2 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
4 * Copyright (c) 2011, 2017, 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 Lustre, https://wiki.whamcloud.com/
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 struct ksock_tx *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.next,
42 struct ksock_tx, 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_hstatus = LNET_MSG_STATUS_OK;
61 tx->tx_desc_size = size;
63 atomic_inc(&ksocknal_data.ksnd_nactive_txs);
69 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
73 tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
75 CERROR("Can't allocate noop tx desc\n");
80 tx->tx_lnetmsg = NULL;
83 tx->tx_iov = tx->tx_frags.virt.iov;
85 tx->tx_nonblk = nonblk;
87 tx->tx_msg.ksm_csum = 0;
88 tx->tx_msg.ksm_type = KSOCK_MSG_NOOP;
89 tx->tx_msg.ksm_zc_cookies[0] = 0;
90 tx->tx_msg.ksm_zc_cookies[1] = cookie;
97 ksocknal_free_tx(struct ksock_tx *tx)
99 atomic_dec(&ksocknal_data.ksnd_nactive_txs);
101 if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
103 spin_lock(&ksocknal_data.ksnd_tx_lock);
105 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
107 spin_unlock(&ksocknal_data.ksnd_tx_lock);
109 LIBCFS_FREE(tx, tx->tx_desc_size);
114 ksocknal_send_iov(struct ksock_conn *conn, struct ksock_tx *tx,
115 struct kvec *scratch_iov)
117 struct kvec *iov = tx->tx_iov;
121 LASSERT(tx->tx_niov > 0);
123 /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
124 rc = ksocknal_lib_send_iov(conn, tx, scratch_iov);
126 if (rc <= 0) /* sent nothing? */
130 LASSERT(nob <= tx->tx_resid);
135 LASSERT(tx->tx_niov > 0);
137 if (nob < (int) iov->iov_len) {
138 iov->iov_base += nob;
152 ksocknal_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx,
153 struct kvec *scratch_iov)
155 lnet_kiov_t *kiov = tx->tx_kiov;
159 LASSERT(tx->tx_niov == 0);
160 LASSERT(tx->tx_nkiov > 0);
162 /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
163 rc = ksocknal_lib_send_kiov(conn, tx, scratch_iov);
165 if (rc <= 0) /* sent nothing? */
169 LASSERT(nob <= tx->tx_resid);
174 LASSERT(tx->tx_nkiov > 0);
176 if (nob < (int)kiov->kiov_len) {
177 kiov->kiov_offset += nob;
178 kiov->kiov_len -= nob;
182 nob -= (int)kiov->kiov_len;
183 tx->tx_kiov = ++kiov;
191 ksocknal_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
192 struct kvec *scratch_iov)
197 if (ksocknal_data.ksnd_stall_tx != 0)
198 schedule_timeout_uninterruptible(
199 cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
201 LASSERT(tx->tx_resid != 0);
203 rc = ksocknal_connsock_addref(conn);
205 LASSERT(conn->ksnc_closing);
210 if (ksocknal_data.ksnd_enomem_tx > 0) {
212 ksocknal_data.ksnd_enomem_tx--;
214 } else if (tx->tx_niov != 0) {
215 rc = ksocknal_send_iov(conn, tx, scratch_iov);
217 rc = ksocknal_send_kiov(conn, tx, scratch_iov);
220 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
221 if (rc > 0) /* sent something? */
222 conn->ksnc_tx_bufnob += rc; /* account it */
224 if (bufnob < conn->ksnc_tx_bufnob) {
225 /* allocated send buffer bytes < computed; infer
226 * something got ACKed */
227 conn->ksnc_tx_deadline = ktime_get_seconds() +
228 lnet_get_lnd_timeout();
229 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
230 conn->ksnc_tx_bufnob = bufnob;
234 if (rc <= 0) { /* Didn't write anything? */
235 /* some stacks return 0 instead of -EAGAIN */
239 /* Check if EAGAIN is due to memory pressure */
240 if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
246 /* socket's wmem_queued now includes 'rc' bytes */
247 atomic_sub (rc, &conn->ksnc_tx_nob);
250 } while (tx->tx_resid != 0);
252 ksocknal_connsock_decref(conn);
257 ksocknal_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov)
259 struct kvec *iov = conn->ksnc_rx_iov;
263 LASSERT(conn->ksnc_rx_niov > 0);
265 /* Never touch conn->ksnc_rx_iov or change connection
266 * status inside ksocknal_lib_recv_iov */
267 rc = ksocknal_lib_recv_iov(conn, scratchiov);
272 /* received something... */
275 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
276 conn->ksnc_rx_deadline = ktime_get_seconds() +
277 lnet_get_lnd_timeout();
278 smp_mb(); /* order with setting rx_started */
279 conn->ksnc_rx_started = 1;
281 conn->ksnc_rx_nob_wanted -= nob;
282 conn->ksnc_rx_nob_left -= nob;
285 LASSERT(conn->ksnc_rx_niov > 0);
287 if (nob < (int)iov->iov_len) {
289 iov->iov_base += nob;
294 conn->ksnc_rx_iov = ++iov;
295 conn->ksnc_rx_niov--;
302 ksocknal_recv_kiov(struct ksock_conn *conn, struct page **rx_scratch_pgs,
303 struct kvec *scratch_iov)
305 lnet_kiov_t *kiov = conn->ksnc_rx_kiov;
308 LASSERT(conn->ksnc_rx_nkiov > 0);
310 /* Never touch conn->ksnc_rx_kiov or change connection
311 * status inside ksocknal_lib_recv_iov */
312 rc = ksocknal_lib_recv_kiov(conn, rx_scratch_pgs, scratch_iov);
317 /* received something... */
320 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
321 conn->ksnc_rx_deadline = ktime_get_seconds() +
322 lnet_get_lnd_timeout();
323 smp_mb(); /* order with setting rx_started */
324 conn->ksnc_rx_started = 1;
326 conn->ksnc_rx_nob_wanted -= nob;
327 conn->ksnc_rx_nob_left -= nob;
330 LASSERT(conn->ksnc_rx_nkiov > 0);
332 if (nob < (int) kiov->kiov_len) {
333 kiov->kiov_offset += nob;
334 kiov->kiov_len -= nob;
338 nob -= kiov->kiov_len;
339 conn->ksnc_rx_kiov = ++kiov;
340 conn->ksnc_rx_nkiov--;
347 ksocknal_receive(struct ksock_conn *conn, struct page **rx_scratch_pgs,
348 struct kvec *scratch_iov)
350 /* Return 1 on success, 0 on EOF, < 0 on error.
351 * Caller checks ksnc_rx_nob_wanted to determine
352 * progress/completion. */
356 if (ksocknal_data.ksnd_stall_rx != 0)
357 schedule_timeout_uninterruptible(
358 cfs_time_seconds(ksocknal_data.ksnd_stall_rx));
360 rc = ksocknal_connsock_addref(conn);
362 LASSERT(conn->ksnc_closing);
367 if (conn->ksnc_rx_niov != 0)
368 rc = ksocknal_recv_iov(conn, scratch_iov);
370 rc = ksocknal_recv_kiov(conn, rx_scratch_pgs,
374 /* error/EOF or partial receive */
377 } else if (rc == 0 && conn->ksnc_rx_started) {
378 /* EOF in the middle of a message */
384 /* Completed a fragment */
386 if (conn->ksnc_rx_nob_wanted == 0) {
392 ksocknal_connsock_decref(conn);
397 ksocknal_tx_done(struct lnet_ni *ni, struct ksock_tx *tx, int rc)
399 struct lnet_msg *lnetmsg = tx->tx_lnetmsg;
400 enum lnet_msg_hstatus hstatus = tx->tx_hstatus;
403 LASSERT(ni != NULL || tx->tx_conn != NULL);
405 if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted)) {
407 if (hstatus == LNET_MSG_STATUS_OK)
408 hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
411 if (tx->tx_conn != NULL)
412 ksocknal_conn_decref(tx->tx_conn);
414 ksocknal_free_tx(tx);
415 if (lnetmsg != NULL) { /* KSOCK_MSG_NOOP go without lnetmsg */
416 lnetmsg->msg_health_status = hstatus;
417 lnet_finalize(lnetmsg, rc);
424 ksocknal_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int error)
428 while (!list_empty(txlist)) {
429 tx = list_entry(txlist->next, struct ksock_tx, tx_list);
431 if (error && tx->tx_lnetmsg != NULL) {
432 CNETERR("Deleting packet type %d len %d %s->%s\n",
433 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type),
434 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length),
435 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
436 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
438 CNETERR("Deleting noop packet\n");
441 list_del(&tx->tx_list);
443 if (tx->tx_hstatus == LNET_MSG_STATUS_OK) {
444 if (error == -ETIMEDOUT)
446 LNET_MSG_STATUS_LOCAL_TIMEOUT;
447 else if (error == -ENETDOWN ||
448 error == -EHOSTUNREACH ||
449 error == -ENETUNREACH ||
450 error == -ECONNREFUSED ||
451 error == -ECONNRESET)
452 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
454 * for all other errors we don't want to
458 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
461 LASSERT(atomic_read(&tx->tx_refcount) == 1);
462 ksocknal_tx_done(ni, tx, error);
467 ksocknal_check_zc_req(struct ksock_tx *tx)
469 struct ksock_conn *conn = tx->tx_conn;
470 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
472 /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
473 * to ksnp_zc_req_list if some fragment of this message should be sent
474 * zero-copy. Our peer_ni will send an ACK containing this cookie when
475 * she has received this message to tell us we can signal completion.
476 * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
477 * ksnp_zc_req_list. */
478 LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
479 LASSERT (tx->tx_zc_capable);
481 tx->tx_zc_checked = 1;
483 if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
484 !conn->ksnc_zc_capable)
487 /* assign cookie and queue tx to pending list, it will be released when
488 * a matching ack is received. See ksocknal_handle_zcack() */
490 ksocknal_tx_addref(tx);
492 spin_lock(&peer_ni->ksnp_lock);
494 /* ZC_REQ is going to be pinned to the peer_ni */
495 tx->tx_deadline = ktime_get_seconds() +
496 lnet_get_lnd_timeout();
498 LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
500 tx->tx_msg.ksm_zc_cookies[0] = peer_ni->ksnp_zc_next_cookie++;
502 if (peer_ni->ksnp_zc_next_cookie == 0)
503 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
505 list_add_tail(&tx->tx_zc_list, &peer_ni->ksnp_zc_req_list);
507 spin_unlock(&peer_ni->ksnp_lock);
511 ksocknal_uncheck_zc_req(struct ksock_tx *tx)
513 struct ksock_peer_ni *peer_ni = tx->tx_conn->ksnc_peer;
515 LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
516 LASSERT(tx->tx_zc_capable);
518 tx->tx_zc_checked = 0;
520 spin_lock(&peer_ni->ksnp_lock);
522 if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
523 /* Not waiting for an ACK */
524 spin_unlock(&peer_ni->ksnp_lock);
528 tx->tx_msg.ksm_zc_cookies[0] = 0;
529 list_del(&tx->tx_zc_list);
531 spin_unlock(&peer_ni->ksnp_lock);
533 ksocknal_tx_decref(tx);
537 ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
538 struct kvec *scratch_iov)
541 bool error_sim = false;
543 if (lnet_send_error_simulation(tx->tx_lnetmsg, &tx->tx_hstatus)) {
549 if (tx->tx_zc_capable && !tx->tx_zc_checked)
550 ksocknal_check_zc_req(tx);
552 rc = ksocknal_transmit(conn, tx, scratch_iov);
554 CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
556 if (tx->tx_resid == 0) {
557 /* Sent everything OK */
569 counter++; /* exponential backoff warnings */
570 if ((counter & (-counter)) == counter)
571 CWARN("%u ENOMEM tx %p (%u allocated)\n",
572 counter, conn, atomic_read(&libcfs_kmemory));
574 /* Queue on ksnd_enomem_conns for retry after a timeout */
575 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
577 /* enomem list takes over scheduler's ref... */
578 LASSERT(conn->ksnc_tx_scheduled);
579 list_add_tail(&conn->ksnc_tx_list,
580 &ksocknal_data.ksnd_enomem_conns);
581 if (ktime_get_seconds() + SOCKNAL_ENOMEM_RETRY <
582 ksocknal_data.ksnd_reaper_waketime)
583 wake_up(&ksocknal_data.ksnd_reaper_waitq);
585 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
588 * set the health status of the message which determines
589 * whether we should retry the transmit
591 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
602 * set the health status of the message which determines
603 * whether we should retry the transmit
605 if (rc == -ETIMEDOUT)
606 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_TIMEOUT;
608 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
611 if (!conn->ksnc_closing) {
614 LCONSOLE_WARN("Host %pI4h reset our connection "
615 "while we were sending data; it may have "
620 LCONSOLE_WARN("There was an unexpected network error "
621 "while writing to %pI4h: %d.\n",
622 &conn->ksnc_ipaddr, rc);
625 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pI4h:%d\n",
626 conn, rc, libcfs_id2str(conn->ksnc_peer->ksnp_id),
627 &conn->ksnc_ipaddr, conn->ksnc_port);
630 if (tx->tx_zc_checked)
631 ksocknal_uncheck_zc_req(tx);
633 /* it's not an error if conn is being closed */
634 ksocknal_close_conn_and_siblings(conn,
635 (conn->ksnc_closing) ? 0 : rc);
641 ksocknal_launch_connection_locked(struct ksock_route *route)
644 /* called holding write lock on ksnd_global_lock */
646 LASSERT (!route->ksnr_scheduled);
647 LASSERT (!route->ksnr_connecting);
648 LASSERT ((ksocknal_route_mask() & ~route->ksnr_connected) != 0);
650 route->ksnr_scheduled = 1; /* scheduling conn for connd */
651 ksocknal_route_addref(route); /* extra ref for connd */
653 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
655 list_add_tail(&route->ksnr_connd_list,
656 &ksocknal_data.ksnd_connd_routes);
657 wake_up(&ksocknal_data.ksnd_connd_waitq);
659 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
663 ksocknal_launch_all_connections_locked(struct ksock_peer_ni *peer_ni)
665 struct ksock_route *route;
667 /* called holding write lock on ksnd_global_lock */
669 /* launch any/all connections that need it */
670 route = ksocknal_find_connectable_route_locked(peer_ni);
674 ksocknal_launch_connection_locked(route);
679 ksocknal_find_conn_locked(struct ksock_peer_ni *peer_ni, struct ksock_tx *tx, int nonblk)
681 struct list_head *tmp;
682 struct ksock_conn *conn;
683 struct ksock_conn *typed = NULL;
684 struct ksock_conn *fallback = NULL;
688 list_for_each(tmp, &peer_ni->ksnp_conns) {
689 struct ksock_conn *c = list_entry(tmp, struct ksock_conn,
691 int nob = atomic_read(&c->ksnc_tx_nob) +
692 c->ksnc_sock->sk->sk_wmem_queued;
695 LASSERT (!c->ksnc_closing);
696 LASSERT (c->ksnc_proto != NULL &&
697 c->ksnc_proto->pro_match_tx != NULL);
699 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
704 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
707 case SOCKNAL_MATCH_YES: /* typed connection */
708 if (typed == NULL || tnob > nob ||
709 (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
710 typed->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
716 case SOCKNAL_MATCH_MAY: /* fallback connection */
717 if (fallback == NULL || fnob > nob ||
718 (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
719 fallback->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
727 /* prefer the typed selection */
728 conn = (typed != NULL) ? typed : fallback;
731 conn->ksnc_tx_last_post = ktime_get_seconds();
737 ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *tx)
739 conn->ksnc_proto->pro_pack(tx);
741 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
742 ksocknal_conn_addref(conn); /* +1 ref for tx */
747 ksocknal_queue_tx_locked(struct ksock_tx *tx, struct ksock_conn *conn)
749 struct ksock_sched *sched = conn->ksnc_scheduler;
750 struct ksock_msg *msg = &tx->tx_msg;
751 struct ksock_tx *ztx = NULL;
754 /* called holding global lock (read or irq-write) and caller may
755 * not have dropped this lock between finding conn and calling me,
756 * so we don't need the {get,put}connsock dance to deref
758 LASSERT(!conn->ksnc_closing);
760 CDEBUG(D_NET, "Sending to %s ip %pI4h:%d\n",
761 libcfs_id2str(conn->ksnc_peer->ksnp_id),
762 &conn->ksnc_ipaddr, conn->ksnc_port);
764 ksocknal_tx_prep(conn, tx);
766 /* Ensure the frags we've been given EXACTLY match the number of
767 * bytes we want to send. Many TCP/IP stacks disregard any total
768 * size parameters passed to them and just look at the frags.
770 * We always expect at least 1 mapped fragment containing the
771 * complete ksocknal message header. */
772 LASSERT (lnet_iov_nob (tx->tx_niov, tx->tx_iov) +
773 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
774 (unsigned int)tx->tx_nob);
775 LASSERT (tx->tx_niov >= 1);
776 LASSERT (tx->tx_resid == tx->tx_nob);
778 CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
779 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
781 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
783 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
784 spin_lock_bh(&sched->kss_lock);
786 if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
787 /* First packet starts the timeout */
788 conn->ksnc_tx_deadline = ktime_get_seconds() +
789 lnet_get_lnd_timeout();
790 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
791 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
792 conn->ksnc_tx_bufnob = 0;
793 smp_mb(); /* order with adding to tx_queue */
796 if (msg->ksm_type == KSOCK_MSG_NOOP) {
797 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
798 * on a normal packet so I don't need to send it */
799 LASSERT (msg->ksm_zc_cookies[1] != 0);
800 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
802 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
803 ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
806 /* It's a normal packet - can it piggback a noop zc-ack that
807 * has been queued already? */
808 LASSERT (msg->ksm_zc_cookies[1] == 0);
809 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
811 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
812 /* ztx will be released later */
816 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
817 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
820 if (conn->ksnc_tx_ready && /* able to send */
821 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
822 /* +1 ref for scheduler */
823 ksocknal_conn_addref(conn);
824 list_add_tail(&conn->ksnc_tx_list,
825 &sched->kss_tx_conns);
826 conn->ksnc_tx_scheduled = 1;
827 wake_up(&sched->kss_waitq);
830 spin_unlock_bh(&sched->kss_lock);
835 ksocknal_find_connectable_route_locked(struct ksock_peer_ni *peer_ni)
837 time64_t now = ktime_get_seconds();
838 struct list_head *tmp;
839 struct ksock_route *route;
841 list_for_each(tmp, &peer_ni->ksnp_routes) {
842 route = list_entry(tmp, struct ksock_route, ksnr_list);
844 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
846 if (route->ksnr_scheduled) /* connections being established */
849 /* all route types connected ? */
850 if ((ksocknal_route_mask() & ~route->ksnr_connected) == 0)
853 if (!(route->ksnr_retry_interval == 0 || /* first attempt */
854 now >= route->ksnr_timeout)) {
856 "Too soon to retry route %pI4h "
857 "(cnted %d, interval %lld, %lld secs later)\n",
859 route->ksnr_connected,
860 route->ksnr_retry_interval,
861 route->ksnr_timeout - now);
872 ksocknal_find_connecting_route_locked(struct ksock_peer_ni *peer_ni)
874 struct list_head *tmp;
875 struct ksock_route *route;
877 list_for_each(tmp, &peer_ni->ksnp_routes) {
878 route = list_entry(tmp, struct ksock_route, ksnr_list);
880 LASSERT (!route->ksnr_connecting || route->ksnr_scheduled);
882 if (route->ksnr_scheduled)
890 ksocknal_launch_packet(struct lnet_ni *ni, struct ksock_tx *tx,
891 struct lnet_process_id id)
893 struct ksock_peer_ni *peer_ni;
894 struct ksock_conn *conn;
899 LASSERT (tx->tx_conn == NULL);
901 g_lock = &ksocknal_data.ksnd_global_lock;
903 for (retry = 0;; retry = 1) {
905 peer_ni = ksocknal_find_peer_locked(ni, id);
906 if (peer_ni != NULL) {
907 if (ksocknal_find_connectable_route_locked(peer_ni) == NULL) {
908 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
910 /* I've got no routes that need to be
911 * connecting and I do have an actual
913 ksocknal_queue_tx_locked (tx, conn);
920 /* I'll need a write lock... */
923 write_lock_bh(g_lock);
925 peer_ni = ksocknal_find_peer_locked(ni, id);
929 write_unlock_bh(g_lock);
931 if ((id.pid & LNET_PID_USERFLAG) != 0) {
932 CERROR("Refusing to create a connection to "
933 "userspace process %s\n", libcfs_id2str(id));
934 return -EHOSTUNREACH;
938 CERROR("Can't find peer_ni %s\n", libcfs_id2str(id));
939 return -EHOSTUNREACH;
942 rc = ksocknal_add_peer(ni, id,
943 LNET_NIDADDR(id.nid),
944 lnet_acceptor_port());
946 CERROR("Can't add peer_ni %s: %d\n",
947 libcfs_id2str(id), rc);
952 ksocknal_launch_all_connections_locked(peer_ni);
954 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
956 /* Connection exists; queue message on it */
957 ksocknal_queue_tx_locked (tx, conn);
958 write_unlock_bh(g_lock);
962 if (peer_ni->ksnp_accepting > 0 ||
963 ksocknal_find_connecting_route_locked (peer_ni) != NULL) {
964 /* the message is going to be pinned to the peer_ni */
965 tx->tx_deadline = ktime_get_seconds() +
966 lnet_get_lnd_timeout();
968 /* Queue the message until a connection is established */
969 list_add_tail(&tx->tx_list, &peer_ni->ksnp_tx_queue);
970 write_unlock_bh(g_lock);
974 write_unlock_bh(g_lock);
976 /* NB Routes may be ignored if connections to them failed recently */
977 CNETERR("No usable routes to %s\n", libcfs_id2str(id));
978 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
979 return (-EHOSTUNREACH);
983 ksocknal_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
986 int type = lntmsg->msg_type;
987 struct lnet_process_id target = lntmsg->msg_target;
988 unsigned int payload_niov = lntmsg->msg_niov;
989 struct kvec *payload_iov = lntmsg->msg_iov;
990 lnet_kiov_t *payload_kiov = lntmsg->msg_kiov;
991 unsigned int payload_offset = lntmsg->msg_offset;
992 unsigned int payload_nob = lntmsg->msg_len;
997 /* NB 'private' is different depending on what we're sending.
998 * Just ignore it... */
1000 CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
1001 payload_nob, payload_niov, libcfs_id2str(target));
1003 LASSERT (payload_nob == 0 || payload_niov > 0);
1004 LASSERT (payload_niov <= LNET_MAX_IOV);
1005 /* payload is either all vaddrs or all pages */
1006 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1007 LASSERT (!in_interrupt ());
1009 if (payload_iov != NULL)
1010 desc_size = offsetof(struct ksock_tx,
1011 tx_frags.virt.iov[1 + payload_niov]);
1013 desc_size = offsetof(struct ksock_tx,
1014 tx_frags.paged.kiov[payload_niov]);
1016 if (lntmsg->msg_vmflush)
1017 mpflag = cfs_memory_pressure_get_and_set();
1018 tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
1020 CERROR("Can't allocate tx desc type %d size %d\n",
1022 if (lntmsg->msg_vmflush)
1023 cfs_memory_pressure_restore(mpflag);
1027 tx->tx_conn = NULL; /* set when assigned a conn */
1028 tx->tx_lnetmsg = lntmsg;
1030 if (payload_iov != NULL) {
1033 tx->tx_iov = tx->tx_frags.virt.iov;
1035 lnet_extract_iov(payload_niov, &tx->tx_iov[1],
1036 payload_niov, payload_iov,
1037 payload_offset, payload_nob);
1040 tx->tx_iov = &tx->tx_frags.paged.iov;
1041 tx->tx_kiov = tx->tx_frags.paged.kiov;
1042 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
1043 payload_niov, payload_kiov,
1044 payload_offset, payload_nob);
1046 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
1047 tx->tx_zc_capable = 1;
1050 tx->tx_msg.ksm_csum = 0;
1051 tx->tx_msg.ksm_type = KSOCK_MSG_LNET;
1052 tx->tx_msg.ksm_zc_cookies[0] = 0;
1053 tx->tx_msg.ksm_zc_cookies[1] = 0;
1055 /* The first fragment will be set later in pro_pack */
1056 rc = ksocknal_launch_packet(ni, tx, target);
1058 cfs_memory_pressure_restore(mpflag);
1063 lntmsg->msg_health_status = tx->tx_hstatus;
1064 ksocknal_free_tx(tx);
1069 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1071 struct task_struct *task = kthread_run(fn, arg, name);
1074 return PTR_ERR(task);
1076 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1077 ksocknal_data.ksnd_nthreads++;
1078 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1083 ksocknal_thread_fini (void)
1085 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1086 ksocknal_data.ksnd_nthreads--;
1087 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1091 ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip)
1093 static char ksocknal_slop_buffer[4096];
1098 LASSERT(conn->ksnc_proto != NULL);
1100 if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1101 /* Remind the socket to ack eagerly... */
1102 ksocknal_lib_eager_ack(conn);
1105 if (nob_to_skip == 0) { /* right at next packet boundary now */
1106 conn->ksnc_rx_started = 0;
1107 smp_mb(); /* racing with timeout thread */
1109 switch (conn->ksnc_proto->pro_version) {
1110 case KSOCK_PROTO_V2:
1111 case KSOCK_PROTO_V3:
1112 conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1113 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1114 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1116 conn->ksnc_rx_nob_wanted = offsetof(struct ksock_msg, ksm_u);
1117 conn->ksnc_rx_nob_left = offsetof(struct ksock_msg, ksm_u);
1118 conn->ksnc_rx_iov[0].iov_len = offsetof(struct ksock_msg, ksm_u);
1121 case KSOCK_PROTO_V1:
1122 /* Receiving bare struct lnet_hdr */
1123 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1124 conn->ksnc_rx_nob_wanted = sizeof(struct lnet_hdr);
1125 conn->ksnc_rx_nob_left = sizeof(struct lnet_hdr);
1127 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1128 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1129 conn->ksnc_rx_iov[0].iov_len = sizeof(struct lnet_hdr);
1135 conn->ksnc_rx_niov = 1;
1137 conn->ksnc_rx_kiov = NULL;
1138 conn->ksnc_rx_nkiov = 0;
1139 conn->ksnc_rx_csum = ~0;
1143 /* Set up to skip as much as possible now. If there's more left
1144 * (ran out of iov entries) we'll get called again */
1146 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1147 conn->ksnc_rx_nob_left = nob_to_skip;
1148 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1153 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1155 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1156 conn->ksnc_rx_iov[niov].iov_len = nob;
1161 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1162 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct kvec));
1164 conn->ksnc_rx_niov = niov;
1165 conn->ksnc_rx_kiov = NULL;
1166 conn->ksnc_rx_nkiov = 0;
1167 conn->ksnc_rx_nob_wanted = skipped;
1172 ksocknal_process_receive(struct ksock_conn *conn,
1173 struct page **rx_scratch_pgs,
1174 struct kvec *scratch_iov)
1176 struct lnet_hdr *lhdr;
1177 struct lnet_process_id *id;
1180 LASSERT (atomic_read(&conn->ksnc_conn_refcount) > 0);
1182 /* NB: sched lock NOT held */
1183 /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1184 LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1185 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1186 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1187 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1189 if (conn->ksnc_rx_nob_wanted != 0) {
1190 rc = ksocknal_receive(conn, rx_scratch_pgs,
1194 struct lnet_process_id ksnp_id;
1196 ksnp_id = conn->ksnc_peer->ksnp_id;
1198 LASSERT(rc != -EAGAIN);
1200 CDEBUG(D_NET, "[%p] EOF from %s "
1201 "ip %pI4h:%d\n", conn,
1202 libcfs_id2str(ksnp_id),
1205 else if (!conn->ksnc_closing)
1206 CERROR("[%p] Error %d on read from %s "
1207 "ip %pI4h:%d\n", conn, rc,
1208 libcfs_id2str(ksnp_id),
1212 /* it's not an error if conn is being closed */
1213 ksocknal_close_conn_and_siblings (conn,
1214 (conn->ksnc_closing) ? 0 : rc);
1215 return (rc == 0 ? -ESHUTDOWN : rc);
1218 if (conn->ksnc_rx_nob_wanted != 0) {
1223 switch (conn->ksnc_rx_state) {
1224 case SOCKNAL_RX_KSM_HEADER:
1225 if (conn->ksnc_flip) {
1226 __swab32s(&conn->ksnc_msg.ksm_type);
1227 __swab32s(&conn->ksnc_msg.ksm_csum);
1228 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1229 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1232 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1233 conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1234 CERROR("%s: Unknown message type: %x\n",
1235 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1236 conn->ksnc_msg.ksm_type);
1237 ksocknal_new_packet(conn, 0);
1238 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1242 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1243 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1244 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1245 /* NOOP Checksum error */
1246 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1247 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1248 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1249 ksocknal_new_packet(conn, 0);
1250 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1254 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1257 LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1259 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1260 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1262 rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1263 conn->ksnc_msg.ksm_zc_cookies[1]);
1266 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1267 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1268 cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1269 ksocknal_new_packet(conn, 0);
1270 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1275 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1276 ksocknal_new_packet (conn, 0);
1277 return 0; /* NOOP is done and just return */
1280 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1281 conn->ksnc_rx_nob_wanted = sizeof(struct ksock_lnet_msg);
1282 conn->ksnc_rx_nob_left = sizeof(struct ksock_lnet_msg);
1284 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1285 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1286 conn->ksnc_rx_iov[0].iov_len = sizeof(struct ksock_lnet_msg);
1288 conn->ksnc_rx_niov = 1;
1289 conn->ksnc_rx_kiov = NULL;
1290 conn->ksnc_rx_nkiov = 0;
1292 goto again; /* read lnet header now */
1294 case SOCKNAL_RX_LNET_HEADER:
1295 /* unpack message header */
1296 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1298 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1299 /* Userspace peer_ni */
1300 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1301 id = &conn->ksnc_peer->ksnp_id;
1303 /* Substitute process ID assigned at connection time */
1304 lhdr->src_pid = cpu_to_le32(id->pid);
1305 lhdr->src_nid = cpu_to_le64(id->nid);
1308 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1309 ksocknal_conn_addref(conn); /* ++ref while parsing */
1311 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1312 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1313 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1315 /* I just received garbage: give up on this conn */
1316 ksocknal_new_packet(conn, 0);
1317 ksocknal_close_conn_and_siblings (conn, rc);
1318 ksocknal_conn_decref(conn);
1322 /* I'm racing with ksocknal_recv() */
1323 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1324 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1326 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1329 /* ksocknal_recv() got called */
1332 case SOCKNAL_RX_LNET_PAYLOAD:
1333 /* payload all received */
1336 if (conn->ksnc_rx_nob_left == 0 && /* not truncating */
1337 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1338 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1339 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1340 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1341 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1345 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1346 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1348 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1349 id = &conn->ksnc_peer->ksnp_id;
1351 rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1352 conn->ksnc_msg.ksm_zc_cookies[0],
1353 *ksocknal_tunables.ksnd_nonblk_zcack ||
1354 le64_to_cpu(lhdr->src_nid) != id->nid);
1357 if (rc && conn->ksnc_lnet_msg)
1358 conn->ksnc_lnet_msg->msg_health_status =
1359 LNET_MSG_STATUS_REMOTE_ERROR;
1360 lnet_finalize(conn->ksnc_lnet_msg, rc);
1363 ksocknal_new_packet(conn, 0);
1364 ksocknal_close_conn_and_siblings (conn, rc);
1369 case SOCKNAL_RX_SLOP:
1370 /* starting new packet? */
1371 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1372 return 0; /* come back later */
1373 goto again; /* try to finish reading slop now */
1381 return (-EINVAL); /* keep gcc happy */
1385 ksocknal_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
1386 int delayed, unsigned int niov, struct kvec *iov,
1387 lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen,
1390 struct ksock_conn *conn = private;
1391 struct ksock_sched *sched = conn->ksnc_scheduler;
1393 LASSERT (mlen <= rlen);
1394 LASSERT (niov <= LNET_MAX_IOV);
1396 conn->ksnc_lnet_msg = msg;
1397 conn->ksnc_rx_nob_wanted = mlen;
1398 conn->ksnc_rx_nob_left = rlen;
1400 if (mlen == 0 || iov != NULL) {
1401 conn->ksnc_rx_nkiov = 0;
1402 conn->ksnc_rx_kiov = NULL;
1403 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1404 conn->ksnc_rx_niov =
1405 lnet_extract_iov(LNET_MAX_IOV, conn->ksnc_rx_iov,
1406 niov, iov, offset, mlen);
1408 conn->ksnc_rx_niov = 0;
1409 conn->ksnc_rx_iov = NULL;
1410 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1411 conn->ksnc_rx_nkiov =
1412 lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1413 niov, kiov, offset, mlen);
1417 lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1418 lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1420 LASSERT (conn->ksnc_rx_scheduled);
1422 spin_lock_bh(&sched->kss_lock);
1424 switch (conn->ksnc_rx_state) {
1425 case SOCKNAL_RX_PARSE_WAIT:
1426 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1427 wake_up(&sched->kss_waitq);
1428 LASSERT(conn->ksnc_rx_ready);
1431 case SOCKNAL_RX_PARSE:
1432 /* scheduler hasn't noticed I'm parsing yet */
1436 conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1438 spin_unlock_bh(&sched->kss_lock);
1439 ksocknal_conn_decref(conn);
1444 ksocknal_sched_cansleep(struct ksock_sched *sched)
1448 spin_lock_bh(&sched->kss_lock);
1450 rc = (!ksocknal_data.ksnd_shuttingdown &&
1451 list_empty(&sched->kss_rx_conns) &&
1452 list_empty(&sched->kss_tx_conns));
1454 spin_unlock_bh(&sched->kss_lock);
1458 int ksocknal_scheduler(void *arg)
1460 struct ksock_sched *sched;
1461 struct ksock_conn *conn;
1462 struct ksock_tx *tx;
1465 long id = (long)arg;
1466 struct page **rx_scratch_pgs;
1467 struct kvec *scratch_iov;
1469 sched = ksocknal_data.ksnd_schedulers[KSOCK_THREAD_CPT(id)];
1471 LIBCFS_CPT_ALLOC(rx_scratch_pgs, lnet_cpt_table(), sched->kss_cpt,
1472 sizeof(*rx_scratch_pgs) * LNET_MAX_IOV);
1473 if (!rx_scratch_pgs) {
1474 CERROR("Unable to allocate scratch pages\n");
1478 LIBCFS_CPT_ALLOC(scratch_iov, lnet_cpt_table(), sched->kss_cpt,
1479 sizeof(*scratch_iov) * LNET_MAX_IOV);
1481 CERROR("Unable to allocate scratch iov\n");
1485 rc = cfs_cpt_bind(lnet_cpt_table(), sched->kss_cpt);
1487 CWARN("Can't set CPU partition affinity to %d: %d\n",
1488 sched->kss_cpt, rc);
1491 spin_lock_bh(&sched->kss_lock);
1493 while (!ksocknal_data.ksnd_shuttingdown) {
1494 int did_something = 0;
1496 /* Ensure I progress everything semi-fairly */
1498 if (!list_empty(&sched->kss_rx_conns)) {
1499 conn = list_entry(sched->kss_rx_conns.next,
1500 struct ksock_conn, ksnc_rx_list);
1501 list_del(&conn->ksnc_rx_list);
1503 LASSERT(conn->ksnc_rx_scheduled);
1504 LASSERT(conn->ksnc_rx_ready);
1506 /* clear rx_ready in case receive isn't complete.
1507 * Do it BEFORE we call process_recv, since
1508 * data_ready can set it any time after we release
1510 conn->ksnc_rx_ready = 0;
1511 spin_unlock_bh(&sched->kss_lock);
1513 rc = ksocknal_process_receive(conn, rx_scratch_pgs,
1516 spin_lock_bh(&sched->kss_lock);
1518 /* I'm the only one that can clear this flag */
1519 LASSERT(conn->ksnc_rx_scheduled);
1521 /* Did process_receive get everything it wanted? */
1523 conn->ksnc_rx_ready = 1;
1525 if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1526 /* Conn blocked waiting for ksocknal_recv()
1527 * I change its state (under lock) to signal
1528 * it can be rescheduled */
1529 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1530 } else if (conn->ksnc_rx_ready) {
1531 /* reschedule for rx */
1532 list_add_tail(&conn->ksnc_rx_list,
1533 &sched->kss_rx_conns);
1535 conn->ksnc_rx_scheduled = 0;
1537 ksocknal_conn_decref(conn);
1543 if (!list_empty(&sched->kss_tx_conns)) {
1546 list_splice_init(&sched->kss_zombie_noop_txs, &zlist);
1548 conn = list_entry(sched->kss_tx_conns.next,
1549 struct ksock_conn, ksnc_tx_list);
1550 list_del(&conn->ksnc_tx_list);
1552 LASSERT(conn->ksnc_tx_scheduled);
1553 LASSERT(conn->ksnc_tx_ready);
1554 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1556 tx = list_entry(conn->ksnc_tx_queue.next,
1557 struct ksock_tx, tx_list);
1559 if (conn->ksnc_tx_carrier == tx)
1560 ksocknal_next_tx_carrier(conn);
1562 /* dequeue now so empty list => more to send */
1563 list_del(&tx->tx_list);
1565 /* Clear tx_ready in case send isn't complete. Do
1566 * it BEFORE we call process_transmit, since
1567 * write_space can set it any time after we release
1569 conn->ksnc_tx_ready = 0;
1570 spin_unlock_bh(&sched->kss_lock);
1572 if (!list_empty(&zlist)) {
1573 /* free zombie noop txs, it's fast because
1574 * noop txs are just put in freelist */
1575 ksocknal_txlist_done(NULL, &zlist, 0);
1578 rc = ksocknal_process_transmit(conn, tx, scratch_iov);
1580 if (rc == -ENOMEM || rc == -EAGAIN) {
1581 /* Incomplete send: replace tx on HEAD of tx_queue */
1582 spin_lock_bh(&sched->kss_lock);
1583 list_add(&tx->tx_list,
1584 &conn->ksnc_tx_queue);
1586 /* Complete send; tx -ref */
1587 ksocknal_tx_decref(tx);
1589 spin_lock_bh(&sched->kss_lock);
1590 /* assume space for more */
1591 conn->ksnc_tx_ready = 1;
1594 if (rc == -ENOMEM) {
1595 /* Do nothing; after a short timeout, this
1596 * conn will be reposted on kss_tx_conns. */
1597 } else if (conn->ksnc_tx_ready &&
1598 !list_empty(&conn->ksnc_tx_queue)) {
1599 /* reschedule for tx */
1600 list_add_tail(&conn->ksnc_tx_list,
1601 &sched->kss_tx_conns);
1603 conn->ksnc_tx_scheduled = 0;
1605 ksocknal_conn_decref(conn);
1610 if (!did_something || /* nothing to do */
1611 ++nloops == SOCKNAL_RESCHED) { /* hogging CPU? */
1612 spin_unlock_bh(&sched->kss_lock);
1616 if (!did_something) { /* wait for something to do */
1617 rc = wait_event_interruptible_exclusive(
1619 !ksocknal_sched_cansleep(sched));
1625 spin_lock_bh(&sched->kss_lock);
1629 spin_unlock_bh(&sched->kss_lock);
1630 LIBCFS_FREE(rx_scratch_pgs, sizeof(*rx_scratch_pgs) *
1632 LIBCFS_FREE(scratch_iov, sizeof(*scratch_iov) *
1634 ksocknal_thread_fini();
1639 * Add connection to kss_rx_conns of scheduler
1640 * and wakeup the scheduler.
1642 void ksocknal_read_callback(struct ksock_conn *conn)
1644 struct ksock_sched *sched;
1647 sched = conn->ksnc_scheduler;
1649 spin_lock_bh(&sched->kss_lock);
1651 conn->ksnc_rx_ready = 1;
1653 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1654 list_add_tail(&conn->ksnc_rx_list,
1655 &sched->kss_rx_conns);
1656 conn->ksnc_rx_scheduled = 1;
1657 /* extra ref for scheduler */
1658 ksocknal_conn_addref(conn);
1660 wake_up (&sched->kss_waitq);
1662 spin_unlock_bh(&sched->kss_lock);
1668 * Add connection to kss_tx_conns of scheduler
1669 * and wakeup the scheduler.
1671 void ksocknal_write_callback(struct ksock_conn *conn)
1673 struct ksock_sched *sched;
1676 sched = conn->ksnc_scheduler;
1678 spin_lock_bh(&sched->kss_lock);
1680 conn->ksnc_tx_ready = 1;
1682 if (!conn->ksnc_tx_scheduled && /* not being progressed */
1683 !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1684 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1685 conn->ksnc_tx_scheduled = 1;
1686 /* extra ref for scheduler */
1687 ksocknal_conn_addref(conn);
1689 wake_up(&sched->kss_waitq);
1692 spin_unlock_bh(&sched->kss_lock);
1697 static const struct ksock_proto *
1698 ksocknal_parse_proto_version(struct ksock_hello_msg *hello)
1702 if (hello->kshm_magic == LNET_PROTO_MAGIC)
1703 version = hello->kshm_version;
1704 else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1705 version = __swab32(hello->kshm_version);
1708 #if SOCKNAL_VERSION_DEBUG
1709 if (*ksocknal_tunables.ksnd_protocol == 1)
1712 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1713 version == KSOCK_PROTO_V3)
1716 if (version == KSOCK_PROTO_V2)
1717 return &ksocknal_protocol_v2x;
1719 if (version == KSOCK_PROTO_V3)
1720 return &ksocknal_protocol_v3x;
1725 if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1726 struct lnet_magicversion *hmv;
1728 BUILD_BUG_ON(sizeof(struct lnet_magicversion) !=
1729 offsetof(struct ksock_hello_msg, kshm_src_nid));
1731 hmv = (struct lnet_magicversion *)hello;
1733 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1734 hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1735 return &ksocknal_protocol_v1x;
1742 ksocknal_send_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1743 lnet_nid_t peer_nid, struct ksock_hello_msg *hello)
1745 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1746 struct ksock_net *net = (struct ksock_net *)ni->ni_data;
1748 LASSERT(hello->kshm_nips <= LNET_INTERFACES_NUM);
1750 /* rely on caller to hold a ref on socket so it wouldn't disappear */
1751 LASSERT(conn->ksnc_proto != NULL);
1753 hello->kshm_src_nid = ni->ni_nid;
1754 hello->kshm_dst_nid = peer_nid;
1755 hello->kshm_src_pid = the_lnet.ln_pid;
1757 hello->kshm_src_incarnation = net->ksnn_incarnation;
1758 hello->kshm_ctype = conn->ksnc_type;
1760 return conn->ksnc_proto->pro_send_hello(conn, hello);
1764 ksocknal_invert_type(int type)
1768 case SOCKLND_CONN_ANY:
1769 case SOCKLND_CONN_CONTROL:
1771 case SOCKLND_CONN_BULK_IN:
1772 return SOCKLND_CONN_BULK_OUT;
1773 case SOCKLND_CONN_BULK_OUT:
1774 return SOCKLND_CONN_BULK_IN;
1776 return (SOCKLND_CONN_NONE);
1781 ksocknal_recv_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1782 struct ksock_hello_msg *hello,
1783 struct lnet_process_id *peerid,
1786 /* Return < 0 fatal error
1788 * EALREADY lost connection race
1789 * EPROTO protocol version mismatch
1791 struct socket *sock = conn->ksnc_sock;
1792 int active = (conn->ksnc_proto != NULL);
1796 const struct ksock_proto *proto;
1797 struct lnet_process_id recv_id;
1799 /* socket type set on active connections - not set on passive */
1800 LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1802 timeout = active ? lnet_get_lnd_timeout() :
1803 lnet_acceptor_timeout();
1805 rc = lnet_sock_read(sock, &hello->kshm_magic,
1806 sizeof(hello->kshm_magic), timeout);
1808 CERROR("Error %d reading HELLO from %pI4h\n",
1809 rc, &conn->ksnc_ipaddr);
1814 if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1815 hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1816 hello->kshm_magic != le32_to_cpu (LNET_PROTO_TCP_MAGIC)) {
1817 /* Unexpected magic! */
1818 CERROR ("Bad magic(1) %#08x (%#08x expected) from "
1819 "%pI4h\n", __cpu_to_le32 (hello->kshm_magic),
1820 LNET_PROTO_TCP_MAGIC, &conn->ksnc_ipaddr);
1824 rc = lnet_sock_read(sock, &hello->kshm_version,
1825 sizeof(hello->kshm_version), timeout);
1827 CERROR("Error %d reading HELLO from %pI4h\n",
1828 rc, &conn->ksnc_ipaddr);
1833 proto = ksocknal_parse_proto_version(hello);
1834 if (proto == NULL) {
1836 /* unknown protocol from peer_ni, tell peer_ni my protocol */
1837 conn->ksnc_proto = &ksocknal_protocol_v3x;
1838 #if SOCKNAL_VERSION_DEBUG
1839 if (*ksocknal_tunables.ksnd_protocol == 2)
1840 conn->ksnc_proto = &ksocknal_protocol_v2x;
1841 else if (*ksocknal_tunables.ksnd_protocol == 1)
1842 conn->ksnc_proto = &ksocknal_protocol_v1x;
1844 hello->kshm_nips = 0;
1845 ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1848 CERROR("Unknown protocol version (%d.x expected) from %pI4h\n",
1849 conn->ksnc_proto->pro_version, &conn->ksnc_ipaddr);
1854 proto_match = (conn->ksnc_proto == proto);
1855 conn->ksnc_proto = proto;
1857 /* receive the rest of hello message anyway */
1858 rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1860 CERROR("Error %d reading or checking hello from from %pI4h\n",
1861 rc, &conn->ksnc_ipaddr);
1866 *incarnation = hello->kshm_src_incarnation;
1868 if (hello->kshm_src_nid == LNET_NID_ANY) {
1869 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pI4h\n",
1870 &conn->ksnc_ipaddr);
1875 conn->ksnc_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1876 /* Userspace NAL assigns peer_ni process ID from socket */
1877 recv_id.pid = conn->ksnc_port | LNET_PID_USERFLAG;
1878 recv_id.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), conn->ksnc_ipaddr);
1880 recv_id.nid = hello->kshm_src_nid;
1881 recv_id.pid = hello->kshm_src_pid;
1887 /* peer_ni determines type */
1888 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1889 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1890 CERROR("Unexpected type %d from %s ip %pI4h\n",
1891 hello->kshm_ctype, libcfs_id2str(*peerid),
1892 &conn->ksnc_ipaddr);
1898 if (peerid->pid != recv_id.pid ||
1899 peerid->nid != recv_id.nid) {
1900 LCONSOLE_ERROR_MSG(0x130, "Connected successfully to %s on host"
1901 " %pI4h, but they claimed they were "
1902 "%s; please check your Lustre "
1904 libcfs_id2str(*peerid),
1906 libcfs_id2str(recv_id));
1910 if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1911 /* Possible protocol mismatch or I lost the connection race */
1912 return proto_match ? EALREADY : EPROTO;
1915 if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1916 CERROR("Mismatched types: me %d, %s ip %pI4h %d\n",
1917 conn->ksnc_type, libcfs_id2str(*peerid),
1926 ksocknal_connect(struct ksock_route *route)
1929 struct ksock_peer_ni *peer_ni = route->ksnr_peer;
1932 struct socket *sock;
1934 int retry_later = 0;
1937 deadline = ktime_get_seconds() + lnet_get_lnd_timeout();
1939 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1941 LASSERT (route->ksnr_scheduled);
1942 LASSERT (!route->ksnr_connecting);
1944 route->ksnr_connecting = 1;
1947 wanted = ksocknal_route_mask() & ~route->ksnr_connected;
1949 /* stop connecting if peer_ni/route got closed under me, or
1950 * route got connected while queued */
1951 if (peer_ni->ksnp_closing || route->ksnr_deleted ||
1957 /* reschedule if peer_ni is connecting to me */
1958 if (peer_ni->ksnp_accepting > 0) {
1960 "peer_ni %s(%d) already connecting to me, retry later.\n",
1961 libcfs_nid2str(peer_ni->ksnp_id.nid), peer_ni->ksnp_accepting);
1965 if (retry_later) /* needs reschedule */
1968 if ((wanted & (1 << SOCKLND_CONN_ANY)) != 0) {
1969 type = SOCKLND_CONN_ANY;
1970 } else if ((wanted & (1 << SOCKLND_CONN_CONTROL)) != 0) {
1971 type = SOCKLND_CONN_CONTROL;
1972 } else if ((wanted & (1 << SOCKLND_CONN_BULK_IN)) != 0) {
1973 type = SOCKLND_CONN_BULK_IN;
1975 LASSERT ((wanted & (1 << SOCKLND_CONN_BULK_OUT)) != 0);
1976 type = SOCKLND_CONN_BULK_OUT;
1979 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1981 if (ktime_get_seconds() >= deadline) {
1983 lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
1989 rc = lnet_connect(&sock, peer_ni->ksnp_id.nid,
1990 route->ksnr_myipaddr,
1991 route->ksnr_ipaddr, route->ksnr_port,
1992 peer_ni->ksnp_ni->ni_net_ns);
1996 rc = ksocknal_create_conn(peer_ni->ksnp_ni, route, sock, type);
1998 lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
2004 /* A +ve RC means I have to retry because I lost the connection
2005 * race or I have to renegotiate protocol version */
2006 retry_later = (rc != 0);
2008 CDEBUG(D_NET, "peer_ni %s: conn race, retry later.\n",
2009 libcfs_nid2str(peer_ni->ksnp_id.nid));
2011 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2014 route->ksnr_scheduled = 0;
2015 route->ksnr_connecting = 0;
2018 /* re-queue for attention; this frees me up to handle
2019 * the peer_ni's incoming connection request */
2021 if (rc == EALREADY ||
2022 (rc == 0 && peer_ni->ksnp_accepting > 0)) {
2023 /* We want to introduce a delay before next
2024 * attempt to connect if we lost conn race,
2025 * but the race is resolved quickly usually,
2026 * so min_reconnectms should be good heuristic */
2027 route->ksnr_retry_interval = *ksocknal_tunables.ksnd_min_reconnectms / 1000;
2028 route->ksnr_timeout = ktime_get_seconds() +
2029 route->ksnr_retry_interval;
2032 ksocknal_launch_connection_locked(route);
2035 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2039 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2041 route->ksnr_scheduled = 0;
2042 route->ksnr_connecting = 0;
2044 /* This is a retry rather than a new connection */
2045 route->ksnr_retry_interval *= 2;
2046 route->ksnr_retry_interval =
2047 max_t(time64_t, route->ksnr_retry_interval,
2048 *ksocknal_tunables.ksnd_min_reconnectms / 1000);
2049 route->ksnr_retry_interval =
2050 min_t(time64_t, route->ksnr_retry_interval,
2051 *ksocknal_tunables.ksnd_max_reconnectms / 1000);
2053 LASSERT(route->ksnr_retry_interval);
2054 route->ksnr_timeout = ktime_get_seconds() + route->ksnr_retry_interval;
2056 if (!list_empty(&peer_ni->ksnp_tx_queue) &&
2057 peer_ni->ksnp_accepting == 0 &&
2058 ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
2059 struct ksock_conn *conn;
2061 /* ksnp_tx_queue is queued on a conn on successful
2062 * connection for V1.x and V2.x */
2063 if (!list_empty(&peer_ni->ksnp_conns)) {
2064 conn = list_entry(peer_ni->ksnp_conns.next,
2065 struct ksock_conn, ksnc_list);
2066 LASSERT (conn->ksnc_proto == &ksocknal_protocol_v3x);
2069 /* take all the blocked packets while I've got the lock and
2070 * complete below... */
2071 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
2074 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2076 ksocknal_peer_failed(peer_ni);
2077 ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc);
2082 * check whether we need to create more connds.
2083 * It will try to create new thread if it's necessary, @timeout can
2084 * be updated if failed to create, so caller wouldn't keep try while
2085 * running out of resource.
2088 ksocknal_connd_check_start(time64_t sec, long *timeout)
2092 int total = ksocknal_data.ksnd_connd_starting +
2093 ksocknal_data.ksnd_connd_running;
2095 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2096 /* still in initializing */
2100 if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2101 total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2102 /* can't create more connd, or still have enough
2103 * threads to handle more connecting */
2107 if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2108 /* no pending connecting request */
2112 if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2113 /* may run out of resource, retry later */
2114 *timeout = cfs_time_seconds(1);
2118 if (ksocknal_data.ksnd_connd_starting > 0) {
2119 /* serialize starting to avoid flood */
2123 ksocknal_data.ksnd_connd_starting_stamp = sec;
2124 ksocknal_data.ksnd_connd_starting++;
2125 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2127 /* NB: total is the next id */
2128 snprintf(name, sizeof(name), "socknal_cd%02d", total);
2129 rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2131 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2136 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2137 ksocknal_data.ksnd_connd_starting--;
2138 ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2144 * check whether current thread can exit, it will return 1 if there are too
2145 * many threads and no creating in past 120 seconds.
2146 * Also, this function may update @timeout to make caller come back
2147 * again to recheck these conditions.
2150 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2154 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2155 /* still in initializing */
2159 if (ksocknal_data.ksnd_connd_starting > 0) {
2160 /* in progress of starting new thread */
2164 if (ksocknal_data.ksnd_connd_running <=
2165 *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2169 /* created thread in past 120 seconds? */
2170 val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2171 SOCKNAL_CONND_TIMEOUT - sec);
2173 *timeout = (val > 0) ? cfs_time_seconds(val) :
2174 cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2178 /* no creating in past 120 seconds */
2180 return ksocknal_data.ksnd_connd_running >
2181 ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2184 /* Go through connd_routes queue looking for a route that we can process
2185 * right now, @timeout_p can be updated if we need to come back later */
2186 static struct ksock_route *
2187 ksocknal_connd_get_route_locked(signed long *timeout_p)
2189 time64_t now = ktime_get_seconds();
2190 struct ksock_route *route;
2192 /* connd_routes can contain both pending and ordinary routes */
2193 list_for_each_entry(route, &ksocknal_data.ksnd_connd_routes,
2196 if (route->ksnr_retry_interval == 0 ||
2197 now >= route->ksnr_timeout)
2200 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2201 *timeout_p > cfs_time_seconds(route->ksnr_timeout - now))
2202 *timeout_p = cfs_time_seconds(route->ksnr_timeout - now);
2209 ksocknal_connd(void *arg)
2211 spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2212 struct ksock_connreq *cr;
2213 wait_queue_entry_t wait;
2217 init_waitqueue_entry(&wait, current);
2219 spin_lock_bh(connd_lock);
2221 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2222 ksocknal_data.ksnd_connd_starting--;
2223 ksocknal_data.ksnd_connd_running++;
2225 while (!ksocknal_data.ksnd_shuttingdown) {
2226 struct ksock_route *route = NULL;
2227 time64_t sec = ktime_get_real_seconds();
2228 long timeout = MAX_SCHEDULE_TIMEOUT;
2229 int dropped_lock = 0;
2231 if (ksocknal_connd_check_stop(sec, &timeout)) {
2232 /* wakeup another one to check stop */
2233 wake_up(&ksocknal_data.ksnd_connd_waitq);
2237 if (ksocknal_connd_check_start(sec, &timeout)) {
2238 /* created new thread */
2242 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2243 /* Connection accepted by the listener */
2244 cr = list_entry(ksocknal_data.ksnd_connd_connreqs.next,
2245 struct ksock_connreq, ksncr_list);
2247 list_del(&cr->ksncr_list);
2248 spin_unlock_bh(connd_lock);
2251 ksocknal_create_conn(cr->ksncr_ni, NULL,
2252 cr->ksncr_sock, SOCKLND_CONN_NONE);
2253 lnet_ni_decref(cr->ksncr_ni);
2254 LIBCFS_FREE(cr, sizeof(*cr));
2256 spin_lock_bh(connd_lock);
2259 /* Only handle an outgoing connection request if there
2260 * is a thread left to handle incoming connections and
2261 * create new connd */
2262 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2263 ksocknal_data.ksnd_connd_running) {
2264 route = ksocknal_connd_get_route_locked(&timeout);
2266 if (route != NULL) {
2267 list_del(&route->ksnr_connd_list);
2268 ksocknal_data.ksnd_connd_connecting++;
2269 spin_unlock_bh(connd_lock);
2272 if (ksocknal_connect(route)) {
2273 /* consecutive retry */
2274 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2275 CWARN("massive consecutive "
2276 "re-connecting to %pI4h\n",
2277 &route->ksnr_ipaddr);
2284 ksocknal_route_decref(route);
2286 spin_lock_bh(connd_lock);
2287 ksocknal_data.ksnd_connd_connecting--;
2291 if (++nloops < SOCKNAL_RESCHED)
2293 spin_unlock_bh(connd_lock);
2296 spin_lock_bh(connd_lock);
2300 /* Nothing to do for 'timeout' */
2301 set_current_state(TASK_INTERRUPTIBLE);
2302 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq, &wait);
2303 spin_unlock_bh(connd_lock);
2306 schedule_timeout(timeout);
2308 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2309 spin_lock_bh(connd_lock);
2311 ksocknal_data.ksnd_connd_running--;
2312 spin_unlock_bh(connd_lock);
2314 ksocknal_thread_fini();
2318 static struct ksock_conn *
2319 ksocknal_find_timed_out_conn(struct ksock_peer_ni *peer_ni)
2321 /* We're called with a shared lock on ksnd_global_lock */
2322 struct ksock_conn *conn;
2323 struct list_head *ctmp;
2324 struct ksock_tx *tx;
2326 list_for_each(ctmp, &peer_ni->ksnp_conns) {
2329 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
2331 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2332 LASSERT (!conn->ksnc_closing);
2334 error = conn->ksnc_sock->sk->sk_err;
2336 ksocknal_conn_addref(conn);
2340 CNETERR("A connection with %s "
2341 "(%pI4h:%d) was reset; "
2342 "it may have rebooted.\n",
2343 libcfs_id2str(peer_ni->ksnp_id),
2348 CNETERR("A connection with %s "
2349 "(%pI4h:%d) timed out; the "
2350 "network or node may be down.\n",
2351 libcfs_id2str(peer_ni->ksnp_id),
2356 CNETERR("An unexpected network error %d "
2358 "(%pI4h:%d\n", error,
2359 libcfs_id2str(peer_ni->ksnp_id),
2368 if (conn->ksnc_rx_started &&
2369 ktime_get_seconds() >= conn->ksnc_rx_deadline) {
2370 /* Timed out incomplete incoming message */
2371 ksocknal_conn_addref(conn);
2372 CNETERR("Timeout receiving from %s (%pI4h:%d), "
2373 "state %d wanted %d left %d\n",
2374 libcfs_id2str(peer_ni->ksnp_id),
2377 conn->ksnc_rx_state,
2378 conn->ksnc_rx_nob_wanted,
2379 conn->ksnc_rx_nob_left);
2383 if ((!list_empty(&conn->ksnc_tx_queue) ||
2384 conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2385 ktime_get_seconds() >= conn->ksnc_tx_deadline) {
2386 /* Timed out messages queued for sending or
2387 * buffered in the socket's send buffer */
2388 ksocknal_conn_addref(conn);
2389 list_for_each_entry(tx, &conn->ksnc_tx_queue,
2392 LNET_MSG_STATUS_LOCAL_TIMEOUT;
2393 CNETERR("Timeout sending data to %s (%pI4h:%d) "
2394 "the network or that node may be down.\n",
2395 libcfs_id2str(peer_ni->ksnp_id),
2396 &conn->ksnc_ipaddr, conn->ksnc_port);
2405 ksocknal_flush_stale_txs(struct ksock_peer_ni *peer_ni)
2407 struct ksock_tx *tx;
2408 LIST_HEAD(stale_txs);
2410 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2412 while (!list_empty(&peer_ni->ksnp_tx_queue)) {
2413 tx = list_entry(peer_ni->ksnp_tx_queue.next,
2414 struct ksock_tx, tx_list);
2416 if (ktime_get_seconds() < tx->tx_deadline)
2419 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2421 list_move_tail(&tx->tx_list, &stale_txs);
2424 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2426 ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT);
2430 ksocknal_send_keepalive_locked(struct ksock_peer_ni *peer_ni)
2431 __must_hold(&ksocknal_data.ksnd_global_lock)
2433 struct ksock_sched *sched;
2434 struct ksock_conn *conn;
2435 struct ksock_tx *tx;
2437 /* last_alive will be updated by create_conn */
2438 if (list_empty(&peer_ni->ksnp_conns))
2441 if (peer_ni->ksnp_proto != &ksocknal_protocol_v3x)
2444 if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2445 ktime_get_seconds() < peer_ni->ksnp_last_alive +
2446 *ksocknal_tunables.ksnd_keepalive)
2449 if (ktime_get_seconds() < peer_ni->ksnp_send_keepalive)
2452 /* retry 10 secs later, so we wouldn't put pressure
2453 * on this peer_ni if we failed to send keepalive this time */
2454 peer_ni->ksnp_send_keepalive = ktime_get_seconds() + 10;
2456 conn = ksocknal_find_conn_locked(peer_ni, NULL, 1);
2458 sched = conn->ksnc_scheduler;
2460 spin_lock_bh(&sched->kss_lock);
2461 if (!list_empty(&conn->ksnc_tx_queue)) {
2462 spin_unlock_bh(&sched->kss_lock);
2463 /* there is an queued ACK, don't need keepalive */
2467 spin_unlock_bh(&sched->kss_lock);
2470 read_unlock(&ksocknal_data.ksnd_global_lock);
2472 /* cookie = 1 is reserved for keepalive PING */
2473 tx = ksocknal_alloc_tx_noop(1, 1);
2475 read_lock(&ksocknal_data.ksnd_global_lock);
2479 if (ksocknal_launch_packet(peer_ni->ksnp_ni, tx, peer_ni->ksnp_id) == 0) {
2480 read_lock(&ksocknal_data.ksnd_global_lock);
2484 ksocknal_free_tx(tx);
2485 read_lock(&ksocknal_data.ksnd_global_lock);
2492 ksocknal_check_peer_timeouts(int idx)
2494 struct hlist_head *peers = &ksocknal_data.ksnd_peers[idx];
2495 struct ksock_peer_ni *peer_ni;
2496 struct ksock_conn *conn;
2497 struct ksock_tx *tx;
2500 /* NB. We expect to have a look at all the peers and not find any
2501 * connections to time out, so we just use a shared lock while we
2504 read_lock(&ksocknal_data.ksnd_global_lock);
2506 hlist_for_each_entry(peer_ni, peers, ksnp_list) {
2507 struct ksock_tx *tx_stale;
2508 time64_t deadline = 0;
2512 if (ksocknal_send_keepalive_locked(peer_ni) != 0) {
2513 read_unlock(&ksocknal_data.ksnd_global_lock);
2517 conn = ksocknal_find_timed_out_conn(peer_ni);
2520 read_unlock(&ksocknal_data.ksnd_global_lock);
2522 ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2524 /* NB we won't find this one again, but we can't
2525 * just proceed with the next peer_ni, since we dropped
2526 * ksnd_global_lock and it might be dead already!
2528 ksocknal_conn_decref(conn);
2532 /* we can't process stale txs right here because we're
2533 * holding only shared lock
2535 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
2536 struct ksock_tx *tx;
2538 tx = list_entry(peer_ni->ksnp_tx_queue.next,
2539 struct ksock_tx, tx_list);
2540 if (ktime_get_seconds() >= tx->tx_deadline) {
2541 ksocknal_peer_addref(peer_ni);
2542 read_unlock(&ksocknal_data.ksnd_global_lock);
2544 ksocknal_flush_stale_txs(peer_ni);
2546 ksocknal_peer_decref(peer_ni);
2551 if (list_empty(&peer_ni->ksnp_zc_req_list))
2555 spin_lock(&peer_ni->ksnp_lock);
2556 list_for_each_entry(tx, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
2557 if (ktime_get_seconds() < tx->tx_deadline)
2559 /* ignore the TX if connection is being closed */
2560 if (tx->tx_conn->ksnc_closing)
2563 if (tx_stale == NULL)
2567 if (tx_stale == NULL) {
2568 spin_unlock(&peer_ni->ksnp_lock);
2572 deadline = tx_stale->tx_deadline;
2573 resid = tx_stale->tx_resid;
2574 conn = tx_stale->tx_conn;
2575 ksocknal_conn_addref(conn);
2577 spin_unlock(&peer_ni->ksnp_lock);
2578 read_unlock(&ksocknal_data.ksnd_global_lock);
2580 CERROR("Total %d stale ZC_REQs for peer_ni %s detected; the "
2581 "oldest(%p) timed out %lld secs ago, "
2582 "resid: %d, wmem: %d\n",
2583 n, libcfs_nid2str(peer_ni->ksnp_id.nid), tx_stale,
2584 ktime_get_seconds() - deadline,
2585 resid, conn->ksnc_sock->sk->sk_wmem_queued);
2587 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2588 ksocknal_conn_decref(conn);
2592 read_unlock(&ksocknal_data.ksnd_global_lock);
2595 int ksocknal_reaper(void *arg)
2597 wait_queue_entry_t wait;
2598 struct ksock_conn *conn;
2599 struct ksock_sched *sched;
2600 LIST_HEAD(enomem_conns);
2605 time64_t deadline = ktime_get_seconds();
2607 init_waitqueue_entry(&wait, current);
2609 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2611 while (!ksocknal_data.ksnd_shuttingdown) {
2612 if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) {
2613 conn = list_entry(ksocknal_data.ksnd_deathrow_conns.next,
2614 struct ksock_conn, ksnc_list);
2615 list_del(&conn->ksnc_list);
2617 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2619 ksocknal_terminate_conn(conn);
2620 ksocknal_conn_decref(conn);
2622 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2626 if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) {
2627 conn = list_entry(ksocknal_data.ksnd_zombie_conns.next,
2628 struct ksock_conn, ksnc_list);
2629 list_del(&conn->ksnc_list);
2631 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2633 ksocknal_destroy_conn(conn);
2635 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2639 list_splice_init(&ksocknal_data.ksnd_enomem_conns,
2642 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2644 /* reschedule all the connections that stalled with ENOMEM... */
2646 while (!list_empty(&enomem_conns)) {
2647 conn = list_entry(enomem_conns.next,
2648 struct ksock_conn, ksnc_tx_list);
2649 list_del(&conn->ksnc_tx_list);
2651 sched = conn->ksnc_scheduler;
2653 spin_lock_bh(&sched->kss_lock);
2655 LASSERT(conn->ksnc_tx_scheduled);
2656 conn->ksnc_tx_ready = 1;
2657 list_add_tail(&conn->ksnc_tx_list,
2658 &sched->kss_tx_conns);
2659 wake_up(&sched->kss_waitq);
2661 spin_unlock_bh(&sched->kss_lock);
2665 /* careful with the jiffy wrap... */
2666 while ((timeout = deadline - ktime_get_seconds()) <= 0) {
2669 int chunk = HASH_SIZE(ksocknal_data.ksnd_peers);
2670 unsigned int lnd_timeout;
2672 /* Time to check for timeouts on a few more peers: I
2673 * do checks every 'p' seconds on a proportion of the
2674 * peer_ni table and I need to check every connection
2675 * 'n' times within a timeout interval, to ensure I
2676 * detect a timeout on any connection within (n+1)/n
2677 * times the timeout interval.
2680 lnd_timeout = lnet_get_lnd_timeout();
2681 if (lnd_timeout > n * p)
2682 chunk = (chunk * n * p) / lnd_timeout;
2686 for (i = 0; i < chunk; i++) {
2687 ksocknal_check_peer_timeouts(peer_index);
2688 peer_index = (peer_index + 1) %
2689 HASH_SIZE(ksocknal_data.ksnd_peers);
2695 if (nenomem_conns != 0) {
2696 /* Reduce my timeout if I rescheduled ENOMEM conns.
2697 * This also prevents me getting woken immediately
2698 * if any go back on my enomem list. */
2699 timeout = SOCKNAL_ENOMEM_RETRY;
2701 ksocknal_data.ksnd_reaper_waketime = ktime_get_seconds() +
2704 set_current_state(TASK_INTERRUPTIBLE);
2705 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2707 if (!ksocknal_data.ksnd_shuttingdown &&
2708 list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2709 list_empty(&ksocknal_data.ksnd_zombie_conns))
2710 schedule_timeout(cfs_time_seconds(timeout));
2712 set_current_state(TASK_RUNNING);
2713 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2715 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2718 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2720 ksocknal_thread_fini();