1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
6 * Copyright (c) 2011, Whamcloud, Inc.
8 * Author: Zach Brown <zab@zabbo.net>
9 * Author: Peter J. Braam <braam@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
11 * Author: Eric Barton <eric@bartonsoftware.com>
13 * This file is part of Portals, http://www.sf.net/projects/sandiaportals/
15 * Portals is free software; you can redistribute it and/or
16 * modify it under the terms of version 2 of the GNU General Public
17 * License as published by the Free Software Foundation.
19 * Portals is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * GNU General Public License for more details.
24 * You should have received a copy of the GNU General Public License
25 * along with Portals; if not, write to the Free Software
26 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32 ksocknal_alloc_tx(int type, int size)
34 ksock_tx_t *tx = NULL;
36 if (type == KSOCK_MSG_NOOP) {
37 LASSERT (size == KSOCK_NOOP_TX_SIZE);
39 /* searching for a noop tx in free list */
40 cfs_spin_lock(&ksocknal_data.ksnd_tx_lock);
42 if (!cfs_list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
43 tx = cfs_list_entry(ksocknal_data.ksnd_idle_noop_txs. \
44 next, ksock_tx_t, tx_list);
45 LASSERT(tx->tx_desc_size == size);
46 cfs_list_del(&tx->tx_list);
49 cfs_spin_unlock(&ksocknal_data.ksnd_tx_lock);
53 LIBCFS_ALLOC(tx, size);
58 cfs_atomic_set(&tx->tx_refcount, 1);
59 tx->tx_zc_aborted = 0;
60 tx->tx_zc_capable = 0;
61 tx->tx_zc_checked = 0;
62 tx->tx_desc_size = size;
64 cfs_atomic_inc(&ksocknal_data.ksnd_nactive_txs);
70 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
74 tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
76 CERROR("Can't allocate noop tx desc\n");
81 tx->tx_lnetmsg = NULL;
84 tx->tx_iov = tx->tx_frags.virt.iov;
86 tx->tx_nonblk = nonblk;
88 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
89 tx->tx_msg.ksm_zc_cookies[1] = cookie;
96 ksocknal_free_tx (ksock_tx_t *tx)
98 cfs_atomic_dec(&ksocknal_data.ksnd_nactive_txs);
100 if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
102 cfs_spin_lock(&ksocknal_data.ksnd_tx_lock);
104 cfs_list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
106 cfs_spin_unlock(&ksocknal_data.ksnd_tx_lock);
108 LIBCFS_FREE(tx, tx->tx_desc_size);
113 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
115 struct iovec *iov = tx->tx_iov;
119 LASSERT (tx->tx_niov > 0);
121 /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
122 rc = ksocknal_lib_send_iov(conn, tx);
124 if (rc <= 0) /* sent nothing? */
128 LASSERT (nob <= tx->tx_resid);
133 LASSERT (tx->tx_niov > 0);
135 if (nob < (int) iov->iov_len) {
136 iov->iov_base = (void *)((char *)iov->iov_base + nob);
150 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
152 lnet_kiov_t *kiov = tx->tx_kiov;
156 LASSERT (tx->tx_niov == 0);
157 LASSERT (tx->tx_nkiov > 0);
159 /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
160 rc = ksocknal_lib_send_kiov(conn, tx);
162 if (rc <= 0) /* sent nothing? */
166 LASSERT (nob <= tx->tx_resid);
171 LASSERT(tx->tx_nkiov > 0);
173 if (nob < (int)kiov->kiov_len) {
174 kiov->kiov_offset += nob;
175 kiov->kiov_len -= nob;
179 nob -= (int)kiov->kiov_len;
180 tx->tx_kiov = ++kiov;
188 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
193 if (ksocknal_data.ksnd_stall_tx != 0) {
194 cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
197 LASSERT (tx->tx_resid != 0);
199 rc = ksocknal_connsock_addref(conn);
201 LASSERT (conn->ksnc_closing);
206 if (ksocknal_data.ksnd_enomem_tx > 0) {
208 ksocknal_data.ksnd_enomem_tx--;
210 } else if (tx->tx_niov != 0) {
211 rc = ksocknal_send_iov (conn, tx);
213 rc = ksocknal_send_kiov (conn, tx);
216 bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
217 if (rc > 0) /* sent something? */
218 conn->ksnc_tx_bufnob += rc; /* account it */
220 if (bufnob < conn->ksnc_tx_bufnob) {
221 /* allocated send buffer bytes < computed; infer
222 * something got ACKed */
223 conn->ksnc_tx_deadline =
224 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
225 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
226 conn->ksnc_tx_bufnob = bufnob;
230 if (rc <= 0) { /* Didn't write anything? */
232 if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
235 /* Check if EAGAIN is due to memory pressure */
236 if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
242 /* socket's wmem_queued now includes 'rc' bytes */
243 cfs_atomic_sub (rc, &conn->ksnc_tx_nob);
246 } while (tx->tx_resid != 0);
248 ksocknal_connsock_decref(conn);
253 ksocknal_recv_iov (ksock_conn_t *conn)
255 struct iovec *iov = conn->ksnc_rx_iov;
259 LASSERT (conn->ksnc_rx_niov > 0);
261 /* Never touch conn->ksnc_rx_iov or change connection
262 * status inside ksocknal_lib_recv_iov */
263 rc = ksocknal_lib_recv_iov(conn);
268 /* received something... */
271 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
272 conn->ksnc_rx_deadline =
273 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
274 cfs_mb(); /* order with setting rx_started */
275 conn->ksnc_rx_started = 1;
277 conn->ksnc_rx_nob_wanted -= nob;
278 conn->ksnc_rx_nob_left -= nob;
281 LASSERT (conn->ksnc_rx_niov > 0);
283 if (nob < (int)iov->iov_len) {
285 iov->iov_base = (void *)((char *)iov->iov_base + nob);
290 conn->ksnc_rx_iov = ++iov;
291 conn->ksnc_rx_niov--;
298 ksocknal_recv_kiov (ksock_conn_t *conn)
300 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
303 LASSERT (conn->ksnc_rx_nkiov > 0);
305 /* Never touch conn->ksnc_rx_kiov or change connection
306 * status inside ksocknal_lib_recv_iov */
307 rc = ksocknal_lib_recv_kiov(conn);
312 /* received something... */
315 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
316 conn->ksnc_rx_deadline =
317 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
318 cfs_mb(); /* order with setting rx_started */
319 conn->ksnc_rx_started = 1;
321 conn->ksnc_rx_nob_wanted -= nob;
322 conn->ksnc_rx_nob_left -= nob;
325 LASSERT (conn->ksnc_rx_nkiov > 0);
327 if (nob < (int) kiov->kiov_len) {
328 kiov->kiov_offset += nob;
329 kiov->kiov_len -= nob;
333 nob -= kiov->kiov_len;
334 conn->ksnc_rx_kiov = ++kiov;
335 conn->ksnc_rx_nkiov--;
342 ksocknal_receive (ksock_conn_t *conn)
344 /* Return 1 on success, 0 on EOF, < 0 on error.
345 * Caller checks ksnc_rx_nob_wanted to determine
346 * progress/completion. */
350 if (ksocknal_data.ksnd_stall_rx != 0) {
351 cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
354 rc = ksocknal_connsock_addref(conn);
356 LASSERT (conn->ksnc_closing);
361 if (conn->ksnc_rx_niov != 0)
362 rc = ksocknal_recv_iov (conn);
364 rc = ksocknal_recv_kiov (conn);
367 /* error/EOF or partial receive */
370 } else if (rc == 0 && conn->ksnc_rx_started) {
371 /* EOF in the middle of a message */
377 /* Completed a fragment */
379 if (conn->ksnc_rx_nob_wanted == 0) {
385 ksocknal_connsock_decref(conn);
390 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
392 lnet_msg_t *lnetmsg = tx->tx_lnetmsg;
393 int rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO;
396 LASSERT(ni != NULL || tx->tx_conn != NULL);
398 if (tx->tx_conn != NULL)
399 ksocknal_conn_decref(tx->tx_conn);
401 if (ni == NULL && tx->tx_conn != NULL)
402 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
404 ksocknal_free_tx (tx);
405 if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
406 lnet_finalize (ni, lnetmsg, rc);
412 ksocknal_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int error)
416 while (!cfs_list_empty (txlist)) {
417 tx = cfs_list_entry (txlist->next, ksock_tx_t, tx_list);
419 if (error && tx->tx_lnetmsg != NULL) {
420 CNETERR("Deleting packet type %d len %d %s->%s\n",
421 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
422 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
423 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
424 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
426 CNETERR("Deleting noop packet\n");
429 cfs_list_del (&tx->tx_list);
431 LASSERT (cfs_atomic_read(&tx->tx_refcount) == 1);
432 ksocknal_tx_done (ni, tx);
437 ksocknal_check_zc_req(ksock_tx_t *tx)
439 ksock_conn_t *conn = tx->tx_conn;
440 ksock_peer_t *peer = conn->ksnc_peer;
442 /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
443 * to ksnp_zc_req_list if some fragment of this message should be sent
444 * zero-copy. Our peer will send an ACK containing this cookie when
445 * she has received this message to tell us we can signal completion.
446 * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
447 * ksnp_zc_req_list. */
448 LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
449 LASSERT (tx->tx_zc_capable);
451 tx->tx_zc_checked = 1;
453 if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
454 !conn->ksnc_zc_capable)
457 /* assign cookie and queue tx to pending list, it will be released when
458 * a matching ack is received. See ksocknal_handle_zcack() */
460 ksocknal_tx_addref(tx);
462 cfs_spin_lock(&peer->ksnp_lock);
464 /* ZC_REQ is going to be pinned to the peer */
466 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
468 LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
470 tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
472 if (peer->ksnp_zc_next_cookie == 0)
473 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
475 cfs_list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
477 cfs_spin_unlock(&peer->ksnp_lock);
481 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
483 ksock_peer_t *peer = tx->tx_conn->ksnc_peer;
485 LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
486 LASSERT (tx->tx_zc_capable);
488 tx->tx_zc_checked = 0;
490 cfs_spin_lock(&peer->ksnp_lock);
492 if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
493 /* Not waiting for an ACK */
494 cfs_spin_unlock(&peer->ksnp_lock);
498 tx->tx_msg.ksm_zc_cookies[0] = 0;
499 cfs_list_del(&tx->tx_zc_list);
501 cfs_spin_unlock(&peer->ksnp_lock);
503 ksocknal_tx_decref(tx);
507 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
511 if (tx->tx_zc_capable && !tx->tx_zc_checked)
512 ksocknal_check_zc_req(tx);
514 rc = ksocknal_transmit (conn, tx);
516 CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
518 if (tx->tx_resid == 0) {
519 /* Sent everything OK */
531 counter++; /* exponential backoff warnings */
532 if ((counter & (-counter)) == counter)
533 CWARN("%u ENOMEM tx %p (%u allocated)\n",
534 counter, conn, cfs_atomic_read(&libcfs_kmemory));
536 /* Queue on ksnd_enomem_conns for retry after a timeout */
537 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
539 /* enomem list takes over scheduler's ref... */
540 LASSERT (conn->ksnc_tx_scheduled);
541 cfs_list_add_tail(&conn->ksnc_tx_list,
542 &ksocknal_data.ksnd_enomem_conns);
543 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
544 SOCKNAL_ENOMEM_RETRY),
545 ksocknal_data.ksnd_reaper_waketime))
546 cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
548 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
555 if (!conn->ksnc_closing) {
558 LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection "
559 "while we were sending data; it may have "
561 HIPQUAD(conn->ksnc_ipaddr));
564 LCONSOLE_WARN("There was an unexpected network error "
565 "while writing to %u.%u.%u.%u: %d.\n",
566 HIPQUAD(conn->ksnc_ipaddr), rc);
569 CDEBUG(D_NET, "[%p] Error %d on write to %s"
570 " ip %d.%d.%d.%d:%d\n", conn, rc,
571 libcfs_id2str(conn->ksnc_peer->ksnp_id),
572 HIPQUAD(conn->ksnc_ipaddr),
576 if (tx->tx_zc_checked)
577 ksocknal_uncheck_zc_req(tx);
579 /* it's not an error if conn is being closed */
580 ksocknal_close_conn_and_siblings (conn,
581 (conn->ksnc_closing) ? 0 : rc);
587 ksocknal_launch_connection_locked (ksock_route_t *route)
590 /* called holding write lock on ksnd_global_lock */
592 LASSERT (!route->ksnr_scheduled);
593 LASSERT (!route->ksnr_connecting);
594 LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
596 route->ksnr_scheduled = 1; /* scheduling conn for connd */
597 ksocknal_route_addref(route); /* extra ref for connd */
599 cfs_spin_lock_bh (&ksocknal_data.ksnd_connd_lock);
601 cfs_list_add_tail (&route->ksnr_connd_list,
602 &ksocknal_data.ksnd_connd_routes);
603 cfs_waitq_signal (&ksocknal_data.ksnd_connd_waitq);
605 cfs_spin_unlock_bh (&ksocknal_data.ksnd_connd_lock);
609 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
611 ksock_route_t *route;
613 /* called holding write lock on ksnd_global_lock */
615 /* launch any/all connections that need it */
616 route = ksocknal_find_connectable_route_locked(peer);
620 ksocknal_launch_connection_locked(route);
625 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
629 ksock_conn_t *typed = NULL;
630 ksock_conn_t *fallback = NULL;
634 cfs_list_for_each (tmp, &peer->ksnp_conns) {
635 ksock_conn_t *c = cfs_list_entry(tmp, ksock_conn_t, ksnc_list);
636 int nob = cfs_atomic_read(&c->ksnc_tx_nob) +
637 libcfs_sock_wmem_queued(c->ksnc_sock);
640 LASSERT (!c->ksnc_closing);
641 LASSERT (c->ksnc_proto != NULL &&
642 c->ksnc_proto->pro_match_tx != NULL);
644 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
649 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
652 case SOCKNAL_MATCH_YES: /* typed connection */
653 if (typed == NULL || tnob > nob ||
654 (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
655 cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
661 case SOCKNAL_MATCH_MAY: /* fallback connection */
662 if (fallback == NULL || fnob > nob ||
663 (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
664 cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
672 /* prefer the typed selection */
673 conn = (typed != NULL) ? typed : fallback;
676 conn->ksnc_tx_last_post = cfs_time_current();
682 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
684 conn->ksnc_proto->pro_pack(tx);
686 cfs_atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
687 ksocknal_conn_addref(conn); /* +1 ref for tx */
692 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
694 ksock_sched_t *sched = conn->ksnc_scheduler;
695 ksock_msg_t *msg = &tx->tx_msg;
696 ksock_tx_t *ztx = NULL;
699 /* called holding global lock (read or irq-write) and caller may
700 * not have dropped this lock between finding conn and calling me,
701 * so we don't need the {get,put}connsock dance to deref
703 LASSERT(!conn->ksnc_closing);
705 CDEBUG (D_NET, "Sending to %s ip %d.%d.%d.%d:%d\n",
706 libcfs_id2str(conn->ksnc_peer->ksnp_id),
707 HIPQUAD(conn->ksnc_ipaddr),
710 ksocknal_tx_prep(conn, tx);
712 /* Ensure the frags we've been given EXACTLY match the number of
713 * bytes we want to send. Many TCP/IP stacks disregard any total
714 * size parameters passed to them and just look at the frags.
716 * We always expect at least 1 mapped fragment containing the
717 * complete ksocknal message header. */
718 LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
719 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
720 (unsigned int)tx->tx_nob);
721 LASSERT (tx->tx_niov >= 1);
722 LASSERT (tx->tx_resid == tx->tx_nob);
724 CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
725 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
727 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
730 * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
731 * but they're used inside spinlocks a lot.
733 bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
734 cfs_spin_lock_bh (&sched->kss_lock);
736 if (cfs_list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
737 /* First packet starts the timeout */
738 conn->ksnc_tx_deadline =
739 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
740 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
741 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
742 conn->ksnc_tx_bufnob = 0;
743 cfs_mb(); /* order with adding to tx_queue */
746 if (msg->ksm_type == KSOCK_MSG_NOOP) {
747 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
748 * on a normal packet so I don't need to send it */
749 LASSERT (msg->ksm_zc_cookies[1] != 0);
750 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
752 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
753 ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
756 /* It's a normal packet - can it piggback a noop zc-ack that
757 * has been queued already? */
758 LASSERT (msg->ksm_zc_cookies[1] == 0);
759 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
761 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
762 /* ztx will be released later */
766 cfs_atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
767 cfs_list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
770 if (conn->ksnc_tx_ready && /* able to send */
771 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
772 /* +1 ref for scheduler */
773 ksocknal_conn_addref(conn);
774 cfs_list_add_tail (&conn->ksnc_tx_list,
775 &sched->kss_tx_conns);
776 conn->ksnc_tx_scheduled = 1;
777 cfs_waitq_signal (&sched->kss_waitq);
780 cfs_spin_unlock_bh (&sched->kss_lock);
785 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
787 cfs_time_t now = cfs_time_current();
789 ksock_route_t *route;
791 cfs_list_for_each (tmp, &peer->ksnp_routes) {
792 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
794 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
796 if (route->ksnr_scheduled) /* connections being established */
799 /* all route types connected ? */
800 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
803 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
804 cfs_time_aftereq(now, route->ksnr_timeout))) {
806 "Too soon to retry route %u.%u.%u.%u "
807 "(cnted %d, interval %ld, %ld secs later)\n",
808 HIPQUAD(route->ksnr_ipaddr),
809 route->ksnr_connected,
810 route->ksnr_retry_interval,
811 cfs_duration_sec(route->ksnr_timeout - now));
822 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
825 ksock_route_t *route;
827 cfs_list_for_each (tmp, &peer->ksnp_routes) {
828 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
830 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
832 if (route->ksnr_scheduled)
840 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
844 cfs_rwlock_t *g_lock;
848 LASSERT (tx->tx_conn == NULL);
850 g_lock = &ksocknal_data.ksnd_global_lock;
852 for (retry = 0;; retry = 1) {
853 cfs_read_lock (g_lock);
854 peer = ksocknal_find_peer_locked(ni, id);
856 if (ksocknal_find_connectable_route_locked(peer) == NULL) {
857 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
859 /* I've got no routes that need to be
860 * connecting and I do have an actual
862 ksocknal_queue_tx_locked (tx, conn);
863 cfs_read_unlock (g_lock);
869 /* I'll need a write lock... */
870 cfs_read_unlock (g_lock);
872 cfs_write_lock_bh (g_lock);
874 peer = ksocknal_find_peer_locked(ni, id);
878 cfs_write_unlock_bh (g_lock);
880 if ((id.pid & LNET_PID_USERFLAG) != 0) {
881 CERROR("Refusing to create a connection to "
882 "userspace process %s\n", libcfs_id2str(id));
883 return -EHOSTUNREACH;
887 CERROR("Can't find peer %s\n", libcfs_id2str(id));
888 return -EHOSTUNREACH;
891 rc = ksocknal_add_peer(ni, id,
892 LNET_NIDADDR(id.nid),
893 lnet_acceptor_port());
895 CERROR("Can't add peer %s: %d\n",
896 libcfs_id2str(id), rc);
901 ksocknal_launch_all_connections_locked(peer);
903 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
905 /* Connection exists; queue message on it */
906 ksocknal_queue_tx_locked (tx, conn);
907 cfs_write_unlock_bh (g_lock);
911 if (peer->ksnp_accepting > 0 ||
912 ksocknal_find_connecting_route_locked (peer) != NULL) {
913 /* the message is going to be pinned to the peer */
915 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
917 /* Queue the message until a connection is established */
918 cfs_list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
919 cfs_write_unlock_bh (g_lock);
923 cfs_write_unlock_bh (g_lock);
925 /* NB Routes may be ignored if connections to them failed recently */
926 CNETERR("No usable routes to %s\n", libcfs_id2str(id));
927 return (-EHOSTUNREACH);
931 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
934 int type = lntmsg->msg_type;
935 lnet_process_id_t target = lntmsg->msg_target;
936 unsigned int payload_niov = lntmsg->msg_niov;
937 struct iovec *payload_iov = lntmsg->msg_iov;
938 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
939 unsigned int payload_offset = lntmsg->msg_offset;
940 unsigned int payload_nob = lntmsg->msg_len;
945 /* NB 'private' is different depending on what we're sending.
946 * Just ignore it... */
948 CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
949 payload_nob, payload_niov, libcfs_id2str(target));
951 LASSERT (payload_nob == 0 || payload_niov > 0);
952 LASSERT (payload_niov <= LNET_MAX_IOV);
953 /* payload is either all vaddrs or all pages */
954 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
955 LASSERT (!cfs_in_interrupt ());
957 if (payload_iov != NULL)
958 desc_size = offsetof(ksock_tx_t,
959 tx_frags.virt.iov[1 + payload_niov]);
961 desc_size = offsetof(ksock_tx_t,
962 tx_frags.paged.kiov[payload_niov]);
964 if (lntmsg->msg_vmflush)
965 mpflag = cfs_memory_pressure_get_and_set();
966 tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
968 CERROR("Can't allocate tx desc type %d size %d\n",
970 if (lntmsg->msg_vmflush)
971 cfs_memory_pressure_restore(mpflag);
975 tx->tx_conn = NULL; /* set when assigned a conn */
976 tx->tx_lnetmsg = lntmsg;
978 if (payload_iov != NULL) {
981 tx->tx_iov = tx->tx_frags.virt.iov;
983 lnet_extract_iov(payload_niov, &tx->tx_iov[1],
984 payload_niov, payload_iov,
985 payload_offset, payload_nob);
988 tx->tx_iov = &tx->tx_frags.paged.iov;
989 tx->tx_kiov = tx->tx_frags.paged.kiov;
990 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
991 payload_niov, payload_kiov,
992 payload_offset, payload_nob);
994 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
995 tx->tx_zc_capable = 1;
998 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
1000 /* The first fragment will be set later in pro_pack */
1001 rc = ksocknal_launch_packet(ni, tx, target);
1002 if (lntmsg->msg_vmflush)
1003 cfs_memory_pressure_restore(mpflag);
1007 ksocknal_free_tx(tx);
1012 ksocknal_thread_start (int (*fn)(void *arg), void *arg)
1014 long pid = cfs_create_thread (fn, arg, 0);
1019 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1020 ksocknal_data.ksnd_nthreads++;
1021 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1026 ksocknal_thread_fini (void)
1028 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1029 ksocknal_data.ksnd_nthreads--;
1030 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1034 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1036 static char ksocknal_slop_buffer[4096];
1042 LASSERT(conn->ksnc_proto != NULL);
1044 if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1045 /* Remind the socket to ack eagerly... */
1046 ksocknal_lib_eager_ack(conn);
1049 if (nob_to_skip == 0) { /* right at next packet boundary now */
1050 conn->ksnc_rx_started = 0;
1051 cfs_mb(); /* racing with timeout thread */
1053 switch (conn->ksnc_proto->pro_version) {
1054 case KSOCK_PROTO_V2:
1055 case KSOCK_PROTO_V3:
1056 conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1057 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1058 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1060 conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1061 conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1062 conn->ksnc_rx_iov[0].iov_len = offsetof(ksock_msg_t, ksm_u);
1065 case KSOCK_PROTO_V1:
1066 /* Receiving bare lnet_hdr_t */
1067 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1068 conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1069 conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1071 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1072 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1073 conn->ksnc_rx_iov[0].iov_len = sizeof (lnet_hdr_t);
1079 conn->ksnc_rx_niov = 1;
1081 conn->ksnc_rx_kiov = NULL;
1082 conn->ksnc_rx_nkiov = 0;
1083 conn->ksnc_rx_csum = ~0;
1087 /* Set up to skip as much as possible now. If there's more left
1088 * (ran out of iov entries) we'll get called again */
1090 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1091 conn->ksnc_rx_nob_left = nob_to_skip;
1092 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1097 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1099 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1100 conn->ksnc_rx_iov[niov].iov_len = nob;
1105 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1106 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1108 conn->ksnc_rx_niov = niov;
1109 conn->ksnc_rx_kiov = NULL;
1110 conn->ksnc_rx_nkiov = 0;
1111 conn->ksnc_rx_nob_wanted = skipped;
1116 ksocknal_process_receive (ksock_conn_t *conn)
1119 lnet_process_id_t *id;
1122 LASSERT (cfs_atomic_read(&conn->ksnc_conn_refcount) > 0);
1124 /* NB: sched lock NOT held */
1125 /* SOCKNAL_RX_LNET_HEADER is here for backward compatability */
1126 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1127 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1128 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1129 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1131 if (conn->ksnc_rx_nob_wanted != 0) {
1132 rc = ksocknal_receive(conn);
1135 LASSERT (rc != -EAGAIN);
1138 CDEBUG (D_NET, "[%p] EOF from %s"
1139 " ip %d.%d.%d.%d:%d\n", conn,
1140 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1141 HIPQUAD(conn->ksnc_ipaddr),
1143 else if (!conn->ksnc_closing)
1144 CERROR ("[%p] Error %d on read from %s"
1145 " ip %d.%d.%d.%d:%d\n",
1147 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1148 HIPQUAD(conn->ksnc_ipaddr),
1151 /* it's not an error if conn is being closed */
1152 ksocknal_close_conn_and_siblings (conn,
1153 (conn->ksnc_closing) ? 0 : rc);
1154 return (rc == 0 ? -ESHUTDOWN : rc);
1157 if (conn->ksnc_rx_nob_wanted != 0) {
1162 switch (conn->ksnc_rx_state) {
1163 case SOCKNAL_RX_KSM_HEADER:
1164 if (conn->ksnc_flip) {
1165 __swab32s(&conn->ksnc_msg.ksm_type);
1166 __swab32s(&conn->ksnc_msg.ksm_csum);
1167 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1168 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1171 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1172 conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1173 CERROR("%s: Unknown message type: %x\n",
1174 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1175 conn->ksnc_msg.ksm_type);
1176 ksocknal_new_packet(conn, 0);
1177 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1181 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1182 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1183 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1184 /* NOOP Checksum error */
1185 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1186 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1187 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1188 ksocknal_new_packet(conn, 0);
1189 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1193 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1196 LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1198 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1199 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1201 rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1202 conn->ksnc_msg.ksm_zc_cookies[1]);
1205 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1206 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1207 cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1208 ksocknal_new_packet(conn, 0);
1209 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1214 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1215 ksocknal_new_packet (conn, 0);
1216 return 0; /* NOOP is done and just return */
1219 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1220 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1221 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1223 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1224 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1225 conn->ksnc_rx_iov[0].iov_len = sizeof(ksock_lnet_msg_t);
1227 conn->ksnc_rx_niov = 1;
1228 conn->ksnc_rx_kiov = NULL;
1229 conn->ksnc_rx_nkiov = 0;
1231 goto again; /* read lnet header now */
1233 case SOCKNAL_RX_LNET_HEADER:
1234 /* unpack message header */
1235 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1237 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1238 /* Userspace peer */
1239 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1240 id = &conn->ksnc_peer->ksnp_id;
1242 /* Substitute process ID assigned at connection time */
1243 lhdr->src_pid = cpu_to_le32(id->pid);
1244 lhdr->src_nid = cpu_to_le64(id->nid);
1247 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1248 ksocknal_conn_addref(conn); /* ++ref while parsing */
1250 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1251 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1252 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1254 /* I just received garbage: give up on this conn */
1255 ksocknal_new_packet(conn, 0);
1256 ksocknal_close_conn_and_siblings (conn, rc);
1257 ksocknal_conn_decref(conn);
1261 /* I'm racing with ksocknal_recv() */
1262 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1263 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1265 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1268 /* ksocknal_recv() got called */
1271 case SOCKNAL_RX_LNET_PAYLOAD:
1272 /* payload all received */
1275 if (conn->ksnc_rx_nob_left == 0 && /* not truncating */
1276 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1277 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1278 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1279 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1280 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1284 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1285 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1287 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1288 id = &conn->ksnc_peer->ksnp_id;
1290 rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1291 conn->ksnc_msg.ksm_zc_cookies[0],
1292 *ksocknal_tunables.ksnd_nonblk_zcack ||
1293 le64_to_cpu(lhdr->src_nid) != id->nid);
1296 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1299 ksocknal_new_packet(conn, 0);
1300 ksocknal_close_conn_and_siblings (conn, rc);
1305 case SOCKNAL_RX_SLOP:
1306 /* starting new packet? */
1307 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1308 return 0; /* come back later */
1309 goto again; /* try to finish reading slop now */
1317 return (-EINVAL); /* keep gcc happy */
1321 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1322 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1323 unsigned int offset, unsigned int mlen, unsigned int rlen)
1325 ksock_conn_t *conn = (ksock_conn_t *)private;
1326 ksock_sched_t *sched = conn->ksnc_scheduler;
1328 LASSERT (mlen <= rlen);
1329 LASSERT (niov <= LNET_MAX_IOV);
1331 conn->ksnc_cookie = msg;
1332 conn->ksnc_rx_nob_wanted = mlen;
1333 conn->ksnc_rx_nob_left = rlen;
1335 if (mlen == 0 || iov != NULL) {
1336 conn->ksnc_rx_nkiov = 0;
1337 conn->ksnc_rx_kiov = NULL;
1338 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1339 conn->ksnc_rx_niov =
1340 lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1341 niov, iov, offset, mlen);
1343 conn->ksnc_rx_niov = 0;
1344 conn->ksnc_rx_iov = NULL;
1345 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1346 conn->ksnc_rx_nkiov =
1347 lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1348 niov, kiov, offset, mlen);
1352 lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1353 lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1355 LASSERT (conn->ksnc_rx_scheduled);
1357 cfs_spin_lock_bh (&sched->kss_lock);
1359 switch (conn->ksnc_rx_state) {
1360 case SOCKNAL_RX_PARSE_WAIT:
1361 cfs_list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1362 cfs_waitq_signal (&sched->kss_waitq);
1363 LASSERT (conn->ksnc_rx_ready);
1366 case SOCKNAL_RX_PARSE:
1367 /* scheduler hasn't noticed I'm parsing yet */
1371 conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1373 cfs_spin_unlock_bh (&sched->kss_lock);
1374 ksocknal_conn_decref(conn);
1379 ksocknal_sched_cansleep(ksock_sched_t *sched)
1383 cfs_spin_lock_bh (&sched->kss_lock);
1385 rc = (!ksocknal_data.ksnd_shuttingdown &&
1386 cfs_list_empty(&sched->kss_rx_conns) &&
1387 cfs_list_empty(&sched->kss_tx_conns));
1389 cfs_spin_unlock_bh (&sched->kss_lock);
1393 int ksocknal_scheduler (void *arg)
1395 ksock_sched_t *sched = (ksock_sched_t *)arg;
1400 int id = (int)(sched - ksocknal_data.ksnd_schedulers);
1403 snprintf (name, sizeof (name),"socknal_sd%02d", id);
1404 cfs_daemonize (name);
1405 cfs_block_allsigs ();
1407 if (ksocknal_lib_bind_thread_to_cpu(id))
1408 CERROR ("Can't set CPU affinity for %s to %d\n", name, id);
1410 cfs_spin_lock_bh (&sched->kss_lock);
1412 while (!ksocknal_data.ksnd_shuttingdown) {
1413 int did_something = 0;
1415 /* Ensure I progress everything semi-fairly */
1417 if (!cfs_list_empty (&sched->kss_rx_conns)) {
1418 conn = cfs_list_entry(sched->kss_rx_conns.next,
1419 ksock_conn_t, ksnc_rx_list);
1420 cfs_list_del(&conn->ksnc_rx_list);
1422 LASSERT(conn->ksnc_rx_scheduled);
1423 LASSERT(conn->ksnc_rx_ready);
1425 /* clear rx_ready in case receive isn't complete.
1426 * Do it BEFORE we call process_recv, since
1427 * data_ready can set it any time after we release
1429 conn->ksnc_rx_ready = 0;
1430 cfs_spin_unlock_bh (&sched->kss_lock);
1432 rc = ksocknal_process_receive(conn);
1434 cfs_spin_lock_bh (&sched->kss_lock);
1436 /* I'm the only one that can clear this flag */
1437 LASSERT(conn->ksnc_rx_scheduled);
1439 /* Did process_receive get everything it wanted? */
1441 conn->ksnc_rx_ready = 1;
1443 if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1444 /* Conn blocked waiting for ksocknal_recv()
1445 * I change its state (under lock) to signal
1446 * it can be rescheduled */
1447 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1448 } else if (conn->ksnc_rx_ready) {
1449 /* reschedule for rx */
1450 cfs_list_add_tail (&conn->ksnc_rx_list,
1451 &sched->kss_rx_conns);
1453 conn->ksnc_rx_scheduled = 0;
1455 ksocknal_conn_decref(conn);
1461 if (!cfs_list_empty (&sched->kss_tx_conns)) {
1462 CFS_LIST_HEAD (zlist);
1464 if (!cfs_list_empty(&sched->kss_zombie_noop_txs)) {
1465 cfs_list_add(&zlist,
1466 &sched->kss_zombie_noop_txs);
1467 cfs_list_del_init(&sched->kss_zombie_noop_txs);
1470 conn = cfs_list_entry(sched->kss_tx_conns.next,
1471 ksock_conn_t, ksnc_tx_list);
1472 cfs_list_del (&conn->ksnc_tx_list);
1474 LASSERT(conn->ksnc_tx_scheduled);
1475 LASSERT(conn->ksnc_tx_ready);
1476 LASSERT(!cfs_list_empty(&conn->ksnc_tx_queue));
1478 tx = cfs_list_entry(conn->ksnc_tx_queue.next,
1479 ksock_tx_t, tx_list);
1481 if (conn->ksnc_tx_carrier == tx)
1482 ksocknal_next_tx_carrier(conn);
1484 /* dequeue now so empty list => more to send */
1485 cfs_list_del(&tx->tx_list);
1487 /* Clear tx_ready in case send isn't complete. Do
1488 * it BEFORE we call process_transmit, since
1489 * write_space can set it any time after we release
1491 conn->ksnc_tx_ready = 0;
1492 cfs_spin_unlock_bh (&sched->kss_lock);
1494 if (!cfs_list_empty(&zlist)) {
1495 /* free zombie noop txs, it's fast because
1496 * noop txs are just put in freelist */
1497 ksocknal_txlist_done(NULL, &zlist, 0);
1500 rc = ksocknal_process_transmit(conn, tx);
1502 if (rc == -ENOMEM || rc == -EAGAIN) {
1503 /* Incomplete send: replace tx on HEAD of tx_queue */
1504 cfs_spin_lock_bh (&sched->kss_lock);
1505 cfs_list_add (&tx->tx_list,
1506 &conn->ksnc_tx_queue);
1508 /* Complete send; tx -ref */
1509 ksocknal_tx_decref (tx);
1511 cfs_spin_lock_bh (&sched->kss_lock);
1512 /* assume space for more */
1513 conn->ksnc_tx_ready = 1;
1516 if (rc == -ENOMEM) {
1517 /* Do nothing; after a short timeout, this
1518 * conn will be reposted on kss_tx_conns. */
1519 } else if (conn->ksnc_tx_ready &&
1520 !cfs_list_empty (&conn->ksnc_tx_queue)) {
1521 /* reschedule for tx */
1522 cfs_list_add_tail (&conn->ksnc_tx_list,
1523 &sched->kss_tx_conns);
1525 conn->ksnc_tx_scheduled = 0;
1527 ksocknal_conn_decref(conn);
1532 if (!did_something || /* nothing to do */
1533 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1534 cfs_spin_unlock_bh (&sched->kss_lock);
1538 if (!did_something) { /* wait for something to do */
1539 cfs_wait_event_interruptible_exclusive(
1541 !ksocknal_sched_cansleep(sched), rc);
1547 cfs_spin_lock_bh (&sched->kss_lock);
1551 cfs_spin_unlock_bh (&sched->kss_lock);
1552 ksocknal_thread_fini ();
1557 * Add connection to kss_rx_conns of scheduler
1558 * and wakeup the scheduler.
1560 void ksocknal_read_callback (ksock_conn_t *conn)
1562 ksock_sched_t *sched;
1565 sched = conn->ksnc_scheduler;
1567 cfs_spin_lock_bh (&sched->kss_lock);
1569 conn->ksnc_rx_ready = 1;
1571 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1572 cfs_list_add_tail(&conn->ksnc_rx_list,
1573 &sched->kss_rx_conns);
1574 conn->ksnc_rx_scheduled = 1;
1575 /* extra ref for scheduler */
1576 ksocknal_conn_addref(conn);
1578 cfs_waitq_signal (&sched->kss_waitq);
1580 cfs_spin_unlock_bh (&sched->kss_lock);
1586 * Add connection to kss_tx_conns of scheduler
1587 * and wakeup the scheduler.
1589 void ksocknal_write_callback (ksock_conn_t *conn)
1591 ksock_sched_t *sched;
1594 sched = conn->ksnc_scheduler;
1596 cfs_spin_lock_bh (&sched->kss_lock);
1598 conn->ksnc_tx_ready = 1;
1600 if (!conn->ksnc_tx_scheduled && // not being progressed
1601 !cfs_list_empty(&conn->ksnc_tx_queue)){//packets to send
1602 cfs_list_add_tail (&conn->ksnc_tx_list,
1603 &sched->kss_tx_conns);
1604 conn->ksnc_tx_scheduled = 1;
1605 /* extra ref for scheduler */
1606 ksocknal_conn_addref(conn);
1608 cfs_waitq_signal (&sched->kss_waitq);
1611 cfs_spin_unlock_bh (&sched->kss_lock);
1617 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1621 if (hello->kshm_magic == LNET_PROTO_MAGIC)
1622 version = hello->kshm_version;
1623 else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1624 version = __swab32(hello->kshm_version);
1627 #if SOCKNAL_VERSION_DEBUG
1628 if (*ksocknal_tunables.ksnd_protocol == 1)
1631 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1632 version == KSOCK_PROTO_V3)
1635 if (version == KSOCK_PROTO_V2)
1636 return &ksocknal_protocol_v2x;
1638 if (version == KSOCK_PROTO_V3)
1639 return &ksocknal_protocol_v3x;
1644 if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1645 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1647 CLASSERT (sizeof (lnet_magicversion_t) ==
1648 offsetof (ksock_hello_msg_t, kshm_src_nid));
1650 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1651 hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1652 return &ksocknal_protocol_v1x;
1659 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1660 lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1662 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1663 ksock_net_t *net = (ksock_net_t *)ni->ni_data;
1665 LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1667 /* rely on caller to hold a ref on socket so it wouldn't disappear */
1668 LASSERT (conn->ksnc_proto != NULL);
1670 hello->kshm_src_nid = ni->ni_nid;
1671 hello->kshm_dst_nid = peer_nid;
1672 hello->kshm_src_pid = the_lnet.ln_pid;
1674 hello->kshm_src_incarnation = net->ksnn_incarnation;
1675 hello->kshm_ctype = conn->ksnc_type;
1677 return conn->ksnc_proto->pro_send_hello(conn, hello);
1681 ksocknal_invert_type(int type)
1685 case SOCKLND_CONN_ANY:
1686 case SOCKLND_CONN_CONTROL:
1688 case SOCKLND_CONN_BULK_IN:
1689 return SOCKLND_CONN_BULK_OUT;
1690 case SOCKLND_CONN_BULK_OUT:
1691 return SOCKLND_CONN_BULK_IN;
1693 return (SOCKLND_CONN_NONE);
1698 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1699 ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1702 /* Return < 0 fatal error
1704 * EALREADY lost connection race
1705 * EPROTO protocol version mismatch
1707 cfs_socket_t *sock = conn->ksnc_sock;
1708 int active = (conn->ksnc_proto != NULL);
1712 ksock_proto_t *proto;
1713 lnet_process_id_t recv_id;
1715 /* socket type set on active connections - not set on passive */
1716 LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1718 timeout = active ? *ksocknal_tunables.ksnd_timeout :
1719 lnet_acceptor_timeout();
1721 rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1723 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1724 rc, HIPQUAD(conn->ksnc_ipaddr));
1729 if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1730 hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1731 hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1732 /* Unexpected magic! */
1733 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1734 "%u.%u.%u.%u\n", __cpu_to_le32 (hello->kshm_magic),
1735 LNET_PROTO_TCP_MAGIC,
1736 HIPQUAD(conn->ksnc_ipaddr));
1740 rc = libcfs_sock_read(sock, &hello->kshm_version,
1741 sizeof(hello->kshm_version), timeout);
1743 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1744 rc, HIPQUAD(conn->ksnc_ipaddr));
1749 proto = ksocknal_parse_proto_version(hello);
1750 if (proto == NULL) {
1752 /* unknown protocol from peer, tell peer my protocol */
1753 conn->ksnc_proto = &ksocknal_protocol_v3x;
1754 #if SOCKNAL_VERSION_DEBUG
1755 if (*ksocknal_tunables.ksnd_protocol == 2)
1756 conn->ksnc_proto = &ksocknal_protocol_v2x;
1757 else if (*ksocknal_tunables.ksnd_protocol == 1)
1758 conn->ksnc_proto = &ksocknal_protocol_v1x;
1760 hello->kshm_nips = 0;
1761 ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1764 CERROR ("Unknown protocol version (%d.x expected)"
1765 " from %u.%u.%u.%u\n",
1766 conn->ksnc_proto->pro_version,
1767 HIPQUAD(conn->ksnc_ipaddr));
1772 proto_match = (conn->ksnc_proto == proto);
1773 conn->ksnc_proto = proto;
1775 /* receive the rest of hello message anyway */
1776 rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1778 CERROR("Error %d reading or checking hello from from %u.%u.%u.%u\n",
1779 rc, HIPQUAD(conn->ksnc_ipaddr));
1784 *incarnation = hello->kshm_src_incarnation;
1786 if (hello->kshm_src_nid == LNET_NID_ANY) {
1787 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1788 "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1793 conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1794 /* Userspace NAL assigns peer process ID from socket */
1795 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1796 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1798 recv_id.nid = hello->kshm_src_nid;
1799 recv_id.pid = hello->kshm_src_pid;
1805 /* peer determines type */
1806 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1807 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1808 CERROR ("Unexpected type %d from %s ip %u.%u.%u.%u\n",
1809 hello->kshm_ctype, libcfs_id2str(*peerid),
1810 HIPQUAD(conn->ksnc_ipaddr));
1817 if (peerid->pid != recv_id.pid ||
1818 peerid->nid != recv_id.nid) {
1819 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1820 " %u.%u.%u.%u, but they claimed they were "
1821 "%s; please check your Lustre "
1823 libcfs_id2str(*peerid),
1824 HIPQUAD(conn->ksnc_ipaddr),
1825 libcfs_id2str(recv_id));
1829 if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1830 /* Possible protocol mismatch or I lost the connection race */
1831 return proto_match ? EALREADY : EPROTO;
1834 if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1835 CERROR ("Mismatched types: me %d, %s ip %u.%u.%u.%u %d\n",
1836 conn->ksnc_type, libcfs_id2str(*peerid),
1837 HIPQUAD(conn->ksnc_ipaddr),
1846 ksocknal_connect (ksock_route_t *route)
1848 CFS_LIST_HEAD (zombies);
1849 ksock_peer_t *peer = route->ksnr_peer;
1853 cfs_time_t deadline;
1854 int retry_later = 0;
1857 deadline = cfs_time_add(cfs_time_current(),
1858 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1860 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1862 LASSERT (route->ksnr_scheduled);
1863 LASSERT (!route->ksnr_connecting);
1865 route->ksnr_connecting = 1;
1868 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1870 /* stop connecting if peer/route got closed under me, or
1871 * route got connected while queued */
1872 if (peer->ksnp_closing || route->ksnr_deleted ||
1878 /* reschedule if peer is connecting to me */
1879 if (peer->ksnp_accepting > 0) {
1881 "peer %s(%d) already connecting to me, retry later.\n",
1882 libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1886 if (retry_later) /* needs reschedule */
1889 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1890 type = SOCKLND_CONN_ANY;
1891 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1892 type = SOCKLND_CONN_CONTROL;
1893 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1894 type = SOCKLND_CONN_BULK_IN;
1896 LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1897 type = SOCKLND_CONN_BULK_OUT;
1900 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1902 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1904 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1910 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1911 route->ksnr_myipaddr,
1912 route->ksnr_ipaddr, route->ksnr_port);
1916 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1918 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1924 /* A +ve RC means I have to retry because I lost the connection
1925 * race or I have to renegotiate protocol version */
1926 retry_later = (rc != 0);
1928 CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1929 libcfs_nid2str(peer->ksnp_id.nid));
1931 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1934 route->ksnr_scheduled = 0;
1935 route->ksnr_connecting = 0;
1938 /* re-queue for attention; this frees me up to handle
1939 * the peer's incoming connection request */
1941 if (rc == EALREADY ||
1942 (rc == 0 && peer->ksnp_accepting > 0)) {
1943 /* We want to introduce a delay before next
1944 * attempt to connect if we lost conn race,
1945 * but the race is resolved quickly usually,
1946 * so min_reconnectms should be good heuristic */
1947 route->ksnr_retry_interval =
1948 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1949 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1950 route->ksnr_retry_interval);
1953 ksocknal_launch_connection_locked(route);
1956 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
1960 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
1962 route->ksnr_scheduled = 0;
1963 route->ksnr_connecting = 0;
1965 /* This is a retry rather than a new connection */
1966 route->ksnr_retry_interval *= 2;
1967 route->ksnr_retry_interval =
1968 MAX(route->ksnr_retry_interval,
1969 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1970 route->ksnr_retry_interval =
1971 MIN(route->ksnr_retry_interval,
1972 cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1974 LASSERT (route->ksnr_retry_interval != 0);
1975 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1976 route->ksnr_retry_interval);
1978 if (!cfs_list_empty(&peer->ksnp_tx_queue) &&
1979 peer->ksnp_accepting == 0 &&
1980 ksocknal_find_connecting_route_locked(peer) == NULL) {
1983 /* ksnp_tx_queue is queued on a conn on successful
1984 * connection for V1.x and V2.x */
1985 if (!cfs_list_empty (&peer->ksnp_conns)) {
1986 conn = cfs_list_entry(peer->ksnp_conns.next,
1987 ksock_conn_t, ksnc_list);
1988 LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1991 /* take all the blocked packets while I've got the lock and
1992 * complete below... */
1993 cfs_list_splice_init(&peer->ksnp_tx_queue, &zombies);
1996 #if 0 /* irrelevent with only eager routes */
1997 if (!route->ksnr_deleted) {
1998 /* make this route least-favourite for re-selection */
1999 cfs_list_del(&route->ksnr_list);
2000 cfs_list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
2003 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
2005 ksocknal_peer_failed(peer);
2006 ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2011 * check whether we need to create more connds.
2012 * It will try to create new thread if it's necessary, @timeout can
2013 * be updated if failed to create, so caller wouldn't keep try while
2014 * running out of resource.
2017 ksocknal_connd_check_start(long sec, long *timeout)
2020 int total = ksocknal_data.ksnd_connd_starting +
2021 ksocknal_data.ksnd_connd_running;
2023 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2024 /* still in initializing */
2028 if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2029 total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2030 /* can't create more connd, or still have enough
2031 * threads to handle more connecting */
2035 if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2036 /* no pending connecting request */
2040 if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2041 /* may run out of resource, retry later */
2042 *timeout = cfs_time_seconds(1);
2046 if (ksocknal_data.ksnd_connd_starting > 0) {
2047 /* serialize starting to avoid flood */
2051 ksocknal_data.ksnd_connd_starting_stamp = sec;
2052 ksocknal_data.ksnd_connd_starting++;
2053 cfs_spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2055 /* NB: total is the next id */
2056 rc = ksocknal_thread_start(ksocknal_connd, (void *)((long)total));
2058 cfs_spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2063 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2064 ksocknal_data.ksnd_connd_starting--;
2065 ksocknal_data.ksnd_connd_failed_stamp = cfs_time_current_sec();
2071 * check whether current thread can exit, it will return 1 if there are too
2072 * many threads and no creating in past 120 seconds.
2073 * Also, this function may update @timeout to make caller come back
2074 * again to recheck these conditions.
2077 ksocknal_connd_check_stop(long sec, long *timeout)
2081 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2082 /* still in initializing */
2086 if (ksocknal_data.ksnd_connd_starting > 0) {
2087 /* in progress of starting new thread */
2091 if (ksocknal_data.ksnd_connd_running <=
2092 *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2096 /* created thread in past 120 seconds? */
2097 val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2098 SOCKNAL_CONND_TIMEOUT - sec);
2100 *timeout = (val > 0) ? cfs_time_seconds(val) :
2101 cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2105 /* no creating in past 120 seconds */
2107 return ksocknal_data.ksnd_connd_running >
2108 ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2111 /* Go through connd_routes queue looking for a route that we can process
2112 * right now, @timeout_p can be updated if we need to come back later */
2113 static ksock_route_t *
2114 ksocknal_connd_get_route_locked(signed long *timeout_p)
2116 ksock_route_t *route;
2119 now = cfs_time_current();
2121 /* connd_routes can contain both pending and ordinary routes */
2122 cfs_list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2125 if (route->ksnr_retry_interval == 0 ||
2126 cfs_time_aftereq(now, route->ksnr_timeout))
2129 if (*timeout_p == CFS_MAX_SCHEDULE_TIMEOUT ||
2130 (int)*timeout_p > (int)(route->ksnr_timeout - now))
2131 *timeout_p = (int)(route->ksnr_timeout - now);
2138 ksocknal_connd (void *arg)
2140 cfs_spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2141 long id = (long)(long_ptr_t)arg;
2143 ksock_connreq_t *cr;
2144 cfs_waitlink_t wait;
2148 snprintf (name, sizeof (name), "socknal_cd%02ld", id);
2149 cfs_daemonize (name);
2150 cfs_block_allsigs ();
2152 cfs_waitlink_init (&wait);
2154 cfs_spin_lock_bh (connd_lock);
2156 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2157 ksocknal_data.ksnd_connd_starting--;
2158 ksocknal_data.ksnd_connd_running++;
2160 while (!ksocknal_data.ksnd_shuttingdown) {
2161 ksock_route_t *route = NULL;
2162 long sec = cfs_time_current_sec();
2163 long timeout = CFS_MAX_SCHEDULE_TIMEOUT;
2164 int dropped_lock = 0;
2166 if (ksocknal_connd_check_stop(sec, &timeout)) {
2167 /* wakeup another one to check stop */
2168 cfs_waitq_signal(&ksocknal_data.ksnd_connd_waitq);
2172 if (ksocknal_connd_check_start(sec, &timeout)) {
2173 /* created new thread */
2177 if (!cfs_list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2178 /* Connection accepted by the listener */
2179 cr = cfs_list_entry(ksocknal_data.ksnd_connd_connreqs. \
2180 next, ksock_connreq_t, ksncr_list);
2182 cfs_list_del(&cr->ksncr_list);
2183 cfs_spin_unlock_bh(connd_lock);
2186 ksocknal_create_conn(cr->ksncr_ni, NULL,
2187 cr->ksncr_sock, SOCKLND_CONN_NONE);
2188 lnet_ni_decref(cr->ksncr_ni);
2189 LIBCFS_FREE(cr, sizeof(*cr));
2191 cfs_spin_lock_bh(connd_lock);
2194 /* Only handle an outgoing connection request if there
2195 * is a thread left to handle incoming connections and
2196 * create new connd */
2197 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2198 ksocknal_data.ksnd_connd_running) {
2199 route = ksocknal_connd_get_route_locked(&timeout);
2201 if (route != NULL) {
2202 cfs_list_del (&route->ksnr_connd_list);
2203 ksocknal_data.ksnd_connd_connecting++;
2204 cfs_spin_unlock_bh(connd_lock);
2207 if (ksocknal_connect(route)) {
2208 /* consecutive retry */
2209 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2210 CWARN("massive consecutive "
2211 "re-connecting to %u.%u.%u.%u\n",
2212 HIPQUAD(route->ksnr_ipaddr));
2219 ksocknal_route_decref(route);
2221 cfs_spin_lock_bh(connd_lock);
2222 ksocknal_data.ksnd_connd_connecting--;
2226 if (++nloops < SOCKNAL_RESCHED)
2228 cfs_spin_unlock_bh(connd_lock);
2231 cfs_spin_lock_bh(connd_lock);
2235 /* Nothing to do for 'timeout' */
2236 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2237 cfs_waitq_add_exclusive (&ksocknal_data.ksnd_connd_waitq,
2239 cfs_spin_unlock_bh(connd_lock);
2242 cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE, timeout);
2244 cfs_set_current_state (CFS_TASK_RUNNING);
2245 cfs_waitq_del (&ksocknal_data.ksnd_connd_waitq, &wait);
2246 cfs_spin_lock_bh(connd_lock);
2248 ksocknal_data.ksnd_connd_running--;
2249 cfs_spin_unlock_bh(connd_lock);
2251 ksocknal_thread_fini ();
2256 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2258 /* We're called with a shared lock on ksnd_global_lock */
2262 cfs_list_for_each (ctmp, &peer->ksnp_conns) {
2264 conn = cfs_list_entry (ctmp, ksock_conn_t, ksnc_list);
2266 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2267 LASSERT (!conn->ksnc_closing);
2269 /* SOCK_ERROR will reset error code of socket in
2270 * some platform (like Darwin8.x) */
2271 error = libcfs_sock_error(conn->ksnc_sock);
2273 ksocknal_conn_addref(conn);
2277 CNETERR("A connection with %s "
2278 "(%u.%u.%u.%u:%d) was reset; "
2279 "it may have rebooted.\n",
2280 libcfs_id2str(peer->ksnp_id),
2281 HIPQUAD(conn->ksnc_ipaddr),
2285 CNETERR("A connection with %s "
2286 "(%u.%u.%u.%u:%d) timed out; the "
2287 "network or node may be down.\n",
2288 libcfs_id2str(peer->ksnp_id),
2289 HIPQUAD(conn->ksnc_ipaddr),
2293 CNETERR("An unexpected network error %d "
2295 "(%u.%u.%u.%u:%d\n", error,
2296 libcfs_id2str(peer->ksnp_id),
2297 HIPQUAD(conn->ksnc_ipaddr),
2305 if (conn->ksnc_rx_started &&
2306 cfs_time_aftereq(cfs_time_current(),
2307 conn->ksnc_rx_deadline)) {
2308 /* Timed out incomplete incoming message */
2309 ksocknal_conn_addref(conn);
2310 CNETERR("Timeout receiving from %s (%u.%u.%u.%u:%d), "
2311 "state %d wanted %d left %d\n",
2312 libcfs_id2str(peer->ksnp_id),
2313 HIPQUAD(conn->ksnc_ipaddr),
2315 conn->ksnc_rx_state,
2316 conn->ksnc_rx_nob_wanted,
2317 conn->ksnc_rx_nob_left);
2321 if ((!cfs_list_empty(&conn->ksnc_tx_queue) ||
2322 libcfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2323 cfs_time_aftereq(cfs_time_current(),
2324 conn->ksnc_tx_deadline)) {
2325 /* Timed out messages queued for sending or
2326 * buffered in the socket's send buffer */
2327 ksocknal_conn_addref(conn);
2328 CNETERR("Timeout sending data to %s (%u.%u.%u.%u:%d) "
2329 "the network or that node may be down.\n",
2330 libcfs_id2str(peer->ksnp_id),
2331 HIPQUAD(conn->ksnc_ipaddr),
2341 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2344 CFS_LIST_HEAD (stale_txs);
2346 cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
2348 while (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2349 tx = cfs_list_entry (peer->ksnp_tx_queue.next,
2350 ksock_tx_t, tx_list);
2352 if (!cfs_time_aftereq(cfs_time_current(),
2356 cfs_list_del (&tx->tx_list);
2357 cfs_list_add_tail (&tx->tx_list, &stale_txs);
2360 cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
2362 ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2366 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2368 ksock_sched_t *sched;
2372 if (cfs_list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2375 if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2378 if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2379 cfs_time_before(cfs_time_current(),
2380 cfs_time_add(peer->ksnp_last_alive,
2381 cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2384 if (cfs_time_before(cfs_time_current(),
2385 peer->ksnp_send_keepalive))
2388 /* retry 10 secs later, so we wouldn't put pressure
2389 * on this peer if we failed to send keepalive this time */
2390 peer->ksnp_send_keepalive = cfs_time_shift(10);
2392 conn = ksocknal_find_conn_locked(peer, NULL, 1);
2394 sched = conn->ksnc_scheduler;
2396 cfs_spin_lock_bh (&sched->kss_lock);
2397 if (!cfs_list_empty(&conn->ksnc_tx_queue)) {
2398 cfs_spin_unlock_bh(&sched->kss_lock);
2399 /* there is an queued ACK, don't need keepalive */
2403 cfs_spin_unlock_bh(&sched->kss_lock);
2406 cfs_read_unlock(&ksocknal_data.ksnd_global_lock);
2408 /* cookie = 1 is reserved for keepalive PING */
2409 tx = ksocknal_alloc_tx_noop(1, 1);
2411 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2415 if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2416 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2420 ksocknal_free_tx(tx);
2421 cfs_read_lock(&ksocknal_data.ksnd_global_lock);
2428 ksocknal_check_peer_timeouts (int idx)
2430 cfs_list_t *peers = &ksocknal_data.ksnd_peers[idx];
2436 /* NB. We expect to have a look at all the peers and not find any
2437 * connections to time out, so we just use a shared lock while we
2439 cfs_read_lock (&ksocknal_data.ksnd_global_lock);
2441 cfs_list_for_each_entry_typed(peer, peers, ksock_peer_t, ksnp_list) {
2442 cfs_time_t deadline = 0;
2446 if (ksocknal_send_keepalive_locked(peer) != 0) {
2447 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2451 conn = ksocknal_find_timed_out_conn (peer);
2454 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2456 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2458 /* NB we won't find this one again, but we can't
2459 * just proceed with the next peer, since we dropped
2460 * ksnd_global_lock and it might be dead already! */
2461 ksocknal_conn_decref(conn);
2465 /* we can't process stale txs right here because we're
2466 * holding only shared lock */
2467 if (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2469 cfs_list_entry (peer->ksnp_tx_queue.next,
2470 ksock_tx_t, tx_list);
2472 if (cfs_time_aftereq(cfs_time_current(),
2475 ksocknal_peer_addref(peer);
2476 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2478 ksocknal_flush_stale_txs(peer);
2480 ksocknal_peer_decref(peer);
2485 if (cfs_list_empty(&peer->ksnp_zc_req_list))
2488 cfs_spin_lock(&peer->ksnp_lock);
2489 cfs_list_for_each_entry_typed(tx, &peer->ksnp_zc_req_list,
2490 ksock_tx_t, tx_zc_list) {
2491 if (!cfs_time_aftereq(cfs_time_current(),
2494 /* ignore the TX if connection is being closed */
2495 if (tx->tx_conn->ksnc_closing)
2501 cfs_spin_unlock(&peer->ksnp_lock);
2505 tx = cfs_list_entry(peer->ksnp_zc_req_list.next,
2506 ksock_tx_t, tx_zc_list);
2507 deadline = tx->tx_deadline;
2508 resid = tx->tx_resid;
2510 ksocknal_conn_addref(conn);
2512 cfs_spin_unlock(&peer->ksnp_lock);
2513 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2515 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2516 "oldest(%p) timed out %ld secs ago, "
2517 "resid: %d, wmem: %d\n",
2518 n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2519 cfs_duration_sec(cfs_time_current() - deadline),
2520 resid, libcfs_sock_wmem_queued(conn->ksnc_sock));
2522 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2523 ksocknal_conn_decref(conn);
2527 cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
2531 ksocknal_reaper (void *arg)
2533 cfs_waitlink_t wait;
2535 ksock_sched_t *sched;
2536 cfs_list_t enomem_conns;
2538 cfs_duration_t timeout;
2541 cfs_time_t deadline = cfs_time_current();
2543 cfs_daemonize ("socknal_reaper");
2544 cfs_block_allsigs ();
2546 CFS_INIT_LIST_HEAD(&enomem_conns);
2547 cfs_waitlink_init (&wait);
2549 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2551 while (!ksocknal_data.ksnd_shuttingdown) {
2553 if (!cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2554 conn = cfs_list_entry (ksocknal_data. \
2555 ksnd_deathrow_conns.next,
2556 ksock_conn_t, ksnc_list);
2557 cfs_list_del (&conn->ksnc_list);
2559 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2561 ksocknal_terminate_conn (conn);
2562 ksocknal_conn_decref(conn);
2564 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2568 if (!cfs_list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2569 conn = cfs_list_entry (ksocknal_data.ksnd_zombie_conns.\
2570 next, ksock_conn_t, ksnc_list);
2571 cfs_list_del (&conn->ksnc_list);
2573 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2575 ksocknal_destroy_conn (conn);
2577 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2581 if (!cfs_list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2582 cfs_list_add(&enomem_conns,
2583 &ksocknal_data.ksnd_enomem_conns);
2584 cfs_list_del_init(&ksocknal_data.ksnd_enomem_conns);
2587 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2589 /* reschedule all the connections that stalled with ENOMEM... */
2591 while (!cfs_list_empty (&enomem_conns)) {
2592 conn = cfs_list_entry (enomem_conns.next,
2593 ksock_conn_t, ksnc_tx_list);
2594 cfs_list_del (&conn->ksnc_tx_list);
2596 sched = conn->ksnc_scheduler;
2598 cfs_spin_lock_bh (&sched->kss_lock);
2600 LASSERT (conn->ksnc_tx_scheduled);
2601 conn->ksnc_tx_ready = 1;
2602 cfs_list_add_tail(&conn->ksnc_tx_list,
2603 &sched->kss_tx_conns);
2604 cfs_waitq_signal (&sched->kss_waitq);
2606 cfs_spin_unlock_bh (&sched->kss_lock);
2610 /* careful with the jiffy wrap... */
2611 while ((timeout = cfs_time_sub(deadline,
2612 cfs_time_current())) <= 0) {
2615 int chunk = ksocknal_data.ksnd_peer_hash_size;
2617 /* Time to check for timeouts on a few more peers: I do
2618 * checks every 'p' seconds on a proportion of the peer
2619 * table and I need to check every connection 'n' times
2620 * within a timeout interval, to ensure I detect a
2621 * timeout on any connection within (n+1)/n times the
2622 * timeout interval. */
2624 if (*ksocknal_tunables.ksnd_timeout > n * p)
2625 chunk = (chunk * n * p) /
2626 *ksocknal_tunables.ksnd_timeout;
2630 for (i = 0; i < chunk; i++) {
2631 ksocknal_check_peer_timeouts (peer_index);
2632 peer_index = (peer_index + 1) %
2633 ksocknal_data.ksnd_peer_hash_size;
2636 deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2639 if (nenomem_conns != 0) {
2640 /* Reduce my timeout if I rescheduled ENOMEM conns.
2641 * This also prevents me getting woken immediately
2642 * if any go back on my enomem list. */
2643 timeout = SOCKNAL_ENOMEM_RETRY;
2645 ksocknal_data.ksnd_reaper_waketime =
2646 cfs_time_add(cfs_time_current(), timeout);
2648 cfs_set_current_state (CFS_TASK_INTERRUPTIBLE);
2649 cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
2651 if (!ksocknal_data.ksnd_shuttingdown &&
2652 cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2653 cfs_list_empty (&ksocknal_data.ksnd_zombie_conns))
2654 cfs_waitq_timedwait (&wait, CFS_TASK_INTERRUPTIBLE,
2657 cfs_set_current_state (CFS_TASK_RUNNING);
2658 cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
2660 cfs_spin_lock_bh (&ksocknal_data.ksnd_reaper_lock);
2663 cfs_spin_unlock_bh (&ksocknal_data.ksnd_reaper_lock);
2665 ksocknal_thread_fini ();