projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[deliverable/linux.git] / drivers / staging / lustre / lnet / klnds / socklnd / socklnd_cb.c
1 /*
2 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
3 *
4 * Copyright (c) 2011, 2012, Intel Corporation.
5 *
6 * Author: Zach Brown <zab@zabbo.net>
7 * Author: Peter J. Braam <braam@clusterfs.com>
8 * Author: Phil Schwan <phil@clusterfs.com>
9 * Author: Eric Barton <eric@bartonsoftware.com>
10 *
11 * This file is part of Portals, http://www.sf.net/projects/sandiaportals/
12 *
13 * Portals is free software; you can redistribute it and/or
14 * modify it under the terms of version 2 of the GNU General Public
15 * License as published by the Free Software Foundation.
16 *
17 * Portals is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
21 *
22 * You should have received a copy of the GNU General Public License
23 * along with Portals; if not, write to the Free Software
24 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 */
26
27 #include "socklnd.h"
28
29 ksock_tx_t *
30 ksocknal_alloc_tx(int type, int size)
31 {
32 ksock_tx_t *tx = NULL;
33
34 if (type == KSOCK_MSG_NOOP) {
35 LASSERT(size == KSOCK_NOOP_TX_SIZE);
36
37 /* searching for a noop tx in free list */
38 spin_lock(&ksocknal_data.ksnd_tx_lock);
39
40 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
41 tx = list_entry(ksocknal_data.ksnd_idle_noop_txs. \
42 next, ksock_tx_t, tx_list);
43 LASSERT(tx->tx_desc_size == size);
44 list_del(&tx->tx_list);
45 }
46
47 spin_unlock(&ksocknal_data.ksnd_tx_lock);
48 }
49
50 if (tx == NULL)
51 LIBCFS_ALLOC(tx, size);
52
53 if (tx == NULL)
54 return NULL;
55
56 atomic_set(&tx->tx_refcount, 1);
57 tx->tx_zc_aborted = 0;
58 tx->tx_zc_capable = 0;
59 tx->tx_zc_checked = 0;
60 tx->tx_desc_size = size;
61
62 atomic_inc(&ksocknal_data.ksnd_nactive_txs);
63
64 return tx;
65 }
66
67 ksock_tx_t *
68 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
69 {
70 ksock_tx_t *tx;
71
72 tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
73 if (tx == NULL) {
74 CERROR("Can't allocate noop tx desc\n");
75 return NULL;
76 }
77
78 tx->tx_conn = NULL;
79 tx->tx_lnetmsg = NULL;
80 tx->tx_kiov = NULL;
81 tx->tx_nkiov = 0;
82 tx->tx_iov = tx->tx_frags.virt.iov;
83 tx->tx_niov = 1;
84 tx->tx_nonblk = nonblk;
85
86 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
87 tx->tx_msg.ksm_zc_cookies[1] = cookie;
88
89 return tx;
90 }
91
92
93 void
94 ksocknal_free_tx (ksock_tx_t *tx)
95 {
96 atomic_dec(&ksocknal_data.ksnd_nactive_txs);
97
98 if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
99 /* it's a noop tx */
100 spin_lock(&ksocknal_data.ksnd_tx_lock);
101
102 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
103
104 spin_unlock(&ksocknal_data.ksnd_tx_lock);
105 } else {
106 LIBCFS_FREE(tx, tx->tx_desc_size);
107 }
108 }
109
110 int
111 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
112 {
113 struct iovec *iov = tx->tx_iov;
114 int nob;
115 int rc;
116
117 LASSERT (tx->tx_niov > 0);
118
119 /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
120 rc = ksocknal_lib_send_iov(conn, tx);
121
122 if (rc <= 0) /* sent nothing? */
123 return rc;
124
125 nob = rc;
126 LASSERT (nob <= tx->tx_resid);
127 tx->tx_resid -= nob;
128
129 /* "consume" iov */
130 do {
131 LASSERT (tx->tx_niov > 0);
132
133 if (nob < (int) iov->iov_len) {
134 iov->iov_base = (void *)((char *)iov->iov_base + nob);
135 iov->iov_len -= nob;
136 return rc;
137 }
138
139 nob -= iov->iov_len;
140 tx->tx_iov = ++iov;
141 tx->tx_niov--;
142 } while (nob != 0);
143
144 return rc;
145 }
146
147 int
148 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
149 {
150 lnet_kiov_t *kiov = tx->tx_kiov;
151 int nob;
152 int rc;
153
154 LASSERT (tx->tx_niov == 0);
155 LASSERT (tx->tx_nkiov > 0);
156
157 /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
158 rc = ksocknal_lib_send_kiov(conn, tx);
159
160 if (rc <= 0) /* sent nothing? */
161 return rc;
162
163 nob = rc;
164 LASSERT (nob <= tx->tx_resid);
165 tx->tx_resid -= nob;
166
167 /* "consume" kiov */
168 do {
169 LASSERT(tx->tx_nkiov > 0);
170
171 if (nob < (int)kiov->kiov_len) {
172 kiov->kiov_offset += nob;
173 kiov->kiov_len -= nob;
174 return rc;
175 }
176
177 nob -= (int)kiov->kiov_len;
178 tx->tx_kiov = ++kiov;
179 tx->tx_nkiov--;
180 } while (nob != 0);
181
182 return rc;
183 }
184
185 int
186 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
187 {
188 int rc;
189 int bufnob;
190
191 if (ksocknal_data.ksnd_stall_tx != 0) {
192 set_current_state(TASK_UNINTERRUPTIBLE);
193 schedule_timeout(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
194 }
195
196 LASSERT (tx->tx_resid != 0);
197
198 rc = ksocknal_connsock_addref(conn);
199 if (rc != 0) {
200 LASSERT (conn->ksnc_closing);
201 return -ESHUTDOWN;
202 }
203
204 do {
205 if (ksocknal_data.ksnd_enomem_tx > 0) {
206 /* testing... */
207 ksocknal_data.ksnd_enomem_tx--;
208 rc = -EAGAIN;
209 } else if (tx->tx_niov != 0) {
210 rc = ksocknal_send_iov (conn, tx);
211 } else {
212 rc = ksocknal_send_kiov (conn, tx);
213 }
214
215 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
216 if (rc > 0) /* sent something? */
217 conn->ksnc_tx_bufnob += rc; /* account it */
218
219 if (bufnob < conn->ksnc_tx_bufnob) {
220 /* allocated send buffer bytes < computed; infer
221 * something got ACKed */
222 conn->ksnc_tx_deadline =
223 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
224 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
225 conn->ksnc_tx_bufnob = bufnob;
226 mb();
227 }
228
229 if (rc <= 0) { /* Didn't write anything? */
230
231 if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
232 rc = -EAGAIN;
233
234 /* Check if EAGAIN is due to memory pressure */
235 if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
236 rc = -ENOMEM;
237
238 break;
239 }
240
241 /* socket's wmem_queued now includes 'rc' bytes */
242 atomic_sub (rc, &conn->ksnc_tx_nob);
243 rc = 0;
244
245 } while (tx->tx_resid != 0);
246
247 ksocknal_connsock_decref(conn);
248 return rc;
249 }
250
251 int
252 ksocknal_recv_iov (ksock_conn_t *conn)
253 {
254 struct iovec *iov = conn->ksnc_rx_iov;
255 int nob;
256 int rc;
257
258 LASSERT (conn->ksnc_rx_niov > 0);
259
260 /* Never touch conn->ksnc_rx_iov or change connection
261 * status inside ksocknal_lib_recv_iov */
262 rc = ksocknal_lib_recv_iov(conn);
263
264 if (rc <= 0)
265 return rc;
266
267 /* received something... */
268 nob = rc;
269
270 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
271 conn->ksnc_rx_deadline =
272 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
273 mb(); /* order with setting rx_started */
274 conn->ksnc_rx_started = 1;
275
276 conn->ksnc_rx_nob_wanted -= nob;
277 conn->ksnc_rx_nob_left -= nob;
278
279 do {
280 LASSERT (conn->ksnc_rx_niov > 0);
281
282 if (nob < (int)iov->iov_len) {
283 iov->iov_len -= nob;
284 iov->iov_base = (void *)((char *)iov->iov_base + nob);
285 return -EAGAIN;
286 }
287
288 nob -= iov->iov_len;
289 conn->ksnc_rx_iov = ++iov;
290 conn->ksnc_rx_niov--;
291 } while (nob != 0);
292
293 return rc;
294 }
295
296 int
297 ksocknal_recv_kiov (ksock_conn_t *conn)
298 {
299 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
300 int nob;
301 int rc;
302 LASSERT (conn->ksnc_rx_nkiov > 0);
303
304 /* Never touch conn->ksnc_rx_kiov or change connection
305 * status inside ksocknal_lib_recv_iov */
306 rc = ksocknal_lib_recv_kiov(conn);
307
308 if (rc <= 0)
309 return rc;
310
311 /* received something... */
312 nob = rc;
313
314 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
315 conn->ksnc_rx_deadline =
316 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
317 mb(); /* order with setting rx_started */
318 conn->ksnc_rx_started = 1;
319
320 conn->ksnc_rx_nob_wanted -= nob;
321 conn->ksnc_rx_nob_left -= nob;
322
323 do {
324 LASSERT (conn->ksnc_rx_nkiov > 0);
325
326 if (nob < (int) kiov->kiov_len) {
327 kiov->kiov_offset += nob;
328 kiov->kiov_len -= nob;
329 return -EAGAIN;
330 }
331
332 nob -= kiov->kiov_len;
333 conn->ksnc_rx_kiov = ++kiov;
334 conn->ksnc_rx_nkiov--;
335 } while (nob != 0);
336
337 return 1;
338 }
339
340 int
341 ksocknal_receive (ksock_conn_t *conn)
342 {
343 /* Return 1 on success, 0 on EOF, < 0 on error.
344 * Caller checks ksnc_rx_nob_wanted to determine
345 * progress/completion. */
346 int rc;
347
348 if (ksocknal_data.ksnd_stall_rx != 0) {
349 set_current_state(TASK_UNINTERRUPTIBLE);
350 schedule_timeout(cfs_time_seconds(ksocknal_data.ksnd_stall_rx));
351 }
352
353 rc = ksocknal_connsock_addref(conn);
354 if (rc != 0) {
355 LASSERT (conn->ksnc_closing);
356 return -ESHUTDOWN;
357 }
358
359 for (;;) {
360 if (conn->ksnc_rx_niov != 0)
361 rc = ksocknal_recv_iov (conn);
362 else
363 rc = ksocknal_recv_kiov (conn);
364
365 if (rc <= 0) {
366 /* error/EOF or partial receive */
367 if (rc == -EAGAIN) {
368 rc = 1;
369 } else if (rc == 0 && conn->ksnc_rx_started) {
370 /* EOF in the middle of a message */
371 rc = -EPROTO;
372 }
373 break;
374 }
375
376 /* Completed a fragment */
377
378 if (conn->ksnc_rx_nob_wanted == 0) {
379 rc = 1;
380 break;
381 }
382 }
383
384 ksocknal_connsock_decref(conn);
385 return rc;
386 }
387
388 void
389 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
390 {
391 lnet_msg_t *lnetmsg = tx->tx_lnetmsg;
392 int rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO;
393
394 LASSERT(ni != NULL || tx->tx_conn != NULL);
395
396 if (tx->tx_conn != NULL)
397 ksocknal_conn_decref(tx->tx_conn);
398
399 if (ni == NULL && tx->tx_conn != NULL)
400 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
401
402 ksocknal_free_tx (tx);
403 if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
404 lnet_finalize (ni, lnetmsg, rc);
405 }
406
407 void
408 ksocknal_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int error)
409 {
410 ksock_tx_t *tx;
411
412 while (!list_empty (txlist)) {
413 tx = list_entry (txlist->next, ksock_tx_t, tx_list);
414
415 if (error && tx->tx_lnetmsg != NULL) {
416 CNETERR("Deleting packet type %d len %d %s->%s\n",
417 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
418 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
419 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
420 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
421 } else if (error) {
422 CNETERR("Deleting noop packet\n");
423 }
424
425 list_del (&tx->tx_list);
426
427 LASSERT (atomic_read(&tx->tx_refcount) == 1);
428 ksocknal_tx_done (ni, tx);
429 }
430 }
431
432 static void
433 ksocknal_check_zc_req(ksock_tx_t *tx)
434 {
435 ksock_conn_t *conn = tx->tx_conn;
436 ksock_peer_t *peer = conn->ksnc_peer;
437
438 /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
439 * to ksnp_zc_req_list if some fragment of this message should be sent
440 * zero-copy. Our peer will send an ACK containing this cookie when
441 * she has received this message to tell us we can signal completion.
442 * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
443 * ksnp_zc_req_list. */
444 LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
445 LASSERT (tx->tx_zc_capable);
446
447 tx->tx_zc_checked = 1;
448
449 if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
450 !conn->ksnc_zc_capable)
451 return;
452
453 /* assign cookie and queue tx to pending list, it will be released when
454 * a matching ack is received. See ksocknal_handle_zcack() */
455
456 ksocknal_tx_addref(tx);
457
458 spin_lock(&peer->ksnp_lock);
459
460 /* ZC_REQ is going to be pinned to the peer */
461 tx->tx_deadline =
462 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
463
464 LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
465
466 tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
467
468 if (peer->ksnp_zc_next_cookie == 0)
469 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
470
471 list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
472
473 spin_unlock(&peer->ksnp_lock);
474 }
475
476 static void
477 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
478 {
479 ksock_peer_t *peer = tx->tx_conn->ksnc_peer;
480
481 LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
482 LASSERT(tx->tx_zc_capable);
483
484 tx->tx_zc_checked = 0;
485
486 spin_lock(&peer->ksnp_lock);
487
488 if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
489 /* Not waiting for an ACK */
490 spin_unlock(&peer->ksnp_lock);
491 return;
492 }
493
494 tx->tx_msg.ksm_zc_cookies[0] = 0;
495 list_del(&tx->tx_zc_list);
496
497 spin_unlock(&peer->ksnp_lock);
498
499 ksocknal_tx_decref(tx);
500 }
501
502 int
503 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
504 {
505 int rc;
506
507 if (tx->tx_zc_capable && !tx->tx_zc_checked)
508 ksocknal_check_zc_req(tx);
509
510 rc = ksocknal_transmit (conn, tx);
511
512 CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
513
514 if (tx->tx_resid == 0) {
515 /* Sent everything OK */
516 LASSERT (rc == 0);
517
518 return 0;
519 }
520
521 if (rc == -EAGAIN)
522 return rc;
523
524 if (rc == -ENOMEM) {
525 static int counter;
526
527 counter++; /* exponential backoff warnings */
528 if ((counter & (-counter)) == counter)
529 CWARN("%u ENOMEM tx %p (%u allocated)\n",
530 counter, conn, atomic_read(&libcfs_kmemory));
531
532 /* Queue on ksnd_enomem_conns for retry after a timeout */
533 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
534
535 /* enomem list takes over scheduler's ref... */
536 LASSERT (conn->ksnc_tx_scheduled);
537 list_add_tail(&conn->ksnc_tx_list,
538 &ksocknal_data.ksnd_enomem_conns);
539 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
540 SOCKNAL_ENOMEM_RETRY),
541 ksocknal_data.ksnd_reaper_waketime))
542 wake_up (&ksocknal_data.ksnd_reaper_waitq);
543
544 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
545 return rc;
546 }
547
548 /* Actual error */
549 LASSERT (rc < 0);
550
551 if (!conn->ksnc_closing) {
552 switch (rc) {
553 case -ECONNRESET:
554 LCONSOLE_WARN("Host %pI4h reset our connection "
555 "while we were sending data; it may have "
556 "rebooted.\n",
557 &conn->ksnc_ipaddr);
558 break;
559 default:
560 LCONSOLE_WARN("There was an unexpected network error "
561 "while writing to %pI4h: %d.\n",
562 &conn->ksnc_ipaddr, rc);
563 break;
564 }
565 CDEBUG(D_NET, "[%p] Error %d on write to %s"
566 " ip %pI4h:%d\n", conn, rc,
567 libcfs_id2str(conn->ksnc_peer->ksnp_id),
568 &conn->ksnc_ipaddr,
569 conn->ksnc_port);
570 }
571
572 if (tx->tx_zc_checked)
573 ksocknal_uncheck_zc_req(tx);
574
575 /* it's not an error if conn is being closed */
576 ksocknal_close_conn_and_siblings (conn,
577 (conn->ksnc_closing) ? 0 : rc);
578
579 return rc;
580 }
581
582 void
583 ksocknal_launch_connection_locked (ksock_route_t *route)
584 {
585
586 /* called holding write lock on ksnd_global_lock */
587
588 LASSERT (!route->ksnr_scheduled);
589 LASSERT (!route->ksnr_connecting);
590 LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
591
592 route->ksnr_scheduled = 1; /* scheduling conn for connd */
593 ksocknal_route_addref(route); /* extra ref for connd */
594
595 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
596
597 list_add_tail(&route->ksnr_connd_list,
598 &ksocknal_data.ksnd_connd_routes);
599 wake_up(&ksocknal_data.ksnd_connd_waitq);
600
601 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
602 }
603
604 void
605 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
606 {
607 ksock_route_t *route;
608
609 /* called holding write lock on ksnd_global_lock */
610 for (;;) {
611 /* launch any/all connections that need it */
612 route = ksocknal_find_connectable_route_locked(peer);
613 if (route == NULL)
614 return;
615
616 ksocknal_launch_connection_locked(route);
617 }
618 }
619
620 ksock_conn_t *
621 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
622 {
623 struct list_head *tmp;
624 ksock_conn_t *conn;
625 ksock_conn_t *typed = NULL;
626 ksock_conn_t *fallback = NULL;
627 int tnob = 0;
628 int fnob = 0;
629
630 list_for_each (tmp, &peer->ksnp_conns) {
631 ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list);
632 int nob = atomic_read(&c->ksnc_tx_nob) +
633 c->ksnc_sock->sk->sk_wmem_queued;
634 int rc;
635
636 LASSERT (!c->ksnc_closing);
637 LASSERT (c->ksnc_proto != NULL &&
638 c->ksnc_proto->pro_match_tx != NULL);
639
640 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
641
642 switch (rc) {
643 default:
644 LBUG();
645 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
646 continue;
647
648 case SOCKNAL_MATCH_YES: /* typed connection */
649 if (typed == NULL || tnob > nob ||
650 (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
651 cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
652 typed = c;
653 tnob = nob;
654 }
655 break;
656
657 case SOCKNAL_MATCH_MAY: /* fallback connection */
658 if (fallback == NULL || fnob > nob ||
659 (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
660 cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
661 fallback = c;
662 fnob = nob;
663 }
664 break;
665 }
666 }
667
668 /* prefer the typed selection */
669 conn = (typed != NULL) ? typed : fallback;
670
671 if (conn != NULL)
672 conn->ksnc_tx_last_post = cfs_time_current();
673
674 return conn;
675 }
676
677 void
678 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
679 {
680 conn->ksnc_proto->pro_pack(tx);
681
682 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
683 ksocknal_conn_addref(conn); /* +1 ref for tx */
684 tx->tx_conn = conn;
685 }
686
687 void
688 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
689 {
690 ksock_sched_t *sched = conn->ksnc_scheduler;
691 ksock_msg_t *msg = &tx->tx_msg;
692 ksock_tx_t *ztx = NULL;
693 int bufnob = 0;
694
695 /* called holding global lock (read or irq-write) and caller may
696 * not have dropped this lock between finding conn and calling me,
697 * so we don't need the {get,put}connsock dance to deref
698 * ksnc_sock... */
699 LASSERT(!conn->ksnc_closing);
700
701 CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
702 libcfs_id2str(conn->ksnc_peer->ksnp_id),
703 &conn->ksnc_ipaddr,
704 conn->ksnc_port);
705
706 ksocknal_tx_prep(conn, tx);
707
708 /* Ensure the frags we've been given EXACTLY match the number of
709 * bytes we want to send. Many TCP/IP stacks disregard any total
710 * size parameters passed to them and just look at the frags.
711 *
712 * We always expect at least 1 mapped fragment containing the
713 * complete ksocknal message header. */
714 LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
715 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
716 (unsigned int)tx->tx_nob);
717 LASSERT (tx->tx_niov >= 1);
718 LASSERT (tx->tx_resid == tx->tx_nob);
719
720 CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
721 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
722 KSOCK_MSG_NOOP,
723 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
724
725 /*
726 * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
727 * but they're used inside spinlocks a lot.
728 */
729 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
730 spin_lock_bh(&sched->kss_lock);
731
732 if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
733 /* First packet starts the timeout */
734 conn->ksnc_tx_deadline =
735 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
736 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
737 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
738 conn->ksnc_tx_bufnob = 0;
739 mb(); /* order with adding to tx_queue */
740 }
741
742 if (msg->ksm_type == KSOCK_MSG_NOOP) {
743 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
744 * on a normal packet so I don't need to send it */
745 LASSERT (msg->ksm_zc_cookies[1] != 0);
746 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
747
748 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
749 ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
750
751 } else {
752 /* It's a normal packet - can it piggback a noop zc-ack that
753 * has been queued already? */
754 LASSERT (msg->ksm_zc_cookies[1] == 0);
755 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
756
757 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
758 /* ztx will be released later */
759 }
760
761 if (ztx != NULL) {
762 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
763 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
764 }
765
766 if (conn->ksnc_tx_ready && /* able to send */
767 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
768 /* +1 ref for scheduler */
769 ksocknal_conn_addref(conn);
770 list_add_tail (&conn->ksnc_tx_list,
771 &sched->kss_tx_conns);
772 conn->ksnc_tx_scheduled = 1;
773 wake_up (&sched->kss_waitq);
774 }
775
776 spin_unlock_bh(&sched->kss_lock);
777 }
778
779
780 ksock_route_t *
781 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
782 {
783 unsigned long now = cfs_time_current();
784 struct list_head *tmp;
785 ksock_route_t *route;
786
787 list_for_each (tmp, &peer->ksnp_routes) {
788 route = list_entry (tmp, ksock_route_t, ksnr_list);
789
790 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
791
792 if (route->ksnr_scheduled) /* connections being established */
793 continue;
794
795 /* all route types connected ? */
796 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
797 continue;
798
799 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
800 cfs_time_aftereq(now, route->ksnr_timeout))) {
801 CDEBUG(D_NET,
802 "Too soon to retry route %pI4h "
803 "(cnted %d, interval %ld, %ld secs later)\n",
804 &route->ksnr_ipaddr,
805 route->ksnr_connected,
806 route->ksnr_retry_interval,
807 cfs_duration_sec(route->ksnr_timeout - now));
808 continue;
809 }
810
811 return route;
812 }
813
814 return NULL;
815 }
816
817 ksock_route_t *
818 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
819 {
820 struct list_head *tmp;
821 ksock_route_t *route;
822
823 list_for_each (tmp, &peer->ksnp_routes) {
824 route = list_entry (tmp, ksock_route_t, ksnr_list);
825
826 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
827
828 if (route->ksnr_scheduled)
829 return route;
830 }
831
832 return NULL;
833 }
834
835 int
836 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
837 {
838 ksock_peer_t *peer;
839 ksock_conn_t *conn;
840 rwlock_t *g_lock;
841 int retry;
842 int rc;
843
844 LASSERT (tx->tx_conn == NULL);
845
846 g_lock = &ksocknal_data.ksnd_global_lock;
847
848 for (retry = 0;; retry = 1) {
849 read_lock(g_lock);
850 peer = ksocknal_find_peer_locked(ni, id);
851 if (peer != NULL) {
852 if (ksocknal_find_connectable_route_locked(peer) == NULL) {
853 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
854 if (conn != NULL) {
855 /* I've got no routes that need to be
856 * connecting and I do have an actual
857 * connection... */
858 ksocknal_queue_tx_locked (tx, conn);
859 read_unlock(g_lock);
860 return 0;
861 }
862 }
863 }
864
865 /* I'll need a write lock... */
866 read_unlock(g_lock);
867
868 write_lock_bh(g_lock);
869
870 peer = ksocknal_find_peer_locked(ni, id);
871 if (peer != NULL)
872 break;
873
874 write_unlock_bh(g_lock);
875
876 if ((id.pid & LNET_PID_USERFLAG) != 0) {
877 CERROR("Refusing to create a connection to "
878 "userspace process %s\n", libcfs_id2str(id));
879 return -EHOSTUNREACH;
880 }
881
882 if (retry) {
883 CERROR("Can't find peer %s\n", libcfs_id2str(id));
884 return -EHOSTUNREACH;
885 }
886
887 rc = ksocknal_add_peer(ni, id,
888 LNET_NIDADDR(id.nid),
889 lnet_acceptor_port());
890 if (rc != 0) {
891 CERROR("Can't add peer %s: %d\n",
892 libcfs_id2str(id), rc);
893 return rc;
894 }
895 }
896
897 ksocknal_launch_all_connections_locked(peer);
898
899 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
900 if (conn != NULL) {
901 /* Connection exists; queue message on it */
902 ksocknal_queue_tx_locked (tx, conn);
903 write_unlock_bh(g_lock);
904 return 0;
905 }
906
907 if (peer->ksnp_accepting > 0 ||
908 ksocknal_find_connecting_route_locked (peer) != NULL) {
909 /* the message is going to be pinned to the peer */
910 tx->tx_deadline =
911 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
912
913 /* Queue the message until a connection is established */
914 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
915 write_unlock_bh(g_lock);
916 return 0;
917 }
918
919 write_unlock_bh(g_lock);
920
921 /* NB Routes may be ignored if connections to them failed recently */
922 CNETERR("No usable routes to %s\n", libcfs_id2str(id));
923 return -EHOSTUNREACH;
924 }
925
926 int
927 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
928 {
929 int mpflag = 1;
930 int type = lntmsg->msg_type;
931 lnet_process_id_t target = lntmsg->msg_target;
932 unsigned int payload_niov = lntmsg->msg_niov;
933 struct iovec *payload_iov = lntmsg->msg_iov;
934 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
935 unsigned int payload_offset = lntmsg->msg_offset;
936 unsigned int payload_nob = lntmsg->msg_len;
937 ksock_tx_t *tx;
938 int desc_size;
939 int rc;
940
941 /* NB 'private' is different depending on what we're sending.
942 * Just ignore it... */
943
944 CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
945 payload_nob, payload_niov, libcfs_id2str(target));
946
947 LASSERT (payload_nob == 0 || payload_niov > 0);
948 LASSERT (payload_niov <= LNET_MAX_IOV);
949 /* payload is either all vaddrs or all pages */
950 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
951 LASSERT (!in_interrupt ());
952
953 if (payload_iov != NULL)
954 desc_size = offsetof(ksock_tx_t,
955 tx_frags.virt.iov[1 + payload_niov]);
956 else
957 desc_size = offsetof(ksock_tx_t,
958 tx_frags.paged.kiov[payload_niov]);
959
960 if (lntmsg->msg_vmflush)
961 mpflag = cfs_memory_pressure_get_and_set();
962 tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
963 if (tx == NULL) {
964 CERROR("Can't allocate tx desc type %d size %d\n",
965 type, desc_size);
966 if (lntmsg->msg_vmflush)
967 cfs_memory_pressure_restore(mpflag);
968 return -ENOMEM;
969 }
970
971 tx->tx_conn = NULL; /* set when assigned a conn */
972 tx->tx_lnetmsg = lntmsg;
973
974 if (payload_iov != NULL) {
975 tx->tx_kiov = NULL;
976 tx->tx_nkiov = 0;
977 tx->tx_iov = tx->tx_frags.virt.iov;
978 tx->tx_niov = 1 +
979 lnet_extract_iov(payload_niov, &tx->tx_iov[1],
980 payload_niov, payload_iov,
981 payload_offset, payload_nob);
982 } else {
983 tx->tx_niov = 1;
984 tx->tx_iov = &tx->tx_frags.paged.iov;
985 tx->tx_kiov = tx->tx_frags.paged.kiov;
986 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
987 payload_niov, payload_kiov,
988 payload_offset, payload_nob);
989
990 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
991 tx->tx_zc_capable = 1;
992 }
993
994 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
995
996 /* The first fragment will be set later in pro_pack */
997 rc = ksocknal_launch_packet(ni, tx, target);
998 if (!mpflag)
999 cfs_memory_pressure_restore(mpflag);
1000
1001 if (rc == 0)
1002 return 0;
1003
1004 ksocknal_free_tx(tx);
1005 return -EIO;
1006 }
1007
1008 int
1009 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1010 {
1011 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1012
1013 if (IS_ERR(task))
1014 return PTR_ERR(task);
1015
1016 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1017 ksocknal_data.ksnd_nthreads++;
1018 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1019 return 0;
1020 }
1021
1022 void
1023 ksocknal_thread_fini (void)
1024 {
1025 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1026 ksocknal_data.ksnd_nthreads--;
1027 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1028 }
1029
1030 int
1031 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1032 {
1033 static char ksocknal_slop_buffer[4096];
1034
1035 int nob;
1036 unsigned int niov;
1037 int skipped;
1038
1039 LASSERT(conn->ksnc_proto != NULL);
1040
1041 if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1042 /* Remind the socket to ack eagerly... */
1043 ksocknal_lib_eager_ack(conn);
1044 }
1045
1046 if (nob_to_skip == 0) { /* right at next packet boundary now */
1047 conn->ksnc_rx_started = 0;
1048 mb(); /* racing with timeout thread */
1049
1050 switch (conn->ksnc_proto->pro_version) {
1051 case KSOCK_PROTO_V2:
1052 case KSOCK_PROTO_V3:
1053 conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1054 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1055 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1056
1057 conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1058 conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1059 conn->ksnc_rx_iov[0].iov_len = offsetof(ksock_msg_t, ksm_u);
1060 break;
1061
1062 case KSOCK_PROTO_V1:
1063 /* Receiving bare lnet_hdr_t */
1064 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1065 conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1066 conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1067
1068 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1069 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1070 conn->ksnc_rx_iov[0].iov_len = sizeof (lnet_hdr_t);
1071 break;
1072
1073 default:
1074 LBUG ();
1075 }
1076 conn->ksnc_rx_niov = 1;
1077
1078 conn->ksnc_rx_kiov = NULL;
1079 conn->ksnc_rx_nkiov = 0;
1080 conn->ksnc_rx_csum = ~0;
1081 return 1;
1082 }
1083
1084 /* Set up to skip as much as possible now. If there's more left
1085 * (ran out of iov entries) we'll get called again */
1086
1087 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1088 conn->ksnc_rx_nob_left = nob_to_skip;
1089 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1090 skipped = 0;
1091 niov = 0;
1092
1093 do {
1094 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1095
1096 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1097 conn->ksnc_rx_iov[niov].iov_len = nob;
1098 niov++;
1099 skipped += nob;
1100 nob_to_skip -=nob;
1101
1102 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1103 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1104
1105 conn->ksnc_rx_niov = niov;
1106 conn->ksnc_rx_kiov = NULL;
1107 conn->ksnc_rx_nkiov = 0;
1108 conn->ksnc_rx_nob_wanted = skipped;
1109 return 0;
1110 }
1111
1112 int
1113 ksocknal_process_receive (ksock_conn_t *conn)
1114 {
1115 lnet_hdr_t *lhdr;
1116 lnet_process_id_t *id;
1117 int rc;
1118
1119 LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1120
1121 /* NB: sched lock NOT held */
1122 /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1123 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1124 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1125 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1126 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1127 again:
1128 if (conn->ksnc_rx_nob_wanted != 0) {
1129 rc = ksocknal_receive(conn);
1130
1131 if (rc <= 0) {
1132 LASSERT (rc != -EAGAIN);
1133
1134 if (rc == 0)
1135 CDEBUG(D_NET, "[%p] EOF from %s"
1136 " ip %pI4h:%d\n", conn,
1137 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1138 &conn->ksnc_ipaddr,
1139 conn->ksnc_port);
1140 else if (!conn->ksnc_closing)
1141 CERROR("[%p] Error %d on read from %s"
1142 " ip %pI4h:%d\n",
1143 conn, rc,
1144 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1145 &conn->ksnc_ipaddr,
1146 conn->ksnc_port);
1147
1148 /* it's not an error if conn is being closed */
1149 ksocknal_close_conn_and_siblings (conn,
1150 (conn->ksnc_closing) ? 0 : rc);
1151 return (rc == 0 ? -ESHUTDOWN : rc);
1152 }
1153
1154 if (conn->ksnc_rx_nob_wanted != 0) {
1155 /* short read */
1156 return -EAGAIN;
1157 }
1158 }
1159 switch (conn->ksnc_rx_state) {
1160 case SOCKNAL_RX_KSM_HEADER:
1161 if (conn->ksnc_flip) {
1162 __swab32s(&conn->ksnc_msg.ksm_type);
1163 __swab32s(&conn->ksnc_msg.ksm_csum);
1164 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1165 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1166 }
1167
1168 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1169 conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1170 CERROR("%s: Unknown message type: %x\n",
1171 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1172 conn->ksnc_msg.ksm_type);
1173 ksocknal_new_packet(conn, 0);
1174 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1175 return -EPROTO;
1176 }
1177
1178 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1179 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1180 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1181 /* NOOP Checksum error */
1182 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1183 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1184 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1185 ksocknal_new_packet(conn, 0);
1186 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1187 return -EIO;
1188 }
1189
1190 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1191 __u64 cookie = 0;
1192
1193 LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1194
1195 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1196 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1197
1198 rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1199 conn->ksnc_msg.ksm_zc_cookies[1]);
1200
1201 if (rc != 0) {
1202 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1203 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1204 cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1205 ksocknal_new_packet(conn, 0);
1206 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1207 return rc;
1208 }
1209 }
1210
1211 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1212 ksocknal_new_packet (conn, 0);
1213 return 0; /* NOOP is done and just return */
1214 }
1215
1216 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1217 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1218 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1219
1220 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1221 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1222 conn->ksnc_rx_iov[0].iov_len = sizeof(ksock_lnet_msg_t);
1223
1224 conn->ksnc_rx_niov = 1;
1225 conn->ksnc_rx_kiov = NULL;
1226 conn->ksnc_rx_nkiov = 0;
1227
1228 goto again; /* read lnet header now */
1229
1230 case SOCKNAL_RX_LNET_HEADER:
1231 /* unpack message header */
1232 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1233
1234 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1235 /* Userspace peer */
1236 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1237 id = &conn->ksnc_peer->ksnp_id;
1238
1239 /* Substitute process ID assigned at connection time */
1240 lhdr->src_pid = cpu_to_le32(id->pid);
1241 lhdr->src_nid = cpu_to_le64(id->nid);
1242 }
1243
1244 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1245 ksocknal_conn_addref(conn); /* ++ref while parsing */
1246
1247 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1248 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1249 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1250 if (rc < 0) {
1251 /* I just received garbage: give up on this conn */
1252 ksocknal_new_packet(conn, 0);
1253 ksocknal_close_conn_and_siblings (conn, rc);
1254 ksocknal_conn_decref(conn);
1255 return -EPROTO;
1256 }
1257
1258 /* I'm racing with ksocknal_recv() */
1259 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1260 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1261
1262 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1263 return 0;
1264
1265 /* ksocknal_recv() got called */
1266 goto again;
1267
1268 case SOCKNAL_RX_LNET_PAYLOAD:
1269 /* payload all received */
1270 rc = 0;
1271
1272 if (conn->ksnc_rx_nob_left == 0 && /* not truncating */
1273 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1274 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1275 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1276 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1277 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1278 rc = -EIO;
1279 }
1280
1281 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1282 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1283
1284 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1285 id = &conn->ksnc_peer->ksnp_id;
1286
1287 rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1288 conn->ksnc_msg.ksm_zc_cookies[0],
1289 *ksocknal_tunables.ksnd_nonblk_zcack ||
1290 le64_to_cpu(lhdr->src_nid) != id->nid);
1291 }
1292
1293 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1294
1295 if (rc != 0) {
1296 ksocknal_new_packet(conn, 0);
1297 ksocknal_close_conn_and_siblings (conn, rc);
1298 return -EPROTO;
1299 }
1300 /* Fall through */
1301
1302 case SOCKNAL_RX_SLOP:
1303 /* starting new packet? */
1304 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1305 return 0; /* come back later */
1306 goto again; /* try to finish reading slop now */
1307
1308 default:
1309 break;
1310 }
1311
1312 /* Not Reached */
1313 LBUG ();
1314 return -EINVAL; /* keep gcc happy */
1315 }
1316
1317 int
1318 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1319 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1320 unsigned int offset, unsigned int mlen, unsigned int rlen)
1321 {
1322 ksock_conn_t *conn = (ksock_conn_t *)private;
1323 ksock_sched_t *sched = conn->ksnc_scheduler;
1324
1325 LASSERT (mlen <= rlen);
1326 LASSERT (niov <= LNET_MAX_IOV);
1327
1328 conn->ksnc_cookie = msg;
1329 conn->ksnc_rx_nob_wanted = mlen;
1330 conn->ksnc_rx_nob_left = rlen;
1331
1332 if (mlen == 0 || iov != NULL) {
1333 conn->ksnc_rx_nkiov = 0;
1334 conn->ksnc_rx_kiov = NULL;
1335 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1336 conn->ksnc_rx_niov =
1337 lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1338 niov, iov, offset, mlen);
1339 } else {
1340 conn->ksnc_rx_niov = 0;
1341 conn->ksnc_rx_iov = NULL;
1342 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1343 conn->ksnc_rx_nkiov =
1344 lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1345 niov, kiov, offset, mlen);
1346 }
1347
1348 LASSERT (mlen ==
1349 lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1350 lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1351
1352 LASSERT (conn->ksnc_rx_scheduled);
1353
1354 spin_lock_bh(&sched->kss_lock);
1355
1356 switch (conn->ksnc_rx_state) {
1357 case SOCKNAL_RX_PARSE_WAIT:
1358 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1359 wake_up (&sched->kss_waitq);
1360 LASSERT (conn->ksnc_rx_ready);
1361 break;
1362
1363 case SOCKNAL_RX_PARSE:
1364 /* scheduler hasn't noticed I'm parsing yet */
1365 break;
1366 }
1367
1368 conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1369
1370 spin_unlock_bh(&sched->kss_lock);
1371 ksocknal_conn_decref(conn);
1372 return 0;
1373 }
1374
1375 static inline int
1376 ksocknal_sched_cansleep(ksock_sched_t *sched)
1377 {
1378 int rc;
1379
1380 spin_lock_bh(&sched->kss_lock);
1381
1382 rc = (!ksocknal_data.ksnd_shuttingdown &&
1383 list_empty(&sched->kss_rx_conns) &&
1384 list_empty(&sched->kss_tx_conns));
1385
1386 spin_unlock_bh(&sched->kss_lock);
1387 return rc;
1388 }
1389
1390 int ksocknal_scheduler(void *arg)
1391 {
1392 struct ksock_sched_info *info;
1393 ksock_sched_t *sched;
1394 ksock_conn_t *conn;
1395 ksock_tx_t *tx;
1396 int rc;
1397 int nloops = 0;
1398 long id = (long)arg;
1399
1400 info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1401 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1402
1403 cfs_block_allsigs();
1404
1405 rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1406 if (rc != 0) {
1407 CERROR("Can't set CPT affinity to %d: %d\n",
1408 info->ksi_cpt, rc);
1409 }
1410
1411 spin_lock_bh(&sched->kss_lock);
1412
1413 while (!ksocknal_data.ksnd_shuttingdown) {
1414 int did_something = 0;
1415
1416 /* Ensure I progress everything semi-fairly */
1417
1418 if (!list_empty (&sched->kss_rx_conns)) {
1419 conn = list_entry(sched->kss_rx_conns.next,
1420 ksock_conn_t, ksnc_rx_list);
1421 list_del(&conn->ksnc_rx_list);
1422
1423 LASSERT(conn->ksnc_rx_scheduled);
1424 LASSERT(conn->ksnc_rx_ready);
1425
1426 /* clear rx_ready in case receive isn't complete.
1427 * Do it BEFORE we call process_recv, since
1428 * data_ready can set it any time after we release
1429 * kss_lock. */
1430 conn->ksnc_rx_ready = 0;
1431 spin_unlock_bh(&sched->kss_lock);
1432
1433 rc = ksocknal_process_receive(conn);
1434
1435 spin_lock_bh(&sched->kss_lock);
1436
1437 /* I'm the only one that can clear this flag */
1438 LASSERT(conn->ksnc_rx_scheduled);
1439
1440 /* Did process_receive get everything it wanted? */
1441 if (rc == 0)
1442 conn->ksnc_rx_ready = 1;
1443
1444 if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1445 /* Conn blocked waiting for ksocknal_recv()
1446 * I change its state (under lock) to signal
1447 * it can be rescheduled */
1448 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1449 } else if (conn->ksnc_rx_ready) {
1450 /* reschedule for rx */
1451 list_add_tail (&conn->ksnc_rx_list,
1452 &sched->kss_rx_conns);
1453 } else {
1454 conn->ksnc_rx_scheduled = 0;
1455 /* drop my ref */
1456 ksocknal_conn_decref(conn);
1457 }
1458
1459 did_something = 1;
1460 }
1461
1462 if (!list_empty (&sched->kss_tx_conns)) {
1463 LIST_HEAD (zlist);
1464
1465 if (!list_empty(&sched->kss_zombie_noop_txs)) {
1466 list_add(&zlist,
1467 &sched->kss_zombie_noop_txs);
1468 list_del_init(&sched->kss_zombie_noop_txs);
1469 }
1470
1471 conn = list_entry(sched->kss_tx_conns.next,
1472 ksock_conn_t, ksnc_tx_list);
1473 list_del (&conn->ksnc_tx_list);
1474
1475 LASSERT(conn->ksnc_tx_scheduled);
1476 LASSERT(conn->ksnc_tx_ready);
1477 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1478
1479 tx = list_entry(conn->ksnc_tx_queue.next,
1480 ksock_tx_t, tx_list);
1481
1482 if (conn->ksnc_tx_carrier == tx)
1483 ksocknal_next_tx_carrier(conn);
1484
1485 /* dequeue now so empty list => more to send */
1486 list_del(&tx->tx_list);
1487
1488 /* Clear tx_ready in case send isn't complete. Do
1489 * it BEFORE we call process_transmit, since
1490 * write_space can set it any time after we release
1491 * kss_lock. */
1492 conn->ksnc_tx_ready = 0;
1493 spin_unlock_bh(&sched->kss_lock);
1494
1495 if (!list_empty(&zlist)) {
1496 /* free zombie noop txs, it's fast because
1497 * noop txs are just put in freelist */
1498 ksocknal_txlist_done(NULL, &zlist, 0);
1499 }
1500
1501 rc = ksocknal_process_transmit(conn, tx);
1502
1503 if (rc == -ENOMEM || rc == -EAGAIN) {
1504 /* Incomplete send: replace tx on HEAD of tx_queue */
1505 spin_lock_bh(&sched->kss_lock);
1506 list_add(&tx->tx_list,
1507 &conn->ksnc_tx_queue);
1508 } else {
1509 /* Complete send; tx -ref */
1510 ksocknal_tx_decref(tx);
1511
1512 spin_lock_bh(&sched->kss_lock);
1513 /* assume space for more */
1514 conn->ksnc_tx_ready = 1;
1515 }
1516
1517 if (rc == -ENOMEM) {
1518 /* Do nothing; after a short timeout, this
1519 * conn will be reposted on kss_tx_conns. */
1520 } else if (conn->ksnc_tx_ready &&
1521 !list_empty (&conn->ksnc_tx_queue)) {
1522 /* reschedule for tx */
1523 list_add_tail (&conn->ksnc_tx_list,
1524 &sched->kss_tx_conns);
1525 } else {
1526 conn->ksnc_tx_scheduled = 0;
1527 /* drop my ref */
1528 ksocknal_conn_decref(conn);
1529 }
1530
1531 did_something = 1;
1532 }
1533 if (!did_something || /* nothing to do */
1534 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1535 spin_unlock_bh(&sched->kss_lock);
1536
1537 nloops = 0;
1538
1539 if (!did_something) { /* wait for something to do */
1540 cfs_wait_event_interruptible_exclusive(
1541 sched->kss_waitq,
1542 !ksocknal_sched_cansleep(sched), rc);
1543 LASSERT (rc == 0);
1544 } else {
1545 cond_resched();
1546 }
1547
1548 spin_lock_bh(&sched->kss_lock);
1549 }
1550 }
1551
1552 spin_unlock_bh(&sched->kss_lock);
1553 ksocknal_thread_fini();
1554 return 0;
1555 }
1556
1557 /*
1558 * Add connection to kss_rx_conns of scheduler
1559 * and wakeup the scheduler.
1560 */
1561 void ksocknal_read_callback (ksock_conn_t *conn)
1562 {
1563 ksock_sched_t *sched;
1564
1565 sched = conn->ksnc_scheduler;
1566
1567 spin_lock_bh(&sched->kss_lock);
1568
1569 conn->ksnc_rx_ready = 1;
1570
1571 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1572 list_add_tail(&conn->ksnc_rx_list,
1573 &sched->kss_rx_conns);
1574 conn->ksnc_rx_scheduled = 1;
1575 /* extra ref for scheduler */
1576 ksocknal_conn_addref(conn);
1577
1578 wake_up (&sched->kss_waitq);
1579 }
1580 spin_unlock_bh(&sched->kss_lock);
1581 }
1582
1583 /*
1584 * Add connection to kss_tx_conns of scheduler
1585 * and wakeup the scheduler.
1586 */
1587 void ksocknal_write_callback (ksock_conn_t *conn)
1588 {
1589 ksock_sched_t *sched;
1590
1591 sched = conn->ksnc_scheduler;
1592
1593 spin_lock_bh(&sched->kss_lock);
1594
1595 conn->ksnc_tx_ready = 1;
1596
1597 if (!conn->ksnc_tx_scheduled && // not being progressed
1598 !list_empty(&conn->ksnc_tx_queue)){//packets to send
1599 list_add_tail (&conn->ksnc_tx_list,
1600 &sched->kss_tx_conns);
1601 conn->ksnc_tx_scheduled = 1;
1602 /* extra ref for scheduler */
1603 ksocknal_conn_addref(conn);
1604
1605 wake_up (&sched->kss_waitq);
1606 }
1607
1608 spin_unlock_bh(&sched->kss_lock);
1609 }
1610
1611 ksock_proto_t *
1612 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1613 {
1614 __u32 version = 0;
1615
1616 if (hello->kshm_magic == LNET_PROTO_MAGIC)
1617 version = hello->kshm_version;
1618 else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1619 version = __swab32(hello->kshm_version);
1620
1621 if (version != 0) {
1622 #if SOCKNAL_VERSION_DEBUG
1623 if (*ksocknal_tunables.ksnd_protocol == 1)
1624 return NULL;
1625
1626 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1627 version == KSOCK_PROTO_V3)
1628 return NULL;
1629 #endif
1630 if (version == KSOCK_PROTO_V2)
1631 return &ksocknal_protocol_v2x;
1632
1633 if (version == KSOCK_PROTO_V3)
1634 return &ksocknal_protocol_v3x;
1635
1636 return NULL;
1637 }
1638
1639 if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1640 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1641
1642 CLASSERT (sizeof (lnet_magicversion_t) ==
1643 offsetof (ksock_hello_msg_t, kshm_src_nid));
1644
1645 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1646 hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1647 return &ksocknal_protocol_v1x;
1648 }
1649
1650 return NULL;
1651 }
1652
1653 int
1654 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1655 lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1656 {
1657 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1658 ksock_net_t *net = (ksock_net_t *)ni->ni_data;
1659
1660 LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1661
1662 /* rely on caller to hold a ref on socket so it wouldn't disappear */
1663 LASSERT (conn->ksnc_proto != NULL);
1664
1665 hello->kshm_src_nid = ni->ni_nid;
1666 hello->kshm_dst_nid = peer_nid;
1667 hello->kshm_src_pid = the_lnet.ln_pid;
1668
1669 hello->kshm_src_incarnation = net->ksnn_incarnation;
1670 hello->kshm_ctype = conn->ksnc_type;
1671
1672 return conn->ksnc_proto->pro_send_hello(conn, hello);
1673 }
1674
1675 int
1676 ksocknal_invert_type(int type)
1677 {
1678 switch (type)
1679 {
1680 case SOCKLND_CONN_ANY:
1681 case SOCKLND_CONN_CONTROL:
1682 return type;
1683 case SOCKLND_CONN_BULK_IN:
1684 return SOCKLND_CONN_BULK_OUT;
1685 case SOCKLND_CONN_BULK_OUT:
1686 return SOCKLND_CONN_BULK_IN;
1687 default:
1688 return SOCKLND_CONN_NONE;
1689 }
1690 }
1691
1692 int
1693 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1694 ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1695 __u64 *incarnation)
1696 {
1697 /* Return < 0 fatal error
1698 * 0 success
1699 * EALREADY lost connection race
1700 * EPROTO protocol version mismatch
1701 */
1702 struct socket *sock = conn->ksnc_sock;
1703 int active = (conn->ksnc_proto != NULL);
1704 int timeout;
1705 int proto_match;
1706 int rc;
1707 ksock_proto_t *proto;
1708 lnet_process_id_t recv_id;
1709
1710 /* socket type set on active connections - not set on passive */
1711 LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1712
1713 timeout = active ? *ksocknal_tunables.ksnd_timeout :
1714 lnet_acceptor_timeout();
1715
1716 rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1717 if (rc != 0) {
1718 CERROR("Error %d reading HELLO from %pI4h\n",
1719 rc, &conn->ksnc_ipaddr);
1720 LASSERT (rc < 0);
1721 return rc;
1722 }
1723
1724 if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1725 hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1726 hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1727 /* Unexpected magic! */
1728 CERROR("Bad magic(1) %#08x (%#08x expected) from "
1729 "%pI4h\n", __cpu_to_le32 (hello->kshm_magic),
1730 LNET_PROTO_TCP_MAGIC,
1731 &conn->ksnc_ipaddr);
1732 return -EPROTO;
1733 }
1734
1735 rc = libcfs_sock_read(sock, &hello->kshm_version,
1736 sizeof(hello->kshm_version), timeout);
1737 if (rc != 0) {
1738 CERROR("Error %d reading HELLO from %pI4h\n",
1739 rc, &conn->ksnc_ipaddr);
1740 LASSERT (rc < 0);
1741 return rc;
1742 }
1743
1744 proto = ksocknal_parse_proto_version(hello);
1745 if (proto == NULL) {
1746 if (!active) {
1747 /* unknown protocol from peer, tell peer my protocol */
1748 conn->ksnc_proto = &ksocknal_protocol_v3x;
1749 #if SOCKNAL_VERSION_DEBUG
1750 if (*ksocknal_tunables.ksnd_protocol == 2)
1751 conn->ksnc_proto = &ksocknal_protocol_v2x;
1752 else if (*ksocknal_tunables.ksnd_protocol == 1)
1753 conn->ksnc_proto = &ksocknal_protocol_v1x;
1754 #endif
1755 hello->kshm_nips = 0;
1756 ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1757 }
1758
1759 CERROR("Unknown protocol version (%d.x expected)"
1760 " from %pI4h\n",
1761 conn->ksnc_proto->pro_version,
1762 &conn->ksnc_ipaddr);
1763
1764 return -EPROTO;
1765 }
1766
1767 proto_match = (conn->ksnc_proto == proto);
1768 conn->ksnc_proto = proto;
1769
1770 /* receive the rest of hello message anyway */
1771 rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1772 if (rc != 0) {
1773 CERROR("Error %d reading or checking hello from from %pI4h\n",
1774 rc, &conn->ksnc_ipaddr);
1775 LASSERT (rc < 0);
1776 return rc;
1777 }
1778
1779 *incarnation = hello->kshm_src_incarnation;
1780
1781 if (hello->kshm_src_nid == LNET_NID_ANY) {
1782 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1783 "from %pI4h\n", &conn->ksnc_ipaddr);
1784 return -EPROTO;
1785 }
1786
1787 if (!active &&
1788 conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1789 /* Userspace NAL assigns peer process ID from socket */
1790 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1791 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1792 } else {
1793 recv_id.nid = hello->kshm_src_nid;
1794 recv_id.pid = hello->kshm_src_pid;
1795 }
1796
1797 if (!active) {
1798 *peerid = recv_id;
1799
1800 /* peer determines type */
1801 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1802 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1803 CERROR("Unexpected type %d from %s ip %pI4h\n",
1804 hello->kshm_ctype, libcfs_id2str(*peerid),
1805 &conn->ksnc_ipaddr);
1806 return -EPROTO;
1807 }
1808
1809 return 0;
1810 }
1811
1812 if (peerid->pid != recv_id.pid ||
1813 peerid->nid != recv_id.nid) {
1814 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1815 " %pI4h, but they claimed they were "
1816 "%s; please check your Lustre "
1817 "configuration.\n",
1818 libcfs_id2str(*peerid),
1819 &conn->ksnc_ipaddr,
1820 libcfs_id2str(recv_id));
1821 return -EPROTO;
1822 }
1823
1824 if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1825 /* Possible protocol mismatch or I lost the connection race */
1826 return proto_match ? EALREADY : EPROTO;
1827 }
1828
1829 if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1830 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1831 conn->ksnc_type, libcfs_id2str(*peerid),
1832 &conn->ksnc_ipaddr,
1833 hello->kshm_ctype);
1834 return -EPROTO;
1835 }
1836
1837 return 0;
1838 }
1839
1840 int
1841 ksocknal_connect (ksock_route_t *route)
1842 {
1843 LIST_HEAD (zombies);
1844 ksock_peer_t *peer = route->ksnr_peer;
1845 int type;
1846 int wanted;
1847 struct socket *sock;
1848 unsigned long deadline;
1849 int retry_later = 0;
1850 int rc = 0;
1851
1852 deadline = cfs_time_add(cfs_time_current(),
1853 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1854
1855 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1856
1857 LASSERT (route->ksnr_scheduled);
1858 LASSERT (!route->ksnr_connecting);
1859
1860 route->ksnr_connecting = 1;
1861
1862 for (;;) {
1863 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1864
1865 /* stop connecting if peer/route got closed under me, or
1866 * route got connected while queued */
1867 if (peer->ksnp_closing || route->ksnr_deleted ||
1868 wanted == 0) {
1869 retry_later = 0;
1870 break;
1871 }
1872
1873 /* reschedule if peer is connecting to me */
1874 if (peer->ksnp_accepting > 0) {
1875 CDEBUG(D_NET,
1876 "peer %s(%d) already connecting to me, retry later.\n",
1877 libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1878 retry_later = 1;
1879 }
1880
1881 if (retry_later) /* needs reschedule */
1882 break;
1883
1884 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1885 type = SOCKLND_CONN_ANY;
1886 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1887 type = SOCKLND_CONN_CONTROL;
1888 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1889 type = SOCKLND_CONN_BULK_IN;
1890 } else {
1891 LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1892 type = SOCKLND_CONN_BULK_OUT;
1893 }
1894
1895 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1896
1897 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1898 rc = -ETIMEDOUT;
1899 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1900 route->ksnr_ipaddr,
1901 route->ksnr_port);
1902 goto failed;
1903 }
1904
1905 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1906 route->ksnr_myipaddr,
1907 route->ksnr_ipaddr, route->ksnr_port);
1908 if (rc != 0)
1909 goto failed;
1910
1911 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1912 if (rc < 0) {
1913 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1914 route->ksnr_ipaddr,
1915 route->ksnr_port);
1916 goto failed;
1917 }
1918
1919 /* A +ve RC means I have to retry because I lost the connection
1920 * race or I have to renegotiate protocol version */
1921 retry_later = (rc != 0);
1922 if (retry_later)
1923 CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1924 libcfs_nid2str(peer->ksnp_id.nid));
1925
1926 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1927 }
1928
1929 route->ksnr_scheduled = 0;
1930 route->ksnr_connecting = 0;
1931
1932 if (retry_later) {
1933 /* re-queue for attention; this frees me up to handle
1934 * the peer's incoming connection request */
1935
1936 if (rc == EALREADY ||
1937 (rc == 0 && peer->ksnp_accepting > 0)) {
1938 /* We want to introduce a delay before next
1939 * attempt to connect if we lost conn race,
1940 * but the race is resolved quickly usually,
1941 * so min_reconnectms should be good heuristic */
1942 route->ksnr_retry_interval =
1943 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1944 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1945 route->ksnr_retry_interval);
1946 }
1947
1948 ksocknal_launch_connection_locked(route);
1949 }
1950
1951 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1952 return retry_later;
1953
1954 failed:
1955 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1956
1957 route->ksnr_scheduled = 0;
1958 route->ksnr_connecting = 0;
1959
1960 /* This is a retry rather than a new connection */
1961 route->ksnr_retry_interval *= 2;
1962 route->ksnr_retry_interval =
1963 MAX(route->ksnr_retry_interval,
1964 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1965 route->ksnr_retry_interval =
1966 MIN(route->ksnr_retry_interval,
1967 cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1968
1969 LASSERT (route->ksnr_retry_interval != 0);
1970 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1971 route->ksnr_retry_interval);
1972
1973 if (!list_empty(&peer->ksnp_tx_queue) &&
1974 peer->ksnp_accepting == 0 &&
1975 ksocknal_find_connecting_route_locked(peer) == NULL) {
1976 ksock_conn_t *conn;
1977
1978 /* ksnp_tx_queue is queued on a conn on successful
1979 * connection for V1.x and V2.x */
1980 if (!list_empty (&peer->ksnp_conns)) {
1981 conn = list_entry(peer->ksnp_conns.next,
1982 ksock_conn_t, ksnc_list);
1983 LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1984 }
1985
1986 /* take all the blocked packets while I've got the lock and
1987 * complete below... */
1988 list_splice_init(&peer->ksnp_tx_queue, &zombies);
1989 }
1990
1991 #if 0 /* irrelevant with only eager routes */
1992 if (!route->ksnr_deleted) {
1993 /* make this route least-favourite for re-selection */
1994 list_del(&route->ksnr_list);
1995 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1996 }
1997 #endif
1998 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1999
2000 ksocknal_peer_failed(peer);
2001 ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2002 return 0;
2003 }
2004
2005 /*
2006 * check whether we need to create more connds.
2007 * It will try to create new thread if it's necessary, @timeout can
2008 * be updated if failed to create, so caller wouldn't keep try while
2009 * running out of resource.
2010 */
2011 static int
2012 ksocknal_connd_check_start(long sec, long *timeout)
2013 {
2014 char name[16];
2015 int rc;
2016 int total = ksocknal_data.ksnd_connd_starting +
2017 ksocknal_data.ksnd_connd_running;
2018
2019 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2020 /* still in initializing */
2021 return 0;
2022 }
2023
2024 if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2025 total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2026 /* can't create more connd, or still have enough
2027 * threads to handle more connecting */
2028 return 0;
2029 }
2030
2031 if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2032 /* no pending connecting request */
2033 return 0;
2034 }
2035
2036 if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2037 /* may run out of resource, retry later */
2038 *timeout = cfs_time_seconds(1);
2039 return 0;
2040 }
2041
2042 if (ksocknal_data.ksnd_connd_starting > 0) {
2043 /* serialize starting to avoid flood */
2044 return 0;
2045 }
2046
2047 ksocknal_data.ksnd_connd_starting_stamp = sec;
2048 ksocknal_data.ksnd_connd_starting++;
2049 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2050
2051 /* NB: total is the next id */
2052 snprintf(name, sizeof(name), "socknal_cd%02d", total);
2053 rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2054
2055 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2056 if (rc == 0)
2057 return 1;
2058
2059 /* we tried ... */
2060 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2061 ksocknal_data.ksnd_connd_starting--;
2062 ksocknal_data.ksnd_connd_failed_stamp = get_seconds();
2063
2064 return 1;
2065 }
2066
2067 /*
2068 * check whether current thread can exit, it will return 1 if there are too
2069 * many threads and no creating in past 120 seconds.
2070 * Also, this function may update @timeout to make caller come back
2071 * again to recheck these conditions.
2072 */
2073 static int
2074 ksocknal_connd_check_stop(long sec, long *timeout)
2075 {
2076 int val;
2077
2078 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2079 /* still in initializing */
2080 return 0;
2081 }
2082
2083 if (ksocknal_data.ksnd_connd_starting > 0) {
2084 /* in progress of starting new thread */
2085 return 0;
2086 }
2087
2088 if (ksocknal_data.ksnd_connd_running <=
2089 *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2090 return 0;
2091 }
2092
2093 /* created thread in past 120 seconds? */
2094 val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2095 SOCKNAL_CONND_TIMEOUT - sec);
2096
2097 *timeout = (val > 0) ? cfs_time_seconds(val) :
2098 cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2099 if (val > 0)
2100 return 0;
2101
2102 /* no creating in past 120 seconds */
2103
2104 return ksocknal_data.ksnd_connd_running >
2105 ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2106 }
2107
2108 /* Go through connd_routes queue looking for a route that we can process
2109 * right now, @timeout_p can be updated if we need to come back later */
2110 static ksock_route_t *
2111 ksocknal_connd_get_route_locked(signed long *timeout_p)
2112 {
2113 ksock_route_t *route;
2114 unsigned long now;
2115
2116 now = cfs_time_current();
2117
2118 /* connd_routes can contain both pending and ordinary routes */
2119 list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2120 ksnr_connd_list) {
2121
2122 if (route->ksnr_retry_interval == 0 ||
2123 cfs_time_aftereq(now, route->ksnr_timeout))
2124 return route;
2125
2126 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2127 (int)*timeout_p > (int)(route->ksnr_timeout - now))
2128 *timeout_p = (int)(route->ksnr_timeout - now);
2129 }
2130
2131 return NULL;
2132 }
2133
2134 int
2135 ksocknal_connd (void *arg)
2136 {
2137 spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2138 ksock_connreq_t *cr;
2139 wait_queue_t wait;
2140 int nloops = 0;
2141 int cons_retry = 0;
2142
2143 cfs_block_allsigs ();
2144
2145 init_waitqueue_entry(&wait, current);
2146
2147 spin_lock_bh(connd_lock);
2148
2149 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2150 ksocknal_data.ksnd_connd_starting--;
2151 ksocknal_data.ksnd_connd_running++;
2152
2153 while (!ksocknal_data.ksnd_shuttingdown) {
2154 ksock_route_t *route = NULL;
2155 long sec = get_seconds();
2156 long timeout = MAX_SCHEDULE_TIMEOUT;
2157 int dropped_lock = 0;
2158
2159 if (ksocknal_connd_check_stop(sec, &timeout)) {
2160 /* wakeup another one to check stop */
2161 wake_up(&ksocknal_data.ksnd_connd_waitq);
2162 break;
2163 }
2164
2165 if (ksocknal_connd_check_start(sec, &timeout)) {
2166 /* created new thread */
2167 dropped_lock = 1;
2168 }
2169
2170 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2171 /* Connection accepted by the listener */
2172 cr = list_entry(ksocknal_data.ksnd_connd_connreqs. \
2173 next, ksock_connreq_t, ksncr_list);
2174
2175 list_del(&cr->ksncr_list);
2176 spin_unlock_bh(connd_lock);
2177 dropped_lock = 1;
2178
2179 ksocknal_create_conn(cr->ksncr_ni, NULL,
2180 cr->ksncr_sock, SOCKLND_CONN_NONE);
2181 lnet_ni_decref(cr->ksncr_ni);
2182 LIBCFS_FREE(cr, sizeof(*cr));
2183
2184 spin_lock_bh(connd_lock);
2185 }
2186
2187 /* Only handle an outgoing connection request if there
2188 * is a thread left to handle incoming connections and
2189 * create new connd */
2190 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2191 ksocknal_data.ksnd_connd_running) {
2192 route = ksocknal_connd_get_route_locked(&timeout);
2193 }
2194 if (route != NULL) {
2195 list_del (&route->ksnr_connd_list);
2196 ksocknal_data.ksnd_connd_connecting++;
2197 spin_unlock_bh(connd_lock);
2198 dropped_lock = 1;
2199
2200 if (ksocknal_connect(route)) {
2201 /* consecutive retry */
2202 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2203 CWARN("massive consecutive "
2204 "re-connecting to %pI4h\n",
2205 &route->ksnr_ipaddr);
2206 cons_retry = 0;
2207 }
2208 } else {
2209 cons_retry = 0;
2210 }
2211
2212 ksocknal_route_decref(route);
2213
2214 spin_lock_bh(connd_lock);
2215 ksocknal_data.ksnd_connd_connecting--;
2216 }
2217
2218 if (dropped_lock) {
2219 if (++nloops < SOCKNAL_RESCHED)
2220 continue;
2221 spin_unlock_bh(connd_lock);
2222 nloops = 0;
2223 cond_resched();
2224 spin_lock_bh(connd_lock);
2225 continue;
2226 }
2227
2228 /* Nothing to do for 'timeout' */
2229 set_current_state(TASK_INTERRUPTIBLE);
2230 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2231 spin_unlock_bh(connd_lock);
2232
2233 nloops = 0;
2234 schedule_timeout(timeout);
2235
2236 set_current_state(TASK_RUNNING);
2237 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2238 spin_lock_bh(connd_lock);
2239 }
2240 ksocknal_data.ksnd_connd_running--;
2241 spin_unlock_bh(connd_lock);
2242
2243 ksocknal_thread_fini();
2244 return 0;
2245 }
2246
2247 ksock_conn_t *
2248 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2249 {
2250 /* We're called with a shared lock on ksnd_global_lock */
2251 ksock_conn_t *conn;
2252 struct list_head *ctmp;
2253
2254 list_for_each (ctmp, &peer->ksnp_conns) {
2255 int error;
2256 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2257
2258 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2259 LASSERT (!conn->ksnc_closing);
2260
2261 /* SOCK_ERROR will reset error code of socket in
2262 * some platform (like Darwin8.x) */
2263 error = conn->ksnc_sock->sk->sk_err;
2264 if (error != 0) {
2265 ksocknal_conn_addref(conn);
2266
2267 switch (error) {
2268 case ECONNRESET:
2269 CNETERR("A connection with %s "
2270 "(%pI4h:%d) was reset; "
2271 "it may have rebooted.\n",
2272 libcfs_id2str(peer->ksnp_id),
2273 &conn->ksnc_ipaddr,
2274 conn->ksnc_port);
2275 break;
2276 case ETIMEDOUT:
2277 CNETERR("A connection with %s "
2278 "(%pI4h:%d) timed out; the "
2279 "network or node may be down.\n",
2280 libcfs_id2str(peer->ksnp_id),
2281 &conn->ksnc_ipaddr,
2282 conn->ksnc_port);
2283 break;
2284 default:
2285 CNETERR("An unexpected network error %d "
2286 "occurred with %s "
2287 "(%pI4h:%d\n", error,
2288 libcfs_id2str(peer->ksnp_id),
2289 &conn->ksnc_ipaddr,
2290 conn->ksnc_port);
2291 break;
2292 }
2293
2294 return conn;
2295 }
2296
2297 if (conn->ksnc_rx_started &&
2298 cfs_time_aftereq(cfs_time_current(),
2299 conn->ksnc_rx_deadline)) {
2300 /* Timed out incomplete incoming message */
2301 ksocknal_conn_addref(conn);
2302 CNETERR("Timeout receiving from %s (%pI4h:%d), "
2303 "state %d wanted %d left %d\n",
2304 libcfs_id2str(peer->ksnp_id),
2305 &conn->ksnc_ipaddr,
2306 conn->ksnc_port,
2307 conn->ksnc_rx_state,
2308 conn->ksnc_rx_nob_wanted,
2309 conn->ksnc_rx_nob_left);
2310 return conn;
2311 }
2312
2313 if ((!list_empty(&conn->ksnc_tx_queue) ||
2314 conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2315 cfs_time_aftereq(cfs_time_current(),
2316 conn->ksnc_tx_deadline)) {
2317 /* Timed out messages queued for sending or
2318 * buffered in the socket's send buffer */
2319 ksocknal_conn_addref(conn);
2320 CNETERR("Timeout sending data to %s (%pI4h:%d) "
2321 "the network or that node may be down.\n",
2322 libcfs_id2str(peer->ksnp_id),
2323 &conn->ksnc_ipaddr,
2324 conn->ksnc_port);
2325 return conn;
2326 }
2327 }
2328
2329 return NULL;
2330 }
2331
2332 static inline void
2333 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2334 {
2335 ksock_tx_t *tx;
2336 LIST_HEAD (stale_txs);
2337
2338 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2339
2340 while (!list_empty (&peer->ksnp_tx_queue)) {
2341 tx = list_entry (peer->ksnp_tx_queue.next,
2342 ksock_tx_t, tx_list);
2343
2344 if (!cfs_time_aftereq(cfs_time_current(),
2345 tx->tx_deadline))
2346 break;
2347
2348 list_del (&tx->tx_list);
2349 list_add_tail (&tx->tx_list, &stale_txs);
2350 }
2351
2352 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2353
2354 ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2355 }
2356
2357 int
2358 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2359 {
2360 ksock_sched_t *sched;
2361 ksock_conn_t *conn;
2362 ksock_tx_t *tx;
2363
2364 if (list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2365 return 0;
2366
2367 if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2368 return 0;
2369
2370 if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2371 time_before(cfs_time_current(),
2372 cfs_time_add(peer->ksnp_last_alive,
2373 cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2374 return 0;
2375
2376 if (time_before(cfs_time_current(), peer->ksnp_send_keepalive))
2377 return 0;
2378
2379 /* retry 10 secs later, so we wouldn't put pressure
2380 * on this peer if we failed to send keepalive this time */
2381 peer->ksnp_send_keepalive = cfs_time_shift(10);
2382
2383 conn = ksocknal_find_conn_locked(peer, NULL, 1);
2384 if (conn != NULL) {
2385 sched = conn->ksnc_scheduler;
2386
2387 spin_lock_bh(&sched->kss_lock);
2388 if (!list_empty(&conn->ksnc_tx_queue)) {
2389 spin_unlock_bh(&sched->kss_lock);
2390 /* there is an queued ACK, don't need keepalive */
2391 return 0;
2392 }
2393
2394 spin_unlock_bh(&sched->kss_lock);
2395 }
2396
2397 read_unlock(&ksocknal_data.ksnd_global_lock);
2398
2399 /* cookie = 1 is reserved for keepalive PING */
2400 tx = ksocknal_alloc_tx_noop(1, 1);
2401 if (tx == NULL) {
2402 read_lock(&ksocknal_data.ksnd_global_lock);
2403 return -ENOMEM;
2404 }
2405
2406 if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2407 read_lock(&ksocknal_data.ksnd_global_lock);
2408 return 1;
2409 }
2410
2411 ksocknal_free_tx(tx);
2412 read_lock(&ksocknal_data.ksnd_global_lock);
2413
2414 return -EIO;
2415 }
2416
2417
2418 void
2419 ksocknal_check_peer_timeouts (int idx)
2420 {
2421 struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2422 ksock_peer_t *peer;
2423 ksock_conn_t *conn;
2424 ksock_tx_t *tx;
2425
2426 again:
2427 /* NB. We expect to have a look at all the peers and not find any
2428 * connections to time out, so we just use a shared lock while we
2429 * take a look... */
2430 read_lock(&ksocknal_data.ksnd_global_lock);
2431
2432 list_for_each_entry(peer, peers, ksnp_list) {
2433 unsigned long deadline = 0;
2434 int resid = 0;
2435 int n = 0;
2436
2437 if (ksocknal_send_keepalive_locked(peer) != 0) {
2438 read_unlock(&ksocknal_data.ksnd_global_lock);
2439 goto again;
2440 }
2441
2442 conn = ksocknal_find_timed_out_conn (peer);
2443
2444 if (conn != NULL) {
2445 read_unlock(&ksocknal_data.ksnd_global_lock);
2446
2447 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2448
2449 /* NB we won't find this one again, but we can't
2450 * just proceed with the next peer, since we dropped
2451 * ksnd_global_lock and it might be dead already! */
2452 ksocknal_conn_decref(conn);
2453 goto again;
2454 }
2455
2456 /* we can't process stale txs right here because we're
2457 * holding only shared lock */
2458 if (!list_empty (&peer->ksnp_tx_queue)) {
2459 ksock_tx_t *tx =
2460 list_entry (peer->ksnp_tx_queue.next,
2461 ksock_tx_t, tx_list);
2462
2463 if (cfs_time_aftereq(cfs_time_current(),
2464 tx->tx_deadline)) {
2465
2466 ksocknal_peer_addref(peer);
2467 read_unlock(&ksocknal_data.ksnd_global_lock);
2468
2469 ksocknal_flush_stale_txs(peer);
2470
2471 ksocknal_peer_decref(peer);
2472 goto again;
2473 }
2474 }
2475
2476 if (list_empty(&peer->ksnp_zc_req_list))
2477 continue;
2478
2479 spin_lock(&peer->ksnp_lock);
2480 list_for_each_entry(tx, &peer->ksnp_zc_req_list, tx_zc_list) {
2481 if (!cfs_time_aftereq(cfs_time_current(),
2482 tx->tx_deadline))
2483 break;
2484 /* ignore the TX if connection is being closed */
2485 if (tx->tx_conn->ksnc_closing)
2486 continue;
2487 n++;
2488 }
2489
2490 if (n == 0) {
2491 spin_unlock(&peer->ksnp_lock);
2492 continue;
2493 }
2494
2495 tx = list_entry(peer->ksnp_zc_req_list.next,
2496 ksock_tx_t, tx_zc_list);
2497 deadline = tx->tx_deadline;
2498 resid = tx->tx_resid;
2499 conn = tx->tx_conn;
2500 ksocknal_conn_addref(conn);
2501
2502 spin_unlock(&peer->ksnp_lock);
2503 read_unlock(&ksocknal_data.ksnd_global_lock);
2504
2505 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2506 "oldest(%p) timed out %ld secs ago, "
2507 "resid: %d, wmem: %d\n",
2508 n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2509 cfs_duration_sec(cfs_time_current() - deadline),
2510 resid, conn->ksnc_sock->sk->sk_wmem_queued);
2511
2512 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2513 ksocknal_conn_decref(conn);
2514 goto again;
2515 }
2516
2517 read_unlock(&ksocknal_data.ksnd_global_lock);
2518 }
2519
2520 int
2521 ksocknal_reaper (void *arg)
2522 {
2523 wait_queue_t wait;
2524 ksock_conn_t *conn;
2525 ksock_sched_t *sched;
2526 struct list_head enomem_conns;
2527 int nenomem_conns;
2528 long timeout;
2529 int i;
2530 int peer_index = 0;
2531 unsigned long deadline = cfs_time_current();
2532
2533 cfs_block_allsigs ();
2534
2535 INIT_LIST_HEAD(&enomem_conns);
2536 init_waitqueue_entry(&wait, current);
2537
2538 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2539
2540 while (!ksocknal_data.ksnd_shuttingdown) {
2541
2542 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2543 conn = list_entry (ksocknal_data. \
2544 ksnd_deathrow_conns.next,
2545 ksock_conn_t, ksnc_list);
2546 list_del (&conn->ksnc_list);
2547
2548 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2549
2550 ksocknal_terminate_conn(conn);
2551 ksocknal_conn_decref(conn);
2552
2553 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2554 continue;
2555 }
2556
2557 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2558 conn = list_entry (ksocknal_data.ksnd_zombie_conns.\
2559 next, ksock_conn_t, ksnc_list);
2560 list_del (&conn->ksnc_list);
2561
2562 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2563
2564 ksocknal_destroy_conn(conn);
2565
2566 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2567 continue;
2568 }
2569
2570 if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2571 list_add(&enomem_conns,
2572 &ksocknal_data.ksnd_enomem_conns);
2573 list_del_init(&ksocknal_data.ksnd_enomem_conns);
2574 }
2575
2576 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2577
2578 /* reschedule all the connections that stalled with ENOMEM... */
2579 nenomem_conns = 0;
2580 while (!list_empty (&enomem_conns)) {
2581 conn = list_entry (enomem_conns.next,
2582 ksock_conn_t, ksnc_tx_list);
2583 list_del (&conn->ksnc_tx_list);
2584
2585 sched = conn->ksnc_scheduler;
2586
2587 spin_lock_bh(&sched->kss_lock);
2588
2589 LASSERT(conn->ksnc_tx_scheduled);
2590 conn->ksnc_tx_ready = 1;
2591 list_add_tail(&conn->ksnc_tx_list,
2592 &sched->kss_tx_conns);
2593 wake_up(&sched->kss_waitq);
2594
2595 spin_unlock_bh(&sched->kss_lock);
2596 nenomem_conns++;
2597 }
2598
2599 /* careful with the jiffy wrap... */
2600 while ((timeout = cfs_time_sub(deadline,
2601 cfs_time_current())) <= 0) {
2602 const int n = 4;
2603 const int p = 1;
2604 int chunk = ksocknal_data.ksnd_peer_hash_size;
2605
2606 /* Time to check for timeouts on a few more peers: I do
2607 * checks every 'p' seconds on a proportion of the peer
2608 * table and I need to check every connection 'n' times
2609 * within a timeout interval, to ensure I detect a
2610 * timeout on any connection within (n+1)/n times the
2611 * timeout interval. */
2612
2613 if (*ksocknal_tunables.ksnd_timeout > n * p)
2614 chunk = (chunk * n * p) /
2615 *ksocknal_tunables.ksnd_timeout;
2616 if (chunk == 0)
2617 chunk = 1;
2618
2619 for (i = 0; i < chunk; i++) {
2620 ksocknal_check_peer_timeouts (peer_index);
2621 peer_index = (peer_index + 1) %
2622 ksocknal_data.ksnd_peer_hash_size;
2623 }
2624
2625 deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2626 }
2627
2628 if (nenomem_conns != 0) {
2629 /* Reduce my timeout if I rescheduled ENOMEM conns.
2630 * This also prevents me getting woken immediately
2631 * if any go back on my enomem list. */
2632 timeout = SOCKNAL_ENOMEM_RETRY;
2633 }
2634 ksocknal_data.ksnd_reaper_waketime =
2635 cfs_time_add(cfs_time_current(), timeout);
2636
2637 set_current_state (TASK_INTERRUPTIBLE);
2638 add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2639
2640 if (!ksocknal_data.ksnd_shuttingdown &&
2641 list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2642 list_empty (&ksocknal_data.ksnd_zombie_conns))
2643 schedule_timeout(timeout);
2644
2645 set_current_state (TASK_RUNNING);
2646 remove_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2647
2648 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2649 }
2650
2651 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2652
2653 ksocknal_thread_fini();
2654 return 0;
2655 }
Linux kernel - Deliverables
This page took 0.252365 seconds and 6 git commands to generate.