Whamcloud - gitweb
b=21678 Add more debug info to lnd_query code path
[fs/lustre-release.git] / lnet / klnds / socklnd / socklnd_cb.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
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                 cfs_spin_lock(&ksocknal_data.ksnd_tx_lock);
39
40                 if (!cfs_list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
41                         tx = cfs_list_entry(ksocknal_data.ksnd_idle_noop_txs. \
42                                             next, ksock_tx_t, tx_list);
43                         LASSERT(tx->tx_desc_size == size);
44                         cfs_list_del(&tx->tx_list);
45                 }
46
47                 cfs_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         cfs_atomic_set(&tx->tx_refcount, 1);
57         tx->tx_zc_capable = 0;
58         tx->tx_zc_checked = 0;
59         tx->tx_desc_size  = size;
60
61         cfs_atomic_inc(&ksocknal_data.ksnd_nactive_txs);
62
63         return tx;
64 }
65
66 ksock_tx_t *
67 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
68 {
69         ksock_tx_t *tx;
70
71         tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
72         if (tx == NULL) {
73                 CERROR("Can't allocate noop tx desc\n");
74                 return NULL;
75         }
76
77         tx->tx_conn     = NULL;
78         tx->tx_lnetmsg  = NULL;
79         tx->tx_kiov     = NULL;
80         tx->tx_nkiov    = 0;
81         tx->tx_iov      = tx->tx_frags.virt.iov;
82         tx->tx_niov     = 1;
83         tx->tx_nonblk   = nonblk;
84
85         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
86         tx->tx_msg.ksm_zc_cookies[1] = cookie;
87
88         return tx;
89 }
90
91
92 void
93 ksocknal_free_tx (ksock_tx_t *tx)
94 {
95         cfs_atomic_dec(&ksocknal_data.ksnd_nactive_txs);
96
97         if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
98                 /* it's a noop tx */
99                 cfs_spin_lock(&ksocknal_data.ksnd_tx_lock);
100
101                 cfs_list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
102
103                 cfs_spin_unlock(&ksocknal_data.ksnd_tx_lock);
104         } else {
105                 LIBCFS_FREE(tx, tx->tx_desc_size);
106         }
107 }
108
109 int
110 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
111 {
112         struct iovec  *iov = tx->tx_iov;
113         int    nob;
114         int    rc;
115
116         LASSERT (tx->tx_niov > 0);
117
118         /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
119         rc = ksocknal_lib_send_iov(conn, tx);
120
121         if (rc <= 0)                            /* sent nothing? */
122                 return (rc);
123
124         nob = rc;
125         LASSERT (nob <= tx->tx_resid);
126         tx->tx_resid -= nob;
127
128         /* "consume" iov */
129         do {
130                 LASSERT (tx->tx_niov > 0);
131
132                 if (nob < (int) iov->iov_len) {
133                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
134                         iov->iov_len -= nob;
135                         return (rc);
136                 }
137
138                 nob -= iov->iov_len;
139                 tx->tx_iov = ++iov;
140                 tx->tx_niov--;
141         } while (nob != 0);
142
143         return (rc);
144 }
145
146 int
147 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
148 {
149         lnet_kiov_t    *kiov = tx->tx_kiov;
150         int     nob;
151         int     rc;
152
153         LASSERT (tx->tx_niov == 0);
154         LASSERT (tx->tx_nkiov > 0);
155
156         /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
157         rc = ksocknal_lib_send_kiov(conn, tx);
158
159         if (rc <= 0)                            /* sent nothing? */
160                 return (rc);
161
162         nob = rc;
163         LASSERT (nob <= tx->tx_resid);
164         tx->tx_resid -= nob;
165
166         /* "consume" kiov */
167         do {
168                 LASSERT(tx->tx_nkiov > 0);
169
170                 if (nob < (int)kiov->kiov_len) {
171                         kiov->kiov_offset += nob;
172                         kiov->kiov_len -= nob;
173                         return rc;
174                 }
175
176                 nob -= (int)kiov->kiov_len;
177                 tx->tx_kiov = ++kiov;
178                 tx->tx_nkiov--;
179         } while (nob != 0);
180
181         return (rc);
182 }
183
184 int
185 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
186 {
187         int      rc;
188         int      bufnob;
189
190         if (ksocknal_data.ksnd_stall_tx != 0) {
191                 cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
192         }
193
194         LASSERT (tx->tx_resid != 0);
195
196         rc = ksocknal_connsock_addref(conn);
197         if (rc != 0) {
198                 LASSERT (conn->ksnc_closing);
199                 return (-ESHUTDOWN);
200         }
201
202         do {
203                 if (ksocknal_data.ksnd_enomem_tx > 0) {
204                         /* testing... */
205                         ksocknal_data.ksnd_enomem_tx--;
206                         rc = -EAGAIN;
207                 } else if (tx->tx_niov != 0) {
208                         rc = ksocknal_send_iov (conn, tx);
209                 } else {
210                         rc = ksocknal_send_kiov (conn, tx);
211                 }
212
213                 bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
214                 if (rc > 0)                     /* sent something? */
215                         conn->ksnc_tx_bufnob += rc; /* account it */
216
217                 if (bufnob < conn->ksnc_tx_bufnob) {
218                         /* allocated send buffer bytes < computed; infer
219                          * something got ACKed */
220                         conn->ksnc_tx_deadline =
221                                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
222                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
223                         conn->ksnc_tx_bufnob = bufnob;
224                         cfs_mb();
225                 }
226
227                 if (rc <= 0) { /* Didn't write anything? */
228
229                         if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
230                                 rc = -EAGAIN;
231
232                         /* Check if EAGAIN is due to memory pressure */
233                         if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
234                                 rc = -ENOMEM;
235
236                         break;
237                 }
238
239                 /* socket's wmem_queued now includes 'rc' bytes */
240                 cfs_atomic_sub (rc, &conn->ksnc_tx_nob);
241                 rc = 0;
242
243         } while (tx->tx_resid != 0);
244
245         ksocknal_connsock_decref(conn);
246         return (rc);
247 }
248
249 int
250 ksocknal_recv_iov (ksock_conn_t *conn)
251 {
252         struct iovec *iov = conn->ksnc_rx_iov;
253         int     nob;
254         int     rc;
255
256         LASSERT (conn->ksnc_rx_niov > 0);
257
258         /* Never touch conn->ksnc_rx_iov or change connection 
259          * status inside ksocknal_lib_recv_iov */
260         rc = ksocknal_lib_recv_iov(conn);
261
262         if (rc <= 0)
263                 return (rc);
264
265         /* received something... */
266         nob = rc;
267
268         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
269         conn->ksnc_rx_deadline =
270                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
271         cfs_mb();                       /* order with setting rx_started */
272         conn->ksnc_rx_started = 1;
273
274         conn->ksnc_rx_nob_wanted -= nob;
275         conn->ksnc_rx_nob_left -= nob;
276
277         do {
278                 LASSERT (conn->ksnc_rx_niov > 0);
279
280                 if (nob < (int)iov->iov_len) {
281                         iov->iov_len -= nob;
282                         iov->iov_base = (void *)((char *)iov->iov_base + nob);
283                         return (-EAGAIN);
284                 }
285
286                 nob -= iov->iov_len;
287                 conn->ksnc_rx_iov = ++iov;
288                 conn->ksnc_rx_niov--;
289         } while (nob != 0);
290
291         return (rc);
292 }
293
294 int
295 ksocknal_recv_kiov (ksock_conn_t *conn)
296 {
297         lnet_kiov_t   *kiov = conn->ksnc_rx_kiov;
298         int     nob;
299         int     rc;
300         LASSERT (conn->ksnc_rx_nkiov > 0);
301
302         /* Never touch conn->ksnc_rx_kiov or change connection 
303          * status inside ksocknal_lib_recv_iov */
304         rc = ksocknal_lib_recv_kiov(conn);
305
306         if (rc <= 0)
307                 return (rc);
308
309         /* received something... */
310         nob = rc;
311
312         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
313         conn->ksnc_rx_deadline =
314                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
315         cfs_mb();                       /* order with setting rx_started */
316         conn->ksnc_rx_started = 1;
317
318         conn->ksnc_rx_nob_wanted -= nob;
319         conn->ksnc_rx_nob_left -= nob;
320
321         do {
322                 LASSERT (conn->ksnc_rx_nkiov > 0);
323
324                 if (nob < (int) kiov->kiov_len) {
325                         kiov->kiov_offset += nob;
326                         kiov->kiov_len -= nob;
327                         return -EAGAIN;
328                 }
329
330                 nob -= kiov->kiov_len;
331                 conn->ksnc_rx_kiov = ++kiov;
332                 conn->ksnc_rx_nkiov--;
333         } while (nob != 0);
334
335         return 1;
336 }
337
338 int
339 ksocknal_receive (ksock_conn_t *conn)
340 {
341         /* Return 1 on success, 0 on EOF, < 0 on error.
342          * Caller checks ksnc_rx_nob_wanted to determine
343          * progress/completion. */
344         int     rc;
345         ENTRY;
346
347         if (ksocknal_data.ksnd_stall_rx != 0) {
348                 cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
349         }
350
351         rc = ksocknal_connsock_addref(conn);
352         if (rc != 0) {
353                 LASSERT (conn->ksnc_closing);
354                 return (-ESHUTDOWN);
355         }
356
357         for (;;) {
358                 if (conn->ksnc_rx_niov != 0)
359                         rc = ksocknal_recv_iov (conn);
360                 else
361                         rc = ksocknal_recv_kiov (conn);
362
363                 if (rc <= 0) {
364                         /* error/EOF or partial receive */
365                         if (rc == -EAGAIN) {
366                                 rc = 1;
367                         } else if (rc == 0 && conn->ksnc_rx_started) {
368                                 /* EOF in the middle of a message */
369                                 rc = -EPROTO;
370                         }
371                         break;
372                 }
373
374                 /* Completed a fragment */
375
376                 if (conn->ksnc_rx_nob_wanted == 0) {
377                         rc = 1;
378                         break;
379                 }
380         }
381
382         ksocknal_connsock_decref(conn);
383         RETURN (rc);
384 }
385
386 void
387 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
388 {
389         lnet_msg_t  *lnetmsg = tx->tx_lnetmsg;
390         int          rc = (tx->tx_resid == 0) ? 0 : -EIO;
391         ENTRY;
392
393         LASSERT(ni != NULL || tx->tx_conn != NULL);
394
395         if (tx->tx_conn != NULL)
396                 ksocknal_conn_decref(tx->tx_conn);
397
398         if (ni == NULL && tx->tx_conn != NULL)
399                 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
400
401         ksocknal_free_tx (tx);
402         if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
403                 lnet_finalize (ni, lnetmsg, rc);
404
405         EXIT;
406 }
407
408 void
409 ksocknal_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int error)
410 {
411         ksock_tx_t *tx;
412
413         while (!cfs_list_empty (txlist)) {
414                 tx = cfs_list_entry (txlist->next, ksock_tx_t, tx_list);
415
416                 if (error && tx->tx_lnetmsg != NULL) {
417                         CDEBUG (D_NETERROR, "Deleting packet type %d len %d %s->%s\n",
418                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
419                                 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
420                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
421                                 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
422                 } else if (error) {
423                         CDEBUG (D_NETERROR, "Deleting noop packet\n");
424                 }
425
426                 cfs_list_del (&tx->tx_list);
427
428                 LASSERT (cfs_atomic_read(&tx->tx_refcount) == 1);
429                 ksocknal_tx_done (ni, tx);
430         }
431 }
432
433 static void
434 ksocknal_check_zc_req(ksock_tx_t *tx)
435 {
436         ksock_conn_t   *conn = tx->tx_conn;
437         ksock_peer_t   *peer = conn->ksnc_peer;
438
439         /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
440          * to ksnp_zc_req_list if some fragment of this message should be sent
441          * zero-copy.  Our peer will send an ACK containing this cookie when
442          * she has received this message to tell us we can signal completion.
443          * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
444          * ksnp_zc_req_list. */
445         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
446         LASSERT (tx->tx_zc_capable);
447
448         tx->tx_zc_checked = 1;
449
450         if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
451             !conn->ksnc_zc_capable)
452                 return;
453
454         /* assign cookie and queue tx to pending list, it will be released when
455          * a matching ack is received. See ksocknal_handle_zcack() */
456
457         ksocknal_tx_addref(tx);
458
459         cfs_spin_lock(&peer->ksnp_lock);
460
461         /* ZC_REQ is going to be pinned to the peer */
462         tx->tx_deadline =
463                 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
464
465         LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
466
467         tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
468
469         if (peer->ksnp_zc_next_cookie == 0)
470                 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
471
472         cfs_list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
473
474         cfs_spin_unlock(&peer->ksnp_lock);
475 }
476
477 static void
478 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
479 {
480         ksock_peer_t   *peer = tx->tx_conn->ksnc_peer;
481
482         LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
483         LASSERT (tx->tx_zc_capable);
484
485         tx->tx_zc_checked = 0;
486
487         cfs_spin_lock(&peer->ksnp_lock);
488
489         if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
490                 /* Not waiting for an ACK */
491                 cfs_spin_unlock(&peer->ksnp_lock);
492                 return;
493         }
494
495         tx->tx_msg.ksm_zc_cookies[0] = 0;
496         cfs_list_del(&tx->tx_zc_list);
497
498         cfs_spin_unlock(&peer->ksnp_lock);
499
500         ksocknal_tx_decref(tx);
501 }
502
503 int
504 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
505 {
506         int            rc;
507
508         if (tx->tx_zc_capable && !tx->tx_zc_checked)
509                 ksocknal_check_zc_req(tx);
510
511         rc = ksocknal_transmit (conn, tx);
512
513         CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
514
515         if (tx->tx_resid == 0) {
516                 /* Sent everything OK */
517                 LASSERT (rc == 0);
518
519                 return (0);
520         }
521
522         if (rc == -EAGAIN)
523                 return (rc);
524
525         if (rc == -ENOMEM) {
526                 static int counter;
527
528                 counter++;   /* exponential backoff warnings */
529                 if ((counter & (-counter)) == counter)
530                         CWARN("%u ENOMEM tx %p (%u allocated)\n",
531                               counter, conn, cfs_atomic_read(&libcfs_kmemory));
532
533                 /* Queue on ksnd_enomem_conns for retry after a timeout */
534                 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
535
536                 /* enomem list takes over scheduler's ref... */
537                 LASSERT (conn->ksnc_tx_scheduled);
538                 cfs_list_add_tail(&conn->ksnc_tx_list,
539                                   &ksocknal_data.ksnd_enomem_conns);
540                 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
541                                                    SOCKNAL_ENOMEM_RETRY),
542                                    ksocknal_data.ksnd_reaper_waketime))
543                         cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
544
545                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
546                 return (rc);
547         }
548
549         /* Actual error */
550         LASSERT (rc < 0);
551
552         if (!conn->ksnc_closing) {
553                 switch (rc) {
554                 case -ECONNRESET:
555                         LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection "
556                                       "while we were sending data; it may have "
557                                       "rebooted.\n",
558                                       HIPQUAD(conn->ksnc_ipaddr));
559                         break;
560                 default:
561                         LCONSOLE_WARN("There was an unexpected network error "
562                                       "while writing to %u.%u.%u.%u: %d.\n",
563                                       HIPQUAD(conn->ksnc_ipaddr), rc);
564                         break;
565                 }
566                 CDEBUG(D_NET, "[%p] Error %d on write to %s"
567                        " ip %d.%d.%d.%d:%d\n", conn, rc,
568                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
569                        HIPQUAD(conn->ksnc_ipaddr),
570                        conn->ksnc_port);
571         }
572
573         if (tx->tx_zc_checked)
574                 ksocknal_uncheck_zc_req(tx);
575
576         /* it's not an error if conn is being closed */
577         ksocknal_close_conn_and_siblings (conn,
578                                           (conn->ksnc_closing) ? 0 : rc);
579
580         return (rc);
581 }
582
583 void
584 ksocknal_launch_connection_locked (ksock_route_t *route)
585 {
586
587         /* called holding write lock on ksnd_global_lock */
588
589         LASSERT (!route->ksnr_scheduled);
590         LASSERT (!route->ksnr_connecting);
591         LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
592
593         route->ksnr_scheduled = 1;              /* scheduling conn for connd */
594         ksocknal_route_addref(route);           /* extra ref for connd */
595
596         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
597
598         cfs_list_add_tail (&route->ksnr_connd_list,
599                            &ksocknal_data.ksnd_connd_routes);
600         cfs_waitq_signal (&ksocknal_data.ksnd_connd_waitq);
601
602         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
603 }
604
605 void
606 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
607 {
608         ksock_route_t *route;
609
610         /* called holding write lock on ksnd_global_lock */
611         for (;;) {
612                 /* launch any/all connections that need it */
613                 route = ksocknal_find_connectable_route_locked(peer);
614                 if (route == NULL)
615                         return;
616
617                 ksocknal_launch_connection_locked(route);
618         }
619 }
620
621 ksock_conn_t *
622 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
623 {
624         cfs_list_t       *tmp;
625         ksock_conn_t     *conn;
626         ksock_conn_t     *typed = NULL;
627         ksock_conn_t     *fallback = NULL;
628         int               tnob     = 0;
629         int               fnob     = 0;
630
631         cfs_list_for_each (tmp, &peer->ksnp_conns) {
632                 ksock_conn_t *c  = cfs_list_entry(tmp, ksock_conn_t, ksnc_list);
633                 int           nob = cfs_atomic_read(&c->ksnc_tx_nob) +
634                                     libcfs_sock_wmem_queued(c->ksnc_sock);
635                 int           rc;
636
637                 LASSERT (!c->ksnc_closing);
638                 LASSERT (c->ksnc_proto != NULL &&
639                          c->ksnc_proto->pro_match_tx != NULL);
640
641                 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
642
643                 switch (rc) {
644                 default:
645                         LBUG();
646                 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
647                         continue;
648
649                 case SOCKNAL_MATCH_YES: /* typed connection */
650                         if (typed == NULL || tnob > nob ||
651                             (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
652                              cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
653                                 typed = c;
654                                 tnob  = nob;
655                         }
656                         break;
657
658                 case SOCKNAL_MATCH_MAY: /* fallback connection */
659                         if (fallback == NULL || fnob > nob ||
660                             (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
661                              cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
662                                 fallback = c;
663                                 fnob     = nob;
664                         }
665                         break;
666                 }
667         }
668
669         /* prefer the typed selection */
670         conn = (typed != NULL) ? typed : fallback;
671
672         if (conn != NULL)
673                 conn->ksnc_tx_last_post = cfs_time_current();
674
675         return conn;
676 }
677
678 void
679 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
680 {
681         conn->ksnc_proto->pro_pack(tx);
682
683         cfs_atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
684         ksocknal_conn_addref(conn); /* +1 ref for tx */
685         tx->tx_conn = conn;
686 }
687
688 void
689 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
690 {
691         ksock_sched_t *sched = conn->ksnc_scheduler;
692         ksock_msg_t   *msg = &tx->tx_msg;
693         ksock_tx_t    *ztx = NULL;
694         int            bufnob = 0;
695
696         /* called holding global lock (read or irq-write) and caller may
697          * not have dropped this lock between finding conn and calling me,
698          * so we don't need the {get,put}connsock dance to deref
699          * ksnc_sock... */
700         LASSERT(!conn->ksnc_closing);
701
702         CDEBUG (D_NET, "Sending to %s ip %d.%d.%d.%d:%d\n", 
703                 libcfs_id2str(conn->ksnc_peer->ksnp_id),
704                 HIPQUAD(conn->ksnc_ipaddr),
705                 conn->ksnc_port);
706
707         ksocknal_tx_prep(conn, tx);
708
709         /* Ensure the frags we've been given EXACTLY match the number of
710          * bytes we want to send.  Many TCP/IP stacks disregard any total
711          * size parameters passed to them and just look at the frags. 
712          *
713          * We always expect at least 1 mapped fragment containing the
714          * complete ksocknal message header. */
715         LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
716                  lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
717                  (unsigned int)tx->tx_nob);
718         LASSERT (tx->tx_niov >= 1);
719         LASSERT (tx->tx_resid == tx->tx_nob);
720
721         CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
722                 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
723                                                KSOCK_MSG_NOOP,
724                 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
725
726         /*
727          * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
728          * but they're used inside spinlocks a lot.
729          */
730         bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
731         cfs_spin_lock_bh (&sched->kss_lock);
732
733         if (cfs_list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
734                 /* First packet starts the timeout */
735                 conn->ksnc_tx_deadline =
736                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
737                 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
738                         conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
739                 conn->ksnc_tx_bufnob = 0;
740                 cfs_mb(); /* order with adding to tx_queue */
741         }
742
743         if (msg->ksm_type == KSOCK_MSG_NOOP) {
744                 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
745                  * on a normal packet so I don't need to send it */
746                 LASSERT (msg->ksm_zc_cookies[1] != 0);
747                 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
748
749                 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
750                         ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
751
752         } else {
753                 /* It's a normal packet - can it piggback a noop zc-ack that
754                  * has been queued already? */
755                 LASSERT (msg->ksm_zc_cookies[1] == 0);
756                 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
757
758                 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
759                 /* ztx will be released later */
760         }
761
762         if (ztx != NULL) {
763                 cfs_atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
764                 cfs_list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
765         }
766
767         if (conn->ksnc_tx_ready &&      /* able to send */
768             !conn->ksnc_tx_scheduled) { /* not scheduled to send */
769                 /* +1 ref for scheduler */
770                 ksocknal_conn_addref(conn);
771                 cfs_list_add_tail (&conn->ksnc_tx_list,
772                                    &sched->kss_tx_conns);
773                 conn->ksnc_tx_scheduled = 1;
774                 cfs_waitq_signal (&sched->kss_waitq);
775         }
776
777         cfs_spin_unlock_bh (&sched->kss_lock);
778 }
779
780
781 ksock_route_t *
782 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
783 {
784         cfs_time_t     now = cfs_time_current();
785         cfs_list_t    *tmp;
786         ksock_route_t *route;
787
788         cfs_list_for_each (tmp, &peer->ksnp_routes) {
789                 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
790
791                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
792
793                 if (route->ksnr_scheduled)      /* connections being established */
794                         continue;
795
796                 /* all route types connected ? */
797                 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
798                         continue;
799
800                 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
801                       cfs_time_aftereq(now, route->ksnr_timeout))) {
802                         CDEBUG(D_NET,
803                                "Too soon to retry route %u.%u.%u.%u "
804                                "(cnted %d, interval %ld, %ld secs later)\n",
805                                HIPQUAD(route->ksnr_ipaddr),
806                                route->ksnr_connected,
807                                route->ksnr_retry_interval,
808                                cfs_duration_sec(route->ksnr_timeout - now));
809                         continue;
810                 }
811
812                 return (route);
813         }
814
815         return (NULL);
816 }
817
818 ksock_route_t *
819 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
820 {
821         cfs_list_t        *tmp;
822         ksock_route_t     *route;
823
824         cfs_list_for_each (tmp, &peer->ksnp_routes) {
825                 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
826
827                 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
828
829                 if (route->ksnr_scheduled)
830                         return (route);
831         }
832
833         return (NULL);
834 }
835
836 int
837 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
838 {
839         ksock_peer_t     *peer;
840         ksock_conn_t     *conn;
841         cfs_rwlock_t     *g_lock;
842         int               retry;
843         int               rc;
844
845         LASSERT (tx->tx_conn == NULL);
846
847         g_lock = &ksocknal_data.ksnd_global_lock;
848
849         for (retry = 0;; retry = 1) {
850                 cfs_read_lock (g_lock);
851                 peer = ksocknal_find_peer_locked(ni, id);
852                 if (peer != NULL) {
853                         if (ksocknal_find_connectable_route_locked(peer) == NULL) {
854                                 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
855                                 if (conn != NULL) {
856                                         /* I've got no routes that need to be
857                                          * connecting and I do have an actual
858                                          * connection... */
859                                         ksocknal_queue_tx_locked (tx, conn);
860                                         cfs_read_unlock (g_lock);
861                                         return (0);
862                                 }
863                         }
864                 }
865
866                 /* I'll need a write lock... */
867                 cfs_read_unlock (g_lock);
868
869                 cfs_write_lock_bh (g_lock);
870
871                 peer = ksocknal_find_peer_locked(ni, id);
872                 if (peer != NULL)
873                         break;
874
875                 cfs_write_unlock_bh (g_lock);
876
877                 if ((id.pid & LNET_PID_USERFLAG) != 0) {
878                         CERROR("Refusing to create a connection to "
879                                "userspace process %s\n", libcfs_id2str(id));
880                         return -EHOSTUNREACH;
881                 }
882
883                 if (retry) {
884                         CERROR("Can't find peer %s\n", libcfs_id2str(id));
885                         return -EHOSTUNREACH;
886                 }
887
888                 rc = ksocknal_add_peer(ni, id,
889                                        LNET_NIDADDR(id.nid),
890                                        lnet_acceptor_port());
891                 if (rc != 0) {
892                         CERROR("Can't add peer %s: %d\n",
893                                libcfs_id2str(id), rc);
894                         return rc;
895                 }
896         }
897
898         ksocknal_launch_all_connections_locked(peer);
899
900         conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
901         if (conn != NULL) {
902                 /* Connection exists; queue message on it */
903                 ksocknal_queue_tx_locked (tx, conn);
904                 cfs_write_unlock_bh (g_lock);
905                 return (0);
906         }
907
908         if (peer->ksnp_accepting > 0 ||
909             ksocknal_find_connecting_route_locked (peer) != NULL) {
910                 /* the message is going to be pinned to the peer */
911                 tx->tx_deadline =
912                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
913
914                 /* Queue the message until a connection is established */
915                 cfs_list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
916                 cfs_write_unlock_bh (g_lock);
917                 return 0;
918         }
919
920         cfs_write_unlock_bh (g_lock);
921
922         /* NB Routes may be ignored if connections to them failed recently */
923         CDEBUG(D_NETERROR, "No usable routes to %s\n", libcfs_id2str(id));
924         return (-EHOSTUNREACH);
925 }
926
927 int
928 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
929 {
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 (!cfs_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         tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
961         if (tx == NULL) {
962                 CERROR("Can't allocate tx desc type %d size %d\n",
963                        type, desc_size);
964                 return (-ENOMEM);
965         }
966
967         tx->tx_conn = NULL;                     /* set when assigned a conn */
968         tx->tx_lnetmsg = lntmsg;
969
970         if (payload_iov != NULL) {
971                 tx->tx_kiov = NULL;
972                 tx->tx_nkiov = 0;
973                 tx->tx_iov = tx->tx_frags.virt.iov;
974                 tx->tx_niov = 1 +
975                               lnet_extract_iov(payload_niov, &tx->tx_iov[1],
976                                                payload_niov, payload_iov,
977                                                payload_offset, payload_nob);
978         } else {
979                 tx->tx_niov = 1;
980                 tx->tx_iov = &tx->tx_frags.paged.iov;
981                 tx->tx_kiov = tx->tx_frags.paged.kiov;
982                 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
983                                                  payload_niov, payload_kiov,
984                                                  payload_offset, payload_nob);
985
986                 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
987                         tx->tx_zc_capable = 1;
988         }
989
990         socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
991
992         /* The first fragment will be set later in pro_pack */
993         rc = ksocknal_launch_packet(ni, tx, target);
994         if (rc == 0)
995                 return (0);
996
997         ksocknal_free_tx(tx);
998         return (-EIO);
999 }
1000
1001 int
1002 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
1003 {
1004         long          pid = cfs_kernel_thread (fn, arg, 0);
1005
1006         if (pid < 0)
1007                 return ((int)pid);
1008
1009         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1010         ksocknal_data.ksnd_nthreads++;
1011         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1012         return (0);
1013 }
1014
1015 void
1016 ksocknal_thread_fini (void)
1017 {
1018         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1019         ksocknal_data.ksnd_nthreads--;
1020         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1021 }
1022
1023 int
1024 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1025 {
1026         static char ksocknal_slop_buffer[4096];
1027
1028         int            nob;
1029         unsigned int   niov;
1030         int            skipped;
1031
1032         LASSERT(conn->ksnc_proto != NULL);
1033
1034         if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1035                 /* Remind the socket to ack eagerly... */
1036                 ksocknal_lib_eager_ack(conn);
1037         }
1038
1039         if (nob_to_skip == 0) {         /* right at next packet boundary now */
1040                 conn->ksnc_rx_started = 0;
1041                 cfs_mb();                       /* racing with timeout thread */
1042
1043                 switch (conn->ksnc_proto->pro_version) {
1044                 case  KSOCK_PROTO_V2:
1045                 case  KSOCK_PROTO_V3:
1046                         conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1047                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1048                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1049
1050                         conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1051                         conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1052                         conn->ksnc_rx_iov[0].iov_len  = offsetof(ksock_msg_t, ksm_u);
1053                         break;
1054
1055                 case KSOCK_PROTO_V1:
1056                         /* Receiving bare lnet_hdr_t */
1057                         conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1058                         conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1059                         conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1060
1061                         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1062                         conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1063                         conn->ksnc_rx_iov[0].iov_len  = sizeof (lnet_hdr_t);
1064                         break;
1065
1066                 default:
1067                         LBUG ();
1068                 }
1069                 conn->ksnc_rx_niov = 1;
1070
1071                 conn->ksnc_rx_kiov = NULL;
1072                 conn->ksnc_rx_nkiov = 0;
1073                 conn->ksnc_rx_csum = ~0;
1074                 return (1);
1075         }
1076
1077         /* Set up to skip as much as possible now.  If there's more left
1078          * (ran out of iov entries) we'll get called again */
1079
1080         conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1081         conn->ksnc_rx_nob_left = nob_to_skip;
1082         conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1083         skipped = 0;
1084         niov = 0;
1085
1086         do {
1087                 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1088
1089                 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1090                 conn->ksnc_rx_iov[niov].iov_len  = nob;
1091                 niov++;
1092                 skipped += nob;
1093                 nob_to_skip -=nob;
1094
1095         } while (nob_to_skip != 0 &&    /* mustn't overflow conn's rx iov */
1096                  niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1097
1098         conn->ksnc_rx_niov = niov;
1099         conn->ksnc_rx_kiov = NULL;
1100         conn->ksnc_rx_nkiov = 0;
1101         conn->ksnc_rx_nob_wanted = skipped;
1102         return (0);
1103 }
1104
1105 int
1106 ksocknal_process_receive (ksock_conn_t *conn)
1107 {
1108         lnet_hdr_t        *lhdr;
1109         lnet_process_id_t *id;
1110         int                rc;
1111
1112         LASSERT (cfs_atomic_read(&conn->ksnc_conn_refcount) > 0);
1113
1114         /* NB: sched lock NOT held */
1115         /* SOCKNAL_RX_LNET_HEADER is here for backward compatability */
1116         LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1117                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1118                  conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1119                  conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1120  again:
1121         if (conn->ksnc_rx_nob_wanted != 0) {
1122                 rc = ksocknal_receive(conn);
1123
1124                 if (rc <= 0) {
1125                         LASSERT (rc != -EAGAIN);
1126
1127                         if (rc == 0)
1128                                 CDEBUG (D_NET, "[%p] EOF from %s"
1129                                         " ip %d.%d.%d.%d:%d\n", conn,
1130                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1131                                         HIPQUAD(conn->ksnc_ipaddr),
1132                                         conn->ksnc_port);
1133                         else if (!conn->ksnc_closing)
1134                                 CERROR ("[%p] Error %d on read from %s"
1135                                         " ip %d.%d.%d.%d:%d\n",
1136                                         conn, rc,
1137                                         libcfs_id2str(conn->ksnc_peer->ksnp_id),
1138                                         HIPQUAD(conn->ksnc_ipaddr),
1139                                         conn->ksnc_port);
1140
1141                         /* it's not an error if conn is being closed */
1142                         ksocknal_close_conn_and_siblings (conn,
1143                                                           (conn->ksnc_closing) ? 0 : rc);
1144                         return (rc == 0 ? -ESHUTDOWN : rc);
1145                 }
1146
1147                 if (conn->ksnc_rx_nob_wanted != 0) {
1148                         /* short read */
1149                         return (-EAGAIN);
1150                 }
1151         }
1152         switch (conn->ksnc_rx_state) {
1153         case SOCKNAL_RX_KSM_HEADER:
1154                 if (conn->ksnc_flip) {
1155                         __swab32s(&conn->ksnc_msg.ksm_type);
1156                         __swab32s(&conn->ksnc_msg.ksm_csum);
1157                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1158                         __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1159                 }
1160
1161                 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1162                     conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1163                         CERROR("%s: Unknown message type: %x\n",
1164                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1165                                conn->ksnc_msg.ksm_type);
1166                         ksocknal_new_packet(conn, 0);
1167                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1168                         return (-EPROTO);
1169                 }
1170
1171                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1172                     conn->ksnc_msg.ksm_csum != 0 &&     /* has checksum */
1173                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1174                         /* NOOP Checksum error */
1175                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1176                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1177                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1178                         ksocknal_new_packet(conn, 0);
1179                         ksocknal_close_conn_and_siblings(conn, -EPROTO);
1180                         return (-EIO);
1181                 }
1182
1183                 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1184                         __u64 cookie = 0;
1185
1186                         LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1187
1188                         if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1189                                 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1190
1191                         rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1192                                                conn->ksnc_msg.ksm_zc_cookies[1]);
1193
1194                         if (rc != 0) {
1195                                 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1196                                        libcfs_id2str(conn->ksnc_peer->ksnp_id),
1197                                        cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1198                                 ksocknal_new_packet(conn, 0);
1199                                 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1200                                 return (rc);
1201                         }
1202                 }
1203
1204                 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1205                         ksocknal_new_packet (conn, 0);
1206                         return 0;       /* NOOP is done and just return */
1207                 }
1208
1209                 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1210                 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1211                 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1212
1213                 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1214                 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1215                 conn->ksnc_rx_iov[0].iov_len  = sizeof(ksock_lnet_msg_t);
1216
1217                 conn->ksnc_rx_niov = 1;
1218                 conn->ksnc_rx_kiov = NULL;
1219                 conn->ksnc_rx_nkiov = 0;
1220
1221                 goto again;     /* read lnet header now */
1222
1223         case SOCKNAL_RX_LNET_HEADER:
1224                 /* unpack message header */
1225                 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1226
1227                 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1228                         /* Userspace peer */
1229                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1230                         id   = &conn->ksnc_peer->ksnp_id;
1231
1232                         /* Substitute process ID assigned at connection time */
1233                         lhdr->src_pid = cpu_to_le32(id->pid);
1234                         lhdr->src_nid = cpu_to_le64(id->nid);
1235                 }
1236
1237                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1238                 ksocknal_conn_addref(conn);     /* ++ref while parsing */
1239
1240                 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1241                                 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1242                                 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1243                 if (rc < 0) {
1244                         /* I just received garbage: give up on this conn */
1245                         ksocknal_new_packet(conn, 0);
1246                         ksocknal_close_conn_and_siblings (conn, rc);
1247                         ksocknal_conn_decref(conn);
1248                         return (-EPROTO);
1249                 }
1250
1251                 /* I'm racing with ksocknal_recv() */
1252                 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1253                          conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1254
1255                 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1256                         return 0;
1257
1258                 /* ksocknal_recv() got called */
1259                 goto again;
1260
1261         case SOCKNAL_RX_LNET_PAYLOAD:
1262                 /* payload all received */
1263                 rc = 0;
1264
1265                 if (conn->ksnc_rx_nob_left == 0 &&   /* not truncating */
1266                     conn->ksnc_msg.ksm_csum != 0 &&  /* has checksum */
1267                     conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1268                         CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1269                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1270                                conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1271                         rc = -EIO;
1272                 }
1273
1274                 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1275                         LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1276
1277                         lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1278                         id   = &conn->ksnc_peer->ksnp_id;
1279
1280                         rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1281                                         conn->ksnc_msg.ksm_zc_cookies[0],
1282                                         *ksocknal_tunables.ksnd_nonblk_zcack ||
1283                                         le64_to_cpu(lhdr->src_nid) != id->nid);
1284                 }
1285
1286                 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1287
1288                 if (rc != 0) {
1289                         ksocknal_new_packet(conn, 0);
1290                         ksocknal_close_conn_and_siblings (conn, rc);
1291                         return (-EPROTO);
1292                 }
1293                 /* Fall through */
1294
1295         case SOCKNAL_RX_SLOP:
1296                 /* starting new packet? */
1297                 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1298                         return 0;       /* come back later */
1299                 goto again;             /* try to finish reading slop now */
1300
1301         default:
1302                 break;
1303         }
1304
1305         /* Not Reached */
1306         LBUG ();
1307         return (-EINVAL);                       /* keep gcc happy */
1308 }
1309
1310 int
1311 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1312                unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1313                unsigned int offset, unsigned int mlen, unsigned int rlen)
1314 {
1315         ksock_conn_t  *conn = (ksock_conn_t *)private;
1316         ksock_sched_t *sched = conn->ksnc_scheduler;
1317
1318         LASSERT (mlen <= rlen);
1319         LASSERT (niov <= LNET_MAX_IOV);
1320
1321         conn->ksnc_cookie = msg;
1322         conn->ksnc_rx_nob_wanted = mlen;
1323         conn->ksnc_rx_nob_left   = rlen;
1324
1325         if (mlen == 0 || iov != NULL) {
1326                 conn->ksnc_rx_nkiov = 0;
1327                 conn->ksnc_rx_kiov = NULL;
1328                 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1329                 conn->ksnc_rx_niov =
1330                         lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1331                                          niov, iov, offset, mlen);
1332         } else {
1333                 conn->ksnc_rx_niov = 0;
1334                 conn->ksnc_rx_iov  = NULL;
1335                 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1336                 conn->ksnc_rx_nkiov =
1337                         lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1338                                           niov, kiov, offset, mlen);
1339         }
1340
1341         LASSERT (mlen ==
1342                  lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1343                  lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1344
1345         LASSERT (conn->ksnc_rx_scheduled);
1346
1347         cfs_spin_lock_bh (&sched->kss_lock);
1348
1349         switch (conn->ksnc_rx_state) {
1350         case SOCKNAL_RX_PARSE_WAIT:
1351                 cfs_list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1352                 cfs_waitq_signal (&sched->kss_waitq);
1353                 LASSERT (conn->ksnc_rx_ready);
1354                 break;
1355
1356         case SOCKNAL_RX_PARSE:
1357                 /* scheduler hasn't noticed I'm parsing yet */
1358                 break;
1359         }
1360
1361         conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1362
1363         cfs_spin_unlock_bh (&sched->kss_lock);
1364         ksocknal_conn_decref(conn);
1365         return (0);
1366 }
1367
1368 static inline int
1369 ksocknal_sched_cansleep(ksock_sched_t *sched)
1370 {
1371         int           rc;
1372
1373         cfs_spin_lock_bh (&sched->kss_lock);
1374
1375         rc = (!ksocknal_data.ksnd_shuttingdown &&
1376               cfs_list_empty(&sched->kss_rx_conns) &&
1377               cfs_list_empty(&sched->kss_tx_conns));
1378
1379         cfs_spin_unlock_bh (&sched->kss_lock);
1380         return (rc);
1381 }
1382
1383 int ksocknal_scheduler (void *arg)
1384 {
1385         ksock_sched_t     *sched = (ksock_sched_t *)arg;
1386         ksock_conn_t      *conn;
1387         ksock_tx_t        *tx;
1388         int                rc;
1389         int                nloops = 0;
1390         int                id = (int)(sched - ksocknal_data.ksnd_schedulers);
1391         char               name[16];
1392
1393         snprintf (name, sizeof (name),"socknal_sd%02d", id);
1394         cfs_daemonize (name);
1395         cfs_block_allsigs ();
1396
1397         if (ksocknal_lib_bind_thread_to_cpu(id))
1398                 CERROR ("Can't set CPU affinity for %s to %d\n", name, id);
1399
1400         cfs_spin_lock_bh (&sched->kss_lock);
1401
1402         while (!ksocknal_data.ksnd_shuttingdown) {
1403                 int did_something = 0;
1404
1405                 /* Ensure I progress everything semi-fairly */
1406
1407                 if (!cfs_list_empty (&sched->kss_rx_conns)) {
1408                         conn = cfs_list_entry(sched->kss_rx_conns.next,
1409                                               ksock_conn_t, ksnc_rx_list);
1410                         cfs_list_del(&conn->ksnc_rx_list);
1411
1412                         LASSERT(conn->ksnc_rx_scheduled);
1413                         LASSERT(conn->ksnc_rx_ready);
1414
1415                         /* clear rx_ready in case receive isn't complete.
1416                          * Do it BEFORE we call process_recv, since
1417                          * data_ready can set it any time after we release
1418                          * kss_lock. */
1419                         conn->ksnc_rx_ready = 0;
1420                         cfs_spin_unlock_bh (&sched->kss_lock);
1421
1422                         rc = ksocknal_process_receive(conn);
1423
1424                         cfs_spin_lock_bh (&sched->kss_lock);
1425
1426                         /* I'm the only one that can clear this flag */
1427                         LASSERT(conn->ksnc_rx_scheduled);
1428
1429                         /* Did process_receive get everything it wanted? */
1430                         if (rc == 0)
1431                                 conn->ksnc_rx_ready = 1;
1432
1433                         if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1434                                 /* Conn blocked waiting for ksocknal_recv()
1435                                  * I change its state (under lock) to signal
1436                                  * it can be rescheduled */
1437                                 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1438                         } else if (conn->ksnc_rx_ready) {
1439                                 /* reschedule for rx */
1440                                 cfs_list_add_tail (&conn->ksnc_rx_list,
1441                                                    &sched->kss_rx_conns);
1442                         } else {
1443                                 conn->ksnc_rx_scheduled = 0;
1444                                 /* drop my ref */
1445                                 ksocknal_conn_decref(conn);
1446                         }
1447
1448                         did_something = 1;
1449                 }
1450
1451                 if (!cfs_list_empty (&sched->kss_tx_conns)) {
1452                         CFS_LIST_HEAD    (zlist);
1453
1454                         if (!cfs_list_empty(&sched->kss_zombie_noop_txs)) {
1455                                 cfs_list_add(&zlist,
1456                                              &sched->kss_zombie_noop_txs);
1457                                 cfs_list_del_init(&sched->kss_zombie_noop_txs);
1458                         }
1459
1460                         conn = cfs_list_entry(sched->kss_tx_conns.next,
1461                                               ksock_conn_t, ksnc_tx_list);
1462                         cfs_list_del (&conn->ksnc_tx_list);
1463
1464                         LASSERT(conn->ksnc_tx_scheduled);
1465                         LASSERT(conn->ksnc_tx_ready);
1466                         LASSERT(!cfs_list_empty(&conn->ksnc_tx_queue));
1467
1468                         tx = cfs_list_entry(conn->ksnc_tx_queue.next,
1469                                             ksock_tx_t, tx_list);
1470
1471                         if (conn->ksnc_tx_carrier == tx)
1472                                 ksocknal_next_tx_carrier(conn);
1473
1474                         /* dequeue now so empty list => more to send */
1475                         cfs_list_del(&tx->tx_list);
1476
1477                         /* Clear tx_ready in case send isn't complete.  Do
1478                          * it BEFORE we call process_transmit, since
1479                          * write_space can set it any time after we release
1480                          * kss_lock. */
1481                         conn->ksnc_tx_ready = 0;
1482                         cfs_spin_unlock_bh (&sched->kss_lock);
1483
1484                         if (!cfs_list_empty(&zlist)) {
1485                                 /* free zombie noop txs, it's fast because 
1486                                  * noop txs are just put in freelist */
1487                                 ksocknal_txlist_done(NULL, &zlist, 0);
1488                         }
1489
1490                         rc = ksocknal_process_transmit(conn, tx);
1491
1492                         if (rc == -ENOMEM || rc == -EAGAIN) {
1493                                 /* Incomplete send: replace tx on HEAD of tx_queue */
1494                                 cfs_spin_lock_bh (&sched->kss_lock);
1495                                 cfs_list_add (&tx->tx_list,
1496                                               &conn->ksnc_tx_queue);
1497                         } else {
1498                                 /* Complete send; tx -ref */
1499                                 ksocknal_tx_decref (tx);
1500
1501                                 cfs_spin_lock_bh (&sched->kss_lock);
1502                                 /* assume space for more */
1503                                 conn->ksnc_tx_ready = 1;
1504                         }
1505
1506                         if (rc == -ENOMEM) {
1507                                 /* Do nothing; after a short timeout, this
1508                                  * conn will be reposted on kss_tx_conns. */
1509                         } else if (conn->ksnc_tx_ready &&
1510                                    !cfs_list_empty (&conn->ksnc_tx_queue)) {
1511                                 /* reschedule for tx */
1512                                 cfs_list_add_tail (&conn->ksnc_tx_list,
1513                                                    &sched->kss_tx_conns);
1514                         } else {
1515                                 conn->ksnc_tx_scheduled = 0;
1516                                 /* drop my ref */
1517                                 ksocknal_conn_decref(conn);
1518                         }
1519
1520                         did_something = 1;
1521                 }
1522                 if (!did_something ||           /* nothing to do */
1523                     ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1524                         cfs_spin_unlock_bh (&sched->kss_lock);
1525
1526                         nloops = 0;
1527
1528                         if (!did_something) {   /* wait for something to do */
1529                                 cfs_wait_event_interruptible_exclusive(
1530                                         sched->kss_waitq,
1531                                         !ksocknal_sched_cansleep(sched), rc);
1532                                 LASSERT (rc == 0);
1533                         } else {
1534                                 cfs_cond_resched();
1535                         }
1536
1537                         cfs_spin_lock_bh (&sched->kss_lock);
1538                 }
1539         }
1540
1541         cfs_spin_unlock_bh (&sched->kss_lock);
1542         ksocknal_thread_fini ();
1543         return (0);
1544 }
1545
1546 /*
1547  * Add connection to kss_rx_conns of scheduler
1548  * and wakeup the scheduler.
1549  */
1550 void ksocknal_read_callback (ksock_conn_t *conn)
1551 {
1552         ksock_sched_t *sched;
1553         ENTRY;
1554
1555         sched = conn->ksnc_scheduler;
1556
1557         cfs_spin_lock_bh (&sched->kss_lock);
1558
1559         conn->ksnc_rx_ready = 1;
1560
1561         if (!conn->ksnc_rx_scheduled) {  /* not being progressed */
1562                 cfs_list_add_tail(&conn->ksnc_rx_list,
1563                                   &sched->kss_rx_conns);
1564                 conn->ksnc_rx_scheduled = 1;
1565                 /* extra ref for scheduler */
1566                 ksocknal_conn_addref(conn);
1567
1568                 cfs_waitq_signal (&sched->kss_waitq);
1569         }
1570         cfs_spin_unlock_bh (&sched->kss_lock);
1571
1572         EXIT;
1573 }
1574
1575 /*
1576  * Add connection to kss_tx_conns of scheduler
1577  * and wakeup the scheduler.
1578  */
1579 void ksocknal_write_callback (ksock_conn_t *conn)
1580 {
1581         ksock_sched_t *sched;
1582         ENTRY;
1583
1584         sched = conn->ksnc_scheduler;
1585
1586         cfs_spin_lock_bh (&sched->kss_lock);
1587
1588         conn->ksnc_tx_ready = 1;
1589
1590         if (!conn->ksnc_tx_scheduled && // not being progressed
1591             !cfs_list_empty(&conn->ksnc_tx_queue)){//packets to send
1592                 cfs_list_add_tail (&conn->ksnc_tx_list,
1593                                    &sched->kss_tx_conns);
1594                 conn->ksnc_tx_scheduled = 1;
1595                 /* extra ref for scheduler */
1596                 ksocknal_conn_addref(conn);
1597
1598                 cfs_waitq_signal (&sched->kss_waitq);
1599         }
1600
1601         cfs_spin_unlock_bh (&sched->kss_lock);
1602
1603         EXIT;
1604 }
1605
1606 ksock_proto_t *
1607 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1608 {
1609         __u32   version = 0;
1610
1611         if (hello->kshm_magic == LNET_PROTO_MAGIC)
1612                 version = hello->kshm_version;
1613         else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1614                 version = __swab32(hello->kshm_version);
1615
1616         if (version != 0) {
1617 #if SOCKNAL_VERSION_DEBUG
1618                 if (*ksocknal_tunables.ksnd_protocol == 1)
1619                         return NULL;
1620
1621                 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1622                     version == KSOCK_PROTO_V3)
1623                         return NULL;
1624 #endif
1625                 if (version == KSOCK_PROTO_V2)
1626                         return &ksocknal_protocol_v2x;
1627
1628                 if (version == KSOCK_PROTO_V3)
1629                         return &ksocknal_protocol_v3x;
1630
1631                 return NULL;
1632         }
1633
1634         if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1635                 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1636
1637                 CLASSERT (sizeof (lnet_magicversion_t) ==
1638                           offsetof (ksock_hello_msg_t, kshm_src_nid));
1639
1640                 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1641                     hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1642                         return &ksocknal_protocol_v1x;
1643         }
1644
1645         return NULL;
1646 }
1647
1648 int
1649 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1650                      lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1651 {
1652         /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1653         ksock_net_t         *net = (ksock_net_t *)ni->ni_data;
1654
1655         LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1656
1657         /* rely on caller to hold a ref on socket so it wouldn't disappear */
1658         LASSERT (conn->ksnc_proto != NULL);
1659
1660         hello->kshm_src_nid         = ni->ni_nid;
1661         hello->kshm_dst_nid         = peer_nid;
1662         hello->kshm_src_pid         = the_lnet.ln_pid;
1663
1664         hello->kshm_src_incarnation = net->ksnn_incarnation;
1665         hello->kshm_ctype           = conn->ksnc_type;
1666
1667         return conn->ksnc_proto->pro_send_hello(conn, hello);
1668 }
1669
1670 int
1671 ksocknal_invert_type(int type)
1672 {
1673         switch (type)
1674         {
1675         case SOCKLND_CONN_ANY:
1676         case SOCKLND_CONN_CONTROL:
1677                 return (type);
1678         case SOCKLND_CONN_BULK_IN:
1679                 return SOCKLND_CONN_BULK_OUT;
1680         case SOCKLND_CONN_BULK_OUT:
1681                 return SOCKLND_CONN_BULK_IN;
1682         default:
1683                 return (SOCKLND_CONN_NONE);
1684         }
1685 }
1686
1687 int
1688 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1689                      ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1690                      __u64 *incarnation)
1691 {
1692         /* Return < 0        fatal error
1693          *        0          success
1694          *        EALREADY   lost connection race
1695          *        EPROTO     protocol version mismatch
1696          */
1697         cfs_socket_t        *sock = conn->ksnc_sock;
1698         int                  active = (conn->ksnc_proto != NULL);
1699         int                  timeout;
1700         int                  proto_match;
1701         int                  rc;
1702         ksock_proto_t       *proto;
1703         lnet_process_id_t    recv_id;
1704
1705         /* socket type set on active connections - not set on passive */
1706         LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1707
1708         timeout = active ? *ksocknal_tunables.ksnd_timeout :
1709                             lnet_acceptor_timeout();
1710
1711         rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1712         if (rc != 0) {
1713                 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1714                         rc, HIPQUAD(conn->ksnc_ipaddr));
1715                 LASSERT (rc < 0);
1716                 return rc;
1717         }
1718
1719         if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1720             hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1721             hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1722                 /* Unexpected magic! */
1723                 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1724                         "%u.%u.%u.%u\n", __cpu_to_le32 (hello->kshm_magic),
1725                         LNET_PROTO_TCP_MAGIC,
1726                         HIPQUAD(conn->ksnc_ipaddr));
1727                 return -EPROTO;
1728         }
1729
1730         rc = libcfs_sock_read(sock, &hello->kshm_version,
1731                               sizeof(hello->kshm_version), timeout);
1732         if (rc != 0) {
1733                 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1734                         rc, HIPQUAD(conn->ksnc_ipaddr));
1735                 LASSERT (rc < 0);
1736                 return rc;
1737         }
1738
1739         proto = ksocknal_parse_proto_version(hello);
1740         if (proto == NULL) {
1741                 if (!active) {
1742                         /* unknown protocol from peer, tell peer my protocol */
1743                         conn->ksnc_proto = &ksocknal_protocol_v3x;
1744 #if SOCKNAL_VERSION_DEBUG
1745                         if (*ksocknal_tunables.ksnd_protocol == 2)
1746                                 conn->ksnc_proto = &ksocknal_protocol_v2x;
1747                         else if (*ksocknal_tunables.ksnd_protocol == 1)
1748                                 conn->ksnc_proto = &ksocknal_protocol_v1x;
1749 #endif
1750                         hello->kshm_nips = 0;
1751                         ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1752                 }
1753
1754                 CERROR ("Unknown protocol version (%d.x expected)"
1755                         " from %u.%u.%u.%u\n",
1756                         conn->ksnc_proto->pro_version,
1757                         HIPQUAD(conn->ksnc_ipaddr));
1758
1759                 return -EPROTO;
1760         }
1761
1762         proto_match = (conn->ksnc_proto == proto);
1763         conn->ksnc_proto = proto;
1764
1765         /* receive the rest of hello message anyway */
1766         rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1767         if (rc != 0) {
1768                 CERROR("Error %d reading or checking hello from from %u.%u.%u.%u\n",
1769                        rc, HIPQUAD(conn->ksnc_ipaddr));
1770                 LASSERT (rc < 0);
1771                 return rc;
1772         }
1773
1774         *incarnation = hello->kshm_src_incarnation;
1775
1776         if (hello->kshm_src_nid == LNET_NID_ANY) {
1777                 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1778                        "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1779                 return -EPROTO;
1780         }
1781
1782         if (!active &&
1783             conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1784                 /* Userspace NAL assigns peer process ID from socket */
1785                 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1786                 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1787         } else {
1788                 recv_id.nid = hello->kshm_src_nid;
1789                 recv_id.pid = hello->kshm_src_pid;
1790         }
1791
1792         if (!active) {
1793                 *peerid = recv_id;
1794
1795                 /* peer determines type */
1796                 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1797                 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1798                         CERROR ("Unexpected type %d from %s ip %u.%u.%u.%u\n",
1799                                 hello->kshm_ctype, libcfs_id2str(*peerid),
1800                                 HIPQUAD(conn->ksnc_ipaddr));
1801                         return -EPROTO;
1802                 }
1803
1804                 return 0;
1805         }
1806
1807         if (peerid->pid != recv_id.pid ||
1808             peerid->nid != recv_id.nid) {
1809                 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1810                                    " %u.%u.%u.%u, but they claimed they were "
1811                                    "%s; please check your Lustre "
1812                                    "configuration.\n",
1813                                    libcfs_id2str(*peerid),
1814                                    HIPQUAD(conn->ksnc_ipaddr),
1815                                    libcfs_id2str(recv_id));
1816                 return -EPROTO;
1817         }
1818
1819         if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1820                 /* Possible protocol mismatch or I lost the connection race */
1821                 return proto_match ? EALREADY : EPROTO;
1822         }
1823
1824         if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1825                 CERROR ("Mismatched types: me %d, %s ip %u.%u.%u.%u %d\n",
1826                         conn->ksnc_type, libcfs_id2str(*peerid), 
1827                         HIPQUAD(conn->ksnc_ipaddr),
1828                         hello->kshm_ctype);
1829                 return -EPROTO;
1830         }
1831
1832         return 0;
1833 }
1834
1835 int
1836 ksocknal_connect (ksock_route_t *route)
1837 {
1838         CFS_LIST_HEAD    (zombies);
1839         ksock_peer_t     *peer = route->ksnr_peer;
1840         int               type;
1841         int               wanted;
1842         cfs_socket_t     *sock;
1843         cfs_time_t        deadline;
1844         int               retry_later = 0;
1845         int               rc = 0;
1846
1847         deadline = cfs_time_add(cfs_time_current(),
1848                                 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1849
1850         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1851
1852         LASSERT (route->ksnr_scheduled);
1853         LASSERT (!route->ksnr_connecting);
1854
1855         route->ksnr_connecting = 1;
1856
1857         for (;;) {
1858                 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1859
1860                 /* stop connecting if peer/route got closed under me, or
1861                  * route got connected while queued */
1862                 if (peer->ksnp_closing || route->ksnr_deleted ||
1863                     wanted == 0) {
1864                         retry_later = 0;
1865                         break;
1866                 }
1867
1868                 /* reschedule if peer is connecting to me */
1869                 if (peer->ksnp_accepting > 0) {
1870                         CDEBUG(D_NET,
1871                                "peer %s(%d) already connecting to me, retry later.\n",
1872                                libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1873                         retry_later = 1;
1874                 }
1875
1876                 if (retry_later) /* needs reschedule */
1877                         break;
1878
1879                 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1880                         type = SOCKLND_CONN_ANY;
1881                 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1882                         type = SOCKLND_CONN_CONTROL;
1883                 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1884                         type = SOCKLND_CONN_BULK_IN;
1885                 } else {
1886                         LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1887                         type = SOCKLND_CONN_BULK_OUT;
1888                 }
1889
1890                 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1891
1892                 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1893                         rc = -ETIMEDOUT;
1894                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1895                                                    route->ksnr_ipaddr,
1896                                                    route->ksnr_port);
1897                         goto failed;
1898                 }
1899
1900                 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1901                                   route->ksnr_myipaddr,
1902                                   route->ksnr_ipaddr, route->ksnr_port);
1903                 if (rc != 0)
1904                         goto failed;
1905
1906                 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1907                 if (rc < 0) {
1908                         lnet_connect_console_error(rc, peer->ksnp_id.nid,
1909                                                    route->ksnr_ipaddr,
1910                                                    route->ksnr_port);
1911                         goto failed;
1912                 }
1913
1914                 /* A +ve RC means I have to retry because I lost the connection
1915                  * race or I have to renegotiate protocol version */
1916                 retry_later = (rc != 0);
1917                 if (retry_later)
1918                         CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1919                                libcfs_nid2str(peer->ksnp_id.nid));
1920
1921                 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1922         }
1923
1924         route->ksnr_scheduled = 0;
1925         route->ksnr_connecting = 0;
1926
1927         if (retry_later) {
1928                 /* re-queue for attention; this frees me up to handle
1929                  * the peer's incoming connection request */
1930
1931                 if (rc == EALREADY ||
1932                     (rc == 0 && peer->ksnp_accepting > 0)) {
1933                         /* We want to introduce a delay before next
1934                          * attempt to connect if we lost conn race,
1935                          * but the race is resolved quickly usually,
1936                          * so min_reconnectms should be good heuristic */
1937                         route->ksnr_retry_interval =
1938                                 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1939                         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1940                                                            route->ksnr_retry_interval);
1941                 }
1942
1943                 ksocknal_launch_connection_locked(route);
1944         }
1945
1946         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1947         return retry_later;
1948
1949  failed:
1950         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1951
1952         route->ksnr_scheduled = 0;
1953         route->ksnr_connecting = 0;
1954
1955         /* This is a retry rather than a new connection */
1956         route->ksnr_retry_interval *= 2;
1957         route->ksnr_retry_interval =
1958                 MAX(route->ksnr_retry_interval,
1959                     cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1960         route->ksnr_retry_interval =
1961                 MIN(route->ksnr_retry_interval,
1962                     cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1963
1964         LASSERT (route->ksnr_retry_interval != 0);
1965         route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1966                                            route->ksnr_retry_interval);
1967
1968         if (!cfs_list_empty(&peer->ksnp_tx_queue) &&
1969             peer->ksnp_accepting == 0 &&
1970             ksocknal_find_connecting_route_locked(peer) == NULL) {
1971                 ksock_conn_t *conn;
1972
1973                 /* ksnp_tx_queue is queued on a conn on successful
1974                  * connection for V1.x and V2.x */
1975                 if (!cfs_list_empty (&peer->ksnp_conns)) {
1976                         conn = cfs_list_entry(peer->ksnp_conns.next,
1977                                               ksock_conn_t, ksnc_list);
1978                         LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1979                 }
1980
1981                 /* take all the blocked packets while I've got the lock and
1982                  * complete below... */
1983                 cfs_list_splice_init(&peer->ksnp_tx_queue, &zombies);
1984         }
1985
1986 #if 0           /* irrelevent with only eager routes */
1987         if (!route->ksnr_deleted) {
1988                 /* make this route least-favourite for re-selection */
1989                 cfs_list_del(&route->ksnr_list);
1990                 cfs_list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1991         }
1992 #endif
1993         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1994
1995         ksocknal_peer_failed(peer);
1996         ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
1997         return 0;
1998 }
1999
2000 /* Go through connd_routes queue looking for a route that
2001    we can process right now */
2002 static ksock_route_t *
2003 ksocknal_connd_get_route_locked(signed long *timeout_p)
2004 {
2005         ksock_route_t *route;
2006         cfs_time_t     now;
2007
2008         /* Only handle an outgoing connection request if there
2009          * is someone left to handle incoming connections */
2010         if ((ksocknal_data.ksnd_connd_connecting + 1) >=
2011             *ksocknal_tunables.ksnd_nconnds)
2012                 return NULL;
2013
2014         now = cfs_time_current();
2015
2016         /* connd_routes can contain both pending and ordinary routes */
2017         cfs_list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2018                                  ksnr_connd_list) {
2019
2020                 if (route->ksnr_retry_interval == 0 ||
2021                     cfs_time_aftereq(now, route->ksnr_timeout))
2022                         return route;
2023
2024                 if (*timeout_p == CFS_MAX_SCHEDULE_TIMEOUT ||
2025                     (int)*timeout_p > (int)(route->ksnr_timeout - now))
2026                         *timeout_p = (int)(route->ksnr_timeout - now);
2027         }
2028
2029         return NULL;
2030 }
2031
2032 int
2033 ksocknal_connd (void *arg)
2034 {
2035         long               id = (long)(long_ptr_t)arg;
2036         char               name[16];
2037         ksock_connreq_t   *cr;
2038         ksock_route_t     *route;
2039         cfs_waitlink_t     wait;
2040         signed long        timeout;
2041         int                nloops = 0;
2042         int                cons_retry = 0;
2043         int                dropped_lock;
2044
2045         snprintf (name, sizeof (name), "socknal_cd%02ld", id);
2046         cfs_daemonize (name);
2047         cfs_block_allsigs ();
2048
2049         cfs_waitlink_init (&wait);
2050
2051         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2052
2053         while (!ksocknal_data.ksnd_shuttingdown) {
2054
2055                 dropped_lock = 0;
2056
2057                 if (!cfs_list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2058                         /* Connection accepted by the listener */
2059                         cr = cfs_list_entry(ksocknal_data.ksnd_connd_connreqs. \
2060                                             next, ksock_connreq_t, ksncr_list);
2061
2062                         cfs_list_del(&cr->ksncr_list);
2063                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2064                         dropped_lock = 1;
2065
2066                         ksocknal_create_conn(cr->ksncr_ni, NULL,
2067                                              cr->ksncr_sock, SOCKLND_CONN_NONE);
2068                         lnet_ni_decref(cr->ksncr_ni);
2069                         LIBCFS_FREE(cr, sizeof(*cr));
2070
2071                         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2072                 }
2073
2074                 /* Sleep till explicit wake_up if no pending routes present */
2075                 timeout = CFS_MAX_SCHEDULE_TIMEOUT;
2076
2077                 /* Connection request */
2078                 route = ksocknal_connd_get_route_locked(&timeout);
2079
2080                 if (route != NULL) {
2081                         cfs_list_del (&route->ksnr_connd_list);
2082                         ksocknal_data.ksnd_connd_connecting++;
2083                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2084                         dropped_lock = 1;
2085
2086                         if (ksocknal_connect(route)) {
2087                                 /* consecutive retry */
2088                                 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2089                                         CWARN("massive consecutive "
2090                                               "re-connecting to %u.%u.%u.%u\n",
2091                                               HIPQUAD(route->ksnr_ipaddr));
2092                                         cons_retry = 0;
2093                                 }
2094                         } else {
2095                                 cons_retry = 0;
2096                         }
2097
2098                         ksocknal_route_decref(route);
2099
2100                         cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2101                         ksocknal_data.ksnd_connd_connecting--;
2102                 }
2103
2104                 if (dropped_lock) {
2105                         if (++nloops < SOCKNAL_RESCHED)
2106                                 continue;
2107                         cfs_spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2108                         nloops = 0;
2109                         cfs_cond_resched();
2110                         cfs_spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2111                         continue;
2112                 }
2113
2114                 /* Nothing to do for 'timeout'  */
2115                 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2116                 cfs_waitq_add_exclusive (&ksocknal_data.ksnd_connd_waitq,
2117                                          &wait);
2118                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2119
2120                 nloops = 0;
2121                 cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE, timeout);
2122
2123                 cfs_set_current_state (CFS_TASK_RUNNING);
2124                 cfs_waitq_del (&ksocknal_data.ksnd_connd_waitq, &wait);
2125                 cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
2126         }
2127
2128         cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
2129
2130         ksocknal_thread_fini ();
2131         return (0);
2132 }
2133
2134 ksock_conn_t *
2135 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2136 {
2137         /* We're called with a shared lock on ksnd_global_lock */
2138         ksock_conn_t      *conn;
2139         cfs_list_t        *ctmp;
2140
2141         cfs_list_for_each (ctmp, &peer->ksnp_conns) {
2142                 int     error;
2143                 conn = cfs_list_entry (ctmp, ksock_conn_t, ksnc_list);
2144
2145                 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2146                 LASSERT (!conn->ksnc_closing);
2147
2148                 /* SOCK_ERROR will reset error code of socket in
2149                  * some platform (like Darwin8.x) */
2150                 error = libcfs_sock_error(conn->ksnc_sock);
2151                 if (error != 0) {
2152                         ksocknal_conn_addref(conn);
2153
2154                         switch (error) {
2155                         case ECONNRESET:
2156                                 CDEBUG(D_NETERROR, "A connection with %s "
2157                                        "(%u.%u.%u.%u:%d) was reset; "
2158                                        "it may have rebooted.\n",
2159                                        libcfs_id2str(peer->ksnp_id),
2160                                        HIPQUAD(conn->ksnc_ipaddr),
2161                                        conn->ksnc_port);
2162                                 break;
2163                         case ETIMEDOUT:
2164                                 CDEBUG(D_NETERROR, "A connection with %s "
2165                                        "(%u.%u.%u.%u:%d) timed out; the "
2166                                        "network or node may be down.\n",
2167                                        libcfs_id2str(peer->ksnp_id),
2168                                        HIPQUAD(conn->ksnc_ipaddr),
2169                                        conn->ksnc_port);
2170                                 break;
2171                         default:
2172                                 CDEBUG(D_NETERROR, "An unexpected network error %d "
2173                                        "occurred with %s "
2174                                        "(%u.%u.%u.%u:%d\n", error,
2175                                        libcfs_id2str(peer->ksnp_id),
2176                                        HIPQUAD(conn->ksnc_ipaddr),
2177                                        conn->ksnc_port);
2178                                 break;
2179                         }
2180
2181                         return (conn);
2182                 }
2183
2184                 if (conn->ksnc_rx_started &&
2185                     cfs_time_aftereq(cfs_time_current(),
2186                                      conn->ksnc_rx_deadline)) {
2187                         /* Timed out incomplete incoming message */
2188                         ksocknal_conn_addref(conn);
2189                         CDEBUG(D_NETERROR, "Timeout receiving from %s "
2190                                "(%u.%u.%u.%u:%d), state %d wanted %d left %d\n",
2191                                libcfs_id2str(peer->ksnp_id),
2192                                HIPQUAD(conn->ksnc_ipaddr),
2193                                conn->ksnc_port,
2194                                conn->ksnc_rx_state,
2195                                conn->ksnc_rx_nob_wanted,
2196                                conn->ksnc_rx_nob_left);
2197                         return (conn);
2198                 }
2199
2200                 if ((!cfs_list_empty(&conn->ksnc_tx_queue) ||
2201                      libcfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2202                     cfs_time_aftereq(cfs_time_current(),
2203                                      conn->ksnc_tx_deadline)) {
2204                         /* Timed out messages queued for sending or
2205                          * buffered in the socket's send buffer */
2206                         ksocknal_conn_addref(conn);
2207                         CDEBUG(D_NETERROR, "Timeout sending data to %s "
2208                                "(%u.%u.%u.%u:%d) the network or that "
2209                                "node may be down.\n",
2210                                libcfs_id2str(peer->ksnp_id),
2211                                HIPQUAD(conn->ksnc_ipaddr),
2212                                conn->ksnc_port);
2213                         return (conn);
2214                 }
2215         }
2216
2217         return (NULL);
2218 }
2219
2220 static inline void
2221 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2222 {
2223         ksock_tx_t        *tx;
2224         CFS_LIST_HEAD      (stale_txs);
2225
2226         cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
2227
2228         while (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2229                 tx = cfs_list_entry (peer->ksnp_tx_queue.next,
2230                                      ksock_tx_t, tx_list);
2231
2232                 if (!cfs_time_aftereq(cfs_time_current(),
2233                                       tx->tx_deadline))
2234                         break;
2235
2236                 cfs_list_del (&tx->tx_list);
2237                 cfs_list_add_tail (&tx->tx_list, &stale_txs);
2238         }
2239
2240         cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
2241
2242         ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2243 }
2244
2245 int
2246 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2247 {
2248         ksock_sched_t  *sched;
2249         ksock_conn_t   *conn;
2250         ksock_tx_t     *tx;
2251
2252         if (cfs_list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2253                 return 0;
2254
2255         if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2256                 return 0;
2257
2258         if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2259             cfs_time_before(cfs_time_current(),
2260                             cfs_time_add(peer->ksnp_last_alive,
2261                                          cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2262                 return 0;
2263
2264         if (cfs_time_before(cfs_time_current(),
2265                             peer->ksnp_send_keepalive))
2266                 return 0;
2267
2268         /* retry 10 secs later, so we wouldn't put pressure
2269          * on this peer if we failed to send keepalive this time */
2270         peer->ksnp_send_keepalive = cfs_time_shift(10);
2271
2272         conn = ksocknal_find_conn_locked(peer, NULL, 1);
2273         if (conn != NULL) {
2274                 sched = conn->ksnc_scheduler;
2275
2276                 cfs_spin_lock_bh (&sched->kss_lock);
2277                 if (!cfs_list_empty(&conn->ksnc_tx_queue)) {
2278                         cfs_spin_unlock_bh(&sched->kss_lock);
2279                         /* there is an queued ACK, don't need keepalive */
2280                         return 0;
2281                 }
2282
2283                 cfs_spin_unlock_bh(&sched->kss_lock);
2284         }
2285
2286         cfs_read_unlock(&ksocknal_data.ksnd_global_lock);
2287
2288         /* cookie = 1 is reserved for keepalive PING */
2289         tx = ksocknal_alloc_tx_noop(1, 1);
2290         if (tx == NULL) {
2291                 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2292                 return -ENOMEM;
2293         }
2294
2295         if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2296                 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2297                 return 1;
2298         }
2299
2300         ksocknal_free_tx(tx);
2301         cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2302
2303         return -EIO;
2304 }
2305
2306
2307 void
2308 ksocknal_check_peer_timeouts (int idx)
2309 {
2310         cfs_list_t       *peers = &ksocknal_data.ksnd_peers[idx];
2311         ksock_peer_t     *peer;
2312         ksock_conn_t     *conn;
2313         ksock_tx_t       *tx;
2314
2315  again:
2316         /* NB. We expect to have a look at all the peers and not find any
2317          * connections to time out, so we just use a shared lock while we
2318          * take a look... */
2319         cfs_read_lock (&ksocknal_data.ksnd_global_lock);
2320
2321         cfs_list_for_each_entry_typed(peer, peers, ksock_peer_t, ksnp_list) {
2322                 cfs_time_t  deadline = 0;
2323                 int         resid = 0;
2324                 int         n     = 0;
2325
2326                 if (ksocknal_send_keepalive_locked(peer) != 0) {
2327                         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2328                         goto again;
2329                 }
2330
2331                 conn = ksocknal_find_timed_out_conn (peer);
2332
2333                 if (conn != NULL) {
2334                         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2335
2336                         ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2337
2338                         /* NB we won't find this one again, but we can't
2339                          * just proceed with the next peer, since we dropped
2340                          * ksnd_global_lock and it might be dead already! */
2341                         ksocknal_conn_decref(conn);
2342                         goto again;
2343                 }
2344
2345                 /* we can't process stale txs right here because we're
2346                  * holding only shared lock */
2347                 if (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2348                         ksock_tx_t *tx =
2349                                 cfs_list_entry (peer->ksnp_tx_queue.next,
2350                                                 ksock_tx_t, tx_list);
2351
2352                         if (cfs_time_aftereq(cfs_time_current(),
2353                                              tx->tx_deadline)) {
2354
2355                                 ksocknal_peer_addref(peer);
2356                                 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2357
2358                                 ksocknal_flush_stale_txs(peer);
2359
2360                                 ksocknal_peer_decref(peer);
2361                                 goto again;
2362                         }
2363                 }
2364
2365                 if (cfs_list_empty(&peer->ksnp_zc_req_list))
2366                         continue;
2367
2368                 cfs_spin_lock(&peer->ksnp_lock);
2369                 cfs_list_for_each_entry_typed(tx, &peer->ksnp_zc_req_list,
2370                                               ksock_tx_t, tx_zc_list) {
2371                         if (!cfs_time_aftereq(cfs_time_current(),
2372                                               tx->tx_deadline))
2373                                 break;
2374                         /* ignore the TX if connection is being closed */
2375                         if (tx->tx_conn->ksnc_closing)
2376                                 continue;
2377                         n++;
2378                 }
2379
2380                 if (n == 0) {
2381                         cfs_spin_unlock(&peer->ksnp_lock);
2382                         continue;
2383                 }
2384
2385                 tx = cfs_list_entry(peer->ksnp_zc_req_list.next,
2386                                     ksock_tx_t, tx_zc_list);
2387                 deadline = tx->tx_deadline;
2388                 resid    = tx->tx_resid;
2389                 conn     = tx->tx_conn;
2390                 ksocknal_conn_addref(conn);
2391
2392                 cfs_spin_unlock(&peer->ksnp_lock);
2393                 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2394
2395                 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2396                        "oldest(%p) timed out %ld secs ago, "
2397                        "resid: %d, wmem: %d\n",
2398                        n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2399                        cfs_duration_sec(cfs_time_current() - deadline),
2400                        resid, libcfs_sock_wmem_queued(conn->ksnc_sock));
2401
2402                 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2403                 ksocknal_conn_decref(conn);
2404                 goto again;
2405         }
2406
2407         cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2408 }
2409
2410 int
2411 ksocknal_reaper (void *arg)
2412 {
2413         cfs_waitlink_t     wait;
2414         ksock_conn_t      *conn;
2415         ksock_sched_t     *sched;
2416         cfs_list_t         enomem_conns;
2417         int                nenomem_conns;
2418         cfs_duration_t     timeout;
2419         int                i;
2420         int                peer_index = 0;
2421         cfs_time_t         deadline = cfs_time_current();
2422
2423         cfs_daemonize ("socknal_reaper");
2424         cfs_block_allsigs ();
2425
2426         CFS_INIT_LIST_HEAD(&enomem_conns);
2427         cfs_waitlink_init (&wait);
2428
2429         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2430
2431         while (!ksocknal_data.ksnd_shuttingdown) {
2432
2433                 if (!cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2434                         conn = cfs_list_entry (ksocknal_data. \
2435                                                ksnd_deathrow_conns.next,
2436                                                ksock_conn_t, ksnc_list);
2437                         cfs_list_del (&conn->ksnc_list);
2438
2439                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2440
2441                         ksocknal_terminate_conn (conn);
2442                         ksocknal_conn_decref(conn);
2443
2444                         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2445                         continue;
2446                 }
2447
2448                 if (!cfs_list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2449                         conn = cfs_list_entry (ksocknal_data.ksnd_zombie_conns.\
2450                                                next, ksock_conn_t, ksnc_list);
2451                         cfs_list_del (&conn->ksnc_list);
2452
2453                         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2454
2455                         ksocknal_destroy_conn (conn);
2456
2457                         cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2458                         continue;
2459                 }
2460
2461                 if (!cfs_list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2462                         cfs_list_add(&enomem_conns,
2463                                      &ksocknal_data.ksnd_enomem_conns);
2464                         cfs_list_del_init(&ksocknal_data.ksnd_enomem_conns);
2465                 }
2466
2467                 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2468
2469                 /* reschedule all the connections that stalled with ENOMEM... */
2470                 nenomem_conns = 0;
2471                 while (!cfs_list_empty (&enomem_conns)) {
2472                         conn = cfs_list_entry (enomem_conns.next,
2473                                                ksock_conn_t, ksnc_tx_list);
2474                         cfs_list_del (&conn->ksnc_tx_list);
2475
2476                         sched = conn->ksnc_scheduler;
2477
2478                         cfs_spin_lock_bh (&sched->kss_lock);
2479
2480                         LASSERT (conn->ksnc_tx_scheduled);
2481                         conn->ksnc_tx_ready = 1;
2482                         cfs_list_add_tail(&conn->ksnc_tx_list,
2483                                           &sched->kss_tx_conns);
2484                         cfs_waitq_signal (&sched->kss_waitq);
2485
2486                         cfs_spin_unlock_bh (&sched->kss_lock);
2487                         nenomem_conns++;
2488                 }
2489
2490                 /* careful with the jiffy wrap... */
2491                 while ((timeout = cfs_time_sub(deadline,
2492                                                cfs_time_current())) <= 0) {
2493                         const int n = 4;
2494                         const int p = 1;
2495                         int       chunk = ksocknal_data.ksnd_peer_hash_size;
2496
2497                         /* Time to check for timeouts on a few more peers: I do
2498                          * checks every 'p' seconds on a proportion of the peer
2499                          * table and I need to check every connection 'n' times
2500                          * within a timeout interval, to ensure I detect a
2501                          * timeout on any connection within (n+1)/n times the
2502                          * timeout interval. */
2503
2504                         if (*ksocknal_tunables.ksnd_timeout > n * p)
2505                                 chunk = (chunk * n * p) /
2506                                         *ksocknal_tunables.ksnd_timeout;
2507                         if (chunk == 0)
2508                                 chunk = 1;
2509
2510                         for (i = 0; i < chunk; i++) {
2511                                 ksocknal_check_peer_timeouts (peer_index);
2512                                 peer_index = (peer_index + 1) %
2513                                              ksocknal_data.ksnd_peer_hash_size;
2514                         }
2515
2516                         deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2517                 }
2518
2519                 if (nenomem_conns != 0) {
2520                         /* Reduce my timeout if I rescheduled ENOMEM conns.
2521                          * This also prevents me getting woken immediately
2522                          * if any go back on my enomem list. */
2523                         timeout = SOCKNAL_ENOMEM_RETRY;
2524                 }
2525                 ksocknal_data.ksnd_reaper_waketime =
2526                         cfs_time_add(cfs_time_current(), timeout);
2527
2528                 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2529                 cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
2530
2531                 if (!ksocknal_data.ksnd_shuttingdown &&
2532                     cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2533                     cfs_list_empty (&ksocknal_data.ksnd_zombie_conns))
2534                         cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE,
2535                                              timeout);
2536
2537                 cfs_set_current_state (CFS_TASK_RUNNING);
2538                 cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
2539
2540                 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2541         }
2542
2543         cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2544
2545         ksocknal_thread_fini ();
2546         return (0);
2547 }