projects / deliverable / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Merge branch 'acpi-lpss'
[deliverable/linux.git] / drivers / staging / zcache / ramster / tcp.c
CommitLineData
14c9fda5
DM		 1/* -*- mode: c; c-basic-offset: 8; -*-
2 *
3 * vim: noexpandtab sw=8 ts=8 sts=0:
4 *
5 * Copyright (C) 2004 Oracle. All rights reserved.
6 *
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.
11 *
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.
16 *
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.
21 *
22 * ----
23 *
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.
27 *
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
32 * message they send.
33 *
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
39 * doing so.
40 *
41 * Any framing errors (bad magic, large payload lengths) close a connection.
42 *
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.
49 *
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.
54 */
55
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>
62#include <linux/export.h>
63#include <linux/uaccess.h>
64#include <net/tcp.h>
65
66
67#include "heartbeat.h"
68#include "tcp.h"
69#include "nodemanager.h"
70#define MLOG_MASK_PREFIX ML_TCP
71#include "masklog.h"
72
73#include "tcp_internal.h"
74
75#define SC_NODEF_FMT "node %s (num %u) at %pI4:%u"
76
77/*
78 * In the following two log macros, the whitespace after the ',' just
79 * before ##args is intentional. Otherwise, gcc 2.95 will eat the
80 * previous token if args expands to nothing.
81 */
82#define msglog(hdr, fmt, args...) do { \
83 typeof(hdr) __hdr = (hdr); \
84 mlog(ML_MSG, "[mag %u len %u typ %u stat %d sys_stat %d " \
85 "key %08x num %u] " fmt, \
86 be16_to_cpu(__hdr->magic), be16_to_cpu(__hdr->data_len), \
87 be16_to_cpu(__hdr->msg_type), be32_to_cpu(__hdr->status), \
88 be32_to_cpu(__hdr->sys_status), be32_to_cpu(__hdr->key), \
89 be32_to_cpu(__hdr->msg_num) , ##args); \
90} while (0)
91
92#define sclog(sc, fmt, args...) do { \
93 typeof(sc) __sc = (sc); \
94 mlog(ML_SOCKET, "[sc %p refs %d sock %p node %u page %p " \
95 "pg_off %zu] " fmt, __sc, \
96 atomic_read(&__sc->sc_kref.refcount), __sc->sc_sock, \
97 __sc->sc_node->nd_num, __sc->sc_page, __sc->sc_page_off , \
98 ##args); \
99} while (0)
100
101static DEFINE_RWLOCK(r2net_handler_lock);
102static struct rb_root r2net_handler_tree = RB_ROOT;
103
104static struct r2net_node r2net_nodes[R2NM_MAX_NODES];
105
106/* XXX someday we'll need better accounting */
107static struct socket *r2net_listen_sock;
108
109/*
110 * listen work is only queued by the listening socket callbacks on the
111 * r2net_wq. teardown detaches the callbacks before destroying the workqueue.
112 * quorum work is queued as sock containers are shutdown.. stop_listening
113 * tears down all the node's sock containers, preventing future shutdowns
114 * and queued quorum work, before canceling delayed quorum work and
115 * destroying the work queue.
116 */
117static struct workqueue_struct *r2net_wq;
118static struct work_struct r2net_listen_work;
119
120static struct r2hb_callback_func r2net_hb_up, r2net_hb_down;
121#define R2NET_HB_PRI 0x1
122
123static struct r2net_handshake *r2net_hand;
124static struct r2net_msg *r2net_keep_req, *r2net_keep_resp;
125
126static int r2net_sys_err_translations[R2NET_ERR_MAX] = {
127 [R2NET_ERR_NONE] = 0,
128 [R2NET_ERR_NO_HNDLR] = -ENOPROTOOPT,
129 [R2NET_ERR_OVERFLOW] = -EOVERFLOW,
130 [R2NET_ERR_DIED] = -EHOSTDOWN,};
131
132/* can't quite avoid *all* internal declarations :/ */
133static void r2net_sc_connect_completed(struct work_struct *work);
134static void r2net_rx_until_empty(struct work_struct *work);
135static void r2net_shutdown_sc(struct work_struct *work);
136static void r2net_listen_data_ready(struct sock *sk, int bytes);
137static void r2net_sc_send_keep_req(struct work_struct *work);
138static void r2net_idle_timer(unsigned long data);
139static void r2net_sc_postpone_idle(struct r2net_sock_container *sc);
140static void r2net_sc_reset_idle_timer(struct r2net_sock_container *sc);
141
142#ifdef CONFIG_DEBUG_FS
143static void r2net_init_nst(struct r2net_send_tracking *nst, u32 msgtype,
144 u32 msgkey, struct task_struct *task, u8 node)
145{
146 INIT_LIST_HEAD(&nst->st_net_debug_item);
147 nst->st_task = task;
148 nst->st_msg_type = msgtype;
149 nst->st_msg_key = msgkey;
150 nst->st_node = node;
151}
152
153static inline void r2net_set_nst_sock_time(struct r2net_send_tracking *nst)
154{
155 nst->st_sock_time = ktime_get();
156}
157
158static inline void r2net_set_nst_send_time(struct r2net_send_tracking *nst)
159{
160 nst->st_send_time = ktime_get();
161}
162
163static inline void r2net_set_nst_status_time(struct r2net_send_tracking *nst)
164{
165 nst->st_status_time = ktime_get();
166}
167
168static inline void r2net_set_nst_sock_container(struct r2net_send_tracking *nst,
169 struct r2net_sock_container *sc)
170{
171 nst->st_sc = sc;
172}
173
174static inline void r2net_set_nst_msg_id(struct r2net_send_tracking *nst,
175 u32 msg_id)
176{
177 nst->st_id = msg_id;
178}
179
180static inline void r2net_set_sock_timer(struct r2net_sock_container *sc)
181{
182 sc->sc_tv_timer = ktime_get();
183}
184
185static inline void r2net_set_data_ready_time(struct r2net_sock_container *sc)
186{
187 sc->sc_tv_data_ready = ktime_get();
188}
189
190static inline void r2net_set_advance_start_time(struct r2net_sock_container *sc)
191{
192 sc->sc_tv_advance_start = ktime_get();
193}
194
195static inline void r2net_set_advance_stop_time(struct r2net_sock_container *sc)
196{
197 sc->sc_tv_advance_stop = ktime_get();
198}
199
200static inline void r2net_set_func_start_time(struct r2net_sock_container *sc)
201{
202 sc->sc_tv_func_start = ktime_get();
203}
204
205static inline void r2net_set_func_stop_time(struct r2net_sock_container *sc)
206{
207 sc->sc_tv_func_stop = ktime_get();
208}
209
210#else /* CONFIG_DEBUG_FS */
211# define r2net_init_nst(a, b, c, d, e)
212# define r2net_set_nst_sock_time(a)
213# define r2net_set_nst_send_time(a)
214# define r2net_set_nst_status_time(a)
215# define r2net_set_nst_sock_container(a, b)
216# define r2net_set_nst_msg_id(a, b)
217# define r2net_set_sock_timer(a)
218# define r2net_set_data_ready_time(a)
219# define r2net_set_advance_start_time(a)
220# define r2net_set_advance_stop_time(a)
221# define r2net_set_func_start_time(a)
222# define r2net_set_func_stop_time(a)
223#endif /* CONFIG_DEBUG_FS */
224
225#ifdef CONFIG_RAMSTER_FS_STATS
226static ktime_t r2net_get_func_run_time(struct r2net_sock_container *sc)
227{
228 return ktime_sub(sc->sc_tv_func_stop, sc->sc_tv_func_start);
229}
230
231static void r2net_update_send_stats(struct r2net_send_tracking *nst,
232 struct r2net_sock_container *sc)
233{
234 sc->sc_tv_status_total = ktime_add(sc->sc_tv_status_total,
235 ktime_sub(ktime_get(),
236 nst->st_status_time));
237 sc->sc_tv_send_total = ktime_add(sc->sc_tv_send_total,
238 ktime_sub(nst->st_status_time,
239 nst->st_send_time));
240 sc->sc_tv_acquiry_total = ktime_add(sc->sc_tv_acquiry_total,
241 ktime_sub(nst->st_send_time,
242 nst->st_sock_time));
243 sc->sc_send_count++;
244}
245
246static void r2net_update_recv_stats(struct r2net_sock_container *sc)
247{
248 sc->sc_tv_process_total = ktime_add(sc->sc_tv_process_total,
249 r2net_get_func_run_time(sc));
250 sc->sc_recv_count++;
251}
252
253#else
254
255# define r2net_update_send_stats(a, b)
256
257# define r2net_update_recv_stats(sc)
258
259#endif /* CONFIG_RAMSTER_FS_STATS */
260
261static inline int r2net_reconnect_delay(void)
262{
263 return r2nm_single_cluster->cl_reconnect_delay_ms;
264}
265
266static inline int r2net_keepalive_delay(void)
267{
268 return r2nm_single_cluster->cl_keepalive_delay_ms;
269}
270
271static inline int r2net_idle_timeout(void)
272{
273 return r2nm_single_cluster->cl_idle_timeout_ms;
274}
275
276static inline int r2net_sys_err_to_errno(enum r2net_system_error err)
277{
278 int trans;
279 BUG_ON(err >= R2NET_ERR_MAX);
280 trans = r2net_sys_err_translations[err];
281
282 /* Just in case we mess up the translation table above */
283 BUG_ON(err != R2NET_ERR_NONE && trans == 0);
284 return trans;
285}
286
287struct r2net_node *r2net_nn_from_num(u8 node_num)
288{
289 BUG_ON(node_num >= ARRAY_SIZE(r2net_nodes));
290 return &r2net_nodes[node_num];
291}
292
293static u8 r2net_num_from_nn(struct r2net_node *nn)
294{
295 BUG_ON(nn == NULL);
296 return nn - r2net_nodes;
297}
298
299/* ------------------------------------------------------------ */
300
301static int r2net_prep_nsw(struct r2net_node *nn, struct r2net_status_wait *nsw)
302{
a37c3010
TH		 303 int ret;
304
305 spin_lock(&nn->nn_lock);
306 ret = idr_alloc(&nn->nn_status_idr, nsw, 0, 0, GFP_ATOMIC);
307 if (ret >= 0) {
308 nsw->ns_id = ret;
309 list_add_tail(&nsw->ns_node_item, &nn->nn_status_list);
310 }
311 spin_unlock(&nn->nn_lock);
312
313 if (ret >= 0) {
14c9fda5
DM		 314 init_waitqueue_head(&nsw->ns_wq);
315 nsw->ns_sys_status = R2NET_ERR_NONE;
316 nsw->ns_status = 0;
a37c3010 317 return 0;
14c9fda5 318 }
14c9fda5
DM		 319 return ret;
320}
321
322static void r2net_complete_nsw_locked(struct r2net_node *nn,
323 struct r2net_status_wait *nsw,
324 enum r2net_system_error sys_status,
325 s32 status)
326{
327 assert_spin_locked(&nn->nn_lock);
328
329 if (!list_empty(&nsw->ns_node_item)) {
330 list_del_init(&nsw->ns_node_item);
331 nsw->ns_sys_status = sys_status;
332 nsw->ns_status = status;
333 idr_remove(&nn->nn_status_idr, nsw->ns_id);
334 wake_up(&nsw->ns_wq);
335 }
336}
337
338static void r2net_complete_nsw(struct r2net_node *nn,
339 struct r2net_status_wait *nsw,
340 u64 id, enum r2net_system_error sys_status,
341 s32 status)
342{
343 spin_lock(&nn->nn_lock);
344 if (nsw == NULL) {
345 if (id > INT_MAX)
346 goto out;
347
348 nsw = idr_find(&nn->nn_status_idr, id);
349 if (nsw == NULL)
350 goto out;
351 }
352
353 r2net_complete_nsw_locked(nn, nsw, sys_status, status);
354
355out:
356 spin_unlock(&nn->nn_lock);
357 return;
358}
359
360static void r2net_complete_nodes_nsw(struct r2net_node *nn)
361{
362 struct r2net_status_wait *nsw, *tmp;
363 unsigned int num_kills = 0;
364
365 assert_spin_locked(&nn->nn_lock);
366
367 list_for_each_entry_safe(nsw, tmp, &nn->nn_status_list, ns_node_item) {
368 r2net_complete_nsw_locked(nn, nsw, R2NET_ERR_DIED, 0);
369 num_kills++;
370 }
371
372 mlog(0, "completed %d messages for node %u\n", num_kills,
373 r2net_num_from_nn(nn));
374}
375
376static int r2net_nsw_completed(struct r2net_node *nn,
377 struct r2net_status_wait *nsw)
378{
379 int completed;
380 spin_lock(&nn->nn_lock);
381 completed = list_empty(&nsw->ns_node_item);
382 spin_unlock(&nn->nn_lock);
383 return completed;
384}
385
386/* ------------------------------------------------------------ */
387
388static void sc_kref_release(struct kref *kref)
389{
390 struct r2net_sock_container *sc = container_of(kref,
391 struct r2net_sock_container, sc_kref);
392 BUG_ON(timer_pending(&sc->sc_idle_timeout));
393
394 sclog(sc, "releasing\n");
395
396 if (sc->sc_sock) {
397 sock_release(sc->sc_sock);
398 sc->sc_sock = NULL;
399 }
400
401 r2nm_undepend_item(&sc->sc_node->nd_item);
402 r2nm_node_put(sc->sc_node);
403 sc->sc_node = NULL;
404
405 r2net_debug_del_sc(sc);
406 kfree(sc);
407}
408
409static void sc_put(struct r2net_sock_container *sc)
410{
411 sclog(sc, "put\n");
412 kref_put(&sc->sc_kref, sc_kref_release);
413}
414static void sc_get(struct r2net_sock_container *sc)
415{
416 sclog(sc, "get\n");
417 kref_get(&sc->sc_kref);
418}
419static struct r2net_sock_container *sc_alloc(struct r2nm_node *node)
420{
421 struct r2net_sock_container *sc, *ret = NULL;
422 struct page *page = NULL;
423 int status = 0;
424
425 page = alloc_page(GFP_NOFS);
426 sc = kzalloc(sizeof(*sc), GFP_NOFS);
427 if (sc == NULL || page == NULL)
428 goto out;
429
430 kref_init(&sc->sc_kref);
431 r2nm_node_get(node);
432 sc->sc_node = node;
433
434 /* pin the node item of the remote node */
435 status = r2nm_depend_item(&node->nd_item);
436 if (status) {
437 mlog_errno(status);
438 r2nm_node_put(node);
439 goto out;
440 }
441 INIT_WORK(&sc->sc_connect_work, r2net_sc_connect_completed);
442 INIT_WORK(&sc->sc_rx_work, r2net_rx_until_empty);
443 INIT_WORK(&sc->sc_shutdown_work, r2net_shutdown_sc);
444 INIT_DELAYED_WORK(&sc->sc_keepalive_work, r2net_sc_send_keep_req);
445
446 init_timer(&sc->sc_idle_timeout);
447 sc->sc_idle_timeout.function = r2net_idle_timer;
448 sc->sc_idle_timeout.data = (unsigned long)sc;
449
450 sclog(sc, "alloced\n");
451
452 ret = sc;
453 sc->sc_page = page;
454 r2net_debug_add_sc(sc);
455 sc = NULL;
456 page = NULL;
457
458out:
459 if (page)
460 __free_page(page);
461 kfree(sc);
462
463 return ret;
464}
465
466/* ------------------------------------------------------------ */
467
468static void r2net_sc_queue_work(struct r2net_sock_container *sc,
469 struct work_struct *work)
470{
471 sc_get(sc);
472 if (!queue_work(r2net_wq, work))
473 sc_put(sc);
474}
475static void r2net_sc_queue_delayed_work(struct r2net_sock_container *sc,
476 struct delayed_work *work,
477 int delay)
478{
479 sc_get(sc);
480 if (!queue_delayed_work(r2net_wq, work, delay))
481 sc_put(sc);
482}
483static void r2net_sc_cancel_delayed_work(struct r2net_sock_container *sc,
484 struct delayed_work *work)
485{
486 if (cancel_delayed_work(work))
487 sc_put(sc);
488}
489
490static atomic_t r2net_connected_peers = ATOMIC_INIT(0);
491
492int r2net_num_connected_peers(void)
493{
494 return atomic_read(&r2net_connected_peers);
495}
496
497static void r2net_set_nn_state(struct r2net_node *nn,
498 struct r2net_sock_container *sc,
499 unsigned valid, int err)
500{
501 int was_valid = nn->nn_sc_valid;
502 int was_err = nn->nn_persistent_error;
503 struct r2net_sock_container *old_sc = nn->nn_sc;
504
505 assert_spin_locked(&nn->nn_lock);
506
507 if (old_sc && !sc)
508 atomic_dec(&r2net_connected_peers);
509 else if (!old_sc && sc)
510 atomic_inc(&r2net_connected_peers);
511
512 /* the node num comparison and single connect/accept path should stop
513 * an non-null sc from being overwritten with another */
514 BUG_ON(sc && nn->nn_sc && nn->nn_sc != sc);
515 mlog_bug_on_msg(err && valid, "err %d valid %u\n", err, valid);
516 mlog_bug_on_msg(valid && !sc, "valid %u sc %p\n", valid, sc);
517
518 if (was_valid && !valid && err == 0)
519 err = -ENOTCONN;
520
521 mlog(ML_CONN, "node %u sc: %p -> %p, valid %u -> %u, err %d -> %d\n",
522 r2net_num_from_nn(nn), nn->nn_sc, sc, nn->nn_sc_valid, valid,
523 nn->nn_persistent_error, err);
524
525 nn->nn_sc = sc;
526 nn->nn_sc_valid = valid ? 1 : 0;
527 nn->nn_persistent_error = err;
528
529 /* mirrors r2net_tx_can_proceed() */
530 if (nn->nn_persistent_error || nn->nn_sc_valid)
531 wake_up(&nn->nn_sc_wq);
532
533 if (!was_err && nn->nn_persistent_error) {
534 queue_delayed_work(r2net_wq, &nn->nn_still_up,
535 msecs_to_jiffies(R2NET_QUORUM_DELAY_MS));
536 }
537
538 if (was_valid && !valid) {
539 pr_notice("ramster: No longer connected to " SC_NODEF_FMT "\n",
540 old_sc->sc_node->nd_name, old_sc->sc_node->nd_num,
541 &old_sc->sc_node->nd_ipv4_address,
542 ntohs(old_sc->sc_node->nd_ipv4_port));
543 r2net_complete_nodes_nsw(nn);
544 }
545
546 if (!was_valid && valid) {
547 cancel_delayed_work(&nn->nn_connect_expired);
548 pr_notice("ramster: %s " SC_NODEF_FMT "\n",
549 r2nm_this_node() > sc->sc_node->nd_num ?
550 "Connected to" : "Accepted connection from",
551 sc->sc_node->nd_name, sc->sc_node->nd_num,
552 &sc->sc_node->nd_ipv4_address,
553 ntohs(sc->sc_node->nd_ipv4_port));
554 }
555
556 /* trigger the connecting worker func as long as we're not valid,
557 * it will back off if it shouldn't connect. This can be called
558 * from node config teardown and so needs to be careful about
559 * the work queue actually being up. */
560 if (!valid && r2net_wq) {
561 unsigned long delay;
562 /* delay if we're within a RECONNECT_DELAY of the
563 * last attempt */
564 delay = (nn->nn_last_connect_attempt +
565 msecs_to_jiffies(r2net_reconnect_delay()))
566 - jiffies;
567 if (delay > msecs_to_jiffies(r2net_reconnect_delay()))
568 delay = 0;
569 mlog(ML_CONN, "queueing conn attempt in %lu jiffies\n", delay);
570 queue_delayed_work(r2net_wq, &nn->nn_connect_work, delay);
571
572 /*
573 * Delay the expired work after idle timeout.
574 *
575 * We might have lots of failed connection attempts that run
576 * through here but we only cancel the connect_expired work when
577 * a connection attempt succeeds. So only the first enqueue of
578 * the connect_expired work will do anything. The rest will see
579 * that it's already queued and do nothing.
580 */
581 delay += msecs_to_jiffies(r2net_idle_timeout());
582 queue_delayed_work(r2net_wq, &nn->nn_connect_expired, delay);
583 }
584
585 /* keep track of the nn's sc ref for the caller */
586 if ((old_sc == NULL) && sc)
587 sc_get(sc);
588 if (old_sc && (old_sc != sc)) {
589 r2net_sc_queue_work(old_sc, &old_sc->sc_shutdown_work);
590 sc_put(old_sc);
591 }
592}
593
594/* see r2net_register_callbacks() */
595static void r2net_data_ready(struct sock *sk, int bytes)
596{
597 void (*ready)(struct sock *sk, int bytes);
598
599 read_lock(&sk->sk_callback_lock);
600 if (sk->sk_user_data) {
601 struct r2net_sock_container *sc = sk->sk_user_data;
602 sclog(sc, "data_ready hit\n");
603 r2net_set_data_ready_time(sc);
604 r2net_sc_queue_work(sc, &sc->sc_rx_work);
605 ready = sc->sc_data_ready;
606 } else {
607 ready = sk->sk_data_ready;
608 }
609 read_unlock(&sk->sk_callback_lock);
610
611 ready(sk, bytes);
612}
613
614/* see r2net_register_callbacks() */
615static void r2net_state_change(struct sock *sk)
616{
617 void (*state_change)(struct sock *sk);
618 struct r2net_sock_container *sc;
619
620 read_lock(&sk->sk_callback_lock);
621 sc = sk->sk_user_data;
622 if (sc == NULL) {
623 state_change = sk->sk_state_change;
624 goto out;
625 }
626
627 sclog(sc, "state_change to %d\n", sk->sk_state);
628
629 state_change = sc->sc_state_change;
630
631 switch (sk->sk_state) {
632
633 /* ignore connecting sockets as they make progress */
634 case TCP_SYN_SENT:
635 case TCP_SYN_RECV:
636 break;
637 case TCP_ESTABLISHED:
638 r2net_sc_queue_work(sc, &sc->sc_connect_work);
639 break;
640 default:
641 pr_info("ramster: Connection to "
642 SC_NODEF_FMT " shutdown, state %d\n",
643 sc->sc_node->nd_name, sc->sc_node->nd_num,
644 &sc->sc_node->nd_ipv4_address,
645 ntohs(sc->sc_node->nd_ipv4_port), sk->sk_state);
646 r2net_sc_queue_work(sc, &sc->sc_shutdown_work);
647 break;
648
649 }
650out:
651 read_unlock(&sk->sk_callback_lock);
652 state_change(sk);
653}
654
655/*
656 * we register callbacks so we can queue work on events before calling
657 * the original callbacks. our callbacks are careful to test user_data
658 * to discover when they've reaced with r2net_unregister_callbacks().
659 */
660static void r2net_register_callbacks(struct sock *sk,
661 struct r2net_sock_container *sc)
662{
663 write_lock_bh(&sk->sk_callback_lock);
664
665 /* accepted sockets inherit the old listen socket data ready */
666 if (sk->sk_data_ready == r2net_listen_data_ready) {
667 sk->sk_data_ready = sk->sk_user_data;
668 sk->sk_user_data = NULL;
669 }
670
671 BUG_ON(sk->sk_user_data != NULL);
672 sk->sk_user_data = sc;
673 sc_get(sc);
674
675 sc->sc_data_ready = sk->sk_data_ready;
676 sc->sc_state_change = sk->sk_state_change;
677 sk->sk_data_ready = r2net_data_ready;
678 sk->sk_state_change = r2net_state_change;
679
680 mutex_init(&sc->sc_send_lock);
681
682 write_unlock_bh(&sk->sk_callback_lock);
683}
684
685static int r2net_unregister_callbacks(struct sock *sk,
686 struct r2net_sock_container *sc)
687{
688 int ret = 0;
689
690 write_lock_bh(&sk->sk_callback_lock);
691 if (sk->sk_user_data == sc) {
692 ret = 1;
693 sk->sk_user_data = NULL;
694 sk->sk_data_ready = sc->sc_data_ready;
695 sk->sk_state_change = sc->sc_state_change;
696 }
697 write_unlock_bh(&sk->sk_callback_lock);
698
699 return ret;
700}
701
702/*
703 * this is a little helper that is called by callers who have seen a problem
704 * with an sc and want to detach it from the nn if someone already hasn't beat
705 * them to it. if an error is given then the shutdown will be persistent
706 * and pending transmits will be canceled.
707 */
708static void r2net_ensure_shutdown(struct r2net_node *nn,
709 struct r2net_sock_container *sc,
710 int err)
711{
712 spin_lock(&nn->nn_lock);
713 if (nn->nn_sc == sc)
714 r2net_set_nn_state(nn, NULL, 0, err);
715 spin_unlock(&nn->nn_lock);
716}
717
718/*
719 * This work queue function performs the blocking parts of socket shutdown. A
720 * few paths lead here. set_nn_state will trigger this callback if it sees an
721 * sc detached from the nn. state_change will also trigger this callback
722 * directly when it sees errors. In that case we need to call set_nn_state
723 * ourselves as state_change couldn't get the nn_lock and call set_nn_state
724 * itself.
725 */
726static void r2net_shutdown_sc(struct work_struct *work)
727{
728 struct r2net_sock_container *sc =
729 container_of(work, struct r2net_sock_container,
730 sc_shutdown_work);
731 struct r2net_node *nn = r2net_nn_from_num(sc->sc_node->nd_num);
732
733 sclog(sc, "shutting down\n");
734
735 /* drop the callbacks ref and call shutdown only once */
736 if (r2net_unregister_callbacks(sc->sc_sock->sk, sc)) {
737 /* we shouldn't flush as we're in the thread, the
738 * races with pending sc work structs are harmless */
739 del_timer_sync(&sc->sc_idle_timeout);
740 r2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
741 sc_put(sc);
742 kernel_sock_shutdown(sc->sc_sock, SHUT_RDWR);
743 }
744
745 /* not fatal so failed connects before the other guy has our
746 * heartbeat can be retried */
747 r2net_ensure_shutdown(nn, sc, 0);
748 sc_put(sc);
749}
750
751/* ------------------------------------------------------------ */
752
753static int r2net_handler_cmp(struct r2net_msg_handler *nmh, u32 msg_type,
754 u32 key)
755{
756 int ret = memcmp(&nmh->nh_key, &key, sizeof(key));
757
758 if (ret == 0)
759 ret = memcmp(&nmh->nh_msg_type, &msg_type, sizeof(msg_type));
760
761 return ret;
762}
763
764static struct r2net_msg_handler *
765r2net_handler_tree_lookup(u32 msg_type, u32 key, struct rb_node ***ret_p,
766 struct rb_node **ret_parent)
767{
768 struct rb_node **p = &r2net_handler_tree.rb_node;
769 struct rb_node *parent = NULL;
770 struct r2net_msg_handler *nmh, *ret = NULL;
771 int cmp;
772
773 while (*p) {
774 parent = *p;
775 nmh = rb_entry(parent, struct r2net_msg_handler, nh_node);
776 cmp = r2net_handler_cmp(nmh, msg_type, key);
777
778 if (cmp < 0)
779 p = &(*p)->rb_left;
780 else if (cmp > 0)
781 p = &(*p)->rb_right;
782 else {
783 ret = nmh;
784 break;
785 }
786 }
787
788 if (ret_p != NULL)
789 *ret_p = p;
790 if (ret_parent != NULL)
791 *ret_parent = parent;
792
793 return ret;
794}
795
796static void r2net_handler_kref_release(struct kref *kref)
797{
798 struct r2net_msg_handler *nmh;
799 nmh = container_of(kref, struct r2net_msg_handler, nh_kref);
800
801 kfree(nmh);
802}
803
804static void r2net_handler_put(struct r2net_msg_handler *nmh)
805{
806 kref_put(&nmh->nh_kref, r2net_handler_kref_release);
807}
808
809/* max_len is protection for the handler func. incoming messages won't
810 * be given to the handler if their payload is longer than the max. */
811int r2net_register_handler(u32 msg_type, u32 key, u32 max_len,
812 r2net_msg_handler_func *func, void *data,
813 r2net_post_msg_handler_func *post_func,
814 struct list_head *unreg_list)
815{
816 struct r2net_msg_handler *nmh = NULL;
817 struct rb_node **p, *parent;
818 int ret = 0;
819
820 if (max_len > R2NET_MAX_PAYLOAD_BYTES) {
821 mlog(0, "max_len for message handler out of range: %u\n",
822 max_len);
823 ret = -EINVAL;
824 goto out;
825 }
826
827 if (!msg_type) {
828 mlog(0, "no message type provided: %u, %p\n", msg_type, func);
829 ret = -EINVAL;
830 goto out;
831
832 }
833 if (!func) {
834 mlog(0, "no message handler provided: %u, %p\n",
835 msg_type, func);
836 ret = -EINVAL;
837 goto out;
838 }
839
840 nmh = kzalloc(sizeof(struct r2net_msg_handler), GFP_NOFS);
841 if (nmh == NULL) {
842 ret = -ENOMEM;
843 goto out;
844 }
845
846 nmh->nh_func = func;
847 nmh->nh_func_data = data;
848 nmh->nh_post_func = post_func;
849 nmh->nh_msg_type = msg_type;
850 nmh->nh_max_len = max_len;
851 nmh->nh_key = key;
852 /* the tree and list get this ref.. they're both removed in
853 * unregister when this ref is dropped */
854 kref_init(&nmh->nh_kref);
855 INIT_LIST_HEAD(&nmh->nh_unregister_item);
856
857 write_lock(&r2net_handler_lock);
858 if (r2net_handler_tree_lookup(msg_type, key, &p, &parent))
859 ret = -EEXIST;
860 else {
861 rb_link_node(&nmh->nh_node, parent, p);
862 rb_insert_color(&nmh->nh_node, &r2net_handler_tree);
863 list_add_tail(&nmh->nh_unregister_item, unreg_list);
864
865 mlog(ML_TCP, "registered handler func %p type %u key %08x\n",
866 func, msg_type, key);
867 /* we've had some trouble with handlers seemingly vanishing. */
868 mlog_bug_on_msg(r2net_handler_tree_lookup(msg_type, key, &p,
869 &parent) == NULL,
870 "couldn't find handler we *just* registered "
871 "for type %u key %08x\n", msg_type, key);
872 }
873 write_unlock(&r2net_handler_lock);
874 if (ret)
875 goto out;
876
877out:
878 if (ret)
879 kfree(nmh);
880
881 return ret;
882}
883EXPORT_SYMBOL_GPL(r2net_register_handler);
884
885void r2net_unregister_handler_list(struct list_head *list)
886{
887 struct r2net_msg_handler *nmh, *n;
888
889 write_lock(&r2net_handler_lock);
890 list_for_each_entry_safe(nmh, n, list, nh_unregister_item) {
891 mlog(ML_TCP, "unregistering handler func %p type %u key %08x\n",
892 nmh->nh_func, nmh->nh_msg_type, nmh->nh_key);
893 rb_erase(&nmh->nh_node, &r2net_handler_tree);
894 list_del_init(&nmh->nh_unregister_item);
895 kref_put(&nmh->nh_kref, r2net_handler_kref_release);
896 }
897 write_unlock(&r2net_handler_lock);
898}
899EXPORT_SYMBOL_GPL(r2net_unregister_handler_list);
900
901static struct r2net_msg_handler *r2net_handler_get(u32 msg_type, u32 key)
902{
903 struct r2net_msg_handler *nmh;
904
905 read_lock(&r2net_handler_lock);
906 nmh = r2net_handler_tree_lookup(msg_type, key, NULL, NULL);
907 if (nmh)
908 kref_get(&nmh->nh_kref);
909 read_unlock(&r2net_handler_lock);
910
911 return nmh;
912}
913
914/* ------------------------------------------------------------ */
915
916static int r2net_recv_tcp_msg(struct socket *sock, void *data, size_t len)
917{
918 int ret;
919 mm_segment_t oldfs;
920 struct kvec vec = {
921 .iov_len = len,
922 .iov_base = data,
923 };
924 struct msghdr msg = {
925 .msg_iovlen = 1,
926 .msg_iov = (struct iovec *)&vec,
927 .msg_flags = MSG_DONTWAIT,
928 };
929
930 oldfs = get_fs();
931 set_fs(get_ds());
932 ret = sock_recvmsg(sock, &msg, len, msg.msg_flags);
933 set_fs(oldfs);
934
935 return ret;
936}
937
938static int r2net_send_tcp_msg(struct socket *sock, struct kvec *vec,
939 size_t veclen, size_t total)
940{
941 int ret;
942 mm_segment_t oldfs;
943 struct msghdr msg = {
944 .msg_iov = (struct iovec *)vec,
945 .msg_iovlen = veclen,
946 };
947
948 if (sock == NULL) {
949 ret = -EINVAL;
950 goto out;
951 }
952
953 oldfs = get_fs();
954 set_fs(get_ds());
955 ret = sock_sendmsg(sock, &msg, total);
956 set_fs(oldfs);
957 if (ret != total) {
958 mlog(ML_ERROR, "sendmsg returned %d instead of %zu\n", ret,
959 total);
960 if (ret >= 0)
961 ret = -EPIPE; /* should be smarter, I bet */
962 goto out;
963 }
964
965 ret = 0;
966out:
967 if (ret < 0)
968 mlog(0, "returning error: %d\n", ret);
969 return ret;
970}
971
972static void r2net_sendpage(struct r2net_sock_container *sc,
973 void *kmalloced_virt,
974 size_t size)
975{
976 struct r2net_node *nn = r2net_nn_from_num(sc->sc_node->nd_num);
977 ssize_t ret;
978
979 while (1) {
980 mutex_lock(&sc->sc_send_lock);
981 ret = sc->sc_sock->ops->sendpage(sc->sc_sock,
982 virt_to_page(kmalloced_virt),
983 (long)kmalloced_virt & ~PAGE_MASK,
984 size, MSG_DONTWAIT);
985 mutex_unlock(&sc->sc_send_lock);
986 if (ret == size)
987 break;
988 if (ret == (ssize_t)-EAGAIN) {
989 mlog(0, "sendpage of size %zu to " SC_NODEF_FMT
990 " returned EAGAIN\n", size, sc->sc_node->nd_name,
991 sc->sc_node->nd_num,
992 &sc->sc_node->nd_ipv4_address,
993 ntohs(sc->sc_node->nd_ipv4_port));
994 cond_resched();
995 continue;
996 }
997 mlog(ML_ERROR, "sendpage of size %zu to " SC_NODEF_FMT
998 " failed with %zd\n", size, sc->sc_node->nd_name,
999 sc->sc_node->nd_num, &sc->sc_node->nd_ipv4_address,
1000 ntohs(sc->sc_node->nd_ipv4_port), ret);
1001 r2net_ensure_shutdown(nn, sc, 0);
1002 break;
1003 }
1004}
1005
1006static void r2net_init_msg(struct r2net_msg *msg, u16 data_len,
1007 u16 msg_type, u32 key)
1008{
1009 memset(msg, 0, sizeof(struct r2net_msg));
1010 msg->magic = cpu_to_be16(R2NET_MSG_MAGIC);
1011 msg->data_len = cpu_to_be16(data_len);
1012 msg->msg_type = cpu_to_be16(msg_type);
1013 msg->sys_status = cpu_to_be32(R2NET_ERR_NONE);
1014 msg->status = 0;
1015 msg->key = cpu_to_be32(key);
1016}
1017
1018static int r2net_tx_can_proceed(struct r2net_node *nn,
1019 struct r2net_sock_container **sc_ret,
1020 int *error)
1021{
1022 int ret = 0;
1023
1024 spin_lock(&nn->nn_lock);
1025 if (nn->nn_persistent_error) {
1026 ret = 1;
1027 *sc_ret = NULL;
1028 *error = nn->nn_persistent_error;
1029 } else if (nn->nn_sc_valid) {
1030 kref_get(&nn->nn_sc->sc_kref);
1031
1032 ret = 1;
1033 *sc_ret = nn->nn_sc;
1034 *error = 0;
1035 }
1036 spin_unlock(&nn->nn_lock);
1037
1038 return ret;
1039}
1040
1041/* Get a map of all nodes to which this node is currently connected to */
1042void r2net_fill_node_map(unsigned long *map, unsigned bytes)
1043{
1044 struct r2net_sock_container *sc;
1045 int node, ret;
1046
1047 BUG_ON(bytes < (BITS_TO_LONGS(R2NM_MAX_NODES) * sizeof(unsigned long)));
1048
1049 memset(map, 0, bytes);
1050 for (node = 0; node < R2NM_MAX_NODES; ++node) {
1051 r2net_tx_can_proceed(r2net_nn_from_num(node), &sc, &ret);
1052 if (!ret) {
1053 set_bit(node, map);
1054 sc_put(sc);
1055 }
1056 }
1057}
1058EXPORT_SYMBOL_GPL(r2net_fill_node_map);
1059
1060int r2net_send_message_vec(u32 msg_type, u32 key, struct kvec *caller_vec,
1061 size_t caller_veclen, u8 target_node, int *status)
1062{
1063 int ret = 0;
1064 struct r2net_msg *msg = NULL;
1065 size_t veclen, caller_bytes = 0;
1066 struct kvec *vec = NULL;
1067 struct r2net_sock_container *sc = NULL;
1068 struct r2net_node *nn = r2net_nn_from_num(target_node);
1069 struct r2net_status_wait nsw = {
1070 .ns_node_item = LIST_HEAD_INIT(nsw.ns_node_item),
1071 };
1072 struct r2net_send_tracking nst;
1073
1074 /* this may be a general bug fix */
1075 init_waitqueue_head(&nsw.ns_wq);
1076
1077 r2net_init_nst(&nst, msg_type, key, current, target_node);
1078
1079 if (r2net_wq == NULL) {
1080 mlog(0, "attempt to tx without r2netd running\n");
1081 ret = -ESRCH;
1082 goto out;
1083 }
1084
1085 if (caller_veclen == 0) {
1086 mlog(0, "bad kvec array length\n");
1087 ret = -EINVAL;
1088 goto out;
1089 }
1090
1091 caller_bytes = iov_length((struct iovec *)caller_vec, caller_veclen);
1092 if (caller_bytes > R2NET_MAX_PAYLOAD_BYTES) {
1093 mlog(0, "total payload len %zu too large\n", caller_bytes);
1094 ret = -EINVAL;
1095 goto out;
1096 }
1097
1098 if (target_node == r2nm_this_node()) {
1099 ret = -ELOOP;
1100 goto out;
1101 }
1102
1103 r2net_debug_add_nst(&nst);
1104
1105 r2net_set_nst_sock_time(&nst);
1106
1107 wait_event(nn->nn_sc_wq, r2net_tx_can_proceed(nn, &sc, &ret));
1108 if (ret)
1109 goto out;
1110
1111 r2net_set_nst_sock_container(&nst, sc);
1112
1113 veclen = caller_veclen + 1;
1114 vec = kmalloc(sizeof(struct kvec) * veclen, GFP_ATOMIC);
1115 if (vec == NULL) {
1116 mlog(0, "failed to %zu element kvec!\n", veclen);
1117 ret = -ENOMEM;
1118 goto out;
1119 }
1120
1121 msg = kmalloc(sizeof(struct r2net_msg), GFP_ATOMIC);
1122 if (!msg) {
1123 mlog(0, "failed to allocate a r2net_msg!\n");
1124 ret = -ENOMEM;
1125 goto out;
1126 }
1127
1128 r2net_init_msg(msg, caller_bytes, msg_type, key);
1129
1130 vec[0].iov_len = sizeof(struct r2net_msg);
1131 vec[0].iov_base = msg;
1132 memcpy(&vec[1], caller_vec, caller_veclen * sizeof(struct kvec));
1133
1134 ret = r2net_prep_nsw(nn, &nsw);
1135 if (ret)
1136 goto out;
1137
1138 msg->msg_num = cpu_to_be32(nsw.ns_id);
1139 r2net_set_nst_msg_id(&nst, nsw.ns_id);
1140
1141 r2net_set_nst_send_time(&nst);
1142
1143 /* finally, convert the message header to network byte-order
1144 * and send */
1145 mutex_lock(&sc->sc_send_lock);
1146 ret = r2net_send_tcp_msg(sc->sc_sock, vec, veclen,
1147 sizeof(struct r2net_msg) + caller_bytes);
1148 mutex_unlock(&sc->sc_send_lock);
1149 msglog(msg, "sending returned %d\n", ret);
1150 if (ret < 0) {
1151 mlog(0, "error returned from r2net_send_tcp_msg=%d\n", ret);
1152 goto out;
1153 }
1154
1155 /* wait on other node's handler */
1156 r2net_set_nst_status_time(&nst);
1157 wait_event(nsw.ns_wq, r2net_nsw_completed(nn, &nsw) ||
1158 nn->nn_persistent_error || !nn->nn_sc_valid);
1159
1160 r2net_update_send_stats(&nst, sc);
1161
1162 /* Note that we avoid overwriting the callers status return
1163 * variable if a system error was reported on the other
1164 * side. Callers beware. */
1165 ret = r2net_sys_err_to_errno(nsw.ns_sys_status);
1166 if (status && !ret)
1167 *status = nsw.ns_status;
1168
1169 mlog(0, "woken, returning system status %d, user status %d\n",
1170 ret, nsw.ns_status);
1171out:
1172 r2net_debug_del_nst(&nst); /* must be before dropping sc and node */
1173 if (sc)
1174 sc_put(sc);
1175 kfree(vec);
1176 kfree(msg);
1177 r2net_complete_nsw(nn, &nsw, 0, 0, 0);
1178 return ret;
1179}
1180EXPORT_SYMBOL_GPL(r2net_send_message_vec);
1181
1182int r2net_send_message(u32 msg_type, u32 key, void *data, u32 len,
1183 u8 target_node, int *status)
1184{
1185 struct kvec vec = {
1186 .iov_base = data,
1187 .iov_len = len,
1188 };
1189 return r2net_send_message_vec(msg_type, key, &vec, 1,
1190 target_node, status);
1191}
1192EXPORT_SYMBOL_GPL(r2net_send_message);
1193
1194static int r2net_send_status_magic(struct socket *sock, struct r2net_msg *hdr,
1195 enum r2net_system_error syserr, int err)
1196{
1197 struct kvec vec = {
1198 .iov_base = hdr,
1199 .iov_len = sizeof(struct r2net_msg),
1200 };
1201
1202 BUG_ON(syserr >= R2NET_ERR_MAX);
1203
1204 /* leave other fields intact from the incoming message, msg_num
1205 * in particular */
1206 hdr->sys_status = cpu_to_be32(syserr);
1207 hdr->status = cpu_to_be32(err);
1208 /* twiddle the magic */
1209 hdr->magic = cpu_to_be16(R2NET_MSG_STATUS_MAGIC);
1210 hdr->data_len = 0;
1211
1212 msglog(hdr, "about to send status magic %d\n", err);
1213 /* hdr has been in host byteorder this whole time */
1214 return r2net_send_tcp_msg(sock, &vec, 1, sizeof(struct r2net_msg));
1215}
1216
1217/*
1218 * "data magic" is a long version of "status magic" where the message
1219 * payload actually contains data to be passed in reply to certain messages
1220 */
1221static int r2net_send_data_magic(struct r2net_sock_container *sc,
1222 struct r2net_msg *hdr,
1223 void *data, size_t data_len,
1224 enum r2net_system_error syserr, int err)
1225{
1226 struct kvec vec[2];
1227 int ret;
1228
1229 vec[0].iov_base = hdr;
1230 vec[0].iov_len = sizeof(struct r2net_msg);
1231 vec[1].iov_base = data;
1232 vec[1].iov_len = data_len;
1233
1234 BUG_ON(syserr >= R2NET_ERR_MAX);
1235
1236 /* leave other fields intact from the incoming message, msg_num
1237 * in particular */
1238 hdr->sys_status = cpu_to_be32(syserr);
1239 hdr->status = cpu_to_be32(err);
1240 hdr->magic = cpu_to_be16(R2NET_MSG_DATA_MAGIC); /* twiddle magic */
1241 hdr->data_len = cpu_to_be16(data_len);
1242
1243 msglog(hdr, "about to send data magic %d\n", err);
1244 /* hdr has been in host byteorder this whole time */
1245 ret = r2net_send_tcp_msg(sc->sc_sock, vec, 2,
1246 sizeof(struct r2net_msg) + data_len);
1247 return ret;
1248}
1249
1250/*
1251 * called by a message handler to convert an otherwise normal reply
1252 * message into a "data magic" message
1253 */
1254void r2net_force_data_magic(struct r2net_msg *hdr, u16 msgtype, u32 msgkey)
1255{
1256 hdr->magic = cpu_to_be16(R2NET_MSG_DATA_MAGIC);
1257 hdr->msg_type = cpu_to_be16(msgtype);
1258 hdr->key = cpu_to_be32(msgkey);
1259}
1260
1261/* this returns -errno if the header was unknown or too large, etc.
1262 * after this is called the buffer us reused for the next message */
1263static int r2net_process_message(struct r2net_sock_container *sc,
1264 struct r2net_msg *hdr)
1265{
1266 struct r2net_node *nn = r2net_nn_from_num(sc->sc_node->nd_num);
1267 int ret = 0, handler_status;
1268 enum r2net_system_error syserr;
1269 struct r2net_msg_handler *nmh = NULL;
1270 void *ret_data = NULL;
1271 int data_magic = 0;
1272
1273 msglog(hdr, "processing message\n");
1274
1275 r2net_sc_postpone_idle(sc);
1276
1277 switch (be16_to_cpu(hdr->magic)) {
1278
1279 case R2NET_MSG_STATUS_MAGIC:
1280 /* special type for returning message status */
1281 r2net_complete_nsw(nn, NULL, be32_to_cpu(hdr->msg_num),
1282 be32_to_cpu(hdr->sys_status),
1283 be32_to_cpu(hdr->status));
1284 goto out;
1285 case R2NET_MSG_KEEP_REQ_MAGIC:
1286 r2net_sendpage(sc, r2net_keep_resp, sizeof(*r2net_keep_resp));
1287 goto out;
1288 case R2NET_MSG_KEEP_RESP_MAGIC:
1289 goto out;
1290 case R2NET_MSG_MAGIC:
1291 break;
1292 case R2NET_MSG_DATA_MAGIC:
1293 /*
1294 * unlike a normal status magic, a data magic DOES
1295 * (MUST) have a handler, so the control flow is
1296 * a little funky here as a result
1297 */
1298 data_magic = 1;
1299 break;
1300 default:
1301 msglog(hdr, "bad magic\n");
1302 ret = -EINVAL;
1303 goto out;
1304 break;
1305 }
1306
1307 /* find a handler for it */
1308 handler_status = 0;
1309 nmh = r2net_handler_get(be16_to_cpu(hdr->msg_type),
1310 be32_to_cpu(hdr->key));
1311 if (!nmh) {
1312 mlog(ML_TCP, "couldn't find handler for type %u key %08x\n",
1313 be16_to_cpu(hdr->msg_type), be32_to_cpu(hdr->key));
1314 syserr = R2NET_ERR_NO_HNDLR;
1315 goto out_respond;
1316 }
1317
1318 syserr = R2NET_ERR_NONE;
1319
1320 if (be16_to_cpu(hdr->data_len) > nmh->nh_max_len)
1321 syserr = R2NET_ERR_OVERFLOW;
1322
1323 if (syserr != R2NET_ERR_NONE) {
1324 pr_err("ramster_r2net, message length problem\n");
1325 goto out_respond;
1326 }
1327
1328 r2net_set_func_start_time(sc);
1329 sc->sc_msg_key = be32_to_cpu(hdr->key);
1330 sc->sc_msg_type = be16_to_cpu(hdr->msg_type);
1331 handler_status = (nmh->nh_func)(hdr, sizeof(struct r2net_msg) +
1332 be16_to_cpu(hdr->data_len),
1333 nmh->nh_func_data, &ret_data);
1334 if (data_magic) {
1335 /*
1336 * handler handled data sent in reply to request
1337 * so complete the transaction
1338 */
1339 r2net_complete_nsw(nn, NULL, be32_to_cpu(hdr->msg_num),
1340 be32_to_cpu(hdr->sys_status), handler_status);
1341 goto out;
1342 }
1343 /*
1344 * handler changed magic to DATA_MAGIC to reply to request for data,
1345 * implies ret_data points to data to return and handler_status
1346 * is the number of bytes of data
1347 */
1348 if (be16_to_cpu(hdr->magic) == R2NET_MSG_DATA_MAGIC) {
1349 ret = r2net_send_data_magic(sc, hdr,
1350 ret_data, handler_status,
1351 syserr, 0);
1352 hdr = NULL;
1353 mlog(0, "sending data reply %d, syserr %d returned %d\n",
1354 handler_status, syserr, ret);
1355 r2net_set_func_stop_time(sc);
1356
1357 r2net_update_recv_stats(sc);
1358 goto out;
1359 }
1360 r2net_set_func_stop_time(sc);
1361
1362 r2net_update_recv_stats(sc);
1363
1364out_respond:
1365 /* this destroys the hdr, so don't use it after this */
1366 mutex_lock(&sc->sc_send_lock);
1367 ret = r2net_send_status_magic(sc->sc_sock, hdr, syserr,
1368 handler_status);
1369 mutex_unlock(&sc->sc_send_lock);
1370 hdr = NULL;
1371 mlog(0, "sending handler status %d, syserr %d returned %d\n",
1372 handler_status, syserr, ret);
1373
1374 if (nmh) {
1375 BUG_ON(ret_data != NULL && nmh->nh_post_func == NULL);
1376 if (nmh->nh_post_func)
1377 (nmh->nh_post_func)(handler_status, nmh->nh_func_data,
1378 ret_data);
1379 }
1380
1381out:
1382 if (nmh)
1383 r2net_handler_put(nmh);
1384 return ret;
1385}
1386
1387static int r2net_check_handshake(struct r2net_sock_container *sc)
1388{
1389 struct r2net_handshake *hand = page_address(sc->sc_page);
1390 struct r2net_node *nn = r2net_nn_from_num(sc->sc_node->nd_num);
1391
1392 if (hand->protocol_version != cpu_to_be64(R2NET_PROTOCOL_VERSION)) {
1393 pr_notice("ramster: " SC_NODEF_FMT " Advertised net "
1394 "protocol version %llu but %llu is required. "
1395 "Disconnecting.\n", sc->sc_node->nd_name,
1396 sc->sc_node->nd_num, &sc->sc_node->nd_ipv4_address,
1397 ntohs(sc->sc_node->nd_ipv4_port),
1398 (unsigned long long)be64_to_cpu(hand->protocol_version),
1399 R2NET_PROTOCOL_VERSION);
1400
1401 /* don't bother reconnecting if its the wrong version. */
1402 r2net_ensure_shutdown(nn, sc, -ENOTCONN);
1403 return -1;
1404 }
1405
1406 /*
1407 * Ensure timeouts are consistent with other nodes, otherwise
1408 * we can end up with one node thinking that the other must be down,
1409 * but isn't. This can ultimately cause corruption.
1410 */
1411 if (be32_to_cpu(hand->r2net_idle_timeout_ms) !=
1412 r2net_idle_timeout()) {
1413 pr_notice("ramster: " SC_NODEF_FMT " uses a network "
1414 "idle timeout of %u ms, but we use %u ms locally. "
1415 "Disconnecting.\n", sc->sc_node->nd_name,
1416 sc->sc_node->nd_num, &sc->sc_node->nd_ipv4_address,
1417 ntohs(sc->sc_node->nd_ipv4_port),
1418 be32_to_cpu(hand->r2net_idle_timeout_ms),
1419 r2net_idle_timeout());
1420 r2net_ensure_shutdown(nn, sc, -ENOTCONN);
1421 return -1;
1422 }
1423
1424 if (be32_to_cpu(hand->r2net_keepalive_delay_ms) !=
1425 r2net_keepalive_delay()) {
1426 pr_notice("ramster: " SC_NODEF_FMT " uses a keepalive "
1427 "delay of %u ms, but we use %u ms locally. "
1428 "Disconnecting.\n", sc->sc_node->nd_name,
1429 sc->sc_node->nd_num, &sc->sc_node->nd_ipv4_address,
1430 ntohs(sc->sc_node->nd_ipv4_port),
1431 be32_to_cpu(hand->r2net_keepalive_delay_ms),
1432 r2net_keepalive_delay());
1433 r2net_ensure_shutdown(nn, sc, -ENOTCONN);
1434 return -1;
1435 }
1436
1437 if (be32_to_cpu(hand->r2hb_heartbeat_timeout_ms) !=
1438 R2HB_MAX_WRITE_TIMEOUT_MS) {
1439 pr_notice("ramster: " SC_NODEF_FMT " uses a heartbeat "
1440 "timeout of %u ms, but we use %u ms locally. "
1441 "Disconnecting.\n", sc->sc_node->nd_name,
1442 sc->sc_node->nd_num, &sc->sc_node->nd_ipv4_address,
1443 ntohs(sc->sc_node->nd_ipv4_port),
1444 be32_to_cpu(hand->r2hb_heartbeat_timeout_ms),
1445 R2HB_MAX_WRITE_TIMEOUT_MS);
1446 r2net_ensure_shutdown(nn, sc, -ENOTCONN);
1447 return -1;
1448 }
1449
1450 sc->sc_handshake_ok = 1;
1451
1452 spin_lock(&nn->nn_lock);
1453 /* set valid and queue the idle timers only if it hasn't been
1454 * shut down already */
1455 if (nn->nn_sc == sc) {
1456 r2net_sc_reset_idle_timer(sc);
1457 atomic_set(&nn->nn_timeout, 0);
1458 r2net_set_nn_state(nn, sc, 1, 0);
1459 }
1460 spin_unlock(&nn->nn_lock);
1461
1462 /* shift everything up as though it wasn't there */
1463 sc->sc_page_off -= sizeof(struct r2net_handshake);
1464 if (sc->sc_page_off)
1465 memmove(hand, hand + 1, sc->sc_page_off);
1466
1467 return 0;
1468}
1469
1470/* this demuxes the queued rx bytes into header or payload bits and calls
1471 * handlers as each full message is read off the socket. it returns -error,
1472 * == 0 eof, or > 0 for progress made.*/
1473static int r2net_advance_rx(struct r2net_sock_container *sc)
1474{
1475 struct r2net_msg *hdr;
1476 int ret = 0;
1477 void *data;
1478 size_t datalen;
1479
1480 sclog(sc, "receiving\n");
1481 r2net_set_advance_start_time(sc);
1482
1483 if (unlikely(sc->sc_handshake_ok == 0)) {
1484 if (sc->sc_page_off < sizeof(struct r2net_handshake)) {
1485 data = page_address(sc->sc_page) + sc->sc_page_off;
1486 datalen = sizeof(struct r2net_handshake) -
1487 sc->sc_page_off;
1488 ret = r2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1489 if (ret > 0)
1490 sc->sc_page_off += ret;
1491 }
1492
1493 if (sc->sc_page_off == sizeof(struct r2net_handshake)) {
1494 r2net_check_handshake(sc);
1495 if (unlikely(sc->sc_handshake_ok == 0))
1496 ret = -EPROTO;
1497 }
1498 goto out;
1499 }
1500
1501 /* do we need more header? */
1502 if (sc->sc_page_off < sizeof(struct r2net_msg)) {
1503 data = page_address(sc->sc_page) + sc->sc_page_off;
1504 datalen = sizeof(struct r2net_msg) - sc->sc_page_off;
1505 ret = r2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1506 if (ret > 0) {
1507 sc->sc_page_off += ret;
1508 /* only swab incoming here.. we can
1509 * only get here once as we cross from
1510 * being under to over */
1511 if (sc->sc_page_off == sizeof(struct r2net_msg)) {
1512 hdr = page_address(sc->sc_page);
1513 if (be16_to_cpu(hdr->data_len) >
1514 R2NET_MAX_PAYLOAD_BYTES)
1515 ret = -EOVERFLOW;
1516 WARN_ON_ONCE(ret == -EOVERFLOW);
1517 }
1518 }
1519 if (ret <= 0)
1520 goto out;
1521 }
1522
1523 if (sc->sc_page_off < sizeof(struct r2net_msg)) {
1524 /* oof, still don't have a header */
1525 goto out;
1526 }
1527
1528 /* this was swabbed above when we first read it */
1529 hdr = page_address(sc->sc_page);
1530
1531 msglog(hdr, "at page_off %zu\n", sc->sc_page_off);
1532
1533 /* do we need more payload? */
1534 if (sc->sc_page_off - sizeof(struct r2net_msg) <
1535 be16_to_cpu(hdr->data_len)) {
1536 /* need more payload */
1537 data = page_address(sc->sc_page) + sc->sc_page_off;
1538 datalen = (sizeof(struct r2net_msg) +
1539 be16_to_cpu(hdr->data_len)) -
1540 sc->sc_page_off;
1541 ret = r2net_recv_tcp_msg(sc->sc_sock, data, datalen);
1542 if (ret > 0)
1543 sc->sc_page_off += ret;
1544 if (ret <= 0)
1545 goto out;
1546 }
1547
1548 if (sc->sc_page_off - sizeof(struct r2net_msg) ==
1549 be16_to_cpu(hdr->data_len)) {
1550 /* we can only get here once, the first time we read
1551 * the payload.. so set ret to progress if the handler
1552 * works out. after calling this the message is toast */
1553 ret = r2net_process_message(sc, hdr);
1554 if (ret == 0)
1555 ret = 1;
1556 sc->sc_page_off = 0;
1557 }
1558
1559out:
1560 sclog(sc, "ret = %d\n", ret);
1561 r2net_set_advance_stop_time(sc);
1562 return ret;
1563}
1564
1565/* this work func is triggerd by data ready. it reads until it can read no
1566 * more. it interprets 0, eof, as fatal. if data_ready hits while we're doing
1567 * our work the work struct will be marked and we'll be called again. */
1568static void r2net_rx_until_empty(struct work_struct *work)
1569{
1570 struct r2net_sock_container *sc =
1571 container_of(work, struct r2net_sock_container, sc_rx_work);
1572 int ret;
1573
1574 do {
1575 ret = r2net_advance_rx(sc);
1576 } while (ret > 0);
1577
1578 if (ret <= 0 && ret != -EAGAIN) {
1579 struct r2net_node *nn = r2net_nn_from_num(sc->sc_node->nd_num);
1580 sclog(sc, "saw error %d, closing\n", ret);
1581 /* not permanent so read failed handshake can retry */
1582 r2net_ensure_shutdown(nn, sc, 0);
1583 }
1584 sc_put(sc);
1585}
1586
1587static int r2net_set_nodelay(struct socket *sock)
1588{
1589 int ret, val = 1;
1590 mm_segment_t oldfs;
1591
1592 oldfs = get_fs();
1593 set_fs(KERNEL_DS);
1594
1595 /*
1596 * Dear unsuspecting programmer,
1597 *
1598 * Don't use sock_setsockopt() for SOL_TCP. It doesn't check its level
1599 * argument and assumes SOL_SOCKET so, say, your TCP_NODELAY will
1600 * silently turn into SO_DEBUG.
1601 *
1602 * Yours,
1603 * Keeper of hilariously fragile interfaces.
1604 */
1605 ret = sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY,
1606 (char __user *)&val, sizeof(val));
1607
1608 set_fs(oldfs);
1609 return ret;
1610}
1611
1612static void r2net_initialize_handshake(void)
1613{
1614 r2net_hand->r2hb_heartbeat_timeout_ms = cpu_to_be32(
1615 R2HB_MAX_WRITE_TIMEOUT_MS);
1616 r2net_hand->r2net_idle_timeout_ms = cpu_to_be32(r2net_idle_timeout());
1617 r2net_hand->r2net_keepalive_delay_ms = cpu_to_be32(
1618 r2net_keepalive_delay());
1619 r2net_hand->r2net_reconnect_delay_ms = cpu_to_be32(
1620 r2net_reconnect_delay());
1621}
1622
1623/* ------------------------------------------------------------ */
1624
1625/* called when a connect completes and after a sock is accepted. the
1626 * rx path will see the response and mark the sc valid */
1627static void r2net_sc_connect_completed(struct work_struct *work)
1628{
1629 struct r2net_sock_container *sc =
1630 container_of(work, struct r2net_sock_container,
1631 sc_connect_work);
1632
1633 mlog(ML_MSG, "sc sending handshake with ver %llu id %llx\n",
1634 (unsigned long long)R2NET_PROTOCOL_VERSION,
1635 (unsigned long long)be64_to_cpu(r2net_hand->connector_id));
1636
1637 r2net_initialize_handshake();
1638 r2net_sendpage(sc, r2net_hand, sizeof(*r2net_hand));
1639 sc_put(sc);
1640}
1641
1642/* this is called as a work_struct func. */
1643static void r2net_sc_send_keep_req(struct work_struct *work)
1644{
1645 struct r2net_sock_container *sc =
1646 container_of(work, struct r2net_sock_container,
1647 sc_keepalive_work.work);
1648
1649 r2net_sendpage(sc, r2net_keep_req, sizeof(*r2net_keep_req));
1650 sc_put(sc);
1651}
1652
1653/* socket shutdown does a del_timer_sync against this as it tears down.
1654 * we can't start this timer until we've got to the point in sc buildup
1655 * where shutdown is going to be involved */
1656static void r2net_idle_timer(unsigned long data)
1657{
1658 struct r2net_sock_container *sc = (struct r2net_sock_container *)data;
1659 struct r2net_node *nn = r2net_nn_from_num(sc->sc_node->nd_num);
1660#ifdef CONFIG_DEBUG_FS
1661 unsigned long msecs = ktime_to_ms(ktime_get()) -
1662 ktime_to_ms(sc->sc_tv_timer);
1663#else
1664 unsigned long msecs = r2net_idle_timeout();
1665#endif
1666
1667 pr_notice("ramster: Connection to " SC_NODEF_FMT " has been "
1668 "idle for %lu.%lu secs, shutting it down.\n",
1669 sc->sc_node->nd_name, sc->sc_node->nd_num,
1670 &sc->sc_node->nd_ipv4_address, ntohs(sc->sc_node->nd_ipv4_port),
1671 msecs / 1000, msecs % 1000);
1672
1673 /*
1674 * Initialize the nn_timeout so that the next connection attempt
1675 * will continue in r2net_start_connect.
1676 */
1677 atomic_set(&nn->nn_timeout, 1);
1678 r2net_sc_queue_work(sc, &sc->sc_shutdown_work);
1679}
1680
1681static void r2net_sc_reset_idle_timer(struct r2net_sock_container *sc)
1682{
1683 r2net_sc_cancel_delayed_work(sc, &sc->sc_keepalive_work);
1684 r2net_sc_queue_delayed_work(sc, &sc->sc_keepalive_work,
1685 msecs_to_jiffies(r2net_keepalive_delay()));
1686 r2net_set_sock_timer(sc);
1687 mod_timer(&sc->sc_idle_timeout,
1688 jiffies + msecs_to_jiffies(r2net_idle_timeout()));
1689}
1690
1691static void r2net_sc_postpone_idle(struct r2net_sock_container *sc)
1692{
1693 /* Only push out an existing timer */
1694 if (timer_pending(&sc->sc_idle_timeout))
1695 r2net_sc_reset_idle_timer(sc);
1696}
1697
1698/* this work func is kicked whenever a path sets the nn state which doesn't
1699 * have valid set. This includes seeing hb come up, losing a connection,
1700 * having a connect attempt fail, etc. This centralizes the logic which decides
1701 * if a connect attempt should be made or if we should give up and all future
1702 * transmit attempts should fail */
1703static void r2net_start_connect(struct work_struct *work)
1704{
1705 struct r2net_node *nn =
1706 container_of(work, struct r2net_node, nn_connect_work.work);
1707 struct r2net_sock_container *sc = NULL;
1708 struct r2nm_node *node = NULL, *mynode = NULL;
1709 struct socket *sock = NULL;
1710 struct sockaddr_in myaddr = {0, }, remoteaddr = {0, };
1711 int ret = 0, stop;
1712 unsigned int timeout;
1713
1714 /* if we're greater we initiate tx, otherwise we accept */
1715 if (r2nm_this_node() <= r2net_num_from_nn(nn))
1716 goto out;
1717
1718 /* watch for racing with tearing a node down */
1719 node = r2nm_get_node_by_num(r2net_num_from_nn(nn));
1720 if (node == NULL) {
1721 ret = 0;
1722 goto out;
1723 }
1724
1725 mynode = r2nm_get_node_by_num(r2nm_this_node());
1726 if (mynode == NULL) {
1727 ret = 0;
1728 goto out;
1729 }
1730
1731 spin_lock(&nn->nn_lock);
1732 /*
1733 * see if we already have one pending or have given up.
1734 * For nn_timeout, it is set when we close the connection
1735 * because of the idle time out. So it means that we have
1736 * at least connected to that node successfully once,
1737 * now try to connect to it again.
1738 */
1739 timeout = atomic_read(&nn->nn_timeout);
1740 stop = (nn->nn_sc ||
1741 (nn->nn_persistent_error &&
1742 (nn->nn_persistent_error != -ENOTCONN || timeout == 0)));
1743 spin_unlock(&nn->nn_lock);
1744 if (stop)
1745 goto out;
1746
1747 nn->nn_last_connect_attempt = jiffies;
1748
1749 sc = sc_alloc(node);
1750 if (sc == NULL) {
1751 mlog(0, "couldn't allocate sc\n");
1752 ret = -ENOMEM;
1753 goto out;
1754 }
1755
1756 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
1757 if (ret < 0) {
1758 mlog(0, "can't create socket: %d\n", ret);
1759 goto out;
1760 }
1761 sc->sc_sock = sock; /* freed by sc_kref_release */
1762
1763 sock->sk->sk_allocation = GFP_ATOMIC;
1764
1765 myaddr.sin_family = AF_INET;
1766 myaddr.sin_addr.s_addr = mynode->nd_ipv4_address;
1767 myaddr.sin_port = htons(0); /* any port */
1768
1769 ret = sock->ops->bind(sock, (struct sockaddr *)&myaddr,
1770 sizeof(myaddr));
1771 if (ret) {
1772 mlog(ML_ERROR, "bind failed with %d at address %pI4\n",
1773 ret, &mynode->nd_ipv4_address);
1774 goto out;
1775 }
1776
1777 ret = r2net_set_nodelay(sc->sc_sock);
1778 if (ret) {
1779 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1780 goto out;
1781 }
1782
1783 r2net_register_callbacks(sc->sc_sock->sk, sc);
1784
1785 spin_lock(&nn->nn_lock);
1786 /* handshake completion will set nn->nn_sc_valid */
1787 r2net_set_nn_state(nn, sc, 0, 0);
1788 spin_unlock(&nn->nn_lock);
1789
1790 remoteaddr.sin_family = AF_INET;
1791 remoteaddr.sin_addr.s_addr = node->nd_ipv4_address;
1792 remoteaddr.sin_port = node->nd_ipv4_port;
1793
1794 ret = sc->sc_sock->ops->connect(sc->sc_sock,
1795 (struct sockaddr *)&remoteaddr,
1796 sizeof(remoteaddr),
1797 O_NONBLOCK);
1798 if (ret == -EINPROGRESS)
1799 ret = 0;
1800
1801out:
1802 if (ret) {
1803 pr_notice("ramster: Connect attempt to " SC_NODEF_FMT
1804 " failed with errno %d\n", sc->sc_node->nd_name,
1805 sc->sc_node->nd_num, &sc->sc_node->nd_ipv4_address,
1806 ntohs(sc->sc_node->nd_ipv4_port), ret);
1807 /* 0 err so that another will be queued and attempted
1808 * from set_nn_state */
1809 if (sc)
1810 r2net_ensure_shutdown(nn, sc, 0);
1811 }
1812 if (sc)
1813 sc_put(sc);
1814 if (node)
1815 r2nm_node_put(node);
1816 if (mynode)
1817 r2nm_node_put(mynode);
1818
1819 return;
1820}
1821
1822static void r2net_connect_expired(struct work_struct *work)
1823{
1824 struct r2net_node *nn =
1825 container_of(work, struct r2net_node, nn_connect_expired.work);
1826
1827 spin_lock(&nn->nn_lock);
1828 if (!nn->nn_sc_valid) {
1829 pr_notice("ramster: No connection established with "
1830 "node %u after %u.%u seconds, giving up.\n",
1831 r2net_num_from_nn(nn),
1832 r2net_idle_timeout() / 1000,
1833 r2net_idle_timeout() % 1000);
1834
1835 r2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1836 }
1837 spin_unlock(&nn->nn_lock);
1838}
1839
1840static void r2net_still_up(struct work_struct *work)
1841{
1842}
1843
1844/* ------------------------------------------------------------ */
1845
1846void r2net_disconnect_node(struct r2nm_node *node)
1847{
1848 struct r2net_node *nn = r2net_nn_from_num(node->nd_num);
1849
1850 /* don't reconnect until it's heartbeating again */
1851 spin_lock(&nn->nn_lock);
1852 atomic_set(&nn->nn_timeout, 0);
1853 r2net_set_nn_state(nn, NULL, 0, -ENOTCONN);
1854 spin_unlock(&nn->nn_lock);
1855
1856 if (r2net_wq) {
1857 cancel_delayed_work(&nn->nn_connect_expired);
1858 cancel_delayed_work(&nn->nn_connect_work);
1859 cancel_delayed_work(&nn->nn_still_up);
1860 flush_workqueue(r2net_wq);
1861 }
1862}
1863
1864static void r2net_hb_node_down_cb(struct r2nm_node *node, int node_num,
1865 void *data)
1866{
1867 if (!node)
1868 return;
1869
1870 if (node_num != r2nm_this_node())
1871 r2net_disconnect_node(node);
1872
1873 BUG_ON(atomic_read(&r2net_connected_peers) < 0);
1874}
1875
1876static void r2net_hb_node_up_cb(struct r2nm_node *node, int node_num,
1877 void *data)
1878{
1879 struct r2net_node *nn = r2net_nn_from_num(node_num);
1880
1881 BUG_ON(!node);
1882
1883 /* ensure an immediate connect attempt */
1884 nn->nn_last_connect_attempt = jiffies -
1885 (msecs_to_jiffies(r2net_reconnect_delay()) + 1);
1886
1887 if (node_num != r2nm_this_node()) {
1888 /* believe it or not, accept and node hearbeating testing
1889 * can succeed for this node before we got here.. so
1890 * only use set_nn_state to clear the persistent error
1891 * if that hasn't already happened */
1892 spin_lock(&nn->nn_lock);
1893 atomic_set(&nn->nn_timeout, 0);
1894 if (nn->nn_persistent_error)
1895 r2net_set_nn_state(nn, NULL, 0, 0);
1896 spin_unlock(&nn->nn_lock);
1897 }
1898}
1899
1900void r2net_unregister_hb_callbacks(void)
1901{
1902 r2hb_unregister_callback(NULL, &r2net_hb_up);
1903 r2hb_unregister_callback(NULL, &r2net_hb_down);
1904}
1905
1906int r2net_register_hb_callbacks(void)
1907{
1908 int ret;
1909
1910 r2hb_setup_callback(&r2net_hb_down, R2HB_NODE_DOWN_CB,
1911 r2net_hb_node_down_cb, NULL, R2NET_HB_PRI);
1912 r2hb_setup_callback(&r2net_hb_up, R2HB_NODE_UP_CB,
1913 r2net_hb_node_up_cb, NULL, R2NET_HB_PRI);
1914
1915 ret = r2hb_register_callback(NULL, &r2net_hb_up);
1916 if (ret == 0)
1917 ret = r2hb_register_callback(NULL, &r2net_hb_down);
1918
1919 if (ret)
1920 r2net_unregister_hb_callbacks();
1921
1922 return ret;
1923}
1924
1925/* ------------------------------------------------------------ */
1926
1927static int r2net_accept_one(struct socket *sock)
1928{
1929 int ret, slen;
1930 struct sockaddr_in sin;
1931 struct socket *new_sock = NULL;
1932 struct r2nm_node *node = NULL;
1933 struct r2nm_node *local_node = NULL;
1934 struct r2net_sock_container *sc = NULL;
1935 struct r2net_node *nn;
1936
1937 BUG_ON(sock == NULL);
1938 ret = sock_create_lite(sock->sk->sk_family, sock->sk->sk_type,
1939 sock->sk->sk_protocol, &new_sock);
1940 if (ret)
1941 goto out;
1942
1943 new_sock->type = sock->type;
1944 new_sock->ops = sock->ops;
1945 ret = sock->ops->accept(sock, new_sock, O_NONBLOCK);
1946 if (ret < 0)
1947 goto out;
1948
1949 new_sock->sk->sk_allocation = GFP_ATOMIC;
1950
1951 ret = r2net_set_nodelay(new_sock);
1952 if (ret) {
1953 mlog(ML_ERROR, "setting TCP_NODELAY failed with %d\n", ret);
1954 goto out;
1955 }
1956
1957 slen = sizeof(sin);
1958 ret = new_sock->ops->getname(new_sock, (struct sockaddr *) &sin,
1959 &slen, 1);
1960 if (ret < 0)
1961 goto out;
1962
1963 node = r2nm_get_node_by_ip(sin.sin_addr.s_addr);
1964 if (node == NULL) {
1965 pr_notice("ramster: Attempt to connect from unknown "
1966 "node at %pI4:%d\n", &sin.sin_addr.s_addr,
1967 ntohs(sin.sin_port));
1968 ret = -EINVAL;
1969 goto out;
1970 }
1971
1972 if (r2nm_this_node() >= node->nd_num) {
1973 local_node = r2nm_get_node_by_num(r2nm_this_node());
1974 pr_notice("ramster: Unexpected connect attempt seen "
1975 "at node '%s' (%u, %pI4:%d) from node '%s' (%u, "
1976 "%pI4:%d)\n", local_node->nd_name, local_node->nd_num,
1977 &(local_node->nd_ipv4_address),
1978 ntohs(local_node->nd_ipv4_port), node->nd_name,
1979 node->nd_num, &sin.sin_addr.s_addr, ntohs(sin.sin_port));
1980 ret = -EINVAL;
1981 goto out;
1982 }
1983
1984 /* this happens all the time when the other node sees our heartbeat
1985 * and tries to connect before we see their heartbeat */
1986 if (!r2hb_check_node_heartbeating_from_callback(node->nd_num)) {
1987 mlog(ML_CONN, "attempt to connect from node '%s' at "
1988 "%pI4:%d but it isn't heartbeating\n",
1989 node->nd_name, &sin.sin_addr.s_addr,
1990 ntohs(sin.sin_port));
1991 ret = -EINVAL;
1992 goto out;
1993 }
1994
1995 nn = r2net_nn_from_num(node->nd_num);
1996
1997 spin_lock(&nn->nn_lock);
1998 if (nn->nn_sc)
1999 ret = -EBUSY;
2000 else
2001 ret = 0;
2002 spin_unlock(&nn->nn_lock);
2003 if (ret) {
2004 pr_notice("ramster: Attempt to connect from node '%s' "
2005 "at %pI4:%d but it already has an open connection\n",
2006 node->nd_name, &sin.sin_addr.s_addr,
2007 ntohs(sin.sin_port));
2008 goto out;
2009 }
2010
2011 sc = sc_alloc(node);
2012 if (sc == NULL) {
2013 ret = -ENOMEM;
2014 goto out;
2015 }
2016
2017 sc->sc_sock = new_sock;
2018 new_sock = NULL;
2019
2020 spin_lock(&nn->nn_lock);
2021 atomic_set(&nn->nn_timeout, 0);
2022 r2net_set_nn_state(nn, sc, 0, 0);
2023 spin_unlock(&nn->nn_lock);
2024
2025 r2net_register_callbacks(sc->sc_sock->sk, sc);
2026 r2net_sc_queue_work(sc, &sc->sc_rx_work);
2027
2028 r2net_initialize_handshake();
2029 r2net_sendpage(sc, r2net_hand, sizeof(*r2net_hand));
2030
2031out:
2032 if (new_sock)
2033 sock_release(new_sock);
2034 if (node)
2035 r2nm_node_put(node);
2036 if (local_node)
2037 r2nm_node_put(local_node);
2038 if (sc)
2039 sc_put(sc);
2040 return ret;
2041}
2042
2043static void r2net_accept_many(struct work_struct *work)
2044{
2045 struct socket *sock = r2net_listen_sock;
2046 while (r2net_accept_one(sock) == 0)
2047 cond_resched();
2048}
2049
2050static void r2net_listen_data_ready(struct sock *sk, int bytes)
2051{
2052 void (*ready)(struct sock *sk, int bytes);
2053
2054 read_lock(&sk->sk_callback_lock);
2055 ready = sk->sk_user_data;
2056 if (ready == NULL) { /* check for teardown race */
2057 ready = sk->sk_data_ready;
2058 goto out;
2059 }
2060
2061 /* ->sk_data_ready is also called for a newly established child socket
2062 * before it has been accepted and the acceptor has set up their
2063 * data_ready.. we only want to queue listen work for our listening
2064 * socket */
2065 if (sk->sk_state == TCP_LISTEN) {
2066 mlog(ML_TCP, "bytes: %d\n", bytes);
2067 queue_work(r2net_wq, &r2net_listen_work);
2068 }
2069
2070out:
2071 read_unlock(&sk->sk_callback_lock);
2072 ready(sk, bytes);
2073}
2074
2075static int r2net_open_listening_sock(__be32 addr, __be16 port)
2076{
2077 struct socket *sock = NULL;
2078 int ret;
2079 struct sockaddr_in sin = {
2080 .sin_family = PF_INET,
2081 .sin_addr = { .s_addr = addr },
2082 .sin_port = port,
2083 };
2084
2085 ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2086 if (ret < 0) {
2087 pr_err("ramster: Error %d while creating socket\n", ret);
2088 goto out;
2089 }
2090
2091 sock->sk->sk_allocation = GFP_ATOMIC;
2092
2093 write_lock_bh(&sock->sk->sk_callback_lock);
2094 sock->sk->sk_user_data = sock->sk->sk_data_ready;
2095 sock->sk->sk_data_ready = r2net_listen_data_ready;
2096 write_unlock_bh(&sock->sk->sk_callback_lock);
2097
2098 r2net_listen_sock = sock;
2099 INIT_WORK(&r2net_listen_work, r2net_accept_many);
2100
2101 sock->sk->sk_reuse = /* SK_CAN_REUSE FIXME FOR 3.4 */ 1;
2102 ret = sock->ops->bind(sock, (struct sockaddr *)&sin, sizeof(sin));
2103 if (ret < 0) {
2104 pr_err("ramster: Error %d while binding socket at %pI4:%u\n",
2105 ret, &addr, ntohs(port));
2106 goto out;
2107 }
2108
2109 ret = sock->ops->listen(sock, 64);
2110 if (ret < 0)
2111 pr_err("ramster: Error %d while listening on %pI4:%u\n",
2112 ret, &addr, ntohs(port));
2113
2114out:
2115 if (ret) {
2116 r2net_listen_sock = NULL;
2117 if (sock)
2118 sock_release(sock);
2119 }
2120 return ret;
2121}
2122
2123/*
2124 * called from node manager when we should bring up our network listening
2125 * socket. node manager handles all the serialization to only call this
2126 * once and to match it with r2net_stop_listening(). note,
2127 * r2nm_this_node() doesn't work yet as we're being called while it
2128 * is being set up.
2129 */
2130int r2net_start_listening(struct r2nm_node *node)
2131{
2132 int ret = 0;
2133
2134 BUG_ON(r2net_wq != NULL);
2135 BUG_ON(r2net_listen_sock != NULL);
2136
2137 mlog(ML_KTHREAD, "starting r2net thread...\n");
2138 r2net_wq = create_singlethread_workqueue("r2net");
2139 if (r2net_wq == NULL) {
2140 mlog(ML_ERROR, "unable to launch r2net thread\n");
2141 return -ENOMEM; /* ? */
2142 }
2143
2144 ret = r2net_open_listening_sock(node->nd_ipv4_address,
2145 node->nd_ipv4_port);
2146 if (ret) {
2147 destroy_workqueue(r2net_wq);
2148 r2net_wq = NULL;
2149 }
2150
2151 return ret;
2152}
2153
2154/* again, r2nm_this_node() doesn't work here as we're involved in
2155 * tearing it down */
2156void r2net_stop_listening(struct r2nm_node *node)
2157{
2158 struct socket *sock = r2net_listen_sock;
2159 size_t i;
2160
2161 BUG_ON(r2net_wq == NULL);
2162 BUG_ON(r2net_listen_sock == NULL);
2163
2164 /* stop the listening socket from generating work */
2165 write_lock_bh(&sock->sk->sk_callback_lock);
2166 sock->sk->sk_data_ready = sock->sk->sk_user_data;
2167 sock->sk->sk_user_data = NULL;
2168 write_unlock_bh(&sock->sk->sk_callback_lock);
2169
2170 for (i = 0; i < ARRAY_SIZE(r2net_nodes); i++) {
2171 struct r2nm_node *node = r2nm_get_node_by_num(i);
2172 if (node) {
2173 r2net_disconnect_node(node);
2174 r2nm_node_put(node);
2175 }
2176 }
2177
2178 /* finish all work and tear down the work queue */
2179 mlog(ML_KTHREAD, "waiting for r2net thread to exit....\n");
2180 destroy_workqueue(r2net_wq);
2181 r2net_wq = NULL;
2182
2183 sock_release(r2net_listen_sock);
2184 r2net_listen_sock = NULL;
2185}
2186
2187void r2net_hb_node_up_manual(int node_num)
2188{
2189 struct r2nm_node dummy;
2190 if (r2nm_single_cluster == NULL)
2191 pr_err("ramster: cluster not alive, node_up_manual ignored\n");
2192 else {
2193 r2hb_manual_set_node_heartbeating(node_num);
2194 r2net_hb_node_up_cb(&dummy, node_num, NULL);
2195 }
2196}
2197
2198/* ------------------------------------------------------------ */
2199
2200int r2net_init(void)
2201{
2202 unsigned long i;
2203
2204 if (r2net_debugfs_init())
2205 return -ENOMEM;
2206
2207 r2net_hand = kzalloc(sizeof(struct r2net_handshake), GFP_KERNEL);
2208 r2net_keep_req = kzalloc(sizeof(struct r2net_msg), GFP_KERNEL);
2209 r2net_keep_resp = kzalloc(sizeof(struct r2net_msg), GFP_KERNEL);
2210 if (!r2net_hand || !r2net_keep_req || !r2net_keep_resp) {
2211 kfree(r2net_hand);
2212 kfree(r2net_keep_req);
2213 kfree(r2net_keep_resp);
2214 return -ENOMEM;
2215 }
2216
2217 r2net_hand->protocol_version = cpu_to_be64(R2NET_PROTOCOL_VERSION);
2218 r2net_hand->connector_id = cpu_to_be64(1);
2219
2220 r2net_keep_req->magic = cpu_to_be16(R2NET_MSG_KEEP_REQ_MAGIC);
2221 r2net_keep_resp->magic = cpu_to_be16(R2NET_MSG_KEEP_RESP_MAGIC);
2222
2223 for (i = 0; i < ARRAY_SIZE(r2net_nodes); i++) {
2224 struct r2net_node *nn = r2net_nn_from_num(i);
2225
2226 atomic_set(&nn->nn_timeout, 0);
2227 spin_lock_init(&nn->nn_lock);
2228 INIT_DELAYED_WORK(&nn->nn_connect_work, r2net_start_connect);
2229 INIT_DELAYED_WORK(&nn->nn_connect_expired,
2230 r2net_connect_expired);
2231 INIT_DELAYED_WORK(&nn->nn_still_up, r2net_still_up);
2232 /* until we see hb from a node we'll return einval */
2233 nn->nn_persistent_error = -ENOTCONN;
2234 init_waitqueue_head(&nn->nn_sc_wq);
2235 idr_init(&nn->nn_status_idr);
2236 INIT_LIST_HEAD(&nn->nn_status_list);
2237 }
2238
2239 return 0;
2240}
2241
2242void r2net_exit(void)
2243{
2244 kfree(r2net_hand);
2245 kfree(r2net_keep_req);
2246 kfree(r2net_keep_resp);
2247 r2net_debugfs_exit();
2248}
Linux kernel - Deliverables
This page took 0.212804 seconds and 5 git commands to generate.