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.
27 ksocknal_alloc_tx(int type, int size)
29 struct ksock_tx *tx = NULL;
31 if (type == KSOCK_MSG_NOOP) {
32 LASSERT(size == KSOCK_NOOP_TX_SIZE);
34 /* searching for a noop tx in free list */
35 spin_lock(&ksocknal_data.ksnd_tx_lock);
37 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
38 tx = list_entry(ksocknal_data.ksnd_idle_noop_txs.next,
39 struct ksock_tx, tx_list);
40 LASSERT(tx->tx_desc_size == size);
41 list_del(&tx->tx_list);
44 spin_unlock(&ksocknal_data.ksnd_tx_lock);
48 LIBCFS_ALLOC(tx, size);
53 atomic_set(&tx->tx_refcount, 1);
54 tx->tx_zc_aborted = 0;
55 tx->tx_zc_capable = 0;
56 tx->tx_zc_checked = 0;
57 tx->tx_desc_size = size;
59 atomic_inc(&ksocknal_data.ksnd_nactive_txs);
65 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
69 tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
71 CERROR("Can't allocate noop tx desc\n");
76 tx->tx_lnetmsg = NULL;
79 tx->tx_iov = tx->tx_frags.virt.iov;
81 tx->tx_nonblk = nonblk;
83 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
84 tx->tx_msg.ksm_zc_cookies[1] = cookie;
90 ksocknal_free_tx(struct ksock_tx *tx)
92 atomic_dec(&ksocknal_data.ksnd_nactive_txs);
94 if (!tx->tx_lnetmsg && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
96 spin_lock(&ksocknal_data.ksnd_tx_lock);
98 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
100 spin_unlock(&ksocknal_data.ksnd_tx_lock);
102 LIBCFS_FREE(tx, tx->tx_desc_size);
107 ksocknal_send_iov(struct ksock_conn *conn, struct ksock_tx *tx)
109 struct kvec *iov = tx->tx_iov;
113 LASSERT(tx->tx_niov > 0);
115 /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
116 rc = ksocknal_lib_send_iov(conn, tx);
118 if (rc <= 0) /* sent nothing? */
122 LASSERT(nob <= tx->tx_resid);
127 LASSERT(tx->tx_niov > 0);
129 if (nob < (int)iov->iov_len) {
130 iov->iov_base = (void *)((char *)iov->iov_base + nob);
144 ksocknal_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx)
146 lnet_kiov_t *kiov = tx->tx_kiov;
150 LASSERT(!tx->tx_niov);
151 LASSERT(tx->tx_nkiov > 0);
153 /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
154 rc = ksocknal_lib_send_kiov(conn, tx);
156 if (rc <= 0) /* sent nothing? */
160 LASSERT(nob <= tx->tx_resid);
165 LASSERT(tx->tx_nkiov > 0);
167 if (nob < (int)kiov->bv_len) {
168 kiov->bv_offset += nob;
173 nob -= (int)kiov->bv_len;
174 tx->tx_kiov = ++kiov;
182 ksocknal_transmit(struct ksock_conn *conn, struct ksock_tx *tx)
187 if (ksocknal_data.ksnd_stall_tx) {
188 set_current_state(TASK_UNINTERRUPTIBLE);
189 schedule_timeout(cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
192 LASSERT(tx->tx_resid);
194 rc = ksocknal_connsock_addref(conn);
196 LASSERT(conn->ksnc_closing);
201 if (ksocknal_data.ksnd_enomem_tx > 0) {
203 ksocknal_data.ksnd_enomem_tx--;
205 } else if (tx->tx_niov) {
206 rc = ksocknal_send_iov(conn, tx);
208 rc = ksocknal_send_kiov(conn, tx);
211 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
212 if (rc > 0) /* sent something? */
213 conn->ksnc_tx_bufnob += rc; /* account it */
215 if (bufnob < conn->ksnc_tx_bufnob) {
217 * allocated send buffer bytes < computed; infer
218 * something got ACKed
220 conn->ksnc_tx_deadline =
221 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
222 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
223 conn->ksnc_tx_bufnob = bufnob;
227 if (rc <= 0) { /* Didn't write anything? */
229 if (!rc) /* some stacks return 0 instead of -EAGAIN */
232 /* Check if EAGAIN is due to memory pressure */
233 if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
239 /* socket's wmem_queued now includes 'rc' bytes */
240 atomic_sub(rc, &conn->ksnc_tx_nob);
243 } while (tx->tx_resid);
245 ksocknal_connsock_decref(conn);
250 ksocknal_recv_iov(struct ksock_conn *conn)
252 struct kvec *iov = conn->ksnc_rx_iov;
256 LASSERT(conn->ksnc_rx_niov > 0);
259 * Never touch conn->ksnc_rx_iov or change connection
260 * status inside ksocknal_lib_recv_iov
262 rc = ksocknal_lib_recv_iov(conn);
267 /* received something... */
270 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
271 conn->ksnc_rx_deadline =
272 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
273 mb(); /* order with setting rx_started */
274 conn->ksnc_rx_started = 1;
276 conn->ksnc_rx_nob_wanted -= nob;
277 conn->ksnc_rx_nob_left -= nob;
280 LASSERT(conn->ksnc_rx_niov > 0);
282 if (nob < (int)iov->iov_len) {
284 iov->iov_base += nob;
289 conn->ksnc_rx_iov = ++iov;
290 conn->ksnc_rx_niov--;
297 ksocknal_recv_kiov(struct ksock_conn *conn)
299 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
303 LASSERT(conn->ksnc_rx_nkiov > 0);
306 * Never touch conn->ksnc_rx_kiov or change connection
307 * status inside ksocknal_lib_recv_iov
309 rc = ksocknal_lib_recv_kiov(conn);
314 /* received something... */
317 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
318 conn->ksnc_rx_deadline =
319 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
320 mb(); /* order with setting rx_started */
321 conn->ksnc_rx_started = 1;
323 conn->ksnc_rx_nob_wanted -= nob;
324 conn->ksnc_rx_nob_left -= nob;
327 LASSERT(conn->ksnc_rx_nkiov > 0);
329 if (nob < (int)kiov->bv_len) {
330 kiov->bv_offset += nob;
336 conn->ksnc_rx_kiov = ++kiov;
337 conn->ksnc_rx_nkiov--;
344 ksocknal_receive(struct ksock_conn *conn)
347 * Return 1 on success, 0 on EOF, < 0 on error.
348 * Caller checks ksnc_rx_nob_wanted to determine
349 * progress/completion.
353 if (ksocknal_data.ksnd_stall_rx) {
354 set_current_state(TASK_UNINTERRUPTIBLE);
355 schedule_timeout(cfs_time_seconds(ksocknal_data.ksnd_stall_rx));
358 rc = ksocknal_connsock_addref(conn);
360 LASSERT(conn->ksnc_closing);
365 if (conn->ksnc_rx_niov)
366 rc = ksocknal_recv_iov(conn);
368 rc = ksocknal_recv_kiov(conn);
371 /* error/EOF or partial receive */
374 } else if (!rc && conn->ksnc_rx_started) {
375 /* EOF in the middle of a message */
381 /* Completed a fragment */
383 if (!conn->ksnc_rx_nob_wanted) {
389 ksocknal_connsock_decref(conn);
394 ksocknal_tx_done(lnet_ni_t *ni, struct ksock_tx *tx)
396 lnet_msg_t *lnetmsg = tx->tx_lnetmsg;
397 int rc = (!tx->tx_resid && !tx->tx_zc_aborted) ? 0 : -EIO;
399 LASSERT(ni || tx->tx_conn);
402 ksocknal_conn_decref(tx->tx_conn);
404 if (!ni && tx->tx_conn)
405 ni = tx->tx_conn->ksnc_peer->ksnp_ni;
407 ksocknal_free_tx(tx);
408 if (lnetmsg) /* KSOCK_MSG_NOOP go without lnetmsg */
409 lnet_finalize(ni, lnetmsg, rc);
413 ksocknal_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int error)
417 while (!list_empty(txlist)) {
418 tx = list_entry(txlist->next, struct ksock_tx, tx_list);
420 if (error && tx->tx_lnetmsg) {
421 CNETERR("Deleting packet type %d len %d %s->%s\n",
422 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type),
423 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length),
424 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
425 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
427 CNETERR("Deleting noop packet\n");
430 list_del(&tx->tx_list);
432 LASSERT(atomic_read(&tx->tx_refcount) == 1);
433 ksocknal_tx_done(ni, tx);
438 ksocknal_check_zc_req(struct ksock_tx *tx)
440 struct ksock_conn *conn = tx->tx_conn;
441 struct ksock_peer *peer = conn->ksnc_peer;
444 * Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
445 * to ksnp_zc_req_list if some fragment of this message should be sent
446 * zero-copy. Our peer will send an ACK containing this cookie when
447 * she has received this message to tell us we can signal completion.
448 * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
451 LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
452 LASSERT(tx->tx_zc_capable);
454 tx->tx_zc_checked = 1;
456 if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
457 !conn->ksnc_zc_capable)
461 * assign cookie and queue tx to pending list, it will be released when
462 * a matching ack is received. See ksocknal_handle_zcack()
464 ksocknal_tx_addref(tx);
466 spin_lock(&peer->ksnp_lock);
468 /* ZC_REQ is going to be pinned to the peer */
470 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
472 LASSERT(!tx->tx_msg.ksm_zc_cookies[0]);
474 tx->tx_msg.ksm_zc_cookies[0] = peer->ksnp_zc_next_cookie++;
476 if (!peer->ksnp_zc_next_cookie)
477 peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
479 list_add_tail(&tx->tx_zc_list, &peer->ksnp_zc_req_list);
481 spin_unlock(&peer->ksnp_lock);
485 ksocknal_uncheck_zc_req(struct ksock_tx *tx)
487 struct ksock_peer *peer = tx->tx_conn->ksnc_peer;
489 LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
490 LASSERT(tx->tx_zc_capable);
492 tx->tx_zc_checked = 0;
494 spin_lock(&peer->ksnp_lock);
496 if (!tx->tx_msg.ksm_zc_cookies[0]) {
497 /* Not waiting for an ACK */
498 spin_unlock(&peer->ksnp_lock);
502 tx->tx_msg.ksm_zc_cookies[0] = 0;
503 list_del(&tx->tx_zc_list);
505 spin_unlock(&peer->ksnp_lock);
507 ksocknal_tx_decref(tx);
511 ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx)
515 if (tx->tx_zc_capable && !tx->tx_zc_checked)
516 ksocknal_check_zc_req(tx);
518 rc = ksocknal_transmit(conn, tx);
520 CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
523 /* Sent everything OK */
535 counter++; /* exponential backoff warnings */
536 if ((counter & (-counter)) == counter)
537 CWARN("%u ENOMEM tx %p\n", counter, conn);
539 /* Queue on ksnd_enomem_conns for retry after a timeout */
540 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
542 /* enomem list takes over scheduler's ref... */
543 LASSERT(conn->ksnc_tx_scheduled);
544 list_add_tail(&conn->ksnc_tx_list,
545 &ksocknal_data.ksnd_enomem_conns);
546 if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
547 SOCKNAL_ENOMEM_RETRY),
548 ksocknal_data.ksnd_reaper_waketime))
549 wake_up(&ksocknal_data.ksnd_reaper_waitq);
551 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
558 if (!conn->ksnc_closing) {
561 LCONSOLE_WARN("Host %pI4h reset our connection while we were sending data; it may have rebooted.\n",
565 LCONSOLE_WARN("There was an unexpected network error while writing to %pI4h: %d.\n",
566 &conn->ksnc_ipaddr, rc);
569 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pI4h:%d\n",
571 libcfs_id2str(conn->ksnc_peer->ksnp_id),
576 if (tx->tx_zc_checked)
577 ksocknal_uncheck_zc_req(tx);
579 /* it's not an error if conn is being closed */
580 ksocknal_close_conn_and_siblings(conn, (conn->ksnc_closing) ? 0 : rc);
586 ksocknal_launch_connection_locked(struct ksock_route *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);
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 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(struct ksock_peer *peer)
609 struct ksock_route *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(struct ksock_peer *peer, struct ksock_tx *tx, int nonblk)
625 struct list_head *tmp;
626 struct ksock_conn *conn;
627 struct ksock_conn *typed = NULL;
628 struct ksock_conn *fallback = NULL;
632 list_for_each(tmp, &peer->ksnp_conns) {
633 struct ksock_conn *c = list_entry(tmp, struct ksock_conn, ksnc_list);
634 int nob = atomic_read(&c->ksnc_tx_nob) +
635 c->ksnc_sock->sk->sk_wmem_queued;
638 LASSERT(!c->ksnc_closing);
639 LASSERT(c->ksnc_proto &&
640 c->ksnc_proto->pro_match_tx);
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 || 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 || 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) ? typed : fallback;
674 conn->ksnc_tx_last_post = cfs_time_current();
680 ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *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(struct ksock_tx *tx, struct ksock_conn *conn)
692 struct ksock_sched *sched = conn->ksnc_scheduler;
693 ksock_msg_t *msg = &tx->tx_msg;
694 struct ksock_tx *ztx = NULL;
698 * called holding global lock (read or irq-write) and caller may
699 * not have dropped this lock between finding conn and calling me,
700 * so we don't need the {get,put}connsock dance to deref
703 LASSERT(!conn->ksnc_closing);
705 CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
706 libcfs_id2str(conn->ksnc_peer->ksnp_id),
707 &conn->ksnc_ipaddr, conn->ksnc_port);
709 ksocknal_tx_prep(conn, tx);
712 * Ensure the frags we've been given EXACTLY match the number of
713 * bytes we want to send. Many TCP/IP stacks disregard any total
714 * size parameters passed to them and just look at the frags.
716 * We always expect at least 1 mapped fragment containing the
717 * complete ksocknal message header.
719 LASSERT(lnet_iov_nob(tx->tx_niov, tx->tx_iov) +
720 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
721 (unsigned int)tx->tx_nob);
722 LASSERT(tx->tx_niov >= 1);
723 LASSERT(tx->tx_resid == tx->tx_nob);
725 CDEBUG(D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
726 tx, (tx->tx_lnetmsg) ? tx->tx_lnetmsg->msg_hdr.type :
728 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
731 * FIXME: SOCK_WMEM_QUEUED and SOCK_ERROR could block in __DARWIN8__
732 * but they're used inside spinlocks a lot.
734 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
735 spin_lock_bh(&sched->kss_lock);
737 if (list_empty(&conn->ksnc_tx_queue) && !bufnob) {
738 /* First packet starts the timeout */
739 conn->ksnc_tx_deadline =
740 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
741 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
742 conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
743 conn->ksnc_tx_bufnob = 0;
744 mb(); /* order with adding to tx_queue */
747 if (msg->ksm_type == KSOCK_MSG_NOOP) {
749 * The packet is noop ZC ACK, try to piggyback the ack_cookie
750 * on a normal packet so I don't need to send it
752 LASSERT(msg->ksm_zc_cookies[1]);
753 LASSERT(conn->ksnc_proto->pro_queue_tx_zcack);
755 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
756 ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
760 * It's a normal packet - can it piggback a noop zc-ack that
761 * has been queued already?
763 LASSERT(!msg->ksm_zc_cookies[1]);
764 LASSERT(conn->ksnc_proto->pro_queue_tx_msg);
766 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
767 /* ztx will be released later */
771 atomic_sub(ztx->tx_nob, &conn->ksnc_tx_nob);
772 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
775 if (conn->ksnc_tx_ready && /* able to send */
776 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
777 /* +1 ref for scheduler */
778 ksocknal_conn_addref(conn);
779 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
780 conn->ksnc_tx_scheduled = 1;
781 wake_up(&sched->kss_waitq);
784 spin_unlock_bh(&sched->kss_lock);
788 ksocknal_find_connectable_route_locked(struct ksock_peer *peer)
790 unsigned long now = cfs_time_current();
791 struct list_head *tmp;
792 struct ksock_route *route;
794 list_for_each(tmp, &peer->ksnp_routes) {
795 route = list_entry(tmp, struct ksock_route, ksnr_list);
797 LASSERT(!route->ksnr_connecting || route->ksnr_scheduled);
799 if (route->ksnr_scheduled) /* connections being established */
802 /* all route types connected ? */
803 if (!(ksocknal_route_mask() & ~route->ksnr_connected))
806 if (!(!route->ksnr_retry_interval || /* first attempt */
807 cfs_time_aftereq(now, route->ksnr_timeout))) {
809 "Too soon to retry route %pI4h (cnted %d, interval %ld, %ld secs later)\n",
811 route->ksnr_connected,
812 route->ksnr_retry_interval,
813 cfs_duration_sec(route->ksnr_timeout - now));
824 ksocknal_find_connecting_route_locked(struct ksock_peer *peer)
826 struct list_head *tmp;
827 struct ksock_route *route;
829 list_for_each(tmp, &peer->ksnp_routes) {
830 route = list_entry(tmp, struct ksock_route, ksnr_list);
832 LASSERT(!route->ksnr_connecting || route->ksnr_scheduled);
834 if (route->ksnr_scheduled)
842 ksocknal_launch_packet(lnet_ni_t *ni, struct ksock_tx *tx, lnet_process_id_t id)
844 struct ksock_peer *peer;
845 struct ksock_conn *conn;
850 LASSERT(!tx->tx_conn);
852 g_lock = &ksocknal_data.ksnd_global_lock;
854 for (retry = 0;; retry = 1) {
856 peer = ksocknal_find_peer_locked(ni, id);
858 if (!ksocknal_find_connectable_route_locked(peer)) {
859 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
862 * I've got no routes that need to be
863 * connecting and I do have an actual
866 ksocknal_queue_tx_locked(tx, conn);
873 /* I'll need a write lock... */
876 write_lock_bh(g_lock);
878 peer = ksocknal_find_peer_locked(ni, id);
882 write_unlock_bh(g_lock);
884 if (id.pid & LNET_PID_USERFLAG) {
885 CERROR("Refusing to create a connection to userspace process %s\n",
887 return -EHOSTUNREACH;
891 CERROR("Can't find peer %s\n", libcfs_id2str(id));
892 return -EHOSTUNREACH;
895 rc = ksocknal_add_peer(ni, id,
896 LNET_NIDADDR(id.nid),
897 lnet_acceptor_port());
899 CERROR("Can't add peer %s: %d\n",
900 libcfs_id2str(id), rc);
905 ksocknal_launch_all_connections_locked(peer);
907 conn = ksocknal_find_conn_locked(peer, tx, tx->tx_nonblk);
909 /* Connection exists; queue message on it */
910 ksocknal_queue_tx_locked(tx, conn);
911 write_unlock_bh(g_lock);
915 if (peer->ksnp_accepting > 0 ||
916 ksocknal_find_connecting_route_locked(peer)) {
917 /* the message is going to be pinned to the peer */
919 cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
921 /* Queue the message until a connection is established */
922 list_add_tail(&tx->tx_list, &peer->ksnp_tx_queue);
923 write_unlock_bh(g_lock);
927 write_unlock_bh(g_lock);
929 /* NB Routes may be ignored if connections to them failed recently */
930 CNETERR("No usable routes to %s\n", libcfs_id2str(id));
931 return -EHOSTUNREACH;
935 ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
938 int type = lntmsg->msg_type;
939 lnet_process_id_t target = lntmsg->msg_target;
940 unsigned int payload_niov = lntmsg->msg_niov;
941 struct kvec *payload_iov = lntmsg->msg_iov;
942 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
943 unsigned int payload_offset = lntmsg->msg_offset;
944 unsigned int payload_nob = lntmsg->msg_len;
950 * NB 'private' is different depending on what we're sending.
953 CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
954 payload_nob, payload_niov, libcfs_id2str(target));
956 LASSERT(!payload_nob || payload_niov > 0);
957 LASSERT(payload_niov <= LNET_MAX_IOV);
958 /* payload is either all vaddrs or all pages */
959 LASSERT(!(payload_kiov && payload_iov));
960 LASSERT(!in_interrupt());
963 desc_size = offsetof(struct ksock_tx,
964 tx_frags.virt.iov[1 + payload_niov]);
966 desc_size = offsetof(struct ksock_tx,
967 tx_frags.paged.kiov[payload_niov]);
969 if (lntmsg->msg_vmflush)
970 mpflag = cfs_memory_pressure_get_and_set();
971 tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
973 CERROR("Can't allocate tx desc type %d size %d\n",
975 if (lntmsg->msg_vmflush)
976 cfs_memory_pressure_restore(mpflag);
980 tx->tx_conn = NULL; /* set when assigned a conn */
981 tx->tx_lnetmsg = lntmsg;
986 tx->tx_iov = tx->tx_frags.virt.iov;
988 lnet_extract_iov(payload_niov, &tx->tx_iov[1],
989 payload_niov, payload_iov,
990 payload_offset, payload_nob);
993 tx->tx_iov = &tx->tx_frags.paged.iov;
994 tx->tx_kiov = tx->tx_frags.paged.kiov;
995 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
996 payload_niov, payload_kiov,
997 payload_offset, payload_nob);
999 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
1000 tx->tx_zc_capable = 1;
1003 socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
1005 /* The first fragment will be set later in pro_pack */
1006 rc = ksocknal_launch_packet(ni, tx, target);
1008 cfs_memory_pressure_restore(mpflag);
1013 ksocknal_free_tx(tx);
1018 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1020 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1023 return PTR_ERR(task);
1025 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1026 ksocknal_data.ksnd_nthreads++;
1027 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1032 ksocknal_thread_fini(void)
1034 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1035 ksocknal_data.ksnd_nthreads--;
1036 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1040 ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip)
1042 static char ksocknal_slop_buffer[4096];
1048 LASSERT(conn->ksnc_proto);
1050 if (*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) {
1051 /* Remind the socket to ack eagerly... */
1052 ksocknal_lib_eager_ack(conn);
1055 if (!nob_to_skip) { /* right at next packet boundary now */
1056 conn->ksnc_rx_started = 0;
1057 mb(); /* racing with timeout thread */
1059 switch (conn->ksnc_proto->pro_version) {
1060 case KSOCK_PROTO_V2:
1061 case KSOCK_PROTO_V3:
1062 conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1063 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1064 conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg;
1066 conn->ksnc_rx_nob_wanted = offsetof(ksock_msg_t, ksm_u);
1067 conn->ksnc_rx_nob_left = offsetof(ksock_msg_t, ksm_u);
1068 conn->ksnc_rx_iov[0].iov_len = offsetof(ksock_msg_t, ksm_u);
1071 case KSOCK_PROTO_V1:
1072 /* Receiving bare lnet_hdr_t */
1073 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1074 conn->ksnc_rx_nob_wanted = sizeof(lnet_hdr_t);
1075 conn->ksnc_rx_nob_left = sizeof(lnet_hdr_t);
1077 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1078 conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg.ksm_u.lnetmsg;
1079 conn->ksnc_rx_iov[0].iov_len = sizeof(lnet_hdr_t);
1085 conn->ksnc_rx_niov = 1;
1087 conn->ksnc_rx_kiov = NULL;
1088 conn->ksnc_rx_nkiov = 0;
1089 conn->ksnc_rx_csum = ~0;
1094 * Set up to skip as much as possible now. If there's more left
1095 * (ran out of iov entries) we'll get called again
1097 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1098 conn->ksnc_rx_nob_left = nob_to_skip;
1099 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1104 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1106 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1107 conn->ksnc_rx_iov[niov].iov_len = nob;
1112 } while (nob_to_skip && /* mustn't overflow conn's rx iov */
1113 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct iovec));
1115 conn->ksnc_rx_niov = niov;
1116 conn->ksnc_rx_kiov = NULL;
1117 conn->ksnc_rx_nkiov = 0;
1118 conn->ksnc_rx_nob_wanted = skipped;
1123 ksocknal_process_receive(struct ksock_conn *conn)
1126 lnet_process_id_t *id;
1129 LASSERT(atomic_read(&conn->ksnc_conn_refcount) > 0);
1131 /* NB: sched lock NOT held */
1132 /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1133 LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1134 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1135 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1136 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1138 if (conn->ksnc_rx_nob_wanted) {
1139 rc = ksocknal_receive(conn);
1142 LASSERT(rc != -EAGAIN);
1145 CDEBUG(D_NET, "[%p] EOF from %s ip %pI4h:%d\n",
1147 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1150 else if (!conn->ksnc_closing)
1151 CERROR("[%p] Error %d on read from %s ip %pI4h:%d\n",
1153 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1157 /* it's not an error if conn is being closed */
1158 ksocknal_close_conn_and_siblings(conn,
1159 (conn->ksnc_closing) ? 0 : rc);
1160 return (!rc ? -ESHUTDOWN : rc);
1163 if (conn->ksnc_rx_nob_wanted) {
1168 switch (conn->ksnc_rx_state) {
1169 case SOCKNAL_RX_KSM_HEADER:
1170 if (conn->ksnc_flip) {
1171 __swab32s(&conn->ksnc_msg.ksm_type);
1172 __swab32s(&conn->ksnc_msg.ksm_csum);
1173 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1174 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1177 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1178 conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1179 CERROR("%s: Unknown message type: %x\n",
1180 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1181 conn->ksnc_msg.ksm_type);
1182 ksocknal_new_packet(conn, 0);
1183 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1187 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1188 conn->ksnc_msg.ksm_csum && /* has checksum */
1189 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1190 /* NOOP Checksum error */
1191 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1192 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1193 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1194 ksocknal_new_packet(conn, 0);
1195 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1199 if (conn->ksnc_msg.ksm_zc_cookies[1]) {
1202 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1204 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1205 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1207 rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1208 conn->ksnc_msg.ksm_zc_cookies[1]);
1211 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1212 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1213 cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1214 ksocknal_new_packet(conn, 0);
1215 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1220 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1221 ksocknal_new_packet(conn, 0);
1222 return 0; /* NOOP is done and just return */
1225 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1226 conn->ksnc_rx_nob_wanted = sizeof(ksock_lnet_msg_t);
1227 conn->ksnc_rx_nob_left = sizeof(ksock_lnet_msg_t);
1229 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1230 conn->ksnc_rx_iov[0].iov_base = &conn->ksnc_msg.ksm_u.lnetmsg;
1231 conn->ksnc_rx_iov[0].iov_len = sizeof(ksock_lnet_msg_t);
1233 conn->ksnc_rx_niov = 1;
1234 conn->ksnc_rx_kiov = NULL;
1235 conn->ksnc_rx_nkiov = 0;
1237 goto again; /* read lnet header now */
1239 case SOCKNAL_RX_LNET_HEADER:
1240 /* unpack message header */
1241 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1243 if (conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) {
1244 /* Userspace peer */
1245 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1246 id = &conn->ksnc_peer->ksnp_id;
1248 /* Substitute process ID assigned at connection time */
1249 lhdr->src_pid = cpu_to_le32(id->pid);
1250 lhdr->src_nid = cpu_to_le64(id->nid);
1253 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1254 ksocknal_conn_addref(conn); /* ++ref while parsing */
1256 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1257 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1258 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1260 /* I just received garbage: give up on this conn */
1261 ksocknal_new_packet(conn, 0);
1262 ksocknal_close_conn_and_siblings(conn, rc);
1263 ksocknal_conn_decref(conn);
1267 /* I'm racing with ksocknal_recv() */
1268 LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1269 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1271 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1274 /* ksocknal_recv() got called */
1277 case SOCKNAL_RX_LNET_PAYLOAD:
1278 /* payload all received */
1281 if (!conn->ksnc_rx_nob_left && /* not truncating */
1282 conn->ksnc_msg.ksm_csum && /* has checksum */
1283 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1284 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1285 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1286 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1290 if (!rc && conn->ksnc_msg.ksm_zc_cookies[0]) {
1291 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1293 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1294 id = &conn->ksnc_peer->ksnp_id;
1296 rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1297 conn->ksnc_msg.ksm_zc_cookies[0],
1298 *ksocknal_tunables.ksnd_nonblk_zcack ||
1299 le64_to_cpu(lhdr->src_nid) != id->nid);
1302 lnet_finalize(conn->ksnc_peer->ksnp_ni, conn->ksnc_cookie, rc);
1305 ksocknal_new_packet(conn, 0);
1306 ksocknal_close_conn_and_siblings(conn, rc);
1311 case SOCKNAL_RX_SLOP:
1312 /* starting new packet? */
1313 if (ksocknal_new_packet(conn, conn->ksnc_rx_nob_left))
1314 return 0; /* come back later */
1315 goto again; /* try to finish reading slop now */
1323 return -EINVAL; /* keep gcc happy */
1327 ksocknal_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
1328 struct iov_iter *to, unsigned int rlen)
1330 struct ksock_conn *conn = private;
1331 struct ksock_sched *sched = conn->ksnc_scheduler;
1333 LASSERT(iov_iter_count(to) <= rlen);
1334 LASSERT(to->nr_segs <= LNET_MAX_IOV);
1336 conn->ksnc_cookie = msg;
1337 conn->ksnc_rx_nob_wanted = iov_iter_count(to);
1338 conn->ksnc_rx_nob_left = rlen;
1340 if (to->type & ITER_KVEC) {
1341 conn->ksnc_rx_nkiov = 0;
1342 conn->ksnc_rx_kiov = NULL;
1343 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1344 conn->ksnc_rx_niov =
1345 lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1346 to->nr_segs, to->kvec,
1347 to->iov_offset, iov_iter_count(to));
1349 conn->ksnc_rx_niov = 0;
1350 conn->ksnc_rx_iov = NULL;
1351 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1352 conn->ksnc_rx_nkiov =
1353 lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1354 to->nr_segs, to->bvec,
1355 to->iov_offset, iov_iter_count(to));
1358 LASSERT(conn->ksnc_rx_scheduled);
1360 spin_lock_bh(&sched->kss_lock);
1362 switch (conn->ksnc_rx_state) {
1363 case SOCKNAL_RX_PARSE_WAIT:
1364 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1365 wake_up(&sched->kss_waitq);
1366 LASSERT(conn->ksnc_rx_ready);
1369 case SOCKNAL_RX_PARSE:
1370 /* scheduler hasn't noticed I'm parsing yet */
1374 conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1376 spin_unlock_bh(&sched->kss_lock);
1377 ksocknal_conn_decref(conn);
1382 ksocknal_sched_cansleep(struct ksock_sched *sched)
1386 spin_lock_bh(&sched->kss_lock);
1388 rc = !ksocknal_data.ksnd_shuttingdown &&
1389 list_empty(&sched->kss_rx_conns) &&
1390 list_empty(&sched->kss_tx_conns);
1392 spin_unlock_bh(&sched->kss_lock);
1396 int ksocknal_scheduler(void *arg)
1398 struct ksock_sched_info *info;
1399 struct ksock_sched *sched;
1400 struct ksock_conn *conn;
1401 struct ksock_tx *tx;
1404 long id = (long)arg;
1406 info = ksocknal_data.ksnd_sched_info[KSOCK_THREAD_CPT(id)];
1407 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
1409 cfs_block_allsigs();
1411 rc = cfs_cpt_bind(lnet_cpt_table(), info->ksi_cpt);
1413 CERROR("Can't set CPT affinity to %d: %d\n",
1417 spin_lock_bh(&sched->kss_lock);
1419 while (!ksocknal_data.ksnd_shuttingdown) {
1420 int did_something = 0;
1422 /* Ensure I progress everything semi-fairly */
1424 if (!list_empty(&sched->kss_rx_conns)) {
1425 conn = list_entry(sched->kss_rx_conns.next,
1426 struct ksock_conn, ksnc_rx_list);
1427 list_del(&conn->ksnc_rx_list);
1429 LASSERT(conn->ksnc_rx_scheduled);
1430 LASSERT(conn->ksnc_rx_ready);
1433 * clear rx_ready in case receive isn't complete.
1434 * Do it BEFORE we call process_recv, since
1435 * data_ready can set it any time after we release
1438 conn->ksnc_rx_ready = 0;
1439 spin_unlock_bh(&sched->kss_lock);
1441 rc = ksocknal_process_receive(conn);
1443 spin_lock_bh(&sched->kss_lock);
1445 /* I'm the only one that can clear this flag */
1446 LASSERT(conn->ksnc_rx_scheduled);
1448 /* Did process_receive get everything it wanted? */
1450 conn->ksnc_rx_ready = 1;
1452 if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1454 * Conn blocked waiting for ksocknal_recv()
1455 * I change its state (under lock) to signal
1456 * it can be rescheduled
1458 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1459 } else if (conn->ksnc_rx_ready) {
1460 /* reschedule for rx */
1461 list_add_tail(&conn->ksnc_rx_list,
1462 &sched->kss_rx_conns);
1464 conn->ksnc_rx_scheduled = 0;
1466 ksocknal_conn_decref(conn);
1472 if (!list_empty(&sched->kss_tx_conns)) {
1475 if (!list_empty(&sched->kss_zombie_noop_txs)) {
1476 list_add(&zlist, &sched->kss_zombie_noop_txs);
1477 list_del_init(&sched->kss_zombie_noop_txs);
1480 conn = list_entry(sched->kss_tx_conns.next,
1481 struct ksock_conn, ksnc_tx_list);
1482 list_del(&conn->ksnc_tx_list);
1484 LASSERT(conn->ksnc_tx_scheduled);
1485 LASSERT(conn->ksnc_tx_ready);
1486 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1488 tx = list_entry(conn->ksnc_tx_queue.next,
1489 struct ksock_tx, tx_list);
1491 if (conn->ksnc_tx_carrier == tx)
1492 ksocknal_next_tx_carrier(conn);
1494 /* dequeue now so empty list => more to send */
1495 list_del(&tx->tx_list);
1498 * Clear tx_ready in case send isn't complete. Do
1499 * it BEFORE we call process_transmit, since
1500 * write_space can set it any time after we release
1503 conn->ksnc_tx_ready = 0;
1504 spin_unlock_bh(&sched->kss_lock);
1506 if (!list_empty(&zlist)) {
1508 * free zombie noop txs, it's fast because
1509 * noop txs are just put in freelist
1511 ksocknal_txlist_done(NULL, &zlist, 0);
1514 rc = ksocknal_process_transmit(conn, tx);
1516 if (rc == -ENOMEM || rc == -EAGAIN) {
1517 /* Incomplete send: replace tx on HEAD of tx_queue */
1518 spin_lock_bh(&sched->kss_lock);
1519 list_add(&tx->tx_list, &conn->ksnc_tx_queue);
1521 /* Complete send; tx -ref */
1522 ksocknal_tx_decref(tx);
1524 spin_lock_bh(&sched->kss_lock);
1525 /* assume space for more */
1526 conn->ksnc_tx_ready = 1;
1529 if (rc == -ENOMEM) {
1531 * Do nothing; after a short timeout, this
1532 * conn will be reposted on kss_tx_conns.
1534 } else if (conn->ksnc_tx_ready &&
1535 !list_empty(&conn->ksnc_tx_queue)) {
1536 /* reschedule for tx */
1537 list_add_tail(&conn->ksnc_tx_list,
1538 &sched->kss_tx_conns);
1540 conn->ksnc_tx_scheduled = 0;
1542 ksocknal_conn_decref(conn);
1547 if (!did_something || /* nothing to do */
1548 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1549 spin_unlock_bh(&sched->kss_lock);
1553 if (!did_something) { /* wait for something to do */
1554 rc = wait_event_interruptible_exclusive(
1556 !ksocknal_sched_cansleep(sched));
1562 spin_lock_bh(&sched->kss_lock);
1566 spin_unlock_bh(&sched->kss_lock);
1567 ksocknal_thread_fini();
1572 * Add connection to kss_rx_conns of scheduler
1573 * and wakeup the scheduler.
1575 void ksocknal_read_callback(struct ksock_conn *conn)
1577 struct ksock_sched *sched;
1579 sched = conn->ksnc_scheduler;
1581 spin_lock_bh(&sched->kss_lock);
1583 conn->ksnc_rx_ready = 1;
1585 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1586 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1587 conn->ksnc_rx_scheduled = 1;
1588 /* extra ref for scheduler */
1589 ksocknal_conn_addref(conn);
1591 wake_up(&sched->kss_waitq);
1593 spin_unlock_bh(&sched->kss_lock);
1597 * Add connection to kss_tx_conns of scheduler
1598 * and wakeup the scheduler.
1600 void ksocknal_write_callback(struct ksock_conn *conn)
1602 struct ksock_sched *sched;
1604 sched = conn->ksnc_scheduler;
1606 spin_lock_bh(&sched->kss_lock);
1608 conn->ksnc_tx_ready = 1;
1610 if (!conn->ksnc_tx_scheduled && /* not being progressed */
1611 !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1612 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1613 conn->ksnc_tx_scheduled = 1;
1614 /* extra ref for scheduler */
1615 ksocknal_conn_addref(conn);
1617 wake_up(&sched->kss_waitq);
1620 spin_unlock_bh(&sched->kss_lock);
1623 static struct ksock_proto *
1624 ksocknal_parse_proto_version(ksock_hello_msg_t *hello)
1628 if (hello->kshm_magic == LNET_PROTO_MAGIC)
1629 version = hello->kshm_version;
1630 else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1631 version = __swab32(hello->kshm_version);
1634 #if SOCKNAL_VERSION_DEBUG
1635 if (*ksocknal_tunables.ksnd_protocol == 1)
1638 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1639 version == KSOCK_PROTO_V3)
1642 if (version == KSOCK_PROTO_V2)
1643 return &ksocknal_protocol_v2x;
1645 if (version == KSOCK_PROTO_V3)
1646 return &ksocknal_protocol_v3x;
1651 if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1652 lnet_magicversion_t *hmv = (lnet_magicversion_t *)hello;
1654 CLASSERT(sizeof(lnet_magicversion_t) ==
1655 offsetof(ksock_hello_msg_t, kshm_src_nid));
1657 if (hmv->version_major == cpu_to_le16(KSOCK_PROTO_V1_MAJOR) &&
1658 hmv->version_minor == cpu_to_le16(KSOCK_PROTO_V1_MINOR))
1659 return &ksocknal_protocol_v1x;
1666 ksocknal_send_hello(lnet_ni_t *ni, struct ksock_conn *conn,
1667 lnet_nid_t peer_nid, ksock_hello_msg_t *hello)
1669 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1670 struct ksock_net *net = (struct ksock_net *)ni->ni_data;
1672 LASSERT(hello->kshm_nips <= LNET_MAX_INTERFACES);
1674 /* rely on caller to hold a ref on socket so it wouldn't disappear */
1675 LASSERT(conn->ksnc_proto);
1677 hello->kshm_src_nid = ni->ni_nid;
1678 hello->kshm_dst_nid = peer_nid;
1679 hello->kshm_src_pid = the_lnet.ln_pid;
1681 hello->kshm_src_incarnation = net->ksnn_incarnation;
1682 hello->kshm_ctype = conn->ksnc_type;
1684 return conn->ksnc_proto->pro_send_hello(conn, hello);
1688 ksocknal_invert_type(int type)
1691 case SOCKLND_CONN_ANY:
1692 case SOCKLND_CONN_CONTROL:
1694 case SOCKLND_CONN_BULK_IN:
1695 return SOCKLND_CONN_BULK_OUT;
1696 case SOCKLND_CONN_BULK_OUT:
1697 return SOCKLND_CONN_BULK_IN;
1699 return SOCKLND_CONN_NONE;
1704 ksocknal_recv_hello(lnet_ni_t *ni, struct ksock_conn *conn,
1705 ksock_hello_msg_t *hello, lnet_process_id_t *peerid,
1708 /* Return < 0 fatal error
1710 * EALREADY lost connection race
1711 * EPROTO protocol version mismatch
1713 struct socket *sock = conn->ksnc_sock;
1714 int active = !!conn->ksnc_proto;
1718 struct ksock_proto *proto;
1719 lnet_process_id_t recv_id;
1721 /* socket type set on active connections - not set on passive */
1722 LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1724 timeout = active ? *ksocknal_tunables.ksnd_timeout :
1725 lnet_acceptor_timeout();
1727 rc = lnet_sock_read(sock, &hello->kshm_magic, sizeof(hello->kshm_magic), timeout);
1729 CERROR("Error %d reading HELLO from %pI4h\n",
1730 rc, &conn->ksnc_ipaddr);
1735 if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1736 hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1737 hello->kshm_magic != le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1738 /* Unexpected magic! */
1739 CERROR("Bad magic(1) %#08x (%#08x expected) from %pI4h\n",
1740 __cpu_to_le32(hello->kshm_magic),
1741 LNET_PROTO_TCP_MAGIC,
1742 &conn->ksnc_ipaddr);
1746 rc = lnet_sock_read(sock, &hello->kshm_version,
1747 sizeof(hello->kshm_version), timeout);
1749 CERROR("Error %d reading HELLO from %pI4h\n",
1750 rc, &conn->ksnc_ipaddr);
1755 proto = ksocknal_parse_proto_version(hello);
1758 /* unknown protocol from peer, tell peer my protocol */
1759 conn->ksnc_proto = &ksocknal_protocol_v3x;
1760 #if SOCKNAL_VERSION_DEBUG
1761 if (*ksocknal_tunables.ksnd_protocol == 2)
1762 conn->ksnc_proto = &ksocknal_protocol_v2x;
1763 else if (*ksocknal_tunables.ksnd_protocol == 1)
1764 conn->ksnc_proto = &ksocknal_protocol_v1x;
1766 hello->kshm_nips = 0;
1767 ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1770 CERROR("Unknown protocol version (%d.x expected) from %pI4h\n",
1771 conn->ksnc_proto->pro_version,
1772 &conn->ksnc_ipaddr);
1777 proto_match = (conn->ksnc_proto == proto);
1778 conn->ksnc_proto = proto;
1780 /* receive the rest of hello message anyway */
1781 rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1783 CERROR("Error %d reading or checking hello from from %pI4h\n",
1784 rc, &conn->ksnc_ipaddr);
1789 *incarnation = hello->kshm_src_incarnation;
1791 if (hello->kshm_src_nid == LNET_NID_ANY) {
1792 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pI4h\n",
1793 &conn->ksnc_ipaddr);
1798 conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1799 /* Userspace NAL assigns peer process ID from socket */
1800 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1801 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1803 recv_id.nid = hello->kshm_src_nid;
1804 recv_id.pid = hello->kshm_src_pid;
1810 /* peer determines type */
1811 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1812 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1813 CERROR("Unexpected type %d from %s ip %pI4h\n",
1814 hello->kshm_ctype, libcfs_id2str(*peerid),
1815 &conn->ksnc_ipaddr);
1822 if (peerid->pid != recv_id.pid ||
1823 peerid->nid != recv_id.nid) {
1824 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host %pI4h, but they claimed they were %s; please check your Lustre configuration.\n",
1825 libcfs_id2str(*peerid),
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 %pI4h %d\n",
1838 conn->ksnc_type, libcfs_id2str(*peerid),
1839 &conn->ksnc_ipaddr, hello->kshm_ctype);
1847 ksocknal_connect(struct ksock_route *route)
1850 struct ksock_peer *peer = route->ksnr_peer;
1853 struct socket *sock;
1854 unsigned long deadline;
1855 int retry_later = 0;
1858 deadline = cfs_time_add(cfs_time_current(),
1859 cfs_time_seconds(*ksocknal_tunables.ksnd_timeout));
1861 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1863 LASSERT(route->ksnr_scheduled);
1864 LASSERT(!route->ksnr_connecting);
1866 route->ksnr_connecting = 1;
1869 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1872 * stop connecting if peer/route got closed under me, or
1873 * route got connected while queued
1875 if (peer->ksnp_closing || route->ksnr_deleted ||
1881 /* reschedule if peer is connecting to me */
1882 if (peer->ksnp_accepting > 0) {
1884 "peer %s(%d) already connecting to me, retry later.\n",
1885 libcfs_nid2str(peer->ksnp_id.nid), peer->ksnp_accepting);
1889 if (retry_later) /* needs reschedule */
1892 if (wanted & (1 << SOCKLND_CONN_ANY)) {
1893 type = SOCKLND_CONN_ANY;
1894 } else if (wanted & (1 << SOCKLND_CONN_CONTROL)) {
1895 type = SOCKLND_CONN_CONTROL;
1896 } else if (wanted & (1 << SOCKLND_CONN_BULK_IN)) {
1897 type = SOCKLND_CONN_BULK_IN;
1899 LASSERT(wanted & (1 << SOCKLND_CONN_BULK_OUT));
1900 type = SOCKLND_CONN_BULK_OUT;
1903 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1905 if (cfs_time_aftereq(cfs_time_current(), deadline)) {
1907 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1913 rc = lnet_connect(&sock, peer->ksnp_id.nid,
1914 route->ksnr_myipaddr,
1915 route->ksnr_ipaddr, route->ksnr_port);
1919 rc = ksocknal_create_conn(peer->ksnp_ni, route, sock, type);
1921 lnet_connect_console_error(rc, peer->ksnp_id.nid,
1928 * A +ve RC means I have to retry because I lost the connection
1929 * race or I have to renegotiate protocol version
1933 CDEBUG(D_NET, "peer %s: conn race, retry later.\n",
1934 libcfs_nid2str(peer->ksnp_id.nid));
1936 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1939 route->ksnr_scheduled = 0;
1940 route->ksnr_connecting = 0;
1944 * re-queue for attention; this frees me up to handle
1945 * the peer's incoming connection request
1947 if (rc == EALREADY ||
1948 (!rc && peer->ksnp_accepting > 0)) {
1950 * We want to introduce a delay before next
1951 * attempt to connect if we lost conn race,
1952 * but the race is resolved quickly usually,
1953 * so min_reconnectms should be good heuristic
1955 route->ksnr_retry_interval =
1956 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms) / 1000;
1957 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1958 route->ksnr_retry_interval);
1961 ksocknal_launch_connection_locked(route);
1964 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1968 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1970 route->ksnr_scheduled = 0;
1971 route->ksnr_connecting = 0;
1973 /* This is a retry rather than a new connection */
1974 route->ksnr_retry_interval *= 2;
1975 route->ksnr_retry_interval =
1976 max(route->ksnr_retry_interval,
1977 cfs_time_seconds(*ksocknal_tunables.ksnd_min_reconnectms) / 1000);
1978 route->ksnr_retry_interval =
1979 min(route->ksnr_retry_interval,
1980 cfs_time_seconds(*ksocknal_tunables.ksnd_max_reconnectms) / 1000);
1982 LASSERT(route->ksnr_retry_interval);
1983 route->ksnr_timeout = cfs_time_add(cfs_time_current(),
1984 route->ksnr_retry_interval);
1986 if (!list_empty(&peer->ksnp_tx_queue) &&
1987 !peer->ksnp_accepting &&
1988 !ksocknal_find_connecting_route_locked(peer)) {
1989 struct ksock_conn *conn;
1992 * ksnp_tx_queue is queued on a conn on successful
1993 * connection for V1.x and V2.x
1995 if (!list_empty(&peer->ksnp_conns)) {
1996 conn = list_entry(peer->ksnp_conns.next,
1997 struct ksock_conn, ksnc_list);
1998 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
2002 * take all the blocked packets while I've got the lock and
2005 list_splice_init(&peer->ksnp_tx_queue, &zombies);
2008 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2010 ksocknal_peer_failed(peer);
2011 ksocknal_txlist_done(peer->ksnp_ni, &zombies, 1);
2016 * check whether we need to create more connds.
2017 * It will try to create new thread if it's necessary, @timeout can
2018 * be updated if failed to create, so caller wouldn't keep try while
2019 * running out of resource.
2022 ksocknal_connd_check_start(time64_t sec, long *timeout)
2026 int total = ksocknal_data.ksnd_connd_starting +
2027 ksocknal_data.ksnd_connd_running;
2029 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2030 /* still in initializing */
2034 if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2035 total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2037 * can't create more connd, or still have enough
2038 * threads to handle more connecting
2043 if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2044 /* no pending connecting request */
2048 if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2049 /* may run out of resource, retry later */
2050 *timeout = cfs_time_seconds(1);
2054 if (ksocknal_data.ksnd_connd_starting > 0) {
2055 /* serialize starting to avoid flood */
2059 ksocknal_data.ksnd_connd_starting_stamp = sec;
2060 ksocknal_data.ksnd_connd_starting++;
2061 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2063 /* NB: total is the next id */
2064 snprintf(name, sizeof(name), "socknal_cd%02d", total);
2065 rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2067 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2072 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2073 ksocknal_data.ksnd_connd_starting--;
2074 ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2080 * check whether current thread can exit, it will return 1 if there are too
2081 * many threads and no creating in past 120 seconds.
2082 * Also, this function may update @timeout to make caller come back
2083 * again to recheck these conditions.
2086 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2090 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2091 /* still in initializing */
2095 if (ksocknal_data.ksnd_connd_starting > 0) {
2096 /* in progress of starting new thread */
2100 if (ksocknal_data.ksnd_connd_running <=
2101 *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2105 /* created thread in past 120 seconds? */
2106 val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2107 SOCKNAL_CONND_TIMEOUT - sec);
2109 *timeout = (val > 0) ? cfs_time_seconds(val) :
2110 cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2114 /* no creating in past 120 seconds */
2116 return ksocknal_data.ksnd_connd_running >
2117 ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2121 * Go through connd_routes queue looking for a route that we can process
2122 * right now, @timeout_p can be updated if we need to come back later
2124 static struct ksock_route *
2125 ksocknal_connd_get_route_locked(signed long *timeout_p)
2127 struct ksock_route *route;
2130 now = cfs_time_current();
2132 /* connd_routes can contain both pending and ordinary routes */
2133 list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes,
2135 if (!route->ksnr_retry_interval ||
2136 cfs_time_aftereq(now, route->ksnr_timeout))
2139 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2140 (int)*timeout_p > (int)(route->ksnr_timeout - now))
2141 *timeout_p = (int)(route->ksnr_timeout - now);
2148 ksocknal_connd(void *arg)
2150 spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2151 struct ksock_connreq *cr;
2156 cfs_block_allsigs();
2158 init_waitqueue_entry(&wait, current);
2160 spin_lock_bh(connd_lock);
2162 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2163 ksocknal_data.ksnd_connd_starting--;
2164 ksocknal_data.ksnd_connd_running++;
2166 while (!ksocknal_data.ksnd_shuttingdown) {
2167 struct ksock_route *route = NULL;
2168 time64_t sec = ktime_get_real_seconds();
2169 long timeout = MAX_SCHEDULE_TIMEOUT;
2170 int dropped_lock = 0;
2172 if (ksocknal_connd_check_stop(sec, &timeout)) {
2173 /* wakeup another one to check stop */
2174 wake_up(&ksocknal_data.ksnd_connd_waitq);
2178 if (ksocknal_connd_check_start(sec, &timeout)) {
2179 /* created new thread */
2183 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2184 /* Connection accepted by the listener */
2185 cr = list_entry(ksocknal_data.ksnd_connd_connreqs.next,
2186 struct ksock_connreq, ksncr_list);
2188 list_del(&cr->ksncr_list);
2189 spin_unlock_bh(connd_lock);
2192 ksocknal_create_conn(cr->ksncr_ni, NULL,
2193 cr->ksncr_sock, SOCKLND_CONN_NONE);
2194 lnet_ni_decref(cr->ksncr_ni);
2195 LIBCFS_FREE(cr, sizeof(*cr));
2197 spin_lock_bh(connd_lock);
2201 * Only handle an outgoing connection request if there
2202 * is a thread left to handle incoming connections and
2205 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2206 ksocknal_data.ksnd_connd_running) {
2207 route = ksocknal_connd_get_route_locked(&timeout);
2210 list_del(&route->ksnr_connd_list);
2211 ksocknal_data.ksnd_connd_connecting++;
2212 spin_unlock_bh(connd_lock);
2215 if (ksocknal_connect(route)) {
2216 /* consecutive retry */
2217 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2218 CWARN("massive consecutive re-connecting to %pI4h\n",
2219 &route->ksnr_ipaddr);
2226 ksocknal_route_decref(route);
2228 spin_lock_bh(connd_lock);
2229 ksocknal_data.ksnd_connd_connecting--;
2233 if (++nloops < SOCKNAL_RESCHED)
2235 spin_unlock_bh(connd_lock);
2238 spin_lock_bh(connd_lock);
2242 /* Nothing to do for 'timeout' */
2243 set_current_state(TASK_INTERRUPTIBLE);
2244 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2245 spin_unlock_bh(connd_lock);
2248 schedule_timeout(timeout);
2250 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2251 spin_lock_bh(connd_lock);
2253 ksocknal_data.ksnd_connd_running--;
2254 spin_unlock_bh(connd_lock);
2256 ksocknal_thread_fini();
2260 static struct ksock_conn *
2261 ksocknal_find_timed_out_conn(struct ksock_peer *peer)
2263 /* We're called with a shared lock on ksnd_global_lock */
2264 struct ksock_conn *conn;
2265 struct list_head *ctmp;
2267 list_for_each(ctmp, &peer->ksnp_conns) {
2270 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
2272 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2273 LASSERT(!conn->ksnc_closing);
2276 * SOCK_ERROR will reset error code of socket in
2277 * some platform (like Darwin8.x)
2279 error = conn->ksnc_sock->sk->sk_err;
2281 ksocknal_conn_addref(conn);
2285 CNETERR("A connection with %s (%pI4h:%d) was reset; it may have rebooted.\n",
2286 libcfs_id2str(peer->ksnp_id),
2291 CNETERR("A connection with %s (%pI4h:%d) timed out; the network or node may be down.\n",
2292 libcfs_id2str(peer->ksnp_id),
2297 CNETERR("An unexpected network error %d occurred with %s (%pI4h:%d\n",
2299 libcfs_id2str(peer->ksnp_id),
2308 if (conn->ksnc_rx_started &&
2309 cfs_time_aftereq(cfs_time_current(),
2310 conn->ksnc_rx_deadline)) {
2311 /* Timed out incomplete incoming message */
2312 ksocknal_conn_addref(conn);
2313 CNETERR("Timeout receiving from %s (%pI4h:%d), state %d wanted %d left %d\n",
2314 libcfs_id2str(peer->ksnp_id),
2317 conn->ksnc_rx_state,
2318 conn->ksnc_rx_nob_wanted,
2319 conn->ksnc_rx_nob_left);
2323 if ((!list_empty(&conn->ksnc_tx_queue) ||
2324 conn->ksnc_sock->sk->sk_wmem_queued) &&
2325 cfs_time_aftereq(cfs_time_current(),
2326 conn->ksnc_tx_deadline)) {
2328 * Timed out messages queued for sending or
2329 * buffered in the socket's send buffer
2331 ksocknal_conn_addref(conn);
2332 CNETERR("Timeout sending data to %s (%pI4h:%d) the network or that node may be down.\n",
2333 libcfs_id2str(peer->ksnp_id),
2344 ksocknal_flush_stale_txs(struct ksock_peer *peer)
2346 struct ksock_tx *tx;
2347 struct ksock_tx *tmp;
2348 LIST_HEAD(stale_txs);
2350 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2352 list_for_each_entry_safe(tx, tmp, &peer->ksnp_tx_queue, tx_list) {
2353 if (!cfs_time_aftereq(cfs_time_current(),
2357 list_del(&tx->tx_list);
2358 list_add_tail(&tx->tx_list, &stale_txs);
2361 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2363 ksocknal_txlist_done(peer->ksnp_ni, &stale_txs, 1);
2367 ksocknal_send_keepalive_locked(struct ksock_peer *peer)
2368 __must_hold(&ksocknal_data.ksnd_global_lock)
2370 struct ksock_sched *sched;
2371 struct ksock_conn *conn;
2372 struct ksock_tx *tx;
2374 if (list_empty(&peer->ksnp_conns)) /* last_alive will be updated by create_conn */
2377 if (peer->ksnp_proto != &ksocknal_protocol_v3x)
2380 if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2381 time_before(cfs_time_current(),
2382 cfs_time_add(peer->ksnp_last_alive,
2383 cfs_time_seconds(*ksocknal_tunables.ksnd_keepalive))))
2386 if (time_before(cfs_time_current(), peer->ksnp_send_keepalive))
2390 * retry 10 secs later, so we wouldn't put pressure
2391 * on this peer if we failed to send keepalive this time
2393 peer->ksnp_send_keepalive = cfs_time_shift(10);
2395 conn = ksocknal_find_conn_locked(peer, NULL, 1);
2397 sched = conn->ksnc_scheduler;
2399 spin_lock_bh(&sched->kss_lock);
2400 if (!list_empty(&conn->ksnc_tx_queue)) {
2401 spin_unlock_bh(&sched->kss_lock);
2402 /* there is an queued ACK, don't need keepalive */
2406 spin_unlock_bh(&sched->kss_lock);
2409 read_unlock(&ksocknal_data.ksnd_global_lock);
2411 /* cookie = 1 is reserved for keepalive PING */
2412 tx = ksocknal_alloc_tx_noop(1, 1);
2414 read_lock(&ksocknal_data.ksnd_global_lock);
2418 if (!ksocknal_launch_packet(peer->ksnp_ni, tx, peer->ksnp_id)) {
2419 read_lock(&ksocknal_data.ksnd_global_lock);
2423 ksocknal_free_tx(tx);
2424 read_lock(&ksocknal_data.ksnd_global_lock);
2430 ksocknal_check_peer_timeouts(int idx)
2432 struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
2433 struct ksock_peer *peer;
2434 struct ksock_conn *conn;
2435 struct ksock_tx *tx;
2439 * NB. We expect to have a look at all the peers and not find any
2440 * connections to time out, so we just use a shared lock while we
2443 read_lock(&ksocknal_data.ksnd_global_lock);
2445 list_for_each_entry(peer, peers, ksnp_list) {
2446 unsigned long deadline = 0;
2450 if (ksocknal_send_keepalive_locked(peer)) {
2451 read_unlock(&ksocknal_data.ksnd_global_lock);
2455 conn = ksocknal_find_timed_out_conn(peer);
2458 read_unlock(&ksocknal_data.ksnd_global_lock);
2460 ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2463 * NB we won't find this one again, but we can't
2464 * just proceed with the next peer, since we dropped
2465 * ksnd_global_lock and it might be dead already!
2467 ksocknal_conn_decref(conn);
2472 * we can't process stale txs right here because we're
2473 * holding only shared lock
2475 if (!list_empty(&peer->ksnp_tx_queue)) {
2476 struct ksock_tx *tx = list_entry(peer->ksnp_tx_queue.next,
2477 struct ksock_tx, tx_list);
2479 if (cfs_time_aftereq(cfs_time_current(),
2481 ksocknal_peer_addref(peer);
2482 read_unlock(&ksocknal_data.ksnd_global_lock);
2484 ksocknal_flush_stale_txs(peer);
2486 ksocknal_peer_decref(peer);
2491 if (list_empty(&peer->ksnp_zc_req_list))
2494 spin_lock(&peer->ksnp_lock);
2495 list_for_each_entry(tx, &peer->ksnp_zc_req_list, tx_zc_list) {
2496 if (!cfs_time_aftereq(cfs_time_current(),
2499 /* ignore the TX if connection is being closed */
2500 if (tx->tx_conn->ksnc_closing)
2506 spin_unlock(&peer->ksnp_lock);
2510 tx = list_entry(peer->ksnp_zc_req_list.next,
2511 struct ksock_tx, tx_zc_list);
2512 deadline = tx->tx_deadline;
2513 resid = tx->tx_resid;
2515 ksocknal_conn_addref(conn);
2517 spin_unlock(&peer->ksnp_lock);
2518 read_unlock(&ksocknal_data.ksnd_global_lock);
2520 CERROR("Total %d stale ZC_REQs for peer %s detected; the oldest(%p) timed out %ld secs ago, resid: %d, wmem: %d\n",
2521 n, libcfs_nid2str(peer->ksnp_id.nid), tx,
2522 cfs_duration_sec(cfs_time_current() - deadline),
2523 resid, conn->ksnc_sock->sk->sk_wmem_queued);
2525 ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2526 ksocknal_conn_decref(conn);
2530 read_unlock(&ksocknal_data.ksnd_global_lock);
2534 ksocknal_reaper(void *arg)
2537 struct ksock_conn *conn;
2538 struct ksock_sched *sched;
2539 struct list_head enomem_conns;
2544 unsigned long deadline = cfs_time_current();
2546 cfs_block_allsigs();
2548 INIT_LIST_HEAD(&enomem_conns);
2549 init_waitqueue_entry(&wait, current);
2551 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2553 while (!ksocknal_data.ksnd_shuttingdown) {
2554 if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) {
2555 conn = list_entry(ksocknal_data.ksnd_deathrow_conns.next,
2556 struct ksock_conn, ksnc_list);
2557 list_del(&conn->ksnc_list);
2559 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2561 ksocknal_terminate_conn(conn);
2562 ksocknal_conn_decref(conn);
2564 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2568 if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) {
2569 conn = list_entry(ksocknal_data.ksnd_zombie_conns.next,
2570 struct ksock_conn, ksnc_list);
2571 list_del(&conn->ksnc_list);
2573 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2575 ksocknal_destroy_conn(conn);
2577 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2581 if (!list_empty(&ksocknal_data.ksnd_enomem_conns)) {
2582 list_add(&enomem_conns,
2583 &ksocknal_data.ksnd_enomem_conns);
2584 list_del_init(&ksocknal_data.ksnd_enomem_conns);
2587 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2589 /* reschedule all the connections that stalled with ENOMEM... */
2591 while (!list_empty(&enomem_conns)) {
2592 conn = list_entry(enomem_conns.next, struct ksock_conn,
2594 list_del(&conn->ksnc_tx_list);
2596 sched = conn->ksnc_scheduler;
2598 spin_lock_bh(&sched->kss_lock);
2600 LASSERT(conn->ksnc_tx_scheduled);
2601 conn->ksnc_tx_ready = 1;
2602 list_add_tail(&conn->ksnc_tx_list,
2603 &sched->kss_tx_conns);
2604 wake_up(&sched->kss_waitq);
2606 spin_unlock_bh(&sched->kss_lock);
2610 /* careful with the jiffy wrap... */
2611 while ((timeout = cfs_time_sub(deadline,
2612 cfs_time_current())) <= 0) {
2615 int chunk = ksocknal_data.ksnd_peer_hash_size;
2618 * Time to check for timeouts on a few more peers: I do
2619 * checks every 'p' seconds on a proportion of the peer
2620 * table and I need to check every connection 'n' times
2621 * within a timeout interval, to ensure I detect a
2622 * timeout on any connection within (n+1)/n times the
2625 if (*ksocknal_tunables.ksnd_timeout > n * p)
2626 chunk = (chunk * n * p) /
2627 *ksocknal_tunables.ksnd_timeout;
2631 for (i = 0; i < chunk; i++) {
2632 ksocknal_check_peer_timeouts(peer_index);
2633 peer_index = (peer_index + 1) %
2634 ksocknal_data.ksnd_peer_hash_size;
2637 deadline = cfs_time_add(deadline, cfs_time_seconds(p));
2640 if (nenomem_conns) {
2642 * Reduce my timeout if I rescheduled ENOMEM conns.
2643 * This also prevents me getting woken immediately
2644 * if any go back on my enomem list.
2646 timeout = SOCKNAL_ENOMEM_RETRY;
2648 ksocknal_data.ksnd_reaper_waketime =
2649 cfs_time_add(cfs_time_current(), timeout);
2651 set_current_state(TASK_INTERRUPTIBLE);
2652 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2654 if (!ksocknal_data.ksnd_shuttingdown &&
2655 list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2656 list_empty(&ksocknal_data.ksnd_zombie_conns))
2657 schedule_timeout(timeout);
2659 set_current_state(TASK_RUNNING);
2660 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2662 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2665 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2667 ksocknal_thread_fini();