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.
27 #include <libcfs/linux/linux-mem.h>
29 #include <linux/sunrpc/addr.h>
32 ksocknal_alloc_tx(int type, int size)
34 struct ksock_tx *tx = NULL;
36 if (type == KSOCK_MSG_NOOP) {
37 LASSERT(size == KSOCK_NOOP_TX_SIZE);
39 /* searching for a noop tx in free list */
40 spin_lock(&ksocknal_data.ksnd_tx_lock);
42 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
43 tx = list_entry(ksocknal_data.ksnd_idle_noop_txs.next,
44 struct ksock_tx, tx_list);
45 LASSERT(tx->tx_desc_size == size);
46 list_del(&tx->tx_list);
49 spin_unlock(&ksocknal_data.ksnd_tx_lock);
53 LIBCFS_ALLOC(tx, size);
58 refcount_set(&tx->tx_refcount, 1);
59 tx->tx_zc_aborted = 0;
60 tx->tx_zc_capable = 0;
61 tx->tx_zc_checked = 0;
62 tx->tx_hstatus = LNET_MSG_STATUS_OK;
63 tx->tx_desc_size = size;
65 atomic_inc(&ksocknal_data.ksnd_nactive_txs);
71 ksocknal_alloc_tx_noop(__u64 cookie, int nonblk)
75 tx = ksocknal_alloc_tx(KSOCK_MSG_NOOP, KSOCK_NOOP_TX_SIZE);
77 CERROR("Can't allocate noop tx desc\n");
82 tx->tx_lnetmsg = NULL;
86 tx->tx_nonblk = nonblk;
88 tx->tx_msg.ksm_csum = 0;
89 tx->tx_msg.ksm_type = KSOCK_MSG_NOOP;
90 tx->tx_msg.ksm_zc_cookies[0] = 0;
91 tx->tx_msg.ksm_zc_cookies[1] = cookie;
98 ksocknal_free_tx(struct ksock_tx *tx)
100 atomic_dec(&ksocknal_data.ksnd_nactive_txs);
102 if (tx->tx_lnetmsg == NULL && tx->tx_desc_size == KSOCK_NOOP_TX_SIZE) {
104 spin_lock(&ksocknal_data.ksnd_tx_lock);
106 list_add(&tx->tx_list, &ksocknal_data.ksnd_idle_noop_txs);
108 spin_unlock(&ksocknal_data.ksnd_tx_lock);
110 LIBCFS_FREE(tx, tx->tx_desc_size);
115 ksocknal_send_hdr(struct ksock_conn *conn, struct ksock_tx *tx,
116 struct kvec *scratch_iov)
118 struct kvec *iov = &tx->tx_hdr;
122 LASSERT(tx->tx_niov > 0);
124 /* Never touch tx->tx_hdr inside ksocknal_lib_send_hdr() */
125 rc = ksocknal_lib_send_hdr(conn, tx, scratch_iov);
127 if (rc <= 0) /* sent nothing? */
131 LASSERT(nob <= tx->tx_resid);
135 LASSERT(tx->tx_niov == 1);
137 if (nob < (int) iov->iov_len) {
138 iov->iov_base += nob;
143 LASSERT(nob == iov->iov_len);
150 ksocknal_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx,
151 struct kvec *scratch_iov)
153 struct bio_vec *kiov = tx->tx_kiov;
157 LASSERT(tx->tx_niov == 0);
158 LASSERT(tx->tx_nkiov > 0);
160 /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
161 rc = ksocknal_lib_send_kiov(conn, tx, scratch_iov);
163 if (rc <= 0) /* sent nothing? */
167 LASSERT(nob <= tx->tx_resid);
172 LASSERT(tx->tx_nkiov > 0);
174 if (nob < (int)kiov->bv_len) {
175 kiov->bv_offset += nob;
180 nob -= (int)kiov->bv_len;
181 tx->tx_kiov = ++kiov;
189 ksocknal_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
190 struct kvec *scratch_iov)
195 if (ksocknal_data.ksnd_stall_tx != 0)
196 schedule_timeout_uninterruptible(
197 cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
199 LASSERT(tx->tx_resid != 0);
201 rc = ksocknal_connsock_addref(conn);
203 LASSERT(conn->ksnc_closing);
208 if (ksocknal_data.ksnd_enomem_tx > 0) {
210 ksocknal_data.ksnd_enomem_tx--;
212 } else if (tx->tx_niov != 0) {
213 rc = ksocknal_send_hdr(conn, tx, scratch_iov);
215 rc = ksocknal_send_kiov(conn, tx, scratch_iov);
218 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
219 if (rc > 0) /* sent something? */
220 conn->ksnc_tx_bufnob += rc; /* account it */
222 if (bufnob < conn->ksnc_tx_bufnob) {
223 /* allocated send buffer bytes < computed; infer
224 * something got ACKed */
225 conn->ksnc_tx_deadline = ktime_get_seconds() +
227 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
228 conn->ksnc_tx_bufnob = bufnob;
232 if (rc <= 0) { /* Didn't write anything? */
233 /* some stacks return 0 instead of -EAGAIN */
237 /* Check if EAGAIN is due to memory pressure */
238 if (rc == -EAGAIN && ksocknal_lib_memory_pressure(conn))
244 /* socket's wmem_queued now includes 'rc' bytes */
245 atomic_sub (rc, &conn->ksnc_tx_nob);
248 } while (tx->tx_resid != 0);
250 ksocknal_connsock_decref(conn);
255 ksocknal_recv_iov(struct ksock_conn *conn, struct kvec *scratchiov)
257 struct kvec *iov = conn->ksnc_rx_iov;
261 LASSERT(conn->ksnc_rx_niov > 0);
263 /* Never touch conn->ksnc_rx_iov or change connection
264 * status inside ksocknal_lib_recv_iov */
265 rc = ksocknal_lib_recv_iov(conn, scratchiov);
270 /* received something... */
273 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
274 conn->ksnc_rx_deadline = ktime_get_seconds() +
276 smp_mb(); /* order with setting rx_started */
277 conn->ksnc_rx_started = 1;
279 conn->ksnc_rx_nob_wanted -= nob;
280 conn->ksnc_rx_nob_left -= nob;
283 LASSERT(conn->ksnc_rx_niov > 0);
285 if (nob < (int)iov->iov_len) {
287 iov->iov_base += nob;
292 conn->ksnc_rx_iov = ++iov;
293 conn->ksnc_rx_niov--;
300 ksocknal_recv_kiov(struct ksock_conn *conn, struct page **rx_scratch_pgs,
301 struct kvec *scratch_iov)
303 struct bio_vec *kiov = conn->ksnc_rx_kiov;
306 LASSERT(conn->ksnc_rx_nkiov > 0);
308 /* Never touch conn->ksnc_rx_kiov or change connection
309 * status inside ksocknal_lib_recv_iov */
310 rc = ksocknal_lib_recv_kiov(conn, rx_scratch_pgs, scratch_iov);
315 /* received something... */
318 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
319 conn->ksnc_rx_deadline = ktime_get_seconds() +
321 smp_mb(); /* order with setting rx_started */
322 conn->ksnc_rx_started = 1;
324 conn->ksnc_rx_nob_wanted -= nob;
325 conn->ksnc_rx_nob_left -= nob;
328 LASSERT(conn->ksnc_rx_nkiov > 0);
330 if (nob < (int) kiov->bv_len) {
331 kiov->bv_offset += nob;
337 conn->ksnc_rx_kiov = ++kiov;
338 conn->ksnc_rx_nkiov--;
345 ksocknal_receive(struct ksock_conn *conn, struct page **rx_scratch_pgs,
346 struct kvec *scratch_iov)
348 /* Return 1 on success, 0 on EOF, < 0 on error.
349 * Caller checks ksnc_rx_nob_wanted to determine
350 * progress/completion. */
354 if (ksocknal_data.ksnd_stall_rx != 0)
355 schedule_timeout_uninterruptible(
356 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 != 0)
366 rc = ksocknal_recv_iov(conn, scratch_iov);
368 rc = ksocknal_recv_kiov(conn, rx_scratch_pgs,
372 /* error/EOF or partial receive */
375 } else if (rc == 0 && conn->ksnc_rx_started) {
376 /* EOF in the middle of a message */
382 /* Completed a fragment */
384 if (conn->ksnc_rx_nob_wanted == 0) {
390 ksocknal_connsock_decref(conn);
395 ksocknal_tx_done(struct lnet_ni *ni, struct ksock_tx *tx, int rc)
397 struct lnet_msg *lnetmsg = tx->tx_lnetmsg;
398 enum lnet_msg_hstatus hstatus = tx->tx_hstatus;
401 LASSERT(ni != NULL || tx->tx_conn != NULL);
403 if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted)) {
405 if (hstatus == LNET_MSG_STATUS_OK)
406 hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
409 if (tx->tx_conn != NULL)
410 ksocknal_conn_decref(tx->tx_conn);
412 ksocknal_free_tx(tx);
413 if (lnetmsg != NULL) { /* KSOCK_MSG_NOOP go without lnetmsg */
414 lnetmsg->msg_health_status = hstatus;
415 lnet_finalize(lnetmsg, rc);
422 ksocknal_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int error)
426 while (!list_empty(txlist)) {
427 tx = list_entry(txlist->next, struct ksock_tx, tx_list);
429 if (error && tx->tx_lnetmsg != NULL) {
430 CNETERR("Deleting packet type %d len %d %s->%s\n",
431 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type),
432 le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length),
433 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
434 libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
436 CNETERR("Deleting noop packet\n");
439 list_del(&tx->tx_list);
441 if (tx->tx_hstatus == LNET_MSG_STATUS_OK) {
442 if (error == -ETIMEDOUT)
444 LNET_MSG_STATUS_LOCAL_TIMEOUT;
445 else if (error == -ENETDOWN ||
446 error == -EHOSTUNREACH ||
447 error == -ENETUNREACH ||
448 error == -ECONNREFUSED ||
449 error == -ECONNRESET)
450 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
452 * for all other errors we don't want to
456 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
459 LASSERT(refcount_read(&tx->tx_refcount) == 1);
460 ksocknal_tx_done(ni, tx, error);
465 ksocknal_check_zc_req(struct ksock_tx *tx)
467 struct ksock_conn *conn = tx->tx_conn;
468 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
470 /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
471 * to ksnp_zc_req_list if some fragment of this message should be sent
472 * zero-copy. Our peer_ni will send an ACK containing this cookie when
473 * she has received this message to tell us we can signal completion.
474 * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
475 * ksnp_zc_req_list. */
476 LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
477 LASSERT (tx->tx_zc_capable);
479 tx->tx_zc_checked = 1;
481 if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
482 !conn->ksnc_zc_capable)
485 /* assign cookie and queue tx to pending list, it will be released when
486 * a matching ack is received. See ksocknal_handle_zcack() */
488 ksocknal_tx_addref(tx);
490 spin_lock(&peer_ni->ksnp_lock);
492 /* ZC_REQ is going to be pinned to the peer_ni */
493 tx->tx_deadline = ktime_get_seconds() +
496 LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
498 tx->tx_msg.ksm_zc_cookies[0] = peer_ni->ksnp_zc_next_cookie++;
500 if (peer_ni->ksnp_zc_next_cookie == 0)
501 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
503 list_add_tail(&tx->tx_zc_list, &peer_ni->ksnp_zc_req_list);
505 spin_unlock(&peer_ni->ksnp_lock);
509 ksocknal_uncheck_zc_req(struct ksock_tx *tx)
511 struct ksock_peer_ni *peer_ni = tx->tx_conn->ksnc_peer;
513 LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
514 LASSERT(tx->tx_zc_capable);
516 tx->tx_zc_checked = 0;
518 spin_lock(&peer_ni->ksnp_lock);
520 if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
521 /* Not waiting for an ACK */
522 spin_unlock(&peer_ni->ksnp_lock);
526 tx->tx_msg.ksm_zc_cookies[0] = 0;
527 list_del(&tx->tx_zc_list);
529 spin_unlock(&peer_ni->ksnp_lock);
531 ksocknal_tx_decref(tx);
535 ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
536 struct kvec *scratch_iov)
539 bool error_sim = false;
541 if (lnet_send_error_simulation(tx->tx_lnetmsg, &tx->tx_hstatus)) {
547 if (tx->tx_zc_capable && !tx->tx_zc_checked)
548 ksocknal_check_zc_req(tx);
550 rc = ksocknal_transmit(conn, tx, scratch_iov);
552 CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
554 if (tx->tx_resid == 0) {
555 /* Sent everything OK */
567 counter++; /* exponential backoff warnings */
568 if ((counter & (-counter)) == counter)
569 CWARN("%u ENOMEM tx %p (%lld allocated)\n",
570 counter, conn, libcfs_kmem_read());
572 /* Queue on ksnd_enomem_conns for retry after a timeout */
573 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
575 /* enomem list takes over scheduler's ref... */
576 LASSERT(conn->ksnc_tx_scheduled);
577 list_add_tail(&conn->ksnc_tx_list,
578 &ksocknal_data.ksnd_enomem_conns);
579 if (ktime_get_seconds() + SOCKNAL_ENOMEM_RETRY <
580 ksocknal_data.ksnd_reaper_waketime)
581 wake_up(&ksocknal_data.ksnd_reaper_waitq);
583 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
586 * set the health status of the message which determines
587 * whether we should retry the transmit
589 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
600 * set the health status of the message which determines
601 * whether we should retry the transmit
603 if (rc == -ETIMEDOUT)
604 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_TIMEOUT;
606 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
609 if (!conn->ksnc_closing) {
612 LCONSOLE_WARN("Host %pIS reset our connection while we were sending data; it may have rebooted.\n",
613 &conn->ksnc_peeraddr);
616 LCONSOLE_WARN("There was an unexpected network error while writing to %pIS: %d.\n",
617 &conn->ksnc_peeraddr, rc);
620 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pISp\n",
621 conn, rc, libcfs_id2str(conn->ksnc_peer->ksnp_id),
622 &conn->ksnc_peeraddr);
625 if (tx->tx_zc_checked)
626 ksocknal_uncheck_zc_req(tx);
628 /* it's not an error if conn is being closed */
629 ksocknal_close_conn_and_siblings(conn,
630 (conn->ksnc_closing) ? 0 : rc);
636 ksocknal_launch_connection_locked(struct ksock_conn_cb *conn_cb)
638 /* called holding write lock on ksnd_global_lock */
640 LASSERT(!conn_cb->ksnr_scheduled);
641 LASSERT(!conn_cb->ksnr_connecting);
642 LASSERT((ksocknal_conn_cb_mask() & ~conn_cb->ksnr_connected) != 0);
644 /* scheduling conn for connd */
645 conn_cb->ksnr_scheduled = 1;
647 /* extra ref for connd */
648 ksocknal_conn_cb_addref(conn_cb);
650 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
652 list_add_tail(&conn_cb->ksnr_connd_list,
653 &ksocknal_data.ksnd_connd_routes);
654 wake_up(&ksocknal_data.ksnd_connd_waitq);
656 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
660 ksocknal_launch_all_connections_locked(struct ksock_peer_ni *peer_ni)
662 struct ksock_conn_cb *conn_cb;
664 /* called holding write lock on ksnd_global_lock */
666 /* launch any/all connections that need it */
667 conn_cb = ksocknal_find_connectable_conn_cb_locked(peer_ni);
671 ksocknal_launch_connection_locked(conn_cb);
676 ksocknal_find_conn_locked(struct ksock_peer_ni *peer_ni, struct ksock_tx *tx, int nonblk)
678 struct list_head *tmp;
679 struct ksock_conn *conn;
680 struct ksock_conn *typed = NULL;
681 struct ksock_conn *fallback = NULL;
685 list_for_each(tmp, &peer_ni->ksnp_conns) {
686 struct ksock_conn *c = list_entry(tmp, struct ksock_conn,
688 int nob = atomic_read(&c->ksnc_tx_nob) +
689 c->ksnc_sock->sk->sk_wmem_queued;
692 LASSERT (!c->ksnc_closing);
693 LASSERT (c->ksnc_proto != NULL &&
694 c->ksnc_proto->pro_match_tx != NULL);
696 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
701 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
704 case SOCKNAL_MATCH_YES: /* typed connection */
705 if (typed == NULL || tnob > nob ||
706 (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
707 typed->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
713 case SOCKNAL_MATCH_MAY: /* fallback connection */
714 if (fallback == NULL || fnob > nob ||
715 (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
716 fallback->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
724 /* prefer the typed selection */
725 conn = (typed != NULL) ? typed : fallback;
728 conn->ksnc_tx_last_post = ktime_get_seconds();
734 ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *tx)
736 conn->ksnc_proto->pro_pack(tx);
738 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
739 ksocknal_conn_addref(conn); /* +1 ref for tx */
744 ksocknal_queue_tx_locked(struct ksock_tx *tx, struct ksock_conn *conn)
746 struct ksock_sched *sched = conn->ksnc_scheduler;
747 struct ksock_msg *msg = &tx->tx_msg;
748 struct ksock_tx *ztx = NULL;
751 /* called holding global lock (read or irq-write) and caller may
752 * not have dropped this lock between finding conn and calling me,
753 * so we don't need the {get,put}connsock dance to deref
755 LASSERT(!conn->ksnc_closing);
757 CDEBUG(D_NET, "Sending to %s ip %pISp\n",
758 libcfs_id2str(conn->ksnc_peer->ksnp_id),
759 &conn->ksnc_peeraddr);
761 ksocknal_tx_prep(conn, tx);
763 /* Ensure the frags we've been given EXACTLY match the number of
764 * bytes we want to send. Many TCP/IP stacks disregard any total
765 * size parameters passed to them and just look at the frags.
767 * We always expect at least 1 mapped fragment containing the
768 * complete ksocknal message header.
770 LASSERT(lnet_iov_nob(tx->tx_niov, &tx->tx_hdr) +
771 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
772 (unsigned int)tx->tx_nob);
773 LASSERT(tx->tx_niov >= 1);
774 LASSERT(tx->tx_resid == tx->tx_nob);
776 CDEBUG (D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
777 tx, (tx->tx_lnetmsg != NULL) ? tx->tx_lnetmsg->msg_hdr.type:
779 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
781 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
782 spin_lock_bh(&sched->kss_lock);
784 if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
785 /* First packet starts the timeout */
786 conn->ksnc_tx_deadline = ktime_get_seconds() +
788 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
789 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
790 conn->ksnc_tx_bufnob = 0;
791 smp_mb(); /* order with adding to tx_queue */
794 if (msg->ksm_type == KSOCK_MSG_NOOP) {
795 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
796 * on a normal packet so I don't need to send it */
797 LASSERT (msg->ksm_zc_cookies[1] != 0);
798 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
800 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
801 ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
804 /* It's a normal packet - can it piggback a noop zc-ack that
805 * has been queued already? */
806 LASSERT (msg->ksm_zc_cookies[1] == 0);
807 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
809 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
810 /* ztx will be released later */
814 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
815 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
818 if (conn->ksnc_tx_ready && /* able to send */
819 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
820 /* +1 ref for scheduler */
821 ksocknal_conn_addref(conn);
822 list_add_tail(&conn->ksnc_tx_list,
823 &sched->kss_tx_conns);
824 conn->ksnc_tx_scheduled = 1;
825 wake_up(&sched->kss_waitq);
828 spin_unlock_bh(&sched->kss_lock);
832 struct ksock_conn_cb *
833 ksocknal_find_connectable_conn_cb_locked(struct ksock_peer_ni *peer_ni)
835 time64_t now = ktime_get_seconds();
836 struct ksock_conn_cb *conn_cb;
838 conn_cb = peer_ni->ksnp_conn_cb;
842 LASSERT(!conn_cb->ksnr_connecting || conn_cb->ksnr_scheduled);
844 if (conn_cb->ksnr_scheduled) /* connections being established */
847 /* all conn types connected ? */
848 if ((ksocknal_conn_cb_mask() & ~conn_cb->ksnr_connected) == 0)
851 if (!(conn_cb->ksnr_retry_interval == 0 || /* first attempt */
852 now >= conn_cb->ksnr_timeout)) {
854 "Too soon to retry route %pIS (cnted %d, interval %lld, %lld secs later)\n",
856 conn_cb->ksnr_connected,
857 conn_cb->ksnr_retry_interval,
858 conn_cb->ksnr_timeout - now);
865 struct ksock_conn_cb *
866 ksocknal_find_connecting_conn_cb_locked(struct ksock_peer_ni *peer_ni)
868 struct ksock_conn_cb *conn_cb;
870 conn_cb = peer_ni->ksnp_conn_cb;
874 LASSERT(!conn_cb->ksnr_connecting || conn_cb->ksnr_scheduled);
876 return conn_cb->ksnr_scheduled ? conn_cb : NULL;
880 ksocknal_launch_packet(struct lnet_ni *ni, struct ksock_tx *tx,
881 struct lnet_process_id id)
883 struct ksock_peer_ni *peer_ni;
884 struct ksock_conn *conn;
885 struct sockaddr_in sa;
890 LASSERT(tx->tx_conn == NULL);
892 g_lock = &ksocknal_data.ksnd_global_lock;
894 for (retry = 0;; retry = 1) {
896 peer_ni = ksocknal_find_peer_locked(ni, id);
897 if (peer_ni != NULL) {
898 if (ksocknal_find_connectable_conn_cb_locked(peer_ni) == NULL) {
899 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
901 /* I've got nothing that need to be
902 * connecting and I do have an actual
905 ksocknal_queue_tx_locked (tx, conn);
912 /* I'll need a write lock... */
915 write_lock_bh(g_lock);
917 peer_ni = ksocknal_find_peer_locked(ni, id);
921 write_unlock_bh(g_lock);
923 if ((id.pid & LNET_PID_USERFLAG) != 0) {
924 CERROR("Refusing to create a connection to "
925 "userspace process %s\n", libcfs_id2str(id));
926 return -EHOSTUNREACH;
930 CERROR("Can't find peer_ni %s\n", libcfs_id2str(id));
931 return -EHOSTUNREACH;
934 memset(&sa, 0, sizeof(sa));
935 sa.sin_family = AF_INET;
936 sa.sin_addr.s_addr = htonl(LNET_NIDADDR(id.nid));
937 sa.sin_port = htons(lnet_acceptor_port());
938 rc = ksocknal_add_peer(ni, id, (struct sockaddr *)&sa);
940 CERROR("Can't add peer_ni %s: %d\n",
941 libcfs_id2str(id), rc);
946 ksocknal_launch_all_connections_locked(peer_ni);
948 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
950 /* Connection exists; queue message on it */
951 ksocknal_queue_tx_locked (tx, conn);
952 write_unlock_bh(g_lock);
956 if (peer_ni->ksnp_accepting > 0 ||
957 ksocknal_find_connecting_conn_cb_locked(peer_ni) != NULL) {
958 /* the message is going to be pinned to the peer_ni */
959 tx->tx_deadline = ktime_get_seconds() +
962 /* Queue the message until a connection is established */
963 list_add_tail(&tx->tx_list, &peer_ni->ksnp_tx_queue);
964 write_unlock_bh(g_lock);
968 write_unlock_bh(g_lock);
970 /* NB Routes may be ignored if connections to them failed recently */
971 CNETERR("No usable routes to %s\n", libcfs_id2str(id));
972 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
973 return (-EHOSTUNREACH);
977 ksocknal_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
979 /* '1' for consistency with code that checks !mpflag to restore */
980 unsigned int mpflag = 1;
981 int type = lntmsg->msg_type;
982 struct lnet_process_id target = lntmsg->msg_target;
983 unsigned int payload_niov = lntmsg->msg_niov;
984 struct bio_vec *payload_kiov = lntmsg->msg_kiov;
985 unsigned int payload_offset = lntmsg->msg_offset;
986 unsigned int payload_nob = lntmsg->msg_len;
991 /* NB 'private' is different depending on what we're sending.
992 * Just ignore it... */
994 CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
995 payload_nob, payload_niov, libcfs_id2str(target));
997 LASSERT (payload_nob == 0 || payload_niov > 0);
998 LASSERT (payload_niov <= LNET_MAX_IOV);
999 LASSERT (!in_interrupt ());
1001 desc_size = offsetof(struct ksock_tx,
1002 tx_payload[payload_niov]);
1004 if (lntmsg->msg_vmflush)
1005 mpflag = memalloc_noreclaim_save();
1007 tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
1009 CERROR("Can't allocate tx desc type %d size %d\n",
1011 if (lntmsg->msg_vmflush)
1012 memalloc_noreclaim_restore(mpflag);
1016 tx->tx_conn = NULL; /* set when assigned a conn */
1017 tx->tx_lnetmsg = lntmsg;
1020 tx->tx_kiov = tx->tx_payload;
1021 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
1022 payload_niov, payload_kiov,
1023 payload_offset, payload_nob);
1025 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
1026 tx->tx_zc_capable = 1;
1028 tx->tx_msg.ksm_csum = 0;
1029 tx->tx_msg.ksm_type = KSOCK_MSG_LNET;
1030 tx->tx_msg.ksm_zc_cookies[0] = 0;
1031 tx->tx_msg.ksm_zc_cookies[1] = 0;
1033 /* The first fragment will be set later in pro_pack */
1034 rc = ksocknal_launch_packet(ni, tx, target);
1036 * We can't test lntsmg->msg_vmflush again as lntmsg may
1040 memalloc_noreclaim_restore(mpflag);
1045 lntmsg->msg_health_status = tx->tx_hstatus;
1046 ksocknal_free_tx(tx);
1051 ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name)
1053 struct task_struct *task = kthread_run(fn, arg, "%s", name);
1056 return PTR_ERR(task);
1058 atomic_inc(&ksocknal_data.ksnd_nthreads);
1063 ksocknal_thread_fini (void)
1065 if (atomic_dec_and_test(&ksocknal_data.ksnd_nthreads))
1066 wake_up_var(&ksocknal_data.ksnd_nthreads);
1070 ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip)
1072 static char ksocknal_slop_buffer[4096];
1077 LASSERT(conn->ksnc_proto != NULL);
1079 if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1080 /* Remind the socket to ack eagerly... */
1081 ksocknal_lib_eager_ack(conn);
1084 if (nob_to_skip == 0) { /* right at next packet boundary now */
1085 conn->ksnc_rx_started = 0;
1086 smp_mb(); /* racing with timeout thread */
1088 switch (conn->ksnc_proto->pro_version) {
1089 case KSOCK_PROTO_V2:
1090 case KSOCK_PROTO_V3:
1091 conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1092 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1093 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1095 conn->ksnc_rx_nob_wanted = offsetof(struct ksock_msg, ksm_u);
1096 conn->ksnc_rx_nob_left = offsetof(struct ksock_msg, ksm_u);
1097 conn->ksnc_rx_iov[0].iov_len = offsetof(struct ksock_msg, ksm_u);
1100 case KSOCK_PROTO_V1:
1101 /* Receiving bare struct lnet_hdr */
1102 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1103 conn->ksnc_rx_nob_wanted = sizeof(struct lnet_hdr);
1104 conn->ksnc_rx_nob_left = sizeof(struct lnet_hdr);
1106 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1107 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1108 conn->ksnc_rx_iov[0].iov_len = sizeof(struct lnet_hdr);
1114 conn->ksnc_rx_niov = 1;
1116 conn->ksnc_rx_kiov = NULL;
1117 conn->ksnc_rx_nkiov = 0;
1118 conn->ksnc_rx_csum = ~0;
1122 /* Set up to skip as much as possible now. If there's more left
1123 * (ran out of iov entries) we'll get called again */
1125 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1126 conn->ksnc_rx_nob_left = nob_to_skip;
1127 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1132 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1134 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1135 conn->ksnc_rx_iov[niov].iov_len = nob;
1140 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1141 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct kvec));
1143 conn->ksnc_rx_niov = niov;
1144 conn->ksnc_rx_kiov = NULL;
1145 conn->ksnc_rx_nkiov = 0;
1146 conn->ksnc_rx_nob_wanted = skipped;
1151 ksocknal_process_receive(struct ksock_conn *conn,
1152 struct page **rx_scratch_pgs,
1153 struct kvec *scratch_iov)
1155 struct lnet_hdr *lhdr;
1156 struct lnet_process_id *id;
1159 LASSERT(refcount_read(&conn->ksnc_conn_refcount) > 0);
1161 /* NB: sched lock NOT held */
1162 /* SOCKNAL_RX_LNET_HEADER is here for backward compatibility */
1163 LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_KSM_HEADER ||
1164 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD ||
1165 conn->ksnc_rx_state == SOCKNAL_RX_LNET_HEADER ||
1166 conn->ksnc_rx_state == SOCKNAL_RX_SLOP);
1168 if (conn->ksnc_rx_nob_wanted != 0) {
1169 rc = ksocknal_receive(conn, rx_scratch_pgs,
1173 struct lnet_process_id ksnp_id;
1175 ksnp_id = conn->ksnc_peer->ksnp_id;
1177 LASSERT(rc != -EAGAIN);
1179 CDEBUG(D_NET, "[%p] EOF from %s ip %pISp\n",
1180 conn, libcfs_id2str(ksnp_id),
1181 &conn->ksnc_peeraddr);
1182 else if (!conn->ksnc_closing)
1183 CERROR("[%p] Error %d on read from %s ip %pISp\n",
1184 conn, rc, libcfs_id2str(ksnp_id),
1185 &conn->ksnc_peeraddr);
1187 /* it's not an error if conn is being closed */
1188 ksocknal_close_conn_and_siblings (conn,
1189 (conn->ksnc_closing) ? 0 : rc);
1190 return (rc == 0 ? -ESHUTDOWN : rc);
1193 if (conn->ksnc_rx_nob_wanted != 0) {
1198 switch (conn->ksnc_rx_state) {
1199 case SOCKNAL_RX_KSM_HEADER:
1200 if (conn->ksnc_flip) {
1201 __swab32s(&conn->ksnc_msg.ksm_type);
1202 __swab32s(&conn->ksnc_msg.ksm_csum);
1203 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[0]);
1204 __swab64s(&conn->ksnc_msg.ksm_zc_cookies[1]);
1207 if (conn->ksnc_msg.ksm_type != KSOCK_MSG_NOOP &&
1208 conn->ksnc_msg.ksm_type != KSOCK_MSG_LNET) {
1209 CERROR("%s: Unknown message type: %x\n",
1210 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1211 conn->ksnc_msg.ksm_type);
1212 ksocknal_new_packet(conn, 0);
1213 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1217 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP &&
1218 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1219 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1220 /* NOOP Checksum error */
1221 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1222 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1223 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1224 ksocknal_new_packet(conn, 0);
1225 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1229 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1232 LASSERT (conn->ksnc_proto != &ksocknal_protocol_v1x);
1234 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1235 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1237 rc = conn->ksnc_proto->pro_handle_zcack(conn, cookie,
1238 conn->ksnc_msg.ksm_zc_cookies[1]);
1241 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1242 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1243 cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1244 ksocknal_new_packet(conn, 0);
1245 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1250 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP) {
1251 ksocknal_new_packet (conn, 0);
1252 return 0; /* NOOP is done and just return */
1255 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1256 conn->ksnc_rx_nob_wanted = sizeof(struct ksock_lnet_msg);
1257 conn->ksnc_rx_nob_left = sizeof(struct ksock_lnet_msg);
1259 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1260 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg.ksm_u.lnetmsg;
1261 conn->ksnc_rx_iov[0].iov_len = sizeof(struct ksock_lnet_msg);
1263 conn->ksnc_rx_niov = 1;
1264 conn->ksnc_rx_kiov = NULL;
1265 conn->ksnc_rx_nkiov = 0;
1267 goto again; /* read lnet header now */
1269 case SOCKNAL_RX_LNET_HEADER:
1270 /* unpack message header */
1271 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg);
1273 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1274 /* Userspace peer_ni */
1275 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1276 id = &conn->ksnc_peer->ksnp_id;
1278 /* Substitute process ID assigned at connection time */
1279 lhdr->src_pid = cpu_to_le32(id->pid);
1280 lhdr->src_nid = cpu_to_le64(id->nid);
1283 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1284 ksocknal_conn_addref(conn); /* ++ref while parsing */
1286 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1287 &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr,
1288 conn->ksnc_peer->ksnp_id.nid, conn, 0);
1290 /* I just received garbage: give up on this conn */
1291 ksocknal_new_packet(conn, 0);
1292 ksocknal_close_conn_and_siblings (conn, rc);
1293 ksocknal_conn_decref(conn);
1297 /* I'm racing with ksocknal_recv() */
1298 LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1299 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1301 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1304 /* ksocknal_recv() got called */
1307 case SOCKNAL_RX_LNET_PAYLOAD:
1308 /* payload all received */
1311 if (conn->ksnc_rx_nob_left == 0 && /* not truncating */
1312 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1313 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1314 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1315 libcfs_id2str(conn->ksnc_peer->ksnp_id),
1316 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1320 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1321 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1323 lhdr = &conn->ksnc_msg.ksm_u.lnetmsg.ksnm_hdr;
1324 id = &conn->ksnc_peer->ksnp_id;
1326 rc = conn->ksnc_proto->pro_handle_zcreq(conn,
1327 conn->ksnc_msg.ksm_zc_cookies[0],
1328 *ksocknal_tunables.ksnd_nonblk_zcack ||
1329 le64_to_cpu(lhdr->src_nid) != id->nid);
1332 if (rc && conn->ksnc_lnet_msg)
1333 conn->ksnc_lnet_msg->msg_health_status =
1334 LNET_MSG_STATUS_REMOTE_ERROR;
1335 lnet_finalize(conn->ksnc_lnet_msg, rc);
1338 ksocknal_new_packet(conn, 0);
1339 ksocknal_close_conn_and_siblings (conn, rc);
1344 case SOCKNAL_RX_SLOP:
1345 /* starting new packet? */
1346 if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
1347 return 0; /* come back later */
1348 goto again; /* try to finish reading slop now */
1356 return (-EINVAL); /* keep gcc happy */
1360 ksocknal_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
1361 int delayed, unsigned int niov,
1362 struct bio_vec *kiov, unsigned int offset, unsigned int mlen,
1365 struct ksock_conn *conn = private;
1366 struct ksock_sched *sched = conn->ksnc_scheduler;
1368 LASSERT (mlen <= rlen);
1369 LASSERT (niov <= LNET_MAX_IOV);
1371 conn->ksnc_lnet_msg = msg;
1372 conn->ksnc_rx_nob_wanted = mlen;
1373 conn->ksnc_rx_nob_left = rlen;
1376 conn->ksnc_rx_nkiov = 0;
1377 conn->ksnc_rx_kiov = NULL;
1378 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1379 conn->ksnc_rx_niov = 0;
1381 conn->ksnc_rx_niov = 0;
1382 conn->ksnc_rx_iov = NULL;
1383 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1384 conn->ksnc_rx_nkiov =
1385 lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1386 niov, kiov, offset, mlen);
1390 lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1391 lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1393 LASSERT (conn->ksnc_rx_scheduled);
1395 spin_lock_bh(&sched->kss_lock);
1397 switch (conn->ksnc_rx_state) {
1398 case SOCKNAL_RX_PARSE_WAIT:
1399 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1400 wake_up(&sched->kss_waitq);
1401 LASSERT(conn->ksnc_rx_ready);
1404 case SOCKNAL_RX_PARSE:
1405 /* scheduler hasn't noticed I'm parsing yet */
1409 conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1411 spin_unlock_bh(&sched->kss_lock);
1412 ksocknal_conn_decref(conn);
1417 ksocknal_sched_cansleep(struct ksock_sched *sched)
1421 spin_lock_bh(&sched->kss_lock);
1423 rc = (!ksocknal_data.ksnd_shuttingdown &&
1424 list_empty(&sched->kss_rx_conns) &&
1425 list_empty(&sched->kss_tx_conns));
1427 spin_unlock_bh(&sched->kss_lock);
1431 int ksocknal_scheduler(void *arg)
1433 struct ksock_sched *sched;
1434 struct ksock_conn *conn;
1435 struct ksock_tx *tx;
1437 long id = (long)arg;
1438 struct page **rx_scratch_pgs;
1439 struct kvec *scratch_iov;
1441 sched = ksocknal_data.ksnd_schedulers[KSOCK_THREAD_CPT(id)];
1443 LIBCFS_CPT_ALLOC(rx_scratch_pgs, lnet_cpt_table(), sched->kss_cpt,
1444 sizeof(*rx_scratch_pgs) * LNET_MAX_IOV);
1445 if (!rx_scratch_pgs) {
1446 CERROR("Unable to allocate scratch pages\n");
1450 LIBCFS_CPT_ALLOC(scratch_iov, lnet_cpt_table(), sched->kss_cpt,
1451 sizeof(*scratch_iov) * LNET_MAX_IOV);
1453 CERROR("Unable to allocate scratch iov\n");
1457 rc = cfs_cpt_bind(lnet_cpt_table(), sched->kss_cpt);
1459 CWARN("Can't set CPU partition affinity to %d: %d\n",
1460 sched->kss_cpt, rc);
1463 spin_lock_bh(&sched->kss_lock);
1465 while (!ksocknal_data.ksnd_shuttingdown) {
1466 bool did_something = false;
1468 /* Ensure I progress everything semi-fairly */
1470 if (!list_empty(&sched->kss_rx_conns)) {
1471 conn = list_entry(sched->kss_rx_conns.next,
1472 struct ksock_conn, ksnc_rx_list);
1473 list_del(&conn->ksnc_rx_list);
1475 LASSERT(conn->ksnc_rx_scheduled);
1476 LASSERT(conn->ksnc_rx_ready);
1478 /* clear rx_ready in case receive isn't complete.
1479 * Do it BEFORE we call process_recv, since
1480 * data_ready can set it any time after we release
1482 conn->ksnc_rx_ready = 0;
1483 spin_unlock_bh(&sched->kss_lock);
1485 rc = ksocknal_process_receive(conn, rx_scratch_pgs,
1488 spin_lock_bh(&sched->kss_lock);
1490 /* I'm the only one that can clear this flag */
1491 LASSERT(conn->ksnc_rx_scheduled);
1493 /* Did process_receive get everything it wanted? */
1495 conn->ksnc_rx_ready = 1;
1497 if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1498 /* Conn blocked waiting for ksocknal_recv()
1499 * I change its state (under lock) to signal
1500 * it can be rescheduled */
1501 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1502 } else if (conn->ksnc_rx_ready) {
1503 /* reschedule for rx */
1504 list_add_tail(&conn->ksnc_rx_list,
1505 &sched->kss_rx_conns);
1507 conn->ksnc_rx_scheduled = 0;
1509 ksocknal_conn_decref(conn);
1512 did_something = true;
1515 if (!list_empty(&sched->kss_tx_conns)) {
1518 list_splice_init(&sched->kss_zombie_noop_txs, &zlist);
1520 conn = list_entry(sched->kss_tx_conns.next,
1521 struct ksock_conn, ksnc_tx_list);
1522 list_del(&conn->ksnc_tx_list);
1524 LASSERT(conn->ksnc_tx_scheduled);
1525 LASSERT(conn->ksnc_tx_ready);
1526 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1528 tx = list_entry(conn->ksnc_tx_queue.next,
1529 struct ksock_tx, tx_list);
1531 if (conn->ksnc_tx_carrier == tx)
1532 ksocknal_next_tx_carrier(conn);
1534 /* dequeue now so empty list => more to send */
1535 list_del(&tx->tx_list);
1537 /* Clear tx_ready in case send isn't complete. Do
1538 * it BEFORE we call process_transmit, since
1539 * write_space can set it any time after we release
1541 conn->ksnc_tx_ready = 0;
1542 spin_unlock_bh(&sched->kss_lock);
1544 if (!list_empty(&zlist)) {
1545 /* free zombie noop txs, it's fast because
1546 * noop txs are just put in freelist */
1547 ksocknal_txlist_done(NULL, &zlist, 0);
1550 rc = ksocknal_process_transmit(conn, tx, scratch_iov);
1552 if (rc == -ENOMEM || rc == -EAGAIN) {
1553 /* Incomplete send: replace tx on HEAD of tx_queue */
1554 spin_lock_bh(&sched->kss_lock);
1555 list_add(&tx->tx_list,
1556 &conn->ksnc_tx_queue);
1558 /* Complete send; tx -ref */
1559 ksocknal_tx_decref(tx);
1561 spin_lock_bh(&sched->kss_lock);
1562 /* assume space for more */
1563 conn->ksnc_tx_ready = 1;
1566 if (rc == -ENOMEM) {
1567 /* Do nothing; after a short timeout, this
1568 * conn will be reposted on kss_tx_conns. */
1569 } else if (conn->ksnc_tx_ready &&
1570 !list_empty(&conn->ksnc_tx_queue)) {
1571 /* reschedule for tx */
1572 list_add_tail(&conn->ksnc_tx_list,
1573 &sched->kss_tx_conns);
1575 conn->ksnc_tx_scheduled = 0;
1577 ksocknal_conn_decref(conn);
1580 did_something = true;
1582 if (!did_something || /* nothing to do */
1583 need_resched()) { /* hogging CPU? */
1584 spin_unlock_bh(&sched->kss_lock);
1586 if (!did_something) { /* wait for something to do */
1587 rc = wait_event_interruptible_exclusive(
1589 !ksocknal_sched_cansleep(sched));
1595 spin_lock_bh(&sched->kss_lock);
1599 spin_unlock_bh(&sched->kss_lock);
1600 CFS_FREE_PTR_ARRAY(rx_scratch_pgs, LNET_MAX_IOV);
1601 CFS_FREE_PTR_ARRAY(scratch_iov, LNET_MAX_IOV);
1602 ksocknal_thread_fini();
1607 * Add connection to kss_rx_conns of scheduler
1608 * and wakeup the scheduler.
1610 void ksocknal_read_callback(struct ksock_conn *conn)
1612 struct ksock_sched *sched;
1615 sched = conn->ksnc_scheduler;
1617 spin_lock_bh(&sched->kss_lock);
1619 conn->ksnc_rx_ready = 1;
1621 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1622 list_add_tail(&conn->ksnc_rx_list,
1623 &sched->kss_rx_conns);
1624 conn->ksnc_rx_scheduled = 1;
1625 /* extra ref for scheduler */
1626 ksocknal_conn_addref(conn);
1628 wake_up (&sched->kss_waitq);
1630 spin_unlock_bh(&sched->kss_lock);
1636 * Add connection to kss_tx_conns of scheduler
1637 * and wakeup the scheduler.
1639 void ksocknal_write_callback(struct ksock_conn *conn)
1641 struct ksock_sched *sched;
1644 sched = conn->ksnc_scheduler;
1646 spin_lock_bh(&sched->kss_lock);
1648 conn->ksnc_tx_ready = 1;
1650 if (!conn->ksnc_tx_scheduled && /* not being progressed */
1651 !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1652 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1653 conn->ksnc_tx_scheduled = 1;
1654 /* extra ref for scheduler */
1655 ksocknal_conn_addref(conn);
1657 wake_up(&sched->kss_waitq);
1660 spin_unlock_bh(&sched->kss_lock);
1665 static const struct ksock_proto *
1666 ksocknal_parse_proto_version(struct ksock_hello_msg *hello)
1670 if (hello->kshm_magic == LNET_PROTO_MAGIC)
1671 version = hello->kshm_version;
1672 else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1673 version = __swab32(hello->kshm_version);
1676 #if SOCKNAL_VERSION_DEBUG
1677 if (*ksocknal_tunables.ksnd_protocol == 1)
1680 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1681 version == KSOCK_PROTO_V3)
1684 if (version == KSOCK_PROTO_V2)
1685 return &ksocknal_protocol_v2x;
1687 if (version == KSOCK_PROTO_V3)
1688 return &ksocknal_protocol_v3x;
1693 if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1694 struct lnet_magicversion *hmv;
1696 BUILD_BUG_ON(sizeof(struct lnet_magicversion) !=
1697 offsetof(struct ksock_hello_msg, kshm_src_nid));
1699 hmv = (struct lnet_magicversion *)hello;
1701 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1702 hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1703 return &ksocknal_protocol_v1x;
1710 ksocknal_send_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1711 lnet_nid_t peer_nid, struct ksock_hello_msg *hello)
1713 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1714 struct ksock_net *net = (struct ksock_net *)ni->ni_data;
1716 LASSERT(hello->kshm_nips <= LNET_INTERFACES_NUM);
1718 /* rely on caller to hold a ref on socket so it wouldn't disappear */
1719 LASSERT(conn->ksnc_proto != NULL);
1721 hello->kshm_src_nid = ni->ni_nid;
1722 hello->kshm_dst_nid = peer_nid;
1723 hello->kshm_src_pid = the_lnet.ln_pid;
1725 hello->kshm_src_incarnation = net->ksnn_incarnation;
1726 hello->kshm_ctype = conn->ksnc_type;
1728 return conn->ksnc_proto->pro_send_hello(conn, hello);
1732 ksocknal_invert_type(int type)
1736 case SOCKLND_CONN_ANY:
1737 case SOCKLND_CONN_CONTROL:
1739 case SOCKLND_CONN_BULK_IN:
1740 return SOCKLND_CONN_BULK_OUT;
1741 case SOCKLND_CONN_BULK_OUT:
1742 return SOCKLND_CONN_BULK_IN;
1744 return (SOCKLND_CONN_NONE);
1749 ksocknal_recv_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1750 struct ksock_hello_msg *hello,
1751 struct lnet_process_id *peerid,
1754 /* Return < 0 fatal error
1756 * EALREADY lost connection race
1757 * EPROTO protocol version mismatch
1759 struct socket *sock = conn->ksnc_sock;
1760 int active = (conn->ksnc_proto != NULL);
1764 const struct ksock_proto *proto;
1765 struct lnet_process_id recv_id;
1767 /* socket type set on active connections - not set on passive */
1768 LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1770 timeout = active ? ksocknal_timeout() :
1771 lnet_acceptor_timeout();
1773 rc = lnet_sock_read(sock, &hello->kshm_magic,
1774 sizeof(hello->kshm_magic), timeout);
1776 CERROR("Error %d reading HELLO from %pIS\n",
1777 rc, &conn->ksnc_peeraddr);
1782 if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1783 hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1784 hello->kshm_magic != le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1785 /* Unexpected magic! */
1786 CERROR("Bad magic(1) %#08x (%#08x expected) from %pIS\n",
1787 __cpu_to_le32 (hello->kshm_magic),
1788 LNET_PROTO_TCP_MAGIC, &conn->ksnc_peeraddr);
1792 rc = lnet_sock_read(sock, &hello->kshm_version,
1793 sizeof(hello->kshm_version), timeout);
1795 CERROR("Error %d reading HELLO from %pIS\n",
1796 rc, &conn->ksnc_peeraddr);
1801 proto = ksocknal_parse_proto_version(hello);
1802 if (proto == NULL) {
1804 /* unknown protocol from peer_ni, tell peer_ni my protocol */
1805 conn->ksnc_proto = &ksocknal_protocol_v3x;
1806 #if SOCKNAL_VERSION_DEBUG
1807 if (*ksocknal_tunables.ksnd_protocol == 2)
1808 conn->ksnc_proto = &ksocknal_protocol_v2x;
1809 else if (*ksocknal_tunables.ksnd_protocol == 1)
1810 conn->ksnc_proto = &ksocknal_protocol_v1x;
1812 hello->kshm_nips = 0;
1813 ksocknal_send_hello(ni, conn, ni->ni_nid, hello);
1816 CERROR("Unknown protocol version (%d.x expected) from %pIS\n",
1817 conn->ksnc_proto->pro_version, &conn->ksnc_peeraddr);
1822 proto_match = (conn->ksnc_proto == proto);
1823 conn->ksnc_proto = proto;
1825 /* receive the rest of hello message anyway */
1826 rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1828 CERROR("Error %d reading or checking hello from from %pIS\n",
1829 rc, &conn->ksnc_peeraddr);
1834 *incarnation = hello->kshm_src_incarnation;
1836 if (hello->kshm_src_nid == LNET_NID_ANY) {
1837 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pIS\n",
1838 &conn->ksnc_peeraddr);
1843 rpc_get_port((struct sockaddr *)&conn->ksnc_peeraddr) >
1844 LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1845 /* Userspace NAL assigns peer_ni process ID from socket */
1846 recv_id.pid = rpc_get_port((struct sockaddr *)
1847 &conn->ksnc_peeraddr) |
1849 LASSERT(conn->ksnc_peeraddr.ss_family == AF_INET);
1850 recv_id.nid = LNET_MKNID(
1851 LNET_NIDNET(ni->ni_nid),
1852 ntohl(((struct sockaddr_in *)
1853 &conn->ksnc_peeraddr)->sin_addr.s_addr));
1855 recv_id.nid = hello->kshm_src_nid;
1856 recv_id.pid = hello->kshm_src_pid;
1862 /* peer_ni determines type */
1863 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1864 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1865 CERROR("Unexpected type %d from %s ip %pIS\n",
1866 hello->kshm_ctype, libcfs_id2str(*peerid),
1867 &conn->ksnc_peeraddr);
1873 if (peerid->pid != recv_id.pid ||
1874 peerid->nid != recv_id.nid) {
1875 LCONSOLE_ERROR_MSG(0x130,
1876 "Connected successfully to %s on host %pIS, but they claimed they were %s; please check your Lustre configuration.\n",
1877 libcfs_id2str(*peerid),
1878 &conn->ksnc_peeraddr,
1879 libcfs_id2str(recv_id));
1883 if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1884 /* Possible protocol mismatch or I lost the connection race */
1885 return proto_match ? EALREADY : EPROTO;
1888 if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1889 CERROR("Mismatched types: me %d, %s ip %pIS %d\n",
1890 conn->ksnc_type, libcfs_id2str(*peerid),
1891 &conn->ksnc_peeraddr,
1899 ksocknal_connect(struct ksock_conn_cb *conn_cb)
1902 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
1905 struct socket *sock;
1907 bool retry_later = false;
1910 deadline = ktime_get_seconds() + ksocknal_timeout();
1912 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1914 LASSERT(conn_cb->ksnr_scheduled);
1915 LASSERT(!conn_cb->ksnr_connecting);
1917 conn_cb->ksnr_connecting = 1;
1920 wanted = ksocknal_conn_cb_mask() & ~conn_cb->ksnr_connected;
1922 /* stop connecting if peer_ni/cb got closed under me, or
1923 * conn cb got connected while queued
1925 if (peer_ni->ksnp_closing || conn_cb->ksnr_deleted ||
1927 retry_later = false;
1931 /* reschedule if peer_ni is connecting to me */
1932 if (peer_ni->ksnp_accepting > 0) {
1934 "peer_ni %s(%d) already connecting to me, retry later.\n",
1935 libcfs_nid2str(peer_ni->ksnp_id.nid), peer_ni->ksnp_accepting);
1939 if (retry_later) /* needs reschedule */
1942 if ((wanted & BIT(SOCKLND_CONN_ANY)) != 0) {
1943 type = SOCKLND_CONN_ANY;
1944 } else if ((wanted & BIT(SOCKLND_CONN_CONTROL)) != 0) {
1945 type = SOCKLND_CONN_CONTROL;
1946 } else if ((wanted & BIT(SOCKLND_CONN_BULK_IN)) != 0) {
1947 type = SOCKLND_CONN_BULK_IN;
1949 LASSERT ((wanted & BIT(SOCKLND_CONN_BULK_OUT)) != 0);
1950 type = SOCKLND_CONN_BULK_OUT;
1953 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1955 if (ktime_get_seconds() >= deadline) {
1957 lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
1959 &conn_cb->ksnr_addr);
1963 sock = lnet_connect(peer_ni->ksnp_id.nid,
1964 conn_cb->ksnr_myiface,
1965 (struct sockaddr *)&conn_cb->ksnr_addr,
1966 peer_ni->ksnp_ni->ni_net_ns);
1972 rc = ksocknal_create_conn(peer_ni->ksnp_ni, conn_cb, sock,
1975 lnet_connect_console_error(rc, peer_ni->ksnp_id.nid,
1977 &conn_cb->ksnr_addr);
1981 /* A +ve RC means I have to retry because I lost the connection
1982 * race or I have to renegotiate protocol version */
1983 retry_later = (rc != 0);
1985 CDEBUG(D_NET, "peer_ni %s: conn race, retry later.\n",
1986 libcfs_nid2str(peer_ni->ksnp_id.nid));
1988 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1991 conn_cb->ksnr_scheduled = 0;
1992 conn_cb->ksnr_connecting = 0;
1995 /* re-queue for attention; this frees me up to handle
1996 * the peer_ni's incoming connection request
1999 if (rc == EALREADY ||
2000 (rc == 0 && peer_ni->ksnp_accepting > 0)) {
2001 /* We want to introduce a delay before next
2002 * attempt to connect if we lost conn race, but
2003 * the race is resolved quickly usually, so
2004 * min_reconnectms should be good heuristic
2006 conn_cb->ksnr_retry_interval =
2007 *ksocknal_tunables.ksnd_min_reconnectms / 1000;
2008 conn_cb->ksnr_timeout = ktime_get_seconds() +
2009 conn_cb->ksnr_retry_interval;
2012 ksocknal_launch_connection_locked(conn_cb);
2015 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2019 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2021 conn_cb->ksnr_scheduled = 0;
2022 conn_cb->ksnr_connecting = 0;
2024 /* This is a retry rather than a new connection */
2025 conn_cb->ksnr_retry_interval *= 2;
2026 conn_cb->ksnr_retry_interval =
2027 max_t(time64_t, conn_cb->ksnr_retry_interval,
2028 *ksocknal_tunables.ksnd_min_reconnectms / 1000);
2029 conn_cb->ksnr_retry_interval =
2030 min_t(time64_t, conn_cb->ksnr_retry_interval,
2031 *ksocknal_tunables.ksnd_max_reconnectms / 1000);
2033 LASSERT(conn_cb->ksnr_retry_interval);
2034 conn_cb->ksnr_timeout = ktime_get_seconds() +
2035 conn_cb->ksnr_retry_interval;
2037 if (!list_empty(&peer_ni->ksnp_tx_queue) &&
2038 peer_ni->ksnp_accepting == 0 &&
2039 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
2040 struct ksock_conn *conn;
2042 /* ksnp_tx_queue is queued on a conn on successful
2043 * connection for V1.x and V2.x
2045 if (!list_empty(&peer_ni->ksnp_conns)) {
2046 conn = list_entry(peer_ni->ksnp_conns.next,
2047 struct ksock_conn, ksnc_list);
2048 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
2051 /* take all the blocked packets while I've got the lock and
2054 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
2057 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2059 ksocknal_peer_failed(peer_ni);
2060 ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc);
2065 * check whether we need to create more connds.
2066 * It will try to create new thread if it's necessary, @timeout can
2067 * be updated if failed to create, so caller wouldn't keep try while
2068 * running out of resource.
2071 ksocknal_connd_check_start(time64_t sec, long *timeout)
2075 int total = ksocknal_data.ksnd_connd_starting +
2076 ksocknal_data.ksnd_connd_running;
2078 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2079 /* still in initializing */
2083 if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2084 total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2085 /* can't create more connd, or still have enough
2086 * threads to handle more connecting */
2090 if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2091 /* no pending connecting request */
2095 if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2096 /* may run out of resource, retry later */
2097 *timeout = cfs_time_seconds(1);
2101 if (ksocknal_data.ksnd_connd_starting > 0) {
2102 /* serialize starting to avoid flood */
2106 ksocknal_data.ksnd_connd_starting_stamp = sec;
2107 ksocknal_data.ksnd_connd_starting++;
2108 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2110 /* NB: total is the next id */
2111 snprintf(name, sizeof(name), "socknal_cd%02d", total);
2112 rc = ksocknal_thread_start(ksocknal_connd, NULL, name);
2114 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2119 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2120 ksocknal_data.ksnd_connd_starting--;
2121 ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2127 * check whether current thread can exit, it will return 1 if there are too
2128 * many threads and no creating in past 120 seconds.
2129 * Also, this function may update @timeout to make caller come back
2130 * again to recheck these conditions.
2133 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2137 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2138 /* still in initializing */
2142 if (ksocknal_data.ksnd_connd_starting > 0) {
2143 /* in progress of starting new thread */
2147 if (ksocknal_data.ksnd_connd_running <=
2148 *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2152 /* created thread in past 120 seconds? */
2153 val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2154 SOCKNAL_CONND_TIMEOUT - sec);
2156 *timeout = (val > 0) ? cfs_time_seconds(val) :
2157 cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2161 /* no creating in past 120 seconds */
2163 return ksocknal_data.ksnd_connd_running >
2164 ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2167 /* Go through connd_cbs queue looking for a conn_cb that we can process
2168 * right now, @timeout_p can be updated if we need to come back later */
2169 static struct ksock_conn_cb *
2170 ksocknal_connd_get_conn_cb_locked(signed long *timeout_p)
2172 time64_t now = ktime_get_seconds();
2173 time64_t conn_timeout;
2174 struct ksock_conn_cb *conn_cb;
2176 /* connd_routes can contain both pending and ordinary routes */
2177 list_for_each_entry(conn_cb, &ksocknal_data.ksnd_connd_routes,
2180 conn_timeout = conn_cb->ksnr_timeout;
2182 if (conn_cb->ksnr_retry_interval == 0 ||
2183 now >= conn_timeout)
2186 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2187 *timeout_p > cfs_time_seconds(conn_timeout - now))
2188 *timeout_p = cfs_time_seconds(conn_timeout - now);
2195 ksocknal_connd(void *arg)
2197 spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2198 struct ksock_connreq *cr;
2199 wait_queue_entry_t wait;
2204 spin_lock_bh(connd_lock);
2206 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2207 ksocknal_data.ksnd_connd_starting--;
2208 ksocknal_data.ksnd_connd_running++;
2210 while (!ksocknal_data.ksnd_shuttingdown) {
2211 struct ksock_conn_cb *conn_cb = NULL;
2212 time64_t sec = ktime_get_real_seconds();
2213 long timeout = MAX_SCHEDULE_TIMEOUT;
2214 bool dropped_lock = false;
2216 if (ksocknal_connd_check_stop(sec, &timeout)) {
2217 /* wakeup another one to check stop */
2218 wake_up(&ksocknal_data.ksnd_connd_waitq);
2222 if (ksocknal_connd_check_start(sec, &timeout)) {
2223 /* created new thread */
2224 dropped_lock = true;
2227 if (!list_empty(&ksocknal_data.ksnd_connd_connreqs)) {
2228 /* Connection accepted by the listener */
2229 cr = list_entry(ksocknal_data.ksnd_connd_connreqs.next,
2230 struct ksock_connreq, ksncr_list);
2232 list_del(&cr->ksncr_list);
2233 spin_unlock_bh(connd_lock);
2234 dropped_lock = true;
2236 ksocknal_create_conn(cr->ksncr_ni, NULL,
2237 cr->ksncr_sock, SOCKLND_CONN_NONE);
2238 lnet_ni_decref(cr->ksncr_ni);
2239 LIBCFS_FREE(cr, sizeof(*cr));
2241 spin_lock_bh(connd_lock);
2244 /* Only handle an outgoing connection request if there
2245 * is a thread left to handle incoming connections and
2248 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2249 ksocknal_data.ksnd_connd_running)
2250 conn_cb = ksocknal_connd_get_conn_cb_locked(&timeout);
2253 list_del(&conn_cb->ksnr_connd_list);
2254 ksocknal_data.ksnd_connd_connecting++;
2255 spin_unlock_bh(connd_lock);
2256 dropped_lock = true;
2258 if (ksocknal_connect(conn_cb)) {
2259 /* consecutive retry */
2260 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2261 CWARN("massive consecutive re-connecting to %pIS\n",
2262 &conn_cb->ksnr_addr);
2269 ksocknal_conn_cb_decref(conn_cb);
2271 spin_lock_bh(connd_lock);
2272 ksocknal_data.ksnd_connd_connecting--;
2276 if (!need_resched())
2278 spin_unlock_bh(connd_lock);
2280 spin_lock_bh(connd_lock);
2284 /* Nothing to do for 'timeout' */
2285 set_current_state(TASK_INTERRUPTIBLE);
2286 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq,
2288 spin_unlock_bh(connd_lock);
2290 schedule_timeout(timeout);
2292 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2293 spin_lock_bh(connd_lock);
2295 ksocknal_data.ksnd_connd_running--;
2296 spin_unlock_bh(connd_lock);
2298 ksocknal_thread_fini();
2302 static struct ksock_conn *
2303 ksocknal_find_timed_out_conn(struct ksock_peer_ni *peer_ni)
2305 /* We're called with a shared lock on ksnd_global_lock */
2306 struct ksock_conn *conn;
2307 struct list_head *ctmp;
2308 struct ksock_tx *tx;
2310 list_for_each(ctmp, &peer_ni->ksnp_conns) {
2313 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
2315 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2316 LASSERT (!conn->ksnc_closing);
2318 error = conn->ksnc_sock->sk->sk_err;
2320 ksocknal_conn_addref(conn);
2324 CNETERR("A connection with %s (%pISp) was reset; it may have rebooted.\n",
2325 libcfs_id2str(peer_ni->ksnp_id),
2326 &conn->ksnc_peeraddr);
2329 CNETERR("A connection with %s (%pISp) timed out; the network or node may be down.\n",
2330 libcfs_id2str(peer_ni->ksnp_id),
2331 &conn->ksnc_peeraddr);
2334 CNETERR("An unexpected network error %d occurred with %s (%pISp\n",
2336 libcfs_id2str(peer_ni->ksnp_id),
2337 &conn->ksnc_peeraddr);
2344 if (conn->ksnc_rx_started &&
2345 ktime_get_seconds() >= conn->ksnc_rx_deadline) {
2346 /* Timed out incomplete incoming message */
2347 ksocknal_conn_addref(conn);
2348 CNETERR("Timeout receiving from %s (%pISp), state %d wanted %d left %d\n",
2349 libcfs_id2str(peer_ni->ksnp_id),
2350 &conn->ksnc_peeraddr,
2351 conn->ksnc_rx_state,
2352 conn->ksnc_rx_nob_wanted,
2353 conn->ksnc_rx_nob_left);
2357 if ((!list_empty(&conn->ksnc_tx_queue) ||
2358 conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2359 ktime_get_seconds() >= conn->ksnc_tx_deadline) {
2360 /* Timed out messages queued for sending or
2361 * buffered in the socket's send buffer
2363 ksocknal_conn_addref(conn);
2364 list_for_each_entry(tx, &conn->ksnc_tx_queue,
2367 LNET_MSG_STATUS_LOCAL_TIMEOUT;
2368 CNETERR("Timeout sending data to %s (%pISp) the network or that node may be down.\n",
2369 libcfs_id2str(peer_ni->ksnp_id),
2370 &conn->ksnc_peeraddr);
2379 ksocknal_flush_stale_txs(struct ksock_peer_ni *peer_ni)
2381 struct ksock_tx *tx;
2382 LIST_HEAD(stale_txs);
2384 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2386 while (!list_empty(&peer_ni->ksnp_tx_queue)) {
2387 tx = list_entry(peer_ni->ksnp_tx_queue.next,
2388 struct ksock_tx, tx_list);
2390 if (ktime_get_seconds() < tx->tx_deadline)
2393 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2395 list_move_tail(&tx->tx_list, &stale_txs);
2398 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2400 ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT);
2404 ksocknal_send_keepalive_locked(struct ksock_peer_ni *peer_ni)
2405 __must_hold(&ksocknal_data.ksnd_global_lock)
2407 struct ksock_sched *sched;
2408 struct ksock_conn *conn;
2409 struct ksock_tx *tx;
2411 /* last_alive will be updated by create_conn */
2412 if (list_empty(&peer_ni->ksnp_conns))
2415 if (peer_ni->ksnp_proto != &ksocknal_protocol_v3x)
2418 if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2419 ktime_get_seconds() < peer_ni->ksnp_last_alive +
2420 *ksocknal_tunables.ksnd_keepalive)
2423 if (ktime_get_seconds() < peer_ni->ksnp_send_keepalive)
2426 /* retry 10 secs later, so we wouldn't put pressure
2427 * on this peer_ni if we failed to send keepalive this time */
2428 peer_ni->ksnp_send_keepalive = ktime_get_seconds() + 10;
2430 conn = ksocknal_find_conn_locked(peer_ni, NULL, 1);
2432 sched = conn->ksnc_scheduler;
2434 spin_lock_bh(&sched->kss_lock);
2435 if (!list_empty(&conn->ksnc_tx_queue)) {
2436 spin_unlock_bh(&sched->kss_lock);
2437 /* there is an queued ACK, don't need keepalive */
2441 spin_unlock_bh(&sched->kss_lock);
2444 read_unlock(&ksocknal_data.ksnd_global_lock);
2446 /* cookie = 1 is reserved for keepalive PING */
2447 tx = ksocknal_alloc_tx_noop(1, 1);
2449 read_lock(&ksocknal_data.ksnd_global_lock);
2453 if (ksocknal_launch_packet(peer_ni->ksnp_ni, tx, peer_ni->ksnp_id) == 0) {
2454 read_lock(&ksocknal_data.ksnd_global_lock);
2458 ksocknal_free_tx(tx);
2459 read_lock(&ksocknal_data.ksnd_global_lock);
2466 ksocknal_check_peer_timeouts(int idx)
2468 struct hlist_head *peers = &ksocknal_data.ksnd_peers[idx];
2469 struct ksock_peer_ni *peer_ni;
2470 struct ksock_conn *conn;
2471 struct ksock_tx *tx;
2474 /* NB. We expect to have a look at all the peers and not find any
2475 * connections to time out, so we just use a shared lock while we
2478 read_lock(&ksocknal_data.ksnd_global_lock);
2480 hlist_for_each_entry(peer_ni, peers, ksnp_list) {
2481 struct ksock_tx *tx_stale;
2482 time64_t deadline = 0;
2486 if (ksocknal_send_keepalive_locked(peer_ni) != 0) {
2487 read_unlock(&ksocknal_data.ksnd_global_lock);
2491 conn = ksocknal_find_timed_out_conn(peer_ni);
2494 read_unlock(&ksocknal_data.ksnd_global_lock);
2496 ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2498 /* NB we won't find this one again, but we can't
2499 * just proceed with the next peer_ni, since we dropped
2500 * ksnd_global_lock and it might be dead already!
2502 ksocknal_conn_decref(conn);
2506 /* we can't process stale txs right here because we're
2507 * holding only shared lock
2509 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
2510 struct ksock_tx *tx;
2512 tx = list_entry(peer_ni->ksnp_tx_queue.next,
2513 struct ksock_tx, tx_list);
2514 if (ktime_get_seconds() >= tx->tx_deadline) {
2515 ksocknal_peer_addref(peer_ni);
2516 read_unlock(&ksocknal_data.ksnd_global_lock);
2518 ksocknal_flush_stale_txs(peer_ni);
2520 ksocknal_peer_decref(peer_ni);
2525 if (list_empty(&peer_ni->ksnp_zc_req_list))
2529 spin_lock(&peer_ni->ksnp_lock);
2530 list_for_each_entry(tx, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
2531 if (ktime_get_seconds() < tx->tx_deadline)
2533 /* ignore the TX if connection is being closed */
2534 if (tx->tx_conn->ksnc_closing)
2537 if (tx_stale == NULL)
2541 if (tx_stale == NULL) {
2542 spin_unlock(&peer_ni->ksnp_lock);
2546 deadline = tx_stale->tx_deadline;
2547 resid = tx_stale->tx_resid;
2548 conn = tx_stale->tx_conn;
2549 ksocknal_conn_addref(conn);
2551 spin_unlock(&peer_ni->ksnp_lock);
2552 read_unlock(&ksocknal_data.ksnd_global_lock);
2554 CERROR("Total %d stale ZC_REQs for peer_ni %s detected; the "
2555 "oldest(%p) timed out %lld secs ago, "
2556 "resid: %d, wmem: %d\n",
2557 n, libcfs_nid2str(peer_ni->ksnp_id.nid), tx_stale,
2558 ktime_get_seconds() - deadline,
2559 resid, conn->ksnc_sock->sk->sk_wmem_queued);
2561 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2562 ksocknal_conn_decref(conn);
2566 read_unlock(&ksocknal_data.ksnd_global_lock);
2569 int ksocknal_reaper(void *arg)
2571 wait_queue_entry_t wait;
2572 struct ksock_conn *conn;
2573 struct ksock_sched *sched;
2574 LIST_HEAD(enomem_conns);
2579 time64_t deadline = ktime_get_seconds();
2583 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2585 while (!ksocknal_data.ksnd_shuttingdown) {
2586 if (!list_empty(&ksocknal_data.ksnd_deathrow_conns)) {
2587 conn = list_entry(ksocknal_data.ksnd_deathrow_conns.next,
2588 struct ksock_conn, ksnc_list);
2589 list_del(&conn->ksnc_list);
2591 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2593 ksocknal_terminate_conn(conn);
2594 ksocknal_conn_decref(conn);
2596 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2600 if (!list_empty(&ksocknal_data.ksnd_zombie_conns)) {
2601 conn = list_entry(ksocknal_data.ksnd_zombie_conns.next,
2602 struct ksock_conn, ksnc_list);
2603 list_del(&conn->ksnc_list);
2605 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2607 ksocknal_destroy_conn(conn);
2609 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2613 list_splice_init(&ksocknal_data.ksnd_enomem_conns,
2616 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2618 /* reschedule all the connections that stalled with ENOMEM... */
2620 while (!list_empty(&enomem_conns)) {
2621 conn = list_entry(enomem_conns.next,
2622 struct ksock_conn, ksnc_tx_list);
2623 list_del(&conn->ksnc_tx_list);
2625 sched = conn->ksnc_scheduler;
2627 spin_lock_bh(&sched->kss_lock);
2629 LASSERT(conn->ksnc_tx_scheduled);
2630 conn->ksnc_tx_ready = 1;
2631 list_add_tail(&conn->ksnc_tx_list,
2632 &sched->kss_tx_conns);
2633 wake_up(&sched->kss_waitq);
2635 spin_unlock_bh(&sched->kss_lock);
2639 /* careful with the jiffy wrap... */
2640 while ((timeout = deadline - ktime_get_seconds()) <= 0) {
2643 int chunk = HASH_SIZE(ksocknal_data.ksnd_peers);
2644 unsigned int lnd_timeout;
2646 /* Time to check for timeouts on a few more peers: I
2647 * do checks every 'p' seconds on a proportion of the
2648 * peer_ni table and I need to check every connection
2649 * 'n' times within a timeout interval, to ensure I
2650 * detect a timeout on any connection within (n+1)/n
2651 * times the timeout interval.
2654 lnd_timeout = ksocknal_timeout();
2655 if (lnd_timeout > n * p)
2656 chunk = (chunk * n * p) / lnd_timeout;
2660 for (i = 0; i < chunk; i++) {
2661 ksocknal_check_peer_timeouts(peer_index);
2662 peer_index = (peer_index + 1) %
2663 HASH_SIZE(ksocknal_data.ksnd_peers);
2669 if (nenomem_conns != 0) {
2670 /* Reduce my timeout if I rescheduled ENOMEM conns.
2671 * This also prevents me getting woken immediately
2672 * if any go back on my enomem list. */
2673 timeout = SOCKNAL_ENOMEM_RETRY;
2675 ksocknal_data.ksnd_reaper_waketime = ktime_get_seconds() +
2678 set_current_state(TASK_INTERRUPTIBLE);
2679 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2681 if (!ksocknal_data.ksnd_shuttingdown &&
2682 list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2683 list_empty(&ksocknal_data.ksnd_zombie_conns))
2684 schedule_timeout(cfs_time_seconds(timeout));
2686 set_current_state(TASK_RUNNING);
2687 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2689 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2692 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2694 ksocknal_thread_fini();