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 (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
41 tx = list_entry(ksocknal_data.ksnd_idle_noop_txs. \
42 next, ksock_tx_t, tx_list);
43 LASSERT(tx->tx_desc_size == size);
44 list_del(&tx->tx_list);
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 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 = cfs_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 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 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. */
347 if (ksocknal_data.ksnd_stall_rx != 0) {
348 cfs_pause(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
351 rc = ksocknal_connsock_addref(conn);
353 LASSERT (conn->ksnc_closing);
358 if (conn->ksnc_rx_niov != 0)
359 rc = ksocknal_recv_iov (conn);
361 rc = ksocknal_recv_kiov (conn);
364 /* error/EOF or partial receive */
367 } else if (rc == 0 && conn->ksnc_rx_started) {
368 /* EOF in the middle of a message */
374 /* Completed a fragment */
376 if (conn->ksnc_rx_nob_wanted == 0) {
382 ksocknal_connsock_decref(conn);
387 ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
389 lnet_msg_t *lnetmsg = tx->tx_lnetmsg;
390 int rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO;
392 LASSERT(ni != NULL || tx->tx_conn != NULL);
394 if (tx->tx_conn != NULL)
395 ksocknal_conn_decref(tx->tx_conn);
397 if (ni == NULL && tx->tx_conn != NULL)
398 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
400 ksocknal_free_tx (tx);
401 if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
402 lnet_finalize (ni, lnetmsg, rc);
406 ksocknal_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int error)
410 while (!list_empty (txlist)) {
411 tx = list_entry (txlist->next, ksock_tx_t, tx_list);
413 if (error && tx->tx_lnetmsg != NULL) {
414 CNETERR("Deleting packet type %d len %d %s->%s\n",
415 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
416 le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
417 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
418 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
420 CNETERR("Deleting noop packet\n");
423 list_del (&tx->tx_list);
425 LASSERT (atomic_read(&tx->tx_refcount) == 1);
426 ksocknal_tx_done (ni, tx);
431 ksocknal_check_zc_req(ksock_tx_t *tx)
433 ksock_conn_t *conn = tx->tx_conn;
434 ksock_peer_t *peer = conn->ksnc_peer;
436 /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
437 * to ksnp_zc_req_list if some fragment of this message should be sent
438 * zero-copy. Our peer will send an ACK containing this cookie when
439 * she has received this message to tell us we can signal completion.
440 * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
441 * ksnp_zc_req_list. */
442 LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
443 LASSERT (tx->tx_zc_capable);
445 tx->tx_zc_checked = 1;
447 if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
448 !conn->ksnc_zc_capable)
451 /* assign cookie and queue tx to pending list, it will be released when
452 * a matching ack is received. See ksocknal_handle_zcack() */
454 ksocknal_tx_addref(tx);
456 spin_lock(&peer->ksnp_lock);
458 /* ZC_REQ is going to be pinned to the peer */
460 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
462 LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
464 tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
466 if (peer->ksnp_zc_next_cookie == 0)
467 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
469 list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
471 spin_unlock(&peer->ksnp_lock);
475 ksocknal_uncheck_zc_req(ksock_tx_t *tx)
477 ksock_peer_t *peer = tx->tx_conn->ksnc_peer;
479 LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
480 LASSERT(tx->tx_zc_capable);
482 tx->tx_zc_checked = 0;
484 spin_lock(&peer->ksnp_lock);
486 if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
487 /* Not waiting for an ACK */
488 spin_unlock(&peer->ksnp_lock);
492 tx->tx_msg.ksm_zc_cookies[0] = 0;
493 list_del(&tx->tx_zc_list);
495 spin_unlock(&peer->ksnp_lock);
497 ksocknal_tx_decref(tx);
501 ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
505 if (tx->tx_zc_capable && !tx->tx_zc_checked)
506 ksocknal_check_zc_req(tx);
508 rc = ksocknal_transmit (conn, tx);
510 CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
512 if (tx->tx_resid == 0) {
513 /* Sent everything OK */
525 counter++; /* exponential backoff warnings */
526 if ((counter & (-counter)) == counter)
527 CWARN("%u ENOMEM tx %p (%u allocated)\n",
528 counter, conn, atomic_read(&libcfs_kmemory));
530 /* Queue on ksnd_enomem_conns for retry after a timeout */
531 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
533 /* enomem list takes over scheduler's ref... */
534 LASSERT (conn->ksnc_tx_scheduled);
535 list_add_tail(&conn->ksnc_tx_list,
536 &ksocknal_data.ksnd_enomem_conns);
537 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
538 SOCKNAL_ENOMEM_RETRY),
539 ksocknal_data.ksnd_reaper_waketime))
540 wake_up (&ksocknal_data.ksnd_reaper_waitq);
542 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
549 if (!conn->ksnc_closing) {
552 LCONSOLE_WARN("Host %pI4h reset our connection "
553 "while we were sending data; it may have "
558 LCONSOLE_WARN("There was an unexpected network error "
559 "while writing to %pI4h: %d.\n",
560 &conn->ksnc_ipaddr, rc);
563 CDEBUG(D_NET, "[%p] Error %d on write to %s"
564 " ip %pI4h:%d\n", conn, rc,
565 libcfs_id2str(conn->ksnc_peer->ksnp_id),
570 if (tx->tx_zc_checked)
571 ksocknal_uncheck_zc_req(tx);
573 /* it's not an error if conn is being closed */
574 ksocknal_close_conn_and_siblings (conn,
575 (conn->ksnc_closing) ? 0 : rc);
581 ksocknal_launch_connection_locked (ksock_route_t *route)
584 /* called holding write lock on ksnd_global_lock */
586 LASSERT (!route->ksnr_scheduled);
587 LASSERT (!route->ksnr_connecting);
588 LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
590 route->ksnr_scheduled = 1; /* scheduling conn for connd */
591 ksocknal_route_addref(route); /* extra ref for connd */
593 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
595 list_add_tail(&route->ksnr_connd_list,
596 &ksocknal_data.ksnd_connd_routes);
597 wake_up(&ksocknal_data.ksnd_connd_waitq);
599 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
603 ksocknal_launch_all_connections_locked (ksock_peer_t *peer)
605 ksock_route_t *route;
607 /* called holding write lock on ksnd_global_lock */
609 /* launch any/all connections that need it */
610 route = ksocknal_find_connectable_route_locked(peer);
614 ksocknal_launch_connection_locked(route);
619 ksocknal_find_conn_locked(ksock_peer_t *peer, ksock_tx_t *tx, int nonblk)
621 struct list_head *tmp;
623 ksock_conn_t *typed = NULL;
624 ksock_conn_t *fallback = NULL;
628 list_for_each (tmp, &peer->ksnp_conns) {
629 ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list);
630 int nob = atomic_read(&c->ksnc_tx_nob) +
631 cfs_sock_wmem_queued(c->ksnc_sock);
634 LASSERT (!c->ksnc_closing);
635 LASSERT (c->ksnc_proto != NULL &&
636 c->ksnc_proto->pro_match_tx != NULL);
638 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
643 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
646 case SOCKNAL_MATCH_YES: /* typed connection */
647 if (typed == NULL || tnob > nob ||
648 (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
649 cfs_time_after(typed->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
655 case SOCKNAL_MATCH_MAY: /* fallback connection */
656 if (fallback == NULL || fnob > nob ||
657 (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
658 cfs_time_after(fallback->ksnc_tx_last_post, c->ksnc_tx_last_post))) {
666 /* prefer the typed selection */
667 conn = (typed != NULL) ? typed : fallback;
670 conn->ksnc_tx_last_post = cfs_time_current();
676 ksocknal_tx_prep(ksock_conn_t *conn, ksock_tx_t *tx)
678 conn->ksnc_proto->pro_pack(tx);
680 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
681 ksocknal_conn_addref(conn); /* +1 ref for tx */
686 ksocknal_queue_tx_locked (ksock_tx_t *tx, ksock_conn_t *conn)
688 ksock_sched_t *sched = conn->ksnc_scheduler;
689 ksock_msg_t *msg = &tx->tx_msg;
690 ksock_tx_t *ztx = NULL;
693 /* called holding global lock (read or irq-write) and caller may
694 * not have dropped this lock between finding conn and calling me,
695 * so we don't need the {get,put}connsock dance to deref
697 LASSERT(!conn->ksnc_closing);
699 CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
700 libcfs_id2str(conn->ksnc_peer->ksnp_id),
704 ksocknal_tx_prep(conn, tx);
706 /* Ensure the frags we've been given EXACTLY match the number of
707 * bytes we want to send. Many TCP/IP stacks disregard any total
708 * size parameters passed to them and just look at the frags.
710 * We always expect at least 1 mapped fragment containing the
711 * complete ksocknal message header. */
712 LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
713 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
714 (unsigned int)tx->tx_nob);
715 LASSERT (tx->tx_niov >= 1);
716 LASSERT (tx->tx_resid == tx->tx_nob);
718 CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
719 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
721 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
724 * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
725 * but they're used inside spinlocks a lot.
727 bufnob = cfs_sock_wmem_queued(conn->ksnc_sock);
728 spin_lock_bh(&sched->kss_lock);
730 if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
731 /* First packet starts the timeout */
732 conn->ksnc_tx_deadline =
733 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
734 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
735 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
736 conn->ksnc_tx_bufnob = 0;
737 mb(); /* order with adding to tx_queue */
740 if (msg->ksm_type == KSOCK_MSG_NOOP) {
741 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
742 * on a normal packet so I don't need to send it */
743 LASSERT (msg->ksm_zc_cookies[1] != 0);
744 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
746 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
747 ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
750 /* It's a normal packet - can it piggback a noop zc-ack that
751 * has been queued already? */
752 LASSERT (msg->ksm_zc_cookies[1] == 0);
753 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
755 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
756 /* ztx will be released later */
760 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
761 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
764 if (conn->ksnc_tx_ready && /* able to send */
765 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
766 /* +1 ref for scheduler */
767 ksocknal_conn_addref(conn);
768 list_add_tail (&conn->ksnc_tx_list,
769 &sched->kss_tx_conns);
770 conn->ksnc_tx_scheduled = 1;
771 wake_up (&sched->kss_waitq);
774 spin_unlock_bh(&sched->kss_lock);
779 ksocknal_find_connectable_route_locked (ksock_peer_t *peer)
781 cfs_time_t now = cfs_time_current();
782 struct list_head *tmp;
783 ksock_route_t *route;
785 list_for_each (tmp, &peer->ksnp_routes) {
786 route = list_entry (tmp, ksock_route_t, ksnr_list);
788 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
790 if (route->ksnr_scheduled) /* connections being established */
793 /* all route types connected ? */
794 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
797 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
798 cfs_time_aftereq(now, route->ksnr_timeout))) {
800 "Too soon to retry route %pI4h "
801 "(cnted %d, interval %ld, %ld secs later)\n",
803 route->ksnr_connected,
804 route->ksnr_retry_interval,
805 cfs_duration_sec(route->ksnr_timeout - now));
816 ksocknal_find_connecting_route_locked (ksock_peer_t *peer)
818 struct list_head *tmp;
819 ksock_route_t *route;
821 list_for_each (tmp, &peer->ksnp_routes) {
822 route = list_entry (tmp, ksock_route_t, ksnr_list);
824 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
826 if (route->ksnr_scheduled)
834 ksocknal_launch_packet (lnet_ni_t *ni, ksock_tx_t *tx, lnet_process_id_t id)
842 LASSERT (tx->tx_conn == NULL);
844 g_lock = &ksocknal_data.ksnd_global_lock;
846 for (retry = 0;; retry = 1) {
848 peer = ksocknal_find_peer_locked(ni, id);
850 if (ksocknal_find_connectable_route_locked(peer) == NULL) {
851 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
853 /* I've got no routes that need to be
854 * connecting and I do have an actual
856 ksocknal_queue_tx_locked (tx, conn);
863 /* I'll need a write lock... */
866 write_lock_bh(g_lock);
868 peer = ksocknal_find_peer_locked(ni, id);
872 write_unlock_bh(g_lock);
874 if ((id.pid & LNET_PID_USERFLAG) != 0) {
875 CERROR("Refusing to create a connection to "
876 "userspace process %s\n", libcfs_id2str(id));
877 return -EHOSTUNREACH;
881 CERROR("Can't find peer %s\n", libcfs_id2str(id));
882 return -EHOSTUNREACH;
885 rc = ksocknal_add_peer(ni, id,
886 LNET_NIDADDR(id.nid),
887 lnet_acceptor_port());
889 CERROR("Can't add peer %s: %d\n",
890 libcfs_id2str(id), rc);
895 ksocknal_launch_all_connections_locked(peer);
897 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
899 /* Connection exists; queue message on it */
900 ksocknal_queue_tx_locked (tx, conn);
901 write_unlock_bh(g_lock);
905 if (peer->ksnp_accepting > 0 ||
906 ksocknal_find_connecting_route_locked (peer) != NULL) {
907 /* the message is going to be pinned to the peer */
909 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
911 /* Queue the message until a connection is established */
912 list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
913 write_unlock_bh(g_lock);
917 write_unlock_bh(g_lock);
919 /* NB Routes may be ignored if connections to them failed recently */
920 CNETERR("No usable routes to %s\n", libcfs_id2str(id));
921 return (-EHOSTUNREACH);
925 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
928 int type = lntmsg->msg_type;
929 lnet_process_id_t target = lntmsg->msg_target;
930 unsigned int payload_niov = lntmsg->msg_niov;
931 struct iovec *payload_iov = lntmsg->msg_iov;
932 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
933 unsigned int payload_offset = lntmsg->msg_offset;
934 unsigned int payload_nob = lntmsg->msg_len;
939 /* NB 'private' is different depending on what we're sending.
940 * Just ignore it... */
942 CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
943 payload_nob, payload_niov, libcfs_id2str(target));
945 LASSERT (payload_nob == 0 || payload_niov > 0);
946 LASSERT (payload_niov <= LNET_MAX_IOV);
947 /* payload is either all vaddrs or all pages */
948 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
949 LASSERT (!in_interrupt ());
951 if (payload_iov != NULL)
952 desc_size = offsetof(ksock_tx_t,
953 tx_frags.virt.iov[1 + payload_niov]);
955 desc_size = offsetof(ksock_tx_t,
956 tx_frags.paged.kiov[payload_niov]);
958 if (lntmsg->msg_vmflush)
959 mpflag = cfs_memory_pressure_get_and_set();
960 tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
962 CERROR("Can't allocate tx desc type %d size %d\n",
964 if (lntmsg->msg_vmflush)
965 cfs_memory_pressure_restore(mpflag);
969 tx->tx_conn = NULL; /* set when assigned a conn */
970 tx->tx_lnetmsg = lntmsg;
972 if (payload_iov != NULL) {
975 tx->tx_iov = tx->tx_frags.virt.iov;
977 lnet_extract_iov(payload_niov, &tx->tx_iov[1],
978 payload_niov, payload_iov,
979 payload_offset, payload_nob);
982 tx->tx_iov = &tx->tx_frags.paged.iov;
983 tx->tx_kiov = tx->tx_frags.paged.kiov;
984 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
985 payload_niov, payload_kiov,
986 payload_offset, payload_nob);
988 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
989 tx->tx_zc_capable = 1;
992 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
994 /* The first fragment will be set later in pro_pack */
995 rc = ksocknal_launch_packet(ni, tx, target);
997 cfs_memory_pressure_restore(mpflag);
1002 ksocknal_free_tx(tx);
1007 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1009 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1012 return PTR_ERR(task);
1014 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1015 ksocknal_data.ksnd_nthreads++;
1016 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1021 ksocknal_thread_fini (void)
1023 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1024 ksocknal_data.ksnd_nthreads--;
1025 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1029 ksocknal_new_packet (ksock_conn_t *conn, int nob_to_skip)
1031 static char ksocknal_slop_buffer[4096];
1037 LASSERT(conn->ksnc_proto != NULL);
1039 if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1040 /* Remind the socket to ack eagerly... */
1041 ksocknal_lib_eager_ack(conn);
1044 if (nob_to_skip == 0) { /* right at next packet boundary now */
1045 conn->ksnc_rx_started = 0;
1046 mb(); /* racing with timeout thread */
1048 switch (conn->ksnc_proto->pro_version) {
1049 case KSOCK_PROTO_V2:
1050 case KSOCK_PROTO_V3:
1051 conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1052 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1053 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1055 conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1056 conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1057 conn->ksnc_rx_iov[0].iov_len = offsetof(ksock_msg_t, ksm_u);
1060 case KSOCK_PROTO_V1:
1061 /* Receiving bare lnet_hdr_t */
1062 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1063 conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1064 conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1066 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1067 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1068 conn->ksnc_rx_iov[0].iov_len = sizeof (lnet_hdr_t);
1074 conn->ksnc_rx_niov = 1;
1076 conn->ksnc_rx_kiov = NULL;
1077 conn->ksnc_rx_nkiov = 0;
1078 conn->ksnc_rx_csum = ~0;
1082 /* Set up to skip as much as possible now. If there's more left
1083 * (ran out of iov entries) we'll get called again */
1085 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1086 conn->ksnc_rx_nob_left = nob_to_skip;
1087 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1092 nob = MIN (nob_to_skip, sizeof (ksocknal_slop_buffer));
1094 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1095 conn->ksnc_rx_iov[niov].iov_len = nob;
1100 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1101 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof (struct iovec));
1103 conn->ksnc_rx_niov = niov;
1104 conn->ksnc_rx_kiov = NULL;
1105 conn->ksnc_rx_nkiov = 0;
1106 conn->ksnc_rx_nob_wanted = skipped;
1111 ksocknal_process_receive (ksock_conn_t *conn)
1114 lnet_process_id_t *id;
1117 LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1119 /* NB: sched lock NOT held */
1120 /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1121 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1122 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1123 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1124 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1126 if (conn->ksnc_rx_nob_wanted != 0) {
1127 rc = ksocknal_receive(conn);
1130 LASSERT (rc != -EAGAIN);
1133 CDEBUG(D_NET, "[%p] EOF from %s"
1134 " ip %pI4h:%d\n", conn,
1135 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1138 else if (!conn->ksnc_closing)
1139 CERROR("[%p] Error %d on read from %s"
1142 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1146 /* it's not an error if conn is being closed */
1147 ksocknal_close_conn_and_siblings (conn,
1148 (conn->ksnc_closing) ? 0 : rc);
1149 return (rc == 0 ? -ESHUTDOWN : rc);
1152 if (conn->ksnc_rx_nob_wanted != 0) {
1157 switch (conn->ksnc_rx_state) {
1158 case SOCKNAL_RX_KSM_HEADER:
1159 if (conn->ksnc_flip) {
1160 __swab32s(&conn->ksnc_msg.ksm_type);
1161 __swab32s(&conn->ksnc_msg.ksm_csum);
1162 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1163 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1166 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1167 conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1168 CERROR("%s: Unknown message type: %x\n",
1169 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1170 conn->ksnc_msg.ksm_type);
1171 ksocknal_new_packet(conn, 0);
1172 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1176 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1177 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1178 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1179 /* NOOP Checksum error */
1180 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1181 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1182 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1183 ksocknal_new_packet(conn, 0);
1184 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1188 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1191 LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1193 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1194 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1196 rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1197 conn->ksnc_msg.ksm_zc_cookies[1]);
1200 CERROR("%s: Unknown ZC-ACK cookie: "LPU64", "LPU64"\n",
1201 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1202 cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1203 ksocknal_new_packet(conn, 0);
1204 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1209 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1210 ksocknal_new_packet (conn, 0);
1211 return 0; /* NOOP is done and just return */
1214 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1215 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1216 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1218 conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
1219 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1220 conn->ksnc_rx_iov[0].iov_len = sizeof(ksock_lnet_msg_t);
1222 conn->ksnc_rx_niov = 1;
1223 conn->ksnc_rx_kiov = NULL;
1224 conn->ksnc_rx_nkiov = 0;
1226 goto again; /* read lnet header now */
1228 case SOCKNAL_RX_LNET_HEADER:
1229 /* unpack message header */
1230 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1232 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1233 /* Userspace peer */
1234 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1235 id = &conn->ksnc_peer->ksnp_id;
1237 /* Substitute process ID assigned at connection time */
1238 lhdr->src_pid = cpu_to_le32(id->pid);
1239 lhdr->src_nid = cpu_to_le64(id->nid);
1242 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1243 ksocknal_conn_addref(conn); /* ++ref while parsing */
1245 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1246 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1247 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1249 /* I just received garbage: give up on this conn */
1250 ksocknal_new_packet(conn, 0);
1251 ksocknal_close_conn_and_siblings (conn, rc);
1252 ksocknal_conn_decref(conn);
1256 /* I'm racing with ksocknal_recv() */
1257 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1258 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1260 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1263 /* ksocknal_recv() got called */
1266 case SOCKNAL_RX_LNET_PAYLOAD:
1267 /* payload all received */
1270 if (conn->ksnc_rx_nob_left == 0 && /* not truncating */
1271 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1272 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1273 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1274 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1275 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1279 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1280 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1282 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1283 id = &conn->ksnc_peer->ksnp_id;
1285 rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1286 conn->ksnc_msg.ksm_zc_cookies[0],
1287 *ksocknal_tunables.ksnd_nonblk_zcack ||
1288 le64_to_cpu(lhdr->src_nid) != id->nid);
1291 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1294 ksocknal_new_packet(conn, 0);
1295 ksocknal_close_conn_and_siblings (conn, rc);
1300 case SOCKNAL_RX_SLOP:
1301 /* starting new packet? */
1302 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1303 return 0; /* come back later */
1304 goto again; /* try to finish reading slop now */
1312 return (-EINVAL); /* keep gcc happy */
1316 ksocknal_recv (lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1317 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1318 unsigned int offset, unsigned int mlen, unsigned int rlen)
1320 ksock_conn_t *conn = (ksock_conn_t *)private;
1321 ksock_sched_t *sched = conn->ksnc_scheduler;
1323 LASSERT (mlen <= rlen);
1324 LASSERT (niov <= LNET_MAX_IOV);
1326 conn->ksnc_cookie = msg;
1327 conn->ksnc_rx_nob_wanted = mlen;
1328 conn->ksnc_rx_nob_left = rlen;
1330 if (mlen == 0 || iov != NULL) {
1331 conn->ksnc_rx_nkiov = 0;
1332 conn->ksnc_rx_kiov = NULL;
1333 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1334 conn->ksnc_rx_niov =
1335 lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1336 niov, iov, offset, mlen);
1338 conn->ksnc_rx_niov = 0;
1339 conn->ksnc_rx_iov = NULL;
1340 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1341 conn->ksnc_rx_nkiov =
1342 lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1343 niov, kiov, offset, mlen);
1347 lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1348 lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1350 LASSERT (conn->ksnc_rx_scheduled);
1352 spin_lock_bh(&sched->kss_lock);
1354 switch (conn->ksnc_rx_state) {
1355 case SOCKNAL_RX_PARSE_WAIT:
1356 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1357 wake_up (&sched->kss_waitq);
1358 LASSERT (conn->ksnc_rx_ready);
1361 case SOCKNAL_RX_PARSE:
1362 /* scheduler hasn't noticed I'm parsing yet */
1366 conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1368 spin_unlock_bh(&sched->kss_lock);
1369 ksocknal_conn_decref(conn);
1374 ksocknal_sched_cansleep(ksock_sched_t *sched)
1378 spin_lock_bh(&sched->kss_lock);
1380 rc = (!ksocknal_data.ksnd_shuttingdown &&
1381 list_empty(&sched->kss_rx_conns) &&
1382 list_empty(&sched->kss_tx_conns));
1384 spin_unlock_bh(&sched->kss_lock);
1388 int ksocknal_scheduler(void *arg)
1390 struct ksock_sched_info *info;
1391 ksock_sched_t *sched;
1396 long id = (long)arg;
1398 info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1399 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1401 cfs_block_allsigs();
1403 rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1405 CERROR("Can't set CPT affinity to %d: %d\n",
1409 spin_lock_bh(&sched->kss_lock);
1411 while (!ksocknal_data.ksnd_shuttingdown) {
1412 int did_something = 0;
1414 /* Ensure I progress everything semi-fairly */
1416 if (!list_empty (&sched->kss_rx_conns)) {
1417 conn = list_entry(sched->kss_rx_conns.next,
1418 ksock_conn_t, ksnc_rx_list);
1419 list_del(&conn->ksnc_rx_list);
1421 LASSERT(conn->ksnc_rx_scheduled);
1422 LASSERT(conn->ksnc_rx_ready);
1424 /* clear rx_ready in case receive isn't complete.
1425 * Do it BEFORE we call process_recv, since
1426 * data_ready can set it any time after we release
1428 conn->ksnc_rx_ready = 0;
1429 spin_unlock_bh(&sched->kss_lock);
1431 rc = ksocknal_process_receive(conn);
1433 spin_lock_bh(&sched->kss_lock);
1435 /* I'm the only one that can clear this flag */
1436 LASSERT(conn->ksnc_rx_scheduled);
1438 /* Did process_receive get everything it wanted? */
1440 conn->ksnc_rx_ready = 1;
1442 if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1443 /* Conn blocked waiting for ksocknal_recv()
1444 * I change its state (under lock) to signal
1445 * it can be rescheduled */
1446 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1447 } else if (conn->ksnc_rx_ready) {
1448 /* reschedule for rx */
1449 list_add_tail (&conn->ksnc_rx_list,
1450 &sched->kss_rx_conns);
1452 conn->ksnc_rx_scheduled = 0;
1454 ksocknal_conn_decref(conn);
1460 if (!list_empty (&sched->kss_tx_conns)) {
1463 if (!list_empty(&sched->kss_zombie_noop_txs)) {
1465 &sched->kss_zombie_noop_txs);
1466 list_del_init(&sched->kss_zombie_noop_txs);
1469 conn = list_entry(sched->kss_tx_conns.next,
1470 ksock_conn_t, ksnc_tx_list);
1471 list_del (&conn->ksnc_tx_list);
1473 LASSERT(conn->ksnc_tx_scheduled);
1474 LASSERT(conn->ksnc_tx_ready);
1475 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1477 tx = list_entry(conn->ksnc_tx_queue.next,
1478 ksock_tx_t, tx_list);
1480 if (conn->ksnc_tx_carrier == tx)
1481 ksocknal_next_tx_carrier(conn);
1483 /* dequeue now so empty list => more to send */
1484 list_del(&tx->tx_list);
1486 /* Clear tx_ready in case send isn't complete. Do
1487 * it BEFORE we call process_transmit, since
1488 * write_space can set it any time after we release
1490 conn->ksnc_tx_ready = 0;
1491 spin_unlock_bh(&sched->kss_lock);
1493 if (!list_empty(&zlist)) {
1494 /* free zombie noop txs, it's fast because
1495 * noop txs are just put in freelist */
1496 ksocknal_txlist_done(NULL, &zlist, 0);
1499 rc = ksocknal_process_transmit(conn, tx);
1501 if (rc == -ENOMEM || rc == -EAGAIN) {
1502 /* Incomplete send: replace tx on HEAD of tx_queue */
1503 spin_lock_bh(&sched->kss_lock);
1504 list_add(&tx->tx_list,
1505 &conn->ksnc_tx_queue);
1507 /* Complete send; tx -ref */
1508 ksocknal_tx_decref(tx);
1510 spin_lock_bh(&sched->kss_lock);
1511 /* assume space for more */
1512 conn->ksnc_tx_ready = 1;
1515 if (rc == -ENOMEM) {
1516 /* Do nothing; after a short timeout, this
1517 * conn will be reposted on kss_tx_conns. */
1518 } else if (conn->ksnc_tx_ready &&
1519 !list_empty (&conn->ksnc_tx_queue)) {
1520 /* reschedule for tx */
1521 list_add_tail (&conn->ksnc_tx_list,
1522 &sched->kss_tx_conns);
1524 conn->ksnc_tx_scheduled = 0;
1526 ksocknal_conn_decref(conn);
1531 if (!did_something || /* nothing to do */
1532 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1533 spin_unlock_bh(&sched->kss_lock);
1537 if (!did_something) { /* wait for something to do */
1538 cfs_wait_event_interruptible_exclusive(
1540 !ksocknal_sched_cansleep(sched), rc);
1546 spin_lock_bh(&sched->kss_lock);
1550 spin_unlock_bh(&sched->kss_lock);
1551 ksocknal_thread_fini();
1556 * Add connection to kss_rx_conns of scheduler
1557 * and wakeup the scheduler.
1559 void ksocknal_read_callback (ksock_conn_t *conn)
1561 ksock_sched_t *sched;
1563 sched = conn->ksnc_scheduler;
1565 spin_lock_bh(&sched->kss_lock);
1567 conn->ksnc_rx_ready = 1;
1569 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1570 list_add_tail(&conn->ksnc_rx_list,
1571 &sched->kss_rx_conns);
1572 conn->ksnc_rx_scheduled = 1;
1573 /* extra ref for scheduler */
1574 ksocknal_conn_addref(conn);
1576 wake_up (&sched->kss_waitq);
1578 spin_unlock_bh(&sched->kss_lock);
1582 * Add connection to kss_tx_conns of scheduler
1583 * and wakeup the scheduler.
1585 void ksocknal_write_callback (ksock_conn_t *conn)
1587 ksock_sched_t *sched;
1589 sched = conn->ksnc_scheduler;
1591 spin_lock_bh(&sched->kss_lock);
1593 conn->ksnc_tx_ready = 1;
1595 if (!conn->ksnc_tx_scheduled && // not being progressed
1596 !list_empty(&conn->ksnc_tx_queue)){//packets to send
1597 list_add_tail (&conn->ksnc_tx_list,
1598 &sched->kss_tx_conns);
1599 conn->ksnc_tx_scheduled = 1;
1600 /* extra ref for scheduler */
1601 ksocknal_conn_addref(conn);
1603 wake_up (&sched->kss_waitq);
1606 spin_unlock_bh(&sched->kss_lock);
1610 ksocknal_parse_proto_version (ksock_hello_msg_t *hello)
1614 if (hello->kshm_magic == LNET_PROTO_MAGIC)
1615 version = hello->kshm_version;
1616 else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1617 version = __swab32(hello->kshm_version);
1620 #if SOCKNAL_VERSION_DEBUG
1621 if (*ksocknal_tunables.ksnd_protocol == 1)
1624 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1625 version == KSOCK_PROTO_V3)
1628 if (version == KSOCK_PROTO_V2)
1629 return &ksocknal_protocol_v2x;
1631 if (version == KSOCK_PROTO_V3)
1632 return &ksocknal_protocol_v3x;
1637 if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1638 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1640 CLASSERT (sizeof (lnet_magicversion_t) ==
1641 offsetof (ksock_hello_msg_t, kshm_src_nid));
1643 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1644 hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1645 return &ksocknal_protocol_v1x;
1652 ksocknal_send_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1653 lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1655 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1656 ksock_net_t *net = (ksock_net_t *)ni->ni_data;
1658 LASSERT (hello->kshm_nips <= LNET_MAX_INTERFACES);
1660 /* rely on caller to hold a ref on socket so it wouldn't disappear */
1661 LASSERT (conn->ksnc_proto != NULL);
1663 hello->kshm_src_nid = ni->ni_nid;
1664 hello->kshm_dst_nid = peer_nid;
1665 hello->kshm_src_pid = the_lnet.ln_pid;
1667 hello->kshm_src_incarnation = net->ksnn_incarnation;
1668 hello->kshm_ctype = conn->ksnc_type;
1670 return conn->ksnc_proto->pro_send_hello(conn, hello);
1674 ksocknal_invert_type(int type)
1678 case SOCKLND_CONN_ANY:
1679 case SOCKLND_CONN_CONTROL:
1681 case SOCKLND_CONN_BULK_IN:
1682 return SOCKLND_CONN_BULK_OUT;
1683 case SOCKLND_CONN_BULK_OUT:
1684 return SOCKLND_CONN_BULK_IN;
1686 return (SOCKLND_CONN_NONE);
1691 ksocknal_recv_hello (lnet_ni_t *ni, ksock_conn_t *conn,
1692 ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1695 /* Return < 0 fatal error
1697 * EALREADY lost connection race
1698 * EPROTO protocol version mismatch
1700 socket_t *sock = conn->ksnc_sock;
1701 int active = (conn->ksnc_proto != NULL);
1705 ksock_proto_t *proto;
1706 lnet_process_id_t recv_id;
1708 /* socket type set on active connections - not set on passive */
1709 LASSERT (!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1711 timeout = active ? *ksocknal_tunables.ksnd_timeout :
1712 lnet_acceptor_timeout();
1714 rc = libcfs_sock_read(sock, &hello->kshm_magic, sizeof (hello->kshm_magic), timeout);
1716 CERROR("Error %d reading HELLO from %pI4h\n",
1717 rc, &conn->ksnc_ipaddr);
1722 if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1723 hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1724 hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1725 /* Unexpected magic! */
1726 CERROR("Bad magic(1) %#08x (%#08x expected) from "
1727 "%pI4h\n", __cpu_to_le32 (hello->kshm_magic),
1728 LNET_PROTO_TCP_MAGIC,
1729 &conn->ksnc_ipaddr);
1733 rc = libcfs_sock_read(sock, &hello->kshm_version,
1734 sizeof(hello->kshm_version), timeout);
1736 CERROR("Error %d reading HELLO from %pI4h\n",
1737 rc, &conn->ksnc_ipaddr);
1742 proto = ksocknal_parse_proto_version(hello);
1743 if (proto == NULL) {
1745 /* unknown protocol from peer, tell peer my protocol */
1746 conn->ksnc_proto = &ksocknal_protocol_v3x;
1747 #if SOCKNAL_VERSION_DEBUG
1748 if (*ksocknal_tunables.ksnd_protocol == 2)
1749 conn->ksnc_proto = &ksocknal_protocol_v2x;
1750 else if (*ksocknal_tunables.ksnd_protocol == 1)
1751 conn->ksnc_proto = &ksocknal_protocol_v1x;
1753 hello->kshm_nips = 0;
1754 ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1757 CERROR("Unknown protocol version (%d.x expected)"
1759 conn->ksnc_proto->pro_version,
1760 &conn->ksnc_ipaddr);
1765 proto_match = (conn->ksnc_proto == proto);
1766 conn->ksnc_proto = proto;
1768 /* receive the rest of hello message anyway */
1769 rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1771 CERROR("Error %d reading or checking hello from from %pI4h\n",
1772 rc, &conn->ksnc_ipaddr);
1777 *incarnation = hello->kshm_src_incarnation;
1779 if (hello->kshm_src_nid == LNET_NID_ANY) {
1780 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY"
1781 "from %pI4h\n", &conn->ksnc_ipaddr);
1786 conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1787 /* Userspace NAL assigns peer process ID from socket */
1788 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1789 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1791 recv_id.nid = hello->kshm_src_nid;
1792 recv_id.pid = hello->kshm_src_pid;
1798 /* peer determines type */
1799 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1800 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1801 CERROR("Unexpected type %d from %s ip %pI4h\n",
1802 hello->kshm_ctype, libcfs_id2str(*peerid),
1803 &conn->ksnc_ipaddr);
1810 if (peerid->pid != recv_id.pid ||
1811 peerid->nid != recv_id.nid) {
1812 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1813 " %pI4h, but they claimed they were "
1814 "%s; please check your Lustre "
1816 libcfs_id2str(*peerid),
1818 libcfs_id2str(recv_id));
1822 if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1823 /* Possible protocol mismatch or I lost the connection race */
1824 return proto_match ? EALREADY : EPROTO;
1827 if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1828 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1829 conn->ksnc_type, libcfs_id2str(*peerid),
1839 ksocknal_connect (ksock_route_t *route)
1841 LIST_HEAD (zombies);
1842 ksock_peer_t *peer = route->ksnr_peer;
1846 cfs_time_t deadline;
1847 int retry_later = 0;
1850 deadline = cfs_time_add(cfs_time_current(),
1851 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1853 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1855 LASSERT (route->ksnr_scheduled);
1856 LASSERT (!route->ksnr_connecting);
1858 route->ksnr_connecting = 1;
1861 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1863 /* stop connecting if peer/route got closed under me, or
1864 * route got connected while queued */
1865 if (peer->ksnp_closing || route->ksnr_deleted ||
1871 /* reschedule if peer is connecting to me */
1872 if (peer->ksnp_accepting > 0) {
1874 "peer %s(%d) already connecting to me, retry later.\n",
1875 libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1879 if (retry_later) /* needs reschedule */
1882 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1883 type = SOCKLND_CONN_ANY;
1884 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1885 type = SOCKLND_CONN_CONTROL;
1886 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1887 type = SOCKLND_CONN_BULK_IN;
1889 LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1890 type = SOCKLND_CONN_BULK_OUT;
1893 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1895 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1897 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1903 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1904 route->ksnr_myipaddr,
1905 route->ksnr_ipaddr, route->ksnr_port);
1909 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1911 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1917 /* A +ve RC means I have to retry because I lost the connection
1918 * race or I have to renegotiate protocol version */
1919 retry_later = (rc != 0);
1921 CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1922 libcfs_nid2str(peer->ksnp_id.nid));
1924 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1927 route->ksnr_scheduled = 0;
1928 route->ksnr_connecting = 0;
1931 /* re-queue for attention; this frees me up to handle
1932 * the peer's incoming connection request */
1934 if (rc == EALREADY ||
1935 (rc == 0 && peer->ksnp_accepting > 0)) {
1936 /* We want to introduce a delay before next
1937 * attempt to connect if we lost conn race,
1938 * but the race is resolved quickly usually,
1939 * so min_reconnectms should be good heuristic */
1940 route->ksnr_retry_interval =
1941 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000;
1942 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1943 route->ksnr_retry_interval);
1946 ksocknal_launch_connection_locked(route);
1949 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1953 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1955 route->ksnr_scheduled = 0;
1956 route->ksnr_connecting = 0;
1958 /* This is a retry rather than a new connection */
1959 route->ksnr_retry_interval *= 2;
1960 route->ksnr_retry_interval =
1961 MAX(route->ksnr_retry_interval,
1962 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms)/1000);
1963 route->ksnr_retry_interval =
1964 MIN(route->ksnr_retry_interval,
1965 cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms)/1000);
1967 LASSERT (route->ksnr_retry_interval != 0);
1968 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1969 route->ksnr_retry_interval);
1971 if (!list_empty(&peer->ksnp_tx_queue) &&
1972 peer->ksnp_accepting == 0 &&
1973 ksocknal_find_connecting_route_locked(peer) == NULL) {
1976 /* ksnp_tx_queue is queued on a conn on successful
1977 * connection for V1.x and V2.x */
1978 if (!list_empty (&peer->ksnp_conns)) {
1979 conn = list_entry(peer->ksnp_conns.next,
1980 ksock_conn_t, ksnc_list);
1981 LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
1984 /* take all the blocked packets while I've got the lock and
1985 * complete below... */
1986 list_splice_init(&peer->ksnp_tx_queue, &zombies);
1989 #if 0 /* irrelevant with only eager routes */
1990 if (!route->ksnr_deleted) {
1991 /* make this route least-favourite for re-selection */
1992 list_del(&route->ksnr_list);
1993 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1996 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1998 ksocknal_peer_failed(peer);
1999 ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2004 * check whether we need to create more connds.
2005 * It will try to create new thread if it's necessary, @timeout can
2006 * be updated if failed to create, so caller wouldn't keep try while
2007 * running out of resource.
2010 ksocknal_connd_check_start(long sec, long *timeout)
2014 int total = ksocknal_data.ksnd_connd_starting +
2015 ksocknal_data.ksnd_connd_running;
2017 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2018 /* still in initializing */
2022 if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2023 total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2024 /* can't create more connd, or still have enough
2025 * threads to handle more connecting */
2029 if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2030 /* no pending connecting request */
2034 if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2035 /* may run out of resource, retry later */
2036 *timeout = cfs_time_seconds(1);
2040 if (ksocknal_data.ksnd_connd_starting > 0) {
2041 /* serialize starting to avoid flood */
2045 ksocknal_data.ksnd_connd_starting_stamp = sec;
2046 ksocknal_data.ksnd_connd_starting++;
2047 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2049 /* NB: total is the next id */
2050 snprintf(name, sizeof(name), "socknal_cd%02d", total);
2051 rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2053 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2058 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2059 ksocknal_data.ksnd_connd_starting--;
2060 ksocknal_data.ksnd_connd_failed_stamp = cfs_time_current_sec();
2066 * check whether current thread can exit, it will return 1 if there are too
2067 * many threads and no creating in past 120 seconds.
2068 * Also, this function may update @timeout to make caller come back
2069 * again to recheck these conditions.
2072 ksocknal_connd_check_stop(long sec, long *timeout)
2076 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2077 /* still in initializing */
2081 if (ksocknal_data.ksnd_connd_starting > 0) {
2082 /* in progress of starting new thread */
2086 if (ksocknal_data.ksnd_connd_running <=
2087 *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2091 /* created thread in past 120 seconds? */
2092 val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2093 SOCKNAL_CONND_TIMEOUT - sec);
2095 *timeout = (val > 0) ? cfs_time_seconds(val) :
2096 cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2100 /* no creating in past 120 seconds */
2102 return ksocknal_data.ksnd_connd_running >
2103 ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2106 /* Go through connd_routes queue looking for a route that we can process
2107 * right now, @timeout_p can be updated if we need to come back later */
2108 static ksock_route_t *
2109 ksocknal_connd_get_route_locked(signed long *timeout_p)
2111 ksock_route_t *route;
2114 now = cfs_time_current();
2116 /* connd_routes can contain both pending and ordinary routes */
2117 list_for_each_entry (route, &ksocknal_data.ksnd_connd_routes,
2120 if (route->ksnr_retry_interval == 0 ||
2121 cfs_time_aftereq(now, route->ksnr_timeout))
2124 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2125 (int)*timeout_p > (int)(route->ksnr_timeout - now))
2126 *timeout_p = (int)(route->ksnr_timeout - now);
2133 ksocknal_connd (void *arg)
2135 spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2136 ksock_connreq_t *cr;
2141 cfs_block_allsigs ();
2143 init_waitqueue_entry_current (&wait);
2145 spin_lock_bh(connd_lock);
2147 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2148 ksocknal_data.ksnd_connd_starting--;
2149 ksocknal_data.ksnd_connd_running++;
2151 while (!ksocknal_data.ksnd_shuttingdown) {
2152 ksock_route_t *route = NULL;
2153 long sec = cfs_time_current_sec();
2154 long timeout = MAX_SCHEDULE_TIMEOUT;
2155 int dropped_lock = 0;
2157 if (ksocknal_connd_check_stop(sec, &timeout)) {
2158 /* wakeup another one to check stop */
2159 wake_up(&ksocknal_data.ksnd_connd_waitq);
2163 if (ksocknal_connd_check_start(sec, &timeout)) {
2164 /* created new thread */
2168 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2169 /* Connection accepted by the listener */
2170 cr = list_entry(ksocknal_data.ksnd_connd_connreqs. \
2171 next, ksock_connreq_t, ksncr_list);
2173 list_del(&cr->ksncr_list);
2174 spin_unlock_bh(connd_lock);
2177 ksocknal_create_conn(cr->ksncr_ni, NULL,
2178 cr->ksncr_sock, SOCKLND_CONN_NONE);
2179 lnet_ni_decref(cr->ksncr_ni);
2180 LIBCFS_FREE(cr, sizeof(*cr));
2182 spin_lock_bh(connd_lock);
2185 /* Only handle an outgoing connection request if there
2186 * is a thread left to handle incoming connections and
2187 * create new connd */
2188 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2189 ksocknal_data.ksnd_connd_running) {
2190 route = ksocknal_connd_get_route_locked(&timeout);
2192 if (route != NULL) {
2193 list_del (&route->ksnr_connd_list);
2194 ksocknal_data.ksnd_connd_connecting++;
2195 spin_unlock_bh(connd_lock);
2198 if (ksocknal_connect(route)) {
2199 /* consecutive retry */
2200 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2201 CWARN("massive consecutive "
2202 "re-connecting to %pI4h\n",
2203 &route->ksnr_ipaddr);
2210 ksocknal_route_decref(route);
2212 spin_lock_bh(connd_lock);
2213 ksocknal_data.ksnd_connd_connecting--;
2217 if (++nloops < SOCKNAL_RESCHED)
2219 spin_unlock_bh(connd_lock);
2222 spin_lock_bh(connd_lock);
2226 /* Nothing to do for 'timeout' */
2227 set_current_state(TASK_INTERRUPTIBLE);
2228 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2229 spin_unlock_bh(connd_lock);
2232 waitq_timedwait(&wait, TASK_INTERRUPTIBLE, timeout);
2234 set_current_state(TASK_RUNNING);
2235 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2236 spin_lock_bh(connd_lock);
2238 ksocknal_data.ksnd_connd_running--;
2239 spin_unlock_bh(connd_lock);
2241 ksocknal_thread_fini();
2246 ksocknal_find_timed_out_conn (ksock_peer_t *peer)
2248 /* We're called with a shared lock on ksnd_global_lock */
2250 struct list_head *ctmp;
2252 list_for_each (ctmp, &peer->ksnp_conns) {
2254 conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
2256 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2257 LASSERT (!conn->ksnc_closing);
2259 /* SOCK_ERROR will reset error code of socket in
2260 * some platform (like Darwin8.x) */
2261 error = cfs_sock_error(conn->ksnc_sock);
2263 ksocknal_conn_addref(conn);
2267 CNETERR("A connection with %s "
2268 "(%pI4h:%d) was reset; "
2269 "it may have rebooted.\n",
2270 libcfs_id2str(peer->ksnp_id),
2275 CNETERR("A connection with %s "
2276 "(%pI4h:%d) timed out; the "
2277 "network or node may be down.\n",
2278 libcfs_id2str(peer->ksnp_id),
2283 CNETERR("An unexpected network error %d "
2285 "(%pI4h:%d\n", error,
2286 libcfs_id2str(peer->ksnp_id),
2295 if (conn->ksnc_rx_started &&
2296 cfs_time_aftereq(cfs_time_current(),
2297 conn->ksnc_rx_deadline)) {
2298 /* Timed out incomplete incoming message */
2299 ksocknal_conn_addref(conn);
2300 CNETERR("Timeout receiving from %s (%pI4h:%d), "
2301 "state %d wanted %d left %d\n",
2302 libcfs_id2str(peer->ksnp_id),
2305 conn->ksnc_rx_state,
2306 conn->ksnc_rx_nob_wanted,
2307 conn->ksnc_rx_nob_left);
2311 if ((!list_empty(&conn->ksnc_tx_queue) ||
2312 cfs_sock_wmem_queued(conn->ksnc_sock) != 0) &&
2313 cfs_time_aftereq(cfs_time_current(),
2314 conn->ksnc_tx_deadline)) {
2315 /* Timed out messages queued for sending or
2316 * buffered in the socket's send buffer */
2317 ksocknal_conn_addref(conn);
2318 CNETERR("Timeout sending data to %s (%pI4h:%d) "
2319 "the network or that node may be down.\n",
2320 libcfs_id2str(peer->ksnp_id),
2331 ksocknal_flush_stale_txs(ksock_peer_t *peer)
2334 LIST_HEAD (stale_txs);
2336 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2338 while (!list_empty (&peer->ksnp_tx_queue)) {
2339 tx = list_entry (peer->ksnp_tx_queue.next,
2340 ksock_tx_t, tx_list);
2342 if (!cfs_time_aftereq(cfs_time_current(),
2346 list_del (&tx->tx_list);
2347 list_add_tail (&tx->tx_list, &stale_txs);
2350 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2352 ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2356 ksocknal_send_keepalive_locked(ksock_peer_t *peer)
2358 ksock_sched_t *sched;
2362 if (list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2365 if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2368 if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2369 cfs_time_before(cfs_time_current(),
2370 cfs_time_add(peer->ksnp_last_alive,
2371 cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2374 if (cfs_time_before(cfs_time_current(),
2375 peer->ksnp_send_keepalive))
2378 /* retry 10 secs later, so we wouldn't put pressure
2379 * on this peer if we failed to send keepalive this time */
2380 peer->ksnp_send_keepalive = cfs_time_shift(10);
2382 conn = ksocknal_find_conn_locked(peer, NULL, 1);
2384 sched = conn->ksnc_scheduler;
2386 spin_lock_bh(&sched->kss_lock);
2387 if (!list_empty(&conn->ksnc_tx_queue)) {
2388 spin_unlock_bh(&sched->kss_lock);
2389 /* there is an queued ACK, don't need keepalive */
2393 spin_unlock_bh(&sched->kss_lock);
2396 read_unlock(&ksocknal_data.ksnd_global_lock);
2398 /* cookie = 1 is reserved for keepalive PING */
2399 tx = ksocknal_alloc_tx_noop(1, 1);
2401 read_lock(&ksocknal_data.ksnd_global_lock);
2405 if (ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id) == 0) {
2406 read_lock(&ksocknal_data.ksnd_global_lock);
2410 ksocknal_free_tx(tx);
2411 read_lock(&ksocknal_data.ksnd_global_lock);
2418 ksocknal_check_peer_timeouts (int idx)
2420 struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2426 /* NB. We expect to have a look at all the peers and not find any
2427 * connections to time out, so we just use a shared lock while we
2429 read_lock(&ksocknal_data.ksnd_global_lock);
2431 list_for_each_entry(peer, peers, ksnp_list) {
2432 cfs_time_t deadline = 0;
2436 if (ksocknal_send_keepalive_locked(peer) != 0) {
2437 read_unlock(&ksocknal_data.ksnd_global_lock);
2441 conn = ksocknal_find_timed_out_conn (peer);
2444 read_unlock(&ksocknal_data.ksnd_global_lock);
2446 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2448 /* NB we won't find this one again, but we can't
2449 * just proceed with the next peer, since we dropped
2450 * ksnd_global_lock and it might be dead already! */
2451 ksocknal_conn_decref(conn);
2455 /* we can't process stale txs right here because we're
2456 * holding only shared lock */
2457 if (!list_empty (&peer->ksnp_tx_queue)) {
2459 list_entry (peer->ksnp_tx_queue.next,
2460 ksock_tx_t, tx_list);
2462 if (cfs_time_aftereq(cfs_time_current(),
2465 ksocknal_peer_addref(peer);
2466 read_unlock(&ksocknal_data.ksnd_global_lock);
2468 ksocknal_flush_stale_txs(peer);
2470 ksocknal_peer_decref(peer);
2475 if (list_empty(&peer->ksnp_zc_req_list))
2478 spin_lock(&peer->ksnp_lock);
2479 list_for_each_entry(tx, &peer->ksnp_zc_req_list, tx_zc_list) {
2480 if (!cfs_time_aftereq(cfs_time_current(),
2483 /* ignore the TX if connection is being closed */
2484 if (tx->tx_conn->ksnc_closing)
2490 spin_unlock(&peer->ksnp_lock);
2494 tx = list_entry(peer->ksnp_zc_req_list.next,
2495 ksock_tx_t, tx_zc_list);
2496 deadline = tx->tx_deadline;
2497 resid = tx->tx_resid;
2499 ksocknal_conn_addref(conn);
2501 spin_unlock(&peer->ksnp_lock);
2502 read_unlock(&ksocknal_data.ksnd_global_lock);
2504 CERROR("Total %d stale ZC_REQs for peer %s detected; the "
2505 "oldest(%p) timed out %ld secs ago, "
2506 "resid: %d, wmem: %d\n",
2507 n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2508 cfs_duration_sec(cfs_time_current() - deadline),
2509 resid, cfs_sock_wmem_queued(conn->ksnc_sock));
2511 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2512 ksocknal_conn_decref(conn);
2516 read_unlock(&ksocknal_data.ksnd_global_lock);
2520 ksocknal_reaper (void *arg)
2524 ksock_sched_t *sched;
2525 struct list_head enomem_conns;
2527 cfs_duration_t timeout;
2530 cfs_time_t deadline = cfs_time_current();
2532 cfs_block_allsigs ();
2534 INIT_LIST_HEAD(&enomem_conns);
2535 init_waitqueue_entry_current (&wait);
2537 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2539 while (!ksocknal_data.ksnd_shuttingdown) {
2541 if (!list_empty (&ksocknal_data.ksnd_deathrow_conns)) {
2542 conn = list_entry (ksocknal_data. \
2543 ksnd_deathrow_conns.next,
2544 ksock_conn_t, ksnc_list);
2545 list_del (&conn->ksnc_list);
2547 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2549 ksocknal_terminate_conn(conn);
2550 ksocknal_conn_decref(conn);
2552 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2556 if (!list_empty (&ksocknal_data.ksnd_zombie_conns)) {
2557 conn = list_entry (ksocknal_data.ksnd_zombie_conns.\
2558 next, ksock_conn_t, ksnc_list);
2559 list_del (&conn->ksnc_list);
2561 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2563 ksocknal_destroy_conn(conn);
2565 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2569 if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
2570 list_add(&enomem_conns,
2571 &ksocknal_data.ksnd_enomem_conns);
2572 list_del_init(&ksocknal_data.ksnd_enomem_conns);
2575 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2577 /* reschedule all the connections that stalled with ENOMEM... */
2579 while (!list_empty (&enomem_conns)) {
2580 conn = list_entry (enomem_conns.next,
2581 ksock_conn_t, ksnc_tx_list);
2582 list_del (&conn->ksnc_tx_list);
2584 sched = conn->ksnc_scheduler;
2586 spin_lock_bh(&sched->kss_lock);
2588 LASSERT(conn->ksnc_tx_scheduled);
2589 conn->ksnc_tx_ready = 1;
2590 list_add_tail(&conn->ksnc_tx_list,
2591 &sched->kss_tx_conns);
2592 wake_up(&sched->kss_waitq);
2594 spin_unlock_bh(&sched->kss_lock);
2598 /* careful with the jiffy wrap... */
2599 while ((timeout = cfs_time_sub(deadline,
2600 cfs_time_current())) <= 0) {
2603 int chunk = ksocknal_data.ksnd_peer_hash_size;
2605 /* Time to check for timeouts on a few more peers: I do
2606 * checks every 'p' seconds on a proportion of the peer
2607 * table and I need to check every connection 'n' times
2608 * within a timeout interval, to ensure I detect a
2609 * timeout on any connection within (n+1)/n times the
2610 * timeout interval. */
2612 if (*ksocknal_tunables.ksnd_timeout > n * p)
2613 chunk = (chunk * n * p) /
2614 *ksocknal_tunables.ksnd_timeout;
2618 for (i = 0; i < chunk; i++) {
2619 ksocknal_check_peer_timeouts (peer_index);
2620 peer_index = (peer_index + 1) %
2621 ksocknal_data.ksnd_peer_hash_size;
2624 deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2627 if (nenomem_conns != 0) {
2628 /* Reduce my timeout if I rescheduled ENOMEM conns.
2629 * This also prevents me getting woken immediately
2630 * if any go back on my enomem list. */
2631 timeout = SOCKNAL_ENOMEM_RETRY;
2633 ksocknal_data.ksnd_reaper_waketime =
2634 cfs_time_add(cfs_time_current(), timeout);
2636 set_current_state (TASK_INTERRUPTIBLE);
2637 add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2639 if (!ksocknal_data.ksnd_shuttingdown &&
2640 list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
2641 list_empty (&ksocknal_data.ksnd_zombie_conns))
2642 waitq_timedwait (&wait, TASK_INTERRUPTIBLE,
2645 set_current_state (TASK_RUNNING);
2646 remove_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
2648 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2651 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2653 ksocknal_thread_fini();