sk->sk_callback_lock);
2070 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2071 struct o2nm_node *node = o2nm_get_node_by_num(i);
2073 o2net_disconnect_node(node);
2074 o2nm_node_put(node);
2078 /* finish all work and tear down the work queue */
2079 mlog(ML_KTHREAD, "waiting for o2net thread to exit....\n");
2080 destroy_workqueue(o2net_wq);
2083 sock_release(o2net_listen_sock);
2084 o2net_listen_sock = NULL;
2086 o2quo_conn_err(node->nd_num);
2089 /* ------------------------------------------------------------ */
2091 int o2net_init(void)
2093 struct folio *folio;
2098 o2net_debugfs_init();
2100 folio = folio_alloc(GFP_KERNEL | __GFP_ZERO, 0);
2104 p = folio_address(folio);
2106 p += sizeof(struct o2net_handshake);
2108 p += sizeof(struct o2net_msg);
2109 o2net_keep_resp = p;
2111 o2net_hand->protocol_version = cpu_to_be64(O2NET_PROTOCOL_VERSION);
2112 o2net_hand->connector_id = cpu_to_be64(1);
2114 o2net_keep_req->magic = cpu_to_be16(O2NET_MSG_KEEP_REQ_MAGIC);
2115 o2net_keep_resp->magic = cpu_to_be16(O2NET_MSG_KEEP_RESP_MAGIC);
2117 for (i = 0; i < ARRAY_SIZE(o2net_nodes); i++) {
2118 struct o2net_node *nn = o2net_nn_from_num(i);
2120 atomic_set(&nn->nn_timeout, 0);
2121 spin_lock_init(&nn->nn_lock);
2122 INIT_DELAYED_WORK(&nn->nn_connect_work, o2net_start_connect);
2123 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2124 o2net_connect_expired);
2125 INIT_DELAYED_WORK(&nn->nn_still_up, o2net_still_up);
2126 /* until we see hb from a node we'll return einval */
2127 nn->nn_persistent_error = -ENOTCONN;
2128 init_waitqueue_head(&nn->nn_sc_wq);
2129 idr_init(&nn->nn_status_idr);
2130 INIT_LIST_HEAD(&nn->nn_status_list);
2136 o2net_debugfs_exit();
2141 void o2net_exit(void)
2144 o2net_debugfs_exit();
2145 folio_put(virt_to_folio(o2net_hand));