2 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
4 * Copyright (c) 2011, 2012, Intel Corporation.
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>
11 * This file is part of Portals, http://www.sf.net/projects/sandiaportals/
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.
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.
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.
30 ksocknal_alloc_tx(int type, int size)
32 ksock_tx_t *tx = NULL;
34 if (type == KSOCK_MSG_NOOP) {
35 LASSERT(size == KSOCK_NOOP_TX_SIZE);
37 /* searching for a noop tx in free list */
38 spin_lock(&ksocknal_data.ksnd_tx_lock);
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);
47 spin_unlock(&ksocknal_data.ksnd_tx_lock);
51 LIBCFS_ALLOC(tx, size);
56 atomic_set(&tx->tx_refcount, 1);
57 tx->tx_zc_aborted = 0;
58 tx->tx_zc_capable = 0;
59 tx->tx_zc_checked = 0;
60 tx->tx_desc_size = size;
62 atomic_inc(&ksocknal_data.ksnd_nactive_txs);
68 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
72 tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
74 CERROR("Can't allocate noop tx desc\n");
79 tx->tx_lnetmsg = NULL;
82 tx->tx_iov = tx->tx_frags.virt.iov;
84 tx->tx_nonblk = nonblk;
86 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
87 tx->tx_msg.ksm_zc_cookies[1] = cookie;
94 ksocknal_free_tx (ksock_tx_t *tx)
96 atomic_dec(&ksocknal_data.ksnd_nactive_txs);
98 if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
100 spin_lock(&ksocknal_data.ksnd_tx_lock);
102 cfs_list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
104 spin_unlock(&ksocknal_data.ksnd_tx_lock);
106 LIBCFS_FREE(tx, tx->tx_desc_size);
111 ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
113 struct iovec *iov = tx->tx_iov;
117 LASSERT (tx->tx_niov > 0);
119 /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
120 rc = ksocknal_lib_send_iov(conn, tx);
122 if (rc <= 0) /* sent nothing? */
126 LASSERT (nob <= tx->tx_resid);
131 LASSERT (tx->tx_niov > 0);
133 if (nob < (int) iov->iov_len) {
134 iov->iov_base = (void *)((char *)iov->iov_base + nob);
148 ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
150 lnet_kiov_t *kiov = tx->tx_kiov;
154 LASSERT (tx->tx_niov == 0);
155 LASSERT (tx->tx_nkiov > 0);
157 /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
158 rc = ksocknal_lib_send_kiov(conn, tx);
160 if (rc <= 0) /* sent nothing? */
164 LASSERT (nob <= tx->tx_resid);
169 LASSERT(tx->tx_nkiov > 0);
171 if (nob < (int)kiov->kiov_len) {
172 kiov->kiov_offset += nob;
173 kiov->kiov_len -= nob;
177 nob -= (int)kiov->kiov_len;
178 tx->tx_kiov = ++kiov;
186 ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
191 if (ksocknal_data.ksnd_stall_tx != 0) {
192 cfs_pause(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
195 LASSERT (tx->tx_resid != 0);
197 rc = ksocknal_connsock_addref(conn);
199 LASSERT (conn->ksnc_closing);
204 if (ksocknal_data.ksnd_enomem_tx > 0) {
206 ksocknal_data.ksnd_enomem_tx--;
208 } else if (tx->tx_niov != 0) {
209 rc = ksocknal_send_iov (conn, tx);
211 rc = ksocknal_send_kiov (conn, tx);
214 bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
215 if (rc > 0) /* sent something? */
216 conn->ksnc_tx_bufnob += rc; /* account it */
218 if (bufnob < conn->ksnc_tx_bufnob) {
219 /* allocated send buffer bytes < computed; infer
220 * something got ACKed */
221 conn->ksnc_tx_deadline =
222 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
223 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
224 conn->ksnc_tx_bufnob = bufnob;
228 if (rc <= 0) { /* Didn't write anything? */
230 if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
233 /* Check if EAGAIN is due to memory pressure */
234 if(rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
240 /* socket's wmem_queued now includes 'rc' bytes */
241 atomic_sub (rc, &conn->ksnc_tx_nob);
244 } while (tx->tx_resid != 0);
246 ksocknal_connsock_decref(conn);
251 ksocknal_recv_iov (ksock_conn_t *conn)
253 struct iovec *iov = conn->ksnc_rx_iov;
257 LASSERT (conn->ksnc_rx_niov > 0);
259 /* Never touch conn->ksnc_rx_iov or change connection
260 * status inside ksocknal_lib_recv_iov */
261 rc = ksocknal_lib_recv_iov(conn);
266 /* received something... */
269 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
270 conn->ksnc_rx_deadline =
271 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
272 smp_mb(); /* order with setting rx_started */
273 conn->ksnc_rx_started = 1;
275 conn->ksnc_rx_nob_wanted -= nob;
276 conn->ksnc_rx_nob_left -= nob;
279 LASSERT (conn->ksnc_rx_niov > 0);
281 if (nob < (int)iov->iov_len) {
283 iov->iov_base = (void *)((char *)iov->iov_base + nob);
288 conn->ksnc_rx_iov = ++iov;
289 conn->ksnc_rx_niov--;
296 ksocknal_recv_kiov (ksock_conn_t *conn)
298 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
301 LASSERT (conn->ksnc_rx_nkiov > 0);
303 /* Never touch conn->ksnc_rx_kiov or change connection
304 * status inside ksocknal_lib_recv_iov */
305 rc = ksocknal_lib_recv_kiov(conn);
310 /* received something... */
313 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
314 conn->ksnc_rx_deadline =
315 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
316 smp_mb(); /* order with setting rx_started */
317 conn->ksnc_rx_started = 1;
319 conn->ksnc_rx_nob_wanted -= nob;
320 conn->ksnc_rx_nob_left -= nob;
323 LASSERT (conn->ksnc_rx_nkiov > 0);
325 if (nob < (int) kiov->kiov_len) {
326 kiov->kiov_offset += nob;
327 kiov->kiov_len -= nob;
331 nob -= kiov->kiov_len;
332 conn->ksnc_rx_kiov = ++kiov;
333 conn->ksnc_rx_nkiov--;
340 ksocknal_receive (ksock_conn_t *conn)
342 /* Return 1 on success, 0 on EOF, < 0 on error.
343 * Caller checks ksnc_rx_nob_wanted to determine
344 * progress/completion. */
348 if (ksocknal_data.ksnd_stall_rx != 0) {
349 cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
352 rc = ksocknal_connsock_addref(conn);
354 LASSERT (conn->ksnc_closing);
359 if (conn->ksnc_rx_niov != 0)
360 rc = ksocknal_recv_iov (conn);
362 rc = ksocknal_recv_kiov (conn);
365 /* error/EOF or partial receive */
368 } else if (rc == 0 && conn->ksnc_rx_started) {
369 /* EOF in the middle of a message */
375 /* Completed a fragment */
377 if (conn->ksnc_rx_nob_wanted == 0) {
383 ksocknal_connsock_decref(conn);
388 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
390 lnet_msg_t *lnetmsg = tx->tx_lnetmsg;
391 int rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO;
394 LASSERT(ni != NULL || tx->tx_conn != NULL);
396 if (tx->tx_conn != NULL)
397 ksocknal_conn_decref(tx->tx_conn);
399 if (ni == NULL && tx->tx_conn != NULL)
400 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
402 ksocknal_free_tx (tx);
403 if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
404 lnet_finalize (ni, lnetmsg, rc);
410 ksocknal_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist, int error)
414 while (!cfs_list_empty (txlist)) {
415 tx = cfs_list_entry (txlist->next, ksock_tx_t, tx_list);
417 if (error && tx->tx_lnetmsg != NULL) {
418 CNETERR("Deleting packet type %d len %d %s->%s\n",
419 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
420 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
421 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
422 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
424 CNETERR("Deleting noop packet\n");
427 cfs_list_del (&tx->tx_list);
429 LASSERT (atomic_read(&tx->tx_refcount) == 1);
430 ksocknal_tx_done (ni, tx);
435 ksocknal_check_zc_req(ksock_tx_t *tx)
437 ksock_conn_t *conn = tx->tx_conn;
438 ksock_peer_t *peer = conn->ksnc_peer;
440 /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
441 * to ksnp_zc_req_list if some fragment of this message should be sent
442 * zero-copy. Our peer will send an ACK containing this cookie when
443 * she has received this message to tell us we can signal completion.
444 * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
445 * ksnp_zc_req_list. */
446 LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
447 LASSERT (tx->tx_zc_capable);
449 tx->tx_zc_checked = 1;
451 if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
452 !conn->ksnc_zc_capable)
455 /* assign cookie and queue tx to pending list, it will be released when
456 * a matching ack is received. See ksocknal_handle_zcack() */
458 ksocknal_tx_addref(tx);
460 spin_lock(&peer->ksnp_lock);
462 /* ZC_REQ is going to be pinned to the peer */
464 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
466 LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
468 tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
470 if (peer->ksnp_zc_next_cookie == 0)
471 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
473 cfs_list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
475 spin_unlock(&peer->ksnp_lock);
479 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
481 ksock_peer_t *peer = tx->tx_conn->ksnc_peer;
483 LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
484 LASSERT(tx->tx_zc_capable);
486 tx->tx_zc_checked = 0;
488 spin_lock(&peer->ksnp_lock);
490 if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
491 /* Not waiting for an ACK */
492 spin_unlock(&peer->ksnp_lock);
496 tx->tx_msg.ksm_zc_cookies[0] = 0;
497 cfs_list_del(&tx->tx_zc_list);
499 spin_unlock(&peer->ksnp_lock);
501 ksocknal_tx_decref(tx);
505 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
509 if (tx->tx_zc_capable && !tx->tx_zc_checked)
510 ksocknal_check_zc_req(tx);
512 rc = ksocknal_transmit (conn, tx);
514 CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
516 if (tx->tx_resid == 0) {
517 /* Sent everything OK */
529 counter++; /* exponential backoff warnings */
530 if ((counter & (-counter)) == counter)
531 CWARN("%u ENOMEM tx %p (%u allocated)\n",
532 counter, conn, atomic_read(&libcfs_kmemory));
534 /* Queue on ksnd_enomem_conns for retry after a timeout */
535 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
537 /* enomem list takes over scheduler's ref... */
538 LASSERT (conn->ksnc_tx_scheduled);
539 cfs_list_add_tail(&conn->ksnc_tx_list,
540 &ksocknal_data.ksnd_enomem_conns);
541 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
542 SOCKNAL_ENOMEM_RETRY),
543 ksocknal_data.ksnd_reaper_waketime))
544 wake_up(&ksocknal_data.ksnd_reaper_waitq);
546 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
553 if (!conn->ksnc_closing) {
556 LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection "
557 "while we were sending data; it may have "
559 HIPQUAD(conn->ksnc_ipaddr));
562 LCONSOLE_WARN("There was an unexpected network error "
563 "while writing to %u.%u.%u.%u: %d.\n",
564 HIPQUAD(conn->ksnc_ipaddr), rc);
567 CDEBUG(D_NET, "[%p] Error %d on write to %s"
568 " ip %d.%d.%d.%d:%d\n", conn, rc,
569 libcfs_id2str(conn->ksnc_peer->ksnp_id),
570 HIPQUAD(conn->ksnc_ipaddr),
574 if (tx->tx_zc_checked)
575 ksocknal_uncheck_zc_req(tx);
577 /* it's not an error if conn is being closed */
578 ksocknal_close_conn_and_siblings (conn,
579 (conn->ksnc_closing) ? 0 : rc);
585 ksocknal_launch_connection_locked (ksock_route_t *route)
588 /* called holding write lock on ksnd_global_lock */
590 LASSERT (!route->ksnr_scheduled);
591 LASSERT (!route->ksnr_connecting);
592 LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
594 route->ksnr_scheduled = 1; /* scheduling conn for connd */
595 ksocknal_route_addref(route); /* extra ref for connd */
597 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
599 cfs_list_add_tail(&route->ksnr_connd_list,
600 &ksocknal_data.ksnd_connd_routes);
601 wake_up(&ksocknal_data.ksnd_connd_waitq);
603 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
607 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
609 ksock_route_t *route;
611 /* called holding write lock on ksnd_global_lock */
613 /* launch any/all connections that need it */
614 route = ksocknal_find_connectable_route_locked(peer);
618 ksocknal_launch_connection_locked(route);
623 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
627 ksock_conn_t *typed = NULL;
628 ksock_conn_t *fallback = NULL;
632 cfs_list_for_each (tmp, &peer->ksnp_conns) {
633 ksock_conn_t *c = cfs_list_entry(tmp, ksock_conn_t, ksnc_list);
634 int nob = atomic_read(&c->ksnc_tx_nob) +
635 libcfs_sock_wmem_queued(c->ksnc_sock);
638 LASSERT (!c->ksnc_closing);
639 LASSERT (c->ksnc_proto != NULL &&
640 c->ksnc_proto->pro_match_tx != NULL);
642 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
647 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
650 case SOCKNAL_MATCH_YES: /* typed connection */
651 if (typed == NULL || tnob > nob ||
652 (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
653 cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
659 case SOCKNAL_MATCH_MAY: /* fallback connection */
660 if (fallback == NULL || fnob > nob ||
661 (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
662 cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
670 /* prefer the typed selection */
671 conn = (typed != NULL) ? typed : fallback;
674 conn->ksnc_tx_last_post = cfs_time_current();
680 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
682 conn->ksnc_proto->pro_pack(tx);
684 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
685 ksocknal_conn_addref(conn); /* +1 ref for tx */
690 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
692 ksock_sched_t *sched = conn->ksnc_scheduler;
693 ksock_msg_t *msg = &tx->tx_msg;
694 ksock_tx_t *ztx = NULL;
697 /* called holding global lock (read or irq-write) and caller may
698 * not have dropped this lock between finding conn and calling me,
699 * so we don't need the {get,put}connsock dance to deref
701 LASSERT(!conn->ksnc_closing);
703 CDEBUG (D_NET, "Sending to %s ip %d.%d.%d.%d:%d\n",
704 libcfs_id2str(conn->ksnc_peer->ksnp_id),
705 HIPQUAD(conn->ksnc_ipaddr),
708 ksocknal_tx_prep(conn, tx);
710 /* Ensure the frags we've been given EXACTLY match the number of
711 * bytes we want to send. Many TCP/IP stacks disregard any total
712 * size parameters passed to them and just look at the frags.
714 * We always expect at least 1 mapped fragment containing the
715 * complete ksocknal message header. */
716 LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
717 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
718 (unsigned int)tx->tx_nob);
719 LASSERT (tx->tx_niov >= 1);
720 LASSERT (tx->tx_resid == tx->tx_nob);
722 CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
723 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
725 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
728 * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
729 * but they're used inside spinlocks a lot.
731 bufnob = libcfs_sock_wmem_queued(conn->ksnc_sock);
732 spin_lock_bh(&sched->kss_lock);
734 if (cfs_list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
735 /* First packet starts the timeout */
736 conn->ksnc_tx_deadline =
737 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
738 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
739 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
740 conn->ksnc_tx_bufnob = 0;
741 smp_mb(); /* order with adding to tx_queue */
744 if (msg->ksm_type == KSOCK_MSG_NOOP) {
745 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
746 * on a normal packet so I don't need to send it */
747 LASSERT (msg->ksm_zc_cookies[1] != 0);
748 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
750 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
751 ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
754 /* It's a normal packet - can it piggback a noop zc-ack that
755 * has been queued already? */
756 LASSERT (msg->ksm_zc_cookies[1] == 0);
757 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
759 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
760 /* ztx will be released later */
764 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
765 cfs_list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
768 if (conn->ksnc_tx_ready && /* able to send */
769 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
770 /* +1 ref for scheduler */
771 ksocknal_conn_addref(conn);
772 cfs_list_add_tail (&conn->ksnc_tx_list,
773 &sched->kss_tx_conns);
774 conn->ksnc_tx_scheduled = 1;
775 wake_up(&sched->kss_waitq);
778 spin_unlock_bh(&sched->kss_lock);
783 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
785 cfs_time_t now = cfs_time_current();
787 ksock_route_t *route;
789 cfs_list_for_each (tmp, &peer->ksnp_routes) {
790 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
792 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
794 if (route->ksnr_scheduled) /* connections being established */
797 /* all route types connected ? */
798 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
801 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
802 cfs_time_aftereq(now, route->ksnr_timeout))) {
804 "Too soon to retry route %u.%u.%u.%u "
805 "(cnted %d, interval %ld, %ld secs later)\n",
806 HIPQUAD(route->ksnr_ipaddr),
807 route->ksnr_connected,
808 route->ksnr_retry_interval,
809 cfs_duration_sec(route->ksnr_timeout - now));
820 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
823 ksock_route_t *route;
825 cfs_list_for_each (tmp, &peer->ksnp_routes) {
826 route = cfs_list_entry (tmp, ksock_route_t, ksnr_list);
828 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
830 if (route->ksnr_scheduled)
838 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
846 LASSERT (tx->tx_conn == NULL);
848 g_lock = &ksocknal_data.ksnd_global_lock;
850 for (retry = 0;; retry = 1) {
852 peer = ksocknal_find_peer_locked(ni, id);
854 if (ksocknal_find_connectable_route_locked(peer) == NULL) {
855 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
857 /* I've got no routes that need to be
858 * connecting and I do have an actual
860 ksocknal_queue_tx_locked (tx, conn);
867 /* I'll need a write lock... */
870 write_lock_bh(g_lock);
872 peer = ksocknal_find_peer_locked(ni, id);
876 write_unlock_bh(g_lock);
878 if ((id.pid & LNET_PID_USERFLAG) != 0) {
879 CERROR("Refusing to create a connection to "
880 "userspace process %s\n", libcfs_id2str(id));
881 return -EHOSTUNREACH;
885 CERROR("Can't find peer %s\n", libcfs_id2str(id));
886 return -EHOSTUNREACH;
889 rc = ksocknal_add_peer(ni, id,
890 LNET_NIDADDR(id.nid),
891 lnet_acceptor_port());
893 CERROR("Can't add peer %s: %d\n",
894 libcfs_id2str(id), rc);
899 ksocknal_launch_all_connections_locked(peer);
901 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
903 /* Connection exists; queue message on it */
904 ksocknal_queue_tx_locked (tx, conn);
905 write_unlock_bh(g_lock);
909 if (peer->ksnp_accepting > 0 ||
910 ksocknal_find_connecting_route_locked (peer) != NULL) {
911 /* the message is going to be pinned to the peer */
913 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
915 /* Queue the message until a connection is established */
916 cfs_list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
917 write_unlock_bh(g_lock);
921 write_unlock_bh(g_lock);
923 /* NB Routes may be ignored if connections to them failed recently */
924 CNETERR("No usable routes to %s\n", libcfs_id2str(id));
925 return (-EHOSTUNREACH);
929 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
932 int type = lntmsg->msg_type;
933 lnet_process_id_t target = lntmsg->msg_target;
934 unsigned int payload_niov = lntmsg->msg_niov;
935 struct iovec *payload_iov = lntmsg->msg_iov;
936 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
937 unsigned int payload_offset = lntmsg->msg_offset;
938 unsigned int payload_nob = lntmsg->msg_len;
943 /* NB 'private' is different depending on what we're sending.
944 * Just ignore it... */
946 CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
947 payload_nob, payload_niov, libcfs_id2str(target));
949 LASSERT (payload_nob == 0 || payload_niov > 0);
950 LASSERT (payload_niov <= LNET_MAX_IOV);
951 /* payload is either all vaddrs or all pages */
952 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
953 LASSERT (!in_interrupt ());
955 if (payload_iov != NULL)
956 desc_size = offsetof(ksock_tx_t,
957 tx_frags.virt.iov[1 + payload_niov]);
959 desc_size = offsetof(ksock_tx_t,
960 tx_frags.paged.kiov[payload_niov]);
962 if (lntmsg->msg_vmflush)
963 mpflag = cfs_memory_pressure_get_and_set();
964 tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
966 CERROR("Can't allocate tx desc type %d size %d\n",
968 if (lntmsg->msg_vmflush)
969 cfs_memory_pressure_restore(mpflag);
973 tx->tx_conn = NULL; /* set when assigned a conn */
974 tx->tx_lnetmsg = lntmsg;
976 if (payload_iov != NULL) {
979 tx->tx_iov = tx->tx_frags.virt.iov;
981 lnet_extract_iov(payload_niov, &tx->tx_iov[1],
982 payload_niov, payload_iov,
983 payload_offset, payload_nob);
986 tx->tx_iov = &tx->tx_frags.paged.iov;
987 tx->tx_kiov = tx->tx_frags.paged.kiov;
988 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
989 payload_niov, payload_kiov,
990 payload_offset, payload_nob);
992 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
993 tx->tx_zc_capable = 1;
996 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
998 /* The first fragment will be set later in pro_pack */
999 rc = ksocknal_launch_packet(ni, tx, target);
1001 cfs_memory_pressure_restore(mpflag);
1006 ksocknal_free_tx(tx);
1011 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1013 struct task_struct *task = kthread_run(fn, arg, name);
1016 return PTR_ERR(task);
1018 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1019 ksocknal_data.ksnd_nthreads++;
1020 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1025 ksocknal_thread_fini (void)
1027 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1028 ksocknal_data.ksnd_nthreads--;
1029 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1033 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1035 static char ksocknal_slop_buffer[4096];
1041 LASSERT(conn->ksnc_proto != NULL);
1043 if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1044 /* Remind the socket to ack eagerly... */
1045 ksocknal_lib_eager_ack(conn);
1048 if (nob_to_skip == 0) { /* right at next packet boundary now */
1049 conn->ksnc_rx_started = 0;
1050 smp_mb(); /* racing with timeout thread */
1052 switch (conn->ksnc_proto->pro_version) {
1053 case KSOCK_PROTO_V2:
1054 case KSOCK_PROTO_V3:
1055 conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1056 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1057 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1059 conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1060 conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1061 conn->ksnc_rx_iov[0].iov_len = offsetof(ksock_msg_t, ksm_u);
1064 case KSOCK_PROTO_V1:
1065 /* Receiving bare lnet_hdr_t */
1066 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1067 conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1068 conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1070 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1071 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1072 conn->ksnc_rx_iov[0].iov_len = sizeof (lnet_hdr_t);
1078 conn->ksnc_rx_niov = 1;
1080 conn->ksnc_rx_kiov = NULL;
1081 conn->ksnc_rx_nkiov = 0;
1082 conn->ksnc_rx_csum = ~0;
1086 /* Set up to skip as much as possible now. If there's more left
1087 * (ran out of iov entries) we'll get called again */
1089 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1090 conn->ksnc_rx_nob_left = nob_to_skip;
1091 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1096 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1098 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1099 conn->ksnc_rx_iov[niov].iov_len = nob;
1104 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1105 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1107 conn->ksnc_rx_niov = niov;
1108 conn->ksnc_rx_kiov = NULL;
1109 conn->ksnc_rx_nkiov = 0;
1110 conn->ksnc_rx_nob_wanted = skipped;
1115 ksocknal_process_receive (ksock_conn_t *conn)
1118 lnet_process_id_t *id;
1121 LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1123 /* NB: sched lock NOT held */
1124 /* SOCKNAL_RX_LNET_HEADER is here for backward compatability */
1125 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1126 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1127 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1128 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1130 if (conn->ksnc_rx_nob_wanted != 0) {
1131 rc = ksocknal_receive(conn);
1134 LASSERT (rc != -EAGAIN);
1137 CDEBUG (D_NET, "[%p] EOF from %s"
1138 " ip %d.%d.%d.%d:%d\n", conn,
1139 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1140 HIPQUAD(conn->ksnc_ipaddr),
1142 else if (!conn->ksnc_closing)
1143 CERROR ("[%p] Error %d on read from %s"
1144 " ip %d.%d.%d.%d:%d\n",
1146 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1147 HIPQUAD(conn->ksnc_ipaddr),
1150 /* it's not an error if conn is being closed */
1151 ksocknal_close_conn_and_siblings (conn,
1152 (conn->ksnc_closing) ? 0 : rc);
1153 return (rc == 0 ? -ESHUTDOWN : rc);
1156 if (conn->ksnc_rx_nob_wanted != 0) {
1161 switch (conn->ksnc_rx_state) {
1162 case SOCKNAL_RX_KSM_HEADER:
1163 if (conn->ksnc_flip) {
1164 __swab32s(&conn->ksnc_msg.ksm_type);
1165 __swab32s(&conn->ksnc_msg.ksm_csum);
1166 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1167 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1170 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1171 conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1172 CERROR("%s: Unknown message type: %x\n",
1173 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1174 conn->ksnc_msg.ksm_type);
1175 ksocknal_new_packet(conn, 0);
1176 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1180 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1181 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1182 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1183 /* NOOP Checksum error */
1184 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1185 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1186 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1187 ksocknal_new_packet(conn, 0);
1188 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1192 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1195 LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1197 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1198 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1200 rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1201 conn->ksnc_msg.ksm_zc_cookies[1]);
1204 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1205 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1206 cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1207 ksocknal_new_packet(conn, 0);
1208 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1213 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1214 ksocknal_new_packet (conn, 0);
1215 return 0; /* NOOP is done and just return */
1218 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1219 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1220 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1222 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1223 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1224 conn->ksnc_rx_iov[0].iov_len = sizeof(ksock_lnet_msg_t);
1226 conn->ksnc_rx_niov = 1;
1227 conn->ksnc_rx_kiov = NULL;
1228 conn->ksnc_rx_nkiov = 0;
1230 goto again; /* read lnet header now */
1232 case SOCKNAL_RX_LNET_HEADER:
1233 /* unpack message header */
1234 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1236 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1237 /* Userspace peer */
1238 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1239 id = &conn->ksnc_peer->ksnp_id;
1241 /* Substitute process ID assigned at connection time */
1242 lhdr->src_pid = cpu_to_le32(id->pid);
1243 lhdr->src_nid = cpu_to_le64(id->nid);
1246 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1247 ksocknal_conn_addref(conn); /* ++ref while parsing */
1249 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1250 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1251 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1253 /* I just received garbage: give up on this conn */
1254 ksocknal_new_packet(conn, 0);
1255 ksocknal_close_conn_and_siblings (conn, rc);
1256 ksocknal_conn_decref(conn);
1260 /* I'm racing with ksocknal_recv() */
1261 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1262 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1264 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1267 /* ksocknal_recv() got called */
1270 case SOCKNAL_RX_LNET_PAYLOAD:
1271 /* payload all received */
1274 if (conn->ksnc_rx_nob_left == 0 && /* not truncating */
1275 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1276 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1277 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1278 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1279 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1283 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1284 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1286 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1287 id = &conn->ksnc_peer->ksnp_id;
1289 rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1290 conn->ksnc_msg.ksm_zc_cookies[0],
1291 *ksocknal_tunables.ksnd_nonblk_zcack ||
1292 le64_to_cpu(lhdr->src_nid) != id->nid);
1295 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1298 ksocknal_new_packet(conn, 0);
1299 ksocknal_close_conn_and_siblings (conn, rc);
1304 case SOCKNAL_RX_SLOP:
1305 /* starting new packet? */
1306 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1307 return 0; /* come back later */
1308 goto again; /* try to finish reading slop now */
1316 return (-EINVAL); /* keep gcc happy */
1320 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1321 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1322 unsigned int offset, unsigned int mlen, unsigned int rlen)
1324 ksock_conn_t *conn = (ksock_conn_t *)private;
1325 ksock_sched_t *sched = conn->ksnc_scheduler;
1327 LASSERT (mlen <= rlen);
1328 LASSERT (niov <= LNET_MAX_IOV);
1330 conn->ksnc_cookie = msg;
1331 conn->ksnc_rx_nob_wanted = mlen;
1332 conn->ksnc_rx_nob_left = rlen;
1334 if (mlen == 0 || iov != NULL) {
1335 conn->ksnc_rx_nkiov = 0;
1336 conn->ksnc_rx_kiov = NULL;
1337 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1338 conn->ksnc_rx_niov =
1339 lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1340 niov, iov, offset, mlen);
1342 conn->ksnc_rx_niov = 0;
1343 conn->ksnc_rx_iov = NULL;
1344 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1345 conn->ksnc_rx_nkiov =
1346 lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1347 niov, kiov, offset, mlen);
1351 lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1352 lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1354 LASSERT (conn->ksnc_rx_scheduled);
1356 spin_lock_bh(&sched->kss_lock);
1358 switch (conn->ksnc_rx_state) {
1359 case SOCKNAL_RX_PARSE_WAIT:
1360 cfs_list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1361 wake_up(&sched->kss_waitq);
1362 LASSERT(conn->ksnc_rx_ready);
1365 case SOCKNAL_RX_PARSE:
1366 /* scheduler hasn't noticed I'm parsing yet */
1370 conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1372 spin_unlock_bh(&sched->kss_lock);
1373 ksocknal_conn_decref(conn);
1378 ksocknal_sched_cansleep(ksock_sched_t *sched)
1382 spin_lock_bh(&sched->kss_lock);
1384 rc = (!ksocknal_data.ksnd_shuttingdown &&
1385 cfs_list_empty(&sched->kss_rx_conns) &&
1386 cfs_list_empty(&sched->kss_tx_conns));
1388 spin_unlock_bh(&sched->kss_lock);
1392 int ksocknal_scheduler(void *arg)
1394 struct ksock_sched_info *info;
1395 ksock_sched_t *sched;
1400 long id = (long)arg;
1402 info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1403 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1405 cfs_block_allsigs();
1407 rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1409 CERROR("Can't set CPT affinity to %d: %d\n",
1413 spin_lock_bh(&sched->kss_lock);
1415 while (!ksocknal_data.ksnd_shuttingdown) {
1416 int did_something = 0;
1418 /* Ensure I progress everything semi-fairly */
1420 if (!cfs_list_empty (&sched->kss_rx_conns)) {
1421 conn = cfs_list_entry(sched->kss_rx_conns.next,
1422 ksock_conn_t, ksnc_rx_list);
1423 cfs_list_del(&conn->ksnc_rx_list);
1425 LASSERT(conn->ksnc_rx_scheduled);
1426 LASSERT(conn->ksnc_rx_ready);
1428 /* clear rx_ready in case receive isn't complete.
1429 * Do it BEFORE we call process_recv, since
1430 * data_ready can set it any time after we release
1432 conn->ksnc_rx_ready = 0;
1433 spin_unlock_bh(&sched->kss_lock);
1435 rc = ksocknal_process_receive(conn);
1437 spin_lock_bh(&sched->kss_lock);
1439 /* I'm the only one that can clear this flag */
1440 LASSERT(conn->ksnc_rx_scheduled);
1442 /* Did process_receive get everything it wanted? */
1444 conn->ksnc_rx_ready = 1;
1446 if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1447 /* Conn blocked waiting for ksocknal_recv()
1448 * I change its state (under lock) to signal
1449 * it can be rescheduled */
1450 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1451 } else if (conn->ksnc_rx_ready) {
1452 /* reschedule for rx */
1453 cfs_list_add_tail (&conn->ksnc_rx_list,
1454 &sched->kss_rx_conns);
1456 conn->ksnc_rx_scheduled = 0;
1458 ksocknal_conn_decref(conn);
1464 if (!cfs_list_empty (&sched->kss_tx_conns)) {
1465 CFS_LIST_HEAD (zlist);
1467 if (!cfs_list_empty(&sched->kss_zombie_noop_txs)) {
1468 cfs_list_add(&zlist,
1469 &sched->kss_zombie_noop_txs);
1470 cfs_list_del_init(&sched->kss_zombie_noop_txs);
1473 conn = cfs_list_entry(sched->kss_tx_conns.next,
1474 ksock_conn_t, ksnc_tx_list);
1475 cfs_list_del (&conn->ksnc_tx_list);
1477 LASSERT(conn->ksnc_tx_scheduled);
1478 LASSERT(conn->ksnc_tx_ready);
1479 LASSERT(!cfs_list_empty(&conn->ksnc_tx_queue));
1481 tx = cfs_list_entry(conn->ksnc_tx_queue.next,
1482 ksock_tx_t, tx_list);
1484 if (conn->ksnc_tx_carrier == tx)
1485 ksocknal_next_tx_carrier(conn);
1487 /* dequeue now so empty list => more to send */
1488 cfs_list_del(&tx->tx_list);
1490 /* Clear tx_ready in case send isn't complete. Do
1491 * it BEFORE we call process_transmit, since
1492 * write_space can set it any time after we release
1494 conn->ksnc_tx_ready = 0;
1495 spin_unlock_bh(&sched->kss_lock);
1497 if (!cfs_list_empty(&zlist)) {
1498 /* free zombie noop txs, it's fast because
1499 * noop txs are just put in freelist */
1500 ksocknal_txlist_done(NULL, &zlist, 0);
1503 rc = ksocknal_process_transmit(conn, tx);
1505 if (rc == -ENOMEM || rc == -EAGAIN) {
1506 /* Incomplete send: replace tx on HEAD of tx_queue */
1507 spin_lock_bh(&sched->kss_lock);
1508 cfs_list_add(&tx->tx_list,
1509 &conn->ksnc_tx_queue);
1511 /* Complete send; tx -ref */
1512 ksocknal_tx_decref(tx);
1514 spin_lock_bh(&sched->kss_lock);
1515 /* assume space for more */
1516 conn->ksnc_tx_ready = 1;
1519 if (rc == -ENOMEM) {
1520 /* Do nothing; after a short timeout, this
1521 * conn will be reposted on kss_tx_conns. */
1522 } else if (conn->ksnc_tx_ready &&
1523 !cfs_list_empty (&conn->ksnc_tx_queue)) {
1524 /* reschedule for tx */
1525 cfs_list_add_tail (&conn->ksnc_tx_list,
1526 &sched->kss_tx_conns);
1528 conn->ksnc_tx_scheduled = 0;
1530 ksocknal_conn_decref(conn);
1535 if (!did_something || /* nothing to do */
1536 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1537 spin_unlock_bh(&sched->kss_lock);
1541 if (!did_something) { /* wait for something to do */
1542 rc = wait_event_interruptible_exclusive(
1544 !ksocknal_sched_cansleep(sched));
1550 spin_lock_bh(&sched->kss_lock);
1554 spin_unlock_bh(&sched->kss_lock);
1555 ksocknal_thread_fini();
1560 * Add connection to kss_rx_conns of scheduler
1561 * and wakeup the scheduler.
1563 void ksocknal_read_callback (ksock_conn_t *conn)
1565 ksock_sched_t *sched;
1568 sched = conn->ksnc_scheduler;
1570 spin_lock_bh(&sched->kss_lock);
1572 conn->ksnc_rx_ready = 1;
1574 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1575 cfs_list_add_tail(&conn->ksnc_rx_list,
1576 &sched->kss_rx_conns);
1577 conn->ksnc_rx_scheduled = 1;
1578 /* extra ref for scheduler */
1579 ksocknal_conn_addref(conn);
1581 wake_up (&sched->kss_waitq);
1583 spin_unlock_bh(&sched->kss_lock);
1589 * Add connection to kss_tx_conns of scheduler
1590 * and wakeup the scheduler.
1592 void ksocknal_write_callback(ksock_conn_t *conn)
1594 ksock_sched_t *sched;
1597 sched = conn->ksnc_scheduler;
1599 spin_lock_bh(&sched->kss_lock);
1601 conn->ksnc_tx_ready = 1;
1603 if (!conn->ksnc_tx_scheduled && // not being progressed
1604 !cfs_list_empty(&conn->ksnc_tx_queue)){//packets to send
1605 cfs_list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1606 conn->ksnc_tx_scheduled = 1;
1607 /* extra ref for scheduler */
1608 ksocknal_conn_addref(conn);
1610 wake_up(&sched->kss_waitq);
1613 spin_unlock_bh(&sched->kss_lock);
1619 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1623 if (hello->kshm_magic == LNET_PROTO_MAGIC)
1624 version = hello->kshm_version;
1625 else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1626 version = __swab32(hello->kshm_version);
1629 #if SOCKNAL_VERSION_DEBUG
1630 if (*ksocknal_tunables.ksnd_protocol == 1)
1633 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1634 version == KSOCK_PROTO_V3)
1637 if (version == KSOCK_PROTO_V2)
1638 return &ksocknal_protocol_v2x;
1640 if (version == KSOCK_PROTO_V3)
1641 return &ksocknal_protocol_v3x;
1646 if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1647 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1649 CLASSERT (sizeof (lnet_magicversion_t) ==
1650 offsetof (ksock_hello_msg_t, kshm_src_nid));
1652 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1653 hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1654 return &ksocknal_protocol_v1x;
1661 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1662 lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1664 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1665 ksock_net_t *net = (ksock_net_t *)ni->ni_data;
1667 LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1669 /* rely on caller to hold a ref on socket so it wouldn't disappear */
1670 LASSERT (conn->ksnc_proto != NULL);
1672 hello->kshm_src_nid = ni->ni_nid;
1673 hello->kshm_dst_nid = peer_nid;
1674 hello->kshm_src_pid = the_lnet.ln_pid;
1676 hello->kshm_src_incarnation = net->ksnn_incarnation;
1677 hello->kshm_ctype = conn->ksnc_type;
1679 return conn->ksnc_proto->pro_send_hello(conn, hello);
1683 ksocknal_invert_type(int type)
1687 case SOCKLND_CONN_ANY:
1688 case SOCKLND_CONN_CONTROL:
1690 case SOCKLND_CONN_BULK_IN:
1691 return SOCKLND_CONN_BULK_OUT;
1692 case SOCKLND_CONN_BULK_OUT:
1693 return SOCKLND_CONN_BULK_IN;
1695 return (SOCKLND_CONN_NONE);
1700 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1701 ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1704 /* Return < 0 fatal error
1706 * EALREADY lost connection race
1707 * EPROTO protocol version mismatch
1709 cfs_socket_t *sock = conn->ksnc_sock;
1710 int active = (conn->ksnc_proto != NULL);
1714 ksock_proto_t *proto;
1715 lnet_process_id_t recv_id;
1717 /* socket type set on active connections - not set on passive */
1718 LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1720 timeout = active ? *ksocknal_tunables.ksnd_timeout :
1721 lnet_acceptor_timeout();
1723 rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1725 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1726 rc, HIPQUAD(conn->ksnc_ipaddr));
1731 if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1732 hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1733 hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1734 /* Unexpected magic! */
1735 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1736 "%u.%u.%u.%u\n", __cpu_to_le32 (hello->kshm_magic),
1737 LNET_PROTO_TCP_MAGIC,
1738 HIPQUAD(conn->ksnc_ipaddr));
1742 rc = libcfs_sock_read(sock, &hello->kshm_version,
1743 sizeof(hello->kshm_version), timeout);
1745 CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
1746 rc, HIPQUAD(conn->ksnc_ipaddr));
1751 proto = ksocknal_parse_proto_version(hello);
1752 if (proto == NULL) {
1754 /* unknown protocol from peer, tell peer my protocol */
1755 conn->ksnc_proto = &ksocknal_protocol_v3x;
1756 #if SOCKNAL_VERSION_DEBUG
1757 if (*ksocknal_tunables.ksnd_protocol == 2)
1758 conn->ksnc_proto = &ksocknal_protocol_v2x;
1759 else if (*ksocknal_tunables.ksnd_protocol == 1)
1760 conn->ksnc_proto = &ksocknal_protocol_v1x;
1762 hello->kshm_nips = 0;
1763 ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1766 CERROR ("Unknown protocol version (%d.x expected)"
1767 " from %u.%u.%u.%u\n",
1768 conn->ksnc_proto->pro_version,
1769 HIPQUAD(conn->ksnc_ipaddr));
1774 proto_match = (conn->ksnc_proto == proto);
1775 conn->ksnc_proto = proto;
1777 /* receive the rest of hello message anyway */
1778 rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1780 CERROR("Error %d reading or checking hello from from %u.%u.%u.%u\n",
1781 rc, HIPQUAD(conn->ksnc_ipaddr));
1786 *incarnation = hello->kshm_src_incarnation;
1788 if (hello->kshm_src_nid == LNET_NID_ANY) {
1789 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1790 "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
1795 conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1796 /* Userspace NAL assigns peer process ID from socket */
1797 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1798 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1800 recv_id.nid = hello->kshm_src_nid;
1801 recv_id.pid = hello->kshm_src_pid;
1807 /* peer determines type */
1808 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1809 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1810 CERROR ("Unexpected type %d from %s ip %u.%u.%u.%u\n",
1811 hello->kshm_ctype, libcfs_id2str(*peerid),
1812 HIPQUAD(conn->ksnc_ipaddr));
1819 if (peerid->pid != recv_id.pid ||
1820 peerid->nid != recv_id.nid) {
1821 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1822 " %u.%u.%u.%u, but they claimed they were "
1823 "%s; please check your Lustre "
1825 libcfs_id2str(*peerid),
1826 HIPQUAD(conn->ksnc_ipaddr),
1827 libcfs_id2str(recv_id));
1831 if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1832 /* Possible protocol mismatch or I lost the connection race */
1833 return proto_match ? EALREADY : EPROTO;
1836 if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1837 CERROR ("Mismatched types: me %d, %s ip %u.%u.%u.%u %d\n",
1838 conn->ksnc_type, libcfs_id2str(*peerid),
1839 HIPQUAD(conn->ksnc_ipaddr),
1848 ksocknal_connect (ksock_route_t *route)
1850 CFS_LIST_HEAD (zombies);
1851 ksock_peer_t *peer = route->ksnr_peer;
1855 cfs_time_t deadline;
1856 int retry_later = 0;
1859 deadline = cfs_time_add(cfs_time_current(),
1860 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1862 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1864 LASSERT (route->ksnr_scheduled);
1865 LASSERT (!route->ksnr_connecting);
1867 route->ksnr_connecting = 1;
1870 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1872 /* stop connecting if peer/route got closed under me, or
1873 * route got connected while queued */
1874 if (peer->ksnp_closing || route->ksnr_deleted ||
1880 /* reschedule if peer is connecting to me */
1881 if (peer->ksnp_accepting > 0) {
1883 "peer %s(%d) already connecting to me, retry later.\n",
1884 libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1888 if (retry_later) /* needs reschedule */
1891 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1892 type = SOCKLND_CONN_ANY;
1893 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1894 type = SOCKLND_CONN_CONTROL;
1895 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1896 type = SOCKLND_CONN_BULK_IN;
1898 LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1899 type = SOCKLND_CONN_BULK_OUT;
1902 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1904 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1906 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1912 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1913 route->ksnr_myipaddr,
1914 route->ksnr_ipaddr, route->ksnr_port);
1918 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1920 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1926 /* A +ve RC means I have to retry because I lost the connection
1927 * race or I have to renegotiate protocol version */
1928 retry_later = (rc != 0);
1930 CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1931 libcfs_nid2str(peer->ksnp_id.nid));
1933 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1936 route->ksnr_scheduled = 0;
1937 route->ksnr_connecting = 0;
1940 /* re-queue for attention; this frees me up to handle
1941 * the peer's incoming connection request */
1943 if (rc == EALREADY ||
1944 (rc == 0 && peer->ksnp_accepting > 0)) {
1945 /* We want to introduce a delay before next
1946 * attempt to connect if we lost conn race,
1947 * but the race is resolved quickly usually,
1948 * so min_reconnectms should be good heuristic */
1949 route->ksnr_retry_interval =
1950 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1951 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1952 route->ksnr_retry_interval);
1955 ksocknal_launch_connection_locked(route);
1958 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1962 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1964 route->ksnr_scheduled = 0;
1965 route->ksnr_connecting = 0;
1967 /* This is a retry rather than a new connection */
1968 route->ksnr_retry_interval *= 2;
1969 route->ksnr_retry_interval =
1970 MAX(route->ksnr_retry_interval,
1971 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1972 route->ksnr_retry_interval =
1973 MIN(route->ksnr_retry_interval,
1974 cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1976 LASSERT (route->ksnr_retry_interval != 0);
1977 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1978 route->ksnr_retry_interval);
1980 if (!cfs_list_empty(&peer->ksnp_tx_queue) &&
1981 peer->ksnp_accepting == 0 &&
1982 ksocknal_find_connecting_route_locked(peer) == NULL) {
1985 /* ksnp_tx_queue is queued on a conn on successful
1986 * connection for V1.x and V2.x */
1987 if (!cfs_list_empty (&peer->ksnp_conns)) {
1988 conn = cfs_list_entry(peer->ksnp_conns.next,
1989 ksock_conn_t, ksnc_list);
1990 LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1993 /* take all the blocked packets while I've got the lock and
1994 * complete below... */
1995 cfs_list_splice_init(&peer->ksnp_tx_queue, &zombies);
1998 #if 0 /* irrelevent with only eager routes */
1999 if (!route->ksnr_deleted) {
2000 /* make this route least-favourite for re-selection */
2001 cfs_list_del(&route->ksnr_list);
2002 cfs_list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
2005 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2007 ksocknal_peer_failed(peer);
2008 ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2013 * check whether we need to create more connds.
2014 * It will try to create new thread if it's necessary, @timeout can
2015 * be updated if failed to create, so caller wouldn't keep try while
2016 * running out of resource.
2019 ksocknal_connd_check_start(long sec, long *timeout)
2023 int total = ksocknal_data.ksnd_connd_starting +
2024 ksocknal_data.ksnd_connd_running;
2026 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2027 /* still in initializing */
2031 if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2032 total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2033 /* can't create more connd, or still have enough
2034 * threads to handle more connecting */
2038 if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2039 /* no pending connecting request */
2043 if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2044 /* may run out of resource, retry later */
2045 *timeout = cfs_time_seconds(1);
2049 if (ksocknal_data.ksnd_connd_starting > 0) {
2050 /* serialize starting to avoid flood */
2054 ksocknal_data.ksnd_connd_starting_stamp = sec;
2055 ksocknal_data.ksnd_connd_starting++;
2056 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2058 /* NB: total is the next id */
2059 snprintf(name, sizeof(name), "socknal_cd%02d", total);
2060 rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2062 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2067 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2068 ksocknal_data.ksnd_connd_starting--;
2069 ksocknal_data.ksnd_connd_failed_stamp = cfs_time_current_sec();
2075 * check whether current thread can exit, it will return 1 if there are too
2076 * many threads and no creating in past 120 seconds.
2077 * Also, this function may update @timeout to make caller come back
2078 * again to recheck these conditions.
2081 ksocknal_connd_check_stop(long sec, long *timeout)
2085 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2086 /* still in initializing */
2090 if (ksocknal_data.ksnd_connd_starting > 0) {
2091 /* in progress of starting new thread */
2095 if (ksocknal_data.ksnd_connd_running <=
2096 *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2100 /* created thread in past 120 seconds? */
2101 val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2102 SOCKNAL_CONND_TIMEOUT - sec);
2104 *timeout = (val > 0) ? cfs_time_seconds(val) :
2105 cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2109 /* no creating in past 120 seconds */
2111 return ksocknal_data.ksnd_connd_running >
2112 ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2115 /* Go through connd_routes queue looking for a route that we can process
2116 * right now, @timeout_p can be updated if we need to come back later */
2117 static ksock_route_t *
2118 ksocknal_connd_get_route_locked(signed long *timeout_p)
2120 ksock_route_t *route;
2123 now = cfs_time_current();
2125 /* connd_routes can contain both pending and ordinary routes */
2126 cfs_list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2129 if (route->ksnr_retry_interval == 0 ||
2130 cfs_time_aftereq(now, route->ksnr_timeout))
2133 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2134 (int)*timeout_p > (int)(route->ksnr_timeout - now))
2135 *timeout_p = (int)(route->ksnr_timeout - now);
2142 ksocknal_connd (void *arg)
2144 spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2145 ksock_connreq_t *cr;
2150 cfs_block_allsigs ();
2152 init_waitqueue_entry_current(&wait);
2154 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 = MAX_SCHEDULE_TIMEOUT;
2164 int dropped_lock = 0;
2166 if (ksocknal_connd_check_stop(sec, &timeout)) {
2167 /* wakeup another one to check stop */
2168 wake_up(&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 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 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 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 spin_lock_bh(connd_lock);
2222 ksocknal_data.ksnd_connd_connecting--;
2226 if (++nloops < SOCKNAL_RESCHED)
2228 spin_unlock_bh(connd_lock);
2231 spin_lock_bh(connd_lock);
2235 /* Nothing to do for 'timeout' */
2236 set_current_state(TASK_INTERRUPTIBLE);
2237 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2238 spin_unlock_bh(connd_lock);
2241 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
2243 set_current_state(TASK_RUNNING);
2244 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2245 spin_lock_bh(connd_lock);
2247 ksocknal_data.ksnd_connd_running--;
2248 spin_unlock_bh(connd_lock);
2250 ksocknal_thread_fini();
2255 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2257 /* We're called with a shared lock on ksnd_global_lock */
2261 cfs_list_for_each (ctmp, &peer->ksnp_conns) {
2263 conn = cfs_list_entry (ctmp, ksock_conn_t, ksnc_list);
2265 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2266 LASSERT (!conn->ksnc_closing);
2268 /* SOCK_ERROR will reset error code of socket in
2269 * some platform (like Darwin8.x) */
2270 error = libcfs_sock_error(conn->ksnc_sock);
2272 ksocknal_conn_addref(conn);
2276 CNETERR("A connection with %s "
2277 "(%u.%u.%u.%u:%d) was reset; "
2278 "it may have rebooted.\n",
2279 libcfs_id2str(peer->ksnp_id),
2280 HIPQUAD(conn->ksnc_ipaddr),
2284 CNETERR("A connection with %s "
2285 "(%u.%u.%u.%u:%d) timed out; the "
2286 "network or node may be down.\n",
2287 libcfs_id2str(peer->ksnp_id),
2288 HIPQUAD(conn->ksnc_ipaddr),
2292 CNETERR("An unexpected network error %d "
2294 "(%u.%u.%u.%u:%d\n", error,
2295 libcfs_id2str(peer->ksnp_id),
2296 HIPQUAD(conn->ksnc_ipaddr),
2304 if (conn->ksnc_rx_started &&
2305 cfs_time_aftereq(cfs_time_current(),
2306 conn->ksnc_rx_deadline)) {
2307 /* Timed out incomplete incoming message */
2308 ksocknal_conn_addref(conn);
2309 CNETERR("Timeout receiving from %s (%u.%u.%u.%u:%d), "
2310 "state %d wanted %d left %d\n",
2311 libcfs_id2str(peer->ksnp_id),
2312 HIPQUAD(conn->ksnc_ipaddr),
2314 conn->ksnc_rx_state,
2315 conn->ksnc_rx_nob_wanted,
2316 conn->ksnc_rx_nob_left);
2320 if ((!cfs_list_empty(&conn->ksnc_tx_queue) ||
2321 libcfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2322 cfs_time_aftereq(cfs_time_current(),
2323 conn->ksnc_tx_deadline)) {
2324 /* Timed out messages queued for sending or
2325 * buffered in the socket's send buffer */
2326 ksocknal_conn_addref(conn);
2327 CNETERR("Timeout sending data to %s (%u.%u.%u.%u:%d) "
2328 "the network or that node may be down.\n",
2329 libcfs_id2str(peer->ksnp_id),
2330 HIPQUAD(conn->ksnc_ipaddr),
2340 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2343 CFS_LIST_HEAD (stale_txs);
2345 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2347 while (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2348 tx = cfs_list_entry (peer->ksnp_tx_queue.next,
2349 ksock_tx_t, tx_list);
2351 if (!cfs_time_aftereq(cfs_time_current(),
2355 cfs_list_del (&tx->tx_list);
2356 cfs_list_add_tail (&tx->tx_list, &stale_txs);
2359 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2361 ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2365 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2367 ksock_sched_t *sched;
2371 if (cfs_list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2374 if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2377 if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2378 cfs_time_before(cfs_time_current(),
2379 cfs_time_add(peer->ksnp_last_alive,
2380 cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2383 if (cfs_time_before(cfs_time_current(),
2384 peer->ksnp_send_keepalive))
2387 /* retry 10 secs later, so we wouldn't put pressure
2388 * on this peer if we failed to send keepalive this time */
2389 peer->ksnp_send_keepalive = cfs_time_shift(10);
2391 conn = ksocknal_find_conn_locked(peer, NULL, 1);
2393 sched = conn->ksnc_scheduler;
2395 spin_lock_bh(&sched->kss_lock);
2396 if (!cfs_list_empty(&conn->ksnc_tx_queue)) {
2397 spin_unlock_bh(&sched->kss_lock);
2398 /* there is an queued ACK, don't need keepalive */
2402 spin_unlock_bh(&sched->kss_lock);
2405 read_unlock(&ksocknal_data.ksnd_global_lock);
2407 /* cookie = 1 is reserved for keepalive PING */
2408 tx = ksocknal_alloc_tx_noop(1, 1);
2410 read_lock(&ksocknal_data.ksnd_global_lock);
2414 if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2415 read_lock(&ksocknal_data.ksnd_global_lock);
2419 ksocknal_free_tx(tx);
2420 read_lock(&ksocknal_data.ksnd_global_lock);
2427 ksocknal_check_peer_timeouts (int idx)
2429 cfs_list_t *peers = &ksocknal_data.ksnd_peers[idx];
2435 /* NB. We expect to have a look at all the peers and not find any
2436 * connections to time out, so we just use a shared lock while we
2438 read_lock(&ksocknal_data.ksnd_global_lock);
2440 cfs_list_for_each_entry_typed(peer, peers, ksock_peer_t, ksnp_list) {
2441 cfs_time_t deadline = 0;
2445 if (ksocknal_send_keepalive_locked(peer) != 0) {
2446 read_unlock(&ksocknal_data.ksnd_global_lock);
2450 conn = ksocknal_find_timed_out_conn (peer);
2453 read_unlock(&ksocknal_data.ksnd_global_lock);
2455 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2457 /* NB we won't find this one again, but we can't
2458 * just proceed with the next peer, since we dropped
2459 * ksnd_global_lock and it might be dead already! */
2460 ksocknal_conn_decref(conn);
2464 /* we can't process stale txs right here because we're
2465 * holding only shared lock */
2466 if (!cfs_list_empty (&peer->ksnp_tx_queue)) {
2468 cfs_list_entry (peer->ksnp_tx_queue.next,
2469 ksock_tx_t, tx_list);
2471 if (cfs_time_aftereq(cfs_time_current(),
2474 ksocknal_peer_addref(peer);
2475 read_unlock(&ksocknal_data.ksnd_global_lock);
2477 ksocknal_flush_stale_txs(peer);
2479 ksocknal_peer_decref(peer);
2484 if (cfs_list_empty(&peer->ksnp_zc_req_list))
2487 spin_lock(&peer->ksnp_lock);
2488 cfs_list_for_each_entry_typed(tx, &peer->ksnp_zc_req_list,
2489 ksock_tx_t, tx_zc_list) {
2490 if (!cfs_time_aftereq(cfs_time_current(),
2493 /* ignore the TX if connection is being closed */
2494 if (tx->tx_conn->ksnc_closing)
2500 spin_unlock(&peer->ksnp_lock);
2504 tx = cfs_list_entry(peer->ksnp_zc_req_list.next,
2505 ksock_tx_t, tx_zc_list);
2506 deadline = tx->tx_deadline;
2507 resid = tx->tx_resid;
2509 ksocknal_conn_addref(conn);
2511 spin_unlock(&peer->ksnp_lock);
2512 read_unlock(&ksocknal_data.ksnd_global_lock);
2514 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2515 "oldest(%p) timed out %ld secs ago, "
2516 "resid: %d, wmem: %d\n",
2517 n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2518 cfs_duration_sec(cfs_time_current() - deadline),
2519 resid, libcfs_sock_wmem_queued(conn->ksnc_sock));
2521 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2522 ksocknal_conn_decref(conn);
2526 read_unlock(&ksocknal_data.ksnd_global_lock);
2529 int ksocknal_reaper(void *arg)
2533 ksock_sched_t *sched;
2534 cfs_list_t enomem_conns;
2536 cfs_duration_t timeout;
2539 cfs_time_t deadline = cfs_time_current();
2541 cfs_block_allsigs ();
2543 CFS_INIT_LIST_HEAD(&enomem_conns);
2544 init_waitqueue_entry_current(&wait);
2546 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2548 while (!ksocknal_data.ksnd_shuttingdown) {
2550 if (!cfs_list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2551 conn = cfs_list_entry (ksocknal_data. \
2552 ksnd_deathrow_conns.next,
2553 ksock_conn_t, ksnc_list);
2554 cfs_list_del (&conn->ksnc_list);
2556 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2558 ksocknal_terminate_conn(conn);
2559 ksocknal_conn_decref(conn);
2561 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2565 if (!cfs_list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2566 conn = cfs_list_entry (ksocknal_data.ksnd_zombie_conns.\
2567 next, ksock_conn_t, ksnc_list);
2568 cfs_list_del (&conn->ksnc_list);
2570 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2572 ksocknal_destroy_conn(conn);
2574 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2578 if (!cfs_list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2579 cfs_list_add(&enomem_conns,
2580 &ksocknal_data.ksnd_enomem_conns);
2581 cfs_list_del_init(&ksocknal_data.ksnd_enomem_conns);
2584 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2586 /* reschedule all the connections that stalled with ENOMEM... */
2588 while (!cfs_list_empty (&enomem_conns)) {
2589 conn = cfs_list_entry (enomem_conns.next,
2590 ksock_conn_t, ksnc_tx_list);
2591 cfs_list_del (&conn->ksnc_tx_list);
2593 sched = conn->ksnc_scheduler;
2595 spin_lock_bh(&sched->kss_lock);
2597 LASSERT(conn->ksnc_tx_scheduled);
2598 conn->ksnc_tx_ready = 1;
2599 cfs_list_add_tail(&conn->ksnc_tx_list,
2600 &sched->kss_tx_conns);
2601 wake_up(&sched->kss_waitq);
2603 spin_unlock_bh(&sched->kss_lock);
2607 /* careful with the jiffy wrap... */
2608 while ((timeout = cfs_time_sub(deadline,
2609 cfs_time_current())) <= 0) {
2612 int chunk = ksocknal_data.ksnd_peer_hash_size;
2614 /* Time to check for timeouts on a few more peers: I do
2615 * checks every 'p' seconds on a proportion of the peer
2616 * table and I need to check every connection 'n' times
2617 * within a timeout interval, to ensure I detect a
2618 * timeout on any connection within (n+1)/n times the
2619 * timeout interval. */
2621 if (*ksocknal_tunables.ksnd_timeout > n * p)
2622 chunk = (chunk * n * p) /
2623 *ksocknal_tunables.ksnd_timeout;
2627 for (i = 0; i < chunk; i++) {
2628 ksocknal_check_peer_timeouts (peer_index);
2629 peer_index = (peer_index + 1) %
2630 ksocknal_data.ksnd_peer_hash_size;
2633 deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2636 if (nenomem_conns != 0) {
2637 /* Reduce my timeout if I rescheduled ENOMEM conns.
2638 * This also prevents me getting woken immediately
2639 * if any go back on my enomem list. */
2640 timeout = SOCKNAL_ENOMEM_RETRY;
2642 ksocknal_data.ksnd_reaper_waketime =
2643 cfs_time_add(cfs_time_current(), timeout);
2645 set_current_state(TASK_INTERRUPTIBLE);
2646 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2648 if (!ksocknal_data.ksnd_shuttingdown &&
2649 cfs_list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2650 cfs_list_empty(&ksocknal_data.ksnd_zombie_conns))
2651 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
2653 set_current_state(TASK_RUNNING);
2654 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2656 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2659 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2661 ksocknal_thread_fini();