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 tx = list_first_entry_or_null(&ksocknal_data.ksnd_idle_noop_txs,
43 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;
307 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;
400 LASSERT(ni != NULL || tx->tx_conn != NULL);
402 if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted)) {
404 if (hstatus == LNET_MSG_STATUS_OK)
405 hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
408 if (tx->tx_conn != NULL)
409 ksocknal_conn_decref(tx->tx_conn);
411 ksocknal_free_tx(tx);
412 if (lnetmsg != NULL) { /* KSOCK_MSG_NOOP go without lnetmsg */
413 lnetmsg->msg_health_status = hstatus;
414 lnet_finalize(lnetmsg, rc);
419 ksocknal_txlist_done(struct lnet_ni *ni, struct list_head *txlist, int error)
423 while ((tx = list_first_entry_or_null(txlist, struct ksock_tx,
425 if (error && tx->tx_lnetmsg) {
426 CNETERR("Deleting packet type %d len %d %s->%s\n",
427 tx->tx_lnetmsg->msg_type,
428 tx->tx_lnetmsg->msg_len,
429 libcfs_nidstr(&tx->tx_lnetmsg->msg_initiator),
430 libcfs_nidstr(&tx->tx_lnetmsg->msg_target.nid));
432 CNETERR("Deleting noop packet\n");
435 list_del(&tx->tx_list);
437 if (tx->tx_hstatus == LNET_MSG_STATUS_OK) {
438 if (error == -ETIMEDOUT)
440 LNET_MSG_STATUS_LOCAL_TIMEOUT;
441 else if (error == -ENETDOWN ||
442 error == -EHOSTUNREACH ||
443 error == -ENETUNREACH ||
444 error == -ECONNREFUSED ||
445 error == -ECONNRESET)
446 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_DROPPED;
448 * for all other errors we don't want to
452 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
455 LASSERT(refcount_read(&tx->tx_refcount) == 1);
456 ksocknal_tx_done(ni, tx, error);
461 ksocknal_check_zc_req(struct ksock_tx *tx)
463 struct ksock_conn *conn = tx->tx_conn;
464 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
466 /* Set tx_msg.ksm_zc_cookies[0] to a unique non-zero cookie and add tx
467 * to ksnp_zc_req_list if some fragment of this message should be sent
468 * zero-copy. Our peer_ni will send an ACK containing this cookie when
469 * she has received this message to tell us we can signal completion.
470 * tx_msg.ksm_zc_cookies[0] remains non-zero while tx is on
471 * ksnp_zc_req_list. */
472 LASSERT (tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
473 LASSERT (tx->tx_zc_capable);
475 tx->tx_zc_checked = 1;
477 if (conn->ksnc_proto == &ksocknal_protocol_v1x ||
478 !conn->ksnc_zc_capable)
481 /* assign cookie and queue tx to pending list, it will be released when
482 * a matching ack is received. See ksocknal_handle_zcack() */
484 ksocknal_tx_addref(tx);
486 spin_lock(&peer_ni->ksnp_lock);
488 /* ZC_REQ is going to be pinned to the peer_ni */
489 tx->tx_deadline = ktime_get_seconds() +
492 LASSERT (tx->tx_msg.ksm_zc_cookies[0] == 0);
494 tx->tx_msg.ksm_zc_cookies[0] = peer_ni->ksnp_zc_next_cookie++;
496 if (peer_ni->ksnp_zc_next_cookie == 0)
497 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
499 list_add_tail(&tx->tx_zc_list, &peer_ni->ksnp_zc_req_list);
501 spin_unlock(&peer_ni->ksnp_lock);
505 ksocknal_uncheck_zc_req(struct ksock_tx *tx)
507 struct ksock_peer_ni *peer_ni = tx->tx_conn->ksnc_peer;
509 LASSERT(tx->tx_msg.ksm_type != KSOCK_MSG_NOOP);
510 LASSERT(tx->tx_zc_capable);
512 tx->tx_zc_checked = 0;
514 spin_lock(&peer_ni->ksnp_lock);
516 if (tx->tx_msg.ksm_zc_cookies[0] == 0) {
517 /* Not waiting for an ACK */
518 spin_unlock(&peer_ni->ksnp_lock);
522 tx->tx_msg.ksm_zc_cookies[0] = 0;
523 list_del(&tx->tx_zc_list);
525 spin_unlock(&peer_ni->ksnp_lock);
527 ksocknal_tx_decref(tx);
531 ksocknal_process_transmit(struct ksock_conn *conn, struct ksock_tx *tx,
532 struct kvec *scratch_iov)
535 bool error_sim = false;
537 if (lnet_send_error_simulation(tx->tx_lnetmsg, &tx->tx_hstatus)) {
543 if (tx->tx_zc_capable && !tx->tx_zc_checked)
544 ksocknal_check_zc_req(tx);
546 rc = ksocknal_transmit(conn, tx, scratch_iov);
548 CDEBUG(D_NET, "send(%d) %d\n", tx->tx_resid, rc);
550 if (tx->tx_resid == 0) {
551 /* Sent everything OK */
563 counter++; /* exponential backoff warnings */
564 if ((counter & (-counter)) == counter)
565 CWARN("%u ENOMEM tx %p (%lld allocated)\n",
566 counter, conn, libcfs_kmem_read());
568 /* Queue on ksnd_enomem_conns for retry after a timeout */
569 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
571 /* enomem list takes over scheduler's ref... */
572 LASSERT(conn->ksnc_tx_scheduled);
573 list_add_tail(&conn->ksnc_tx_list,
574 &ksocknal_data.ksnd_enomem_conns);
575 if (ktime_get_seconds() + SOCKNAL_ENOMEM_RETRY <
576 ksocknal_data.ksnd_reaper_waketime)
577 wake_up(&ksocknal_data.ksnd_reaper_waitq);
579 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
582 * set the health status of the message which determines
583 * whether we should retry the transmit
585 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
596 * set the health status of the message which determines
597 * whether we should retry the transmit
599 if (rc == -ETIMEDOUT)
600 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_TIMEOUT;
602 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_ERROR;
605 if (!conn->ksnc_closing) {
608 LCONSOLE_WARN("Host %pIS reset our connection while we were sending data; it may have rebooted.\n",
609 &conn->ksnc_peeraddr);
612 LCONSOLE_WARN("There was an unexpected network error while writing to %pIS: %d.\n",
613 &conn->ksnc_peeraddr, rc);
616 CDEBUG(D_NET, "[%p] Error %d on write to %s ip %pISp\n",
617 conn, rc, libcfs_idstr(&conn->ksnc_peer->ksnp_id),
618 &conn->ksnc_peeraddr);
621 if (tx->tx_zc_checked)
622 ksocknal_uncheck_zc_req(tx);
624 /* it's not an error if conn is being closed */
625 ksocknal_close_conn_and_siblings(conn,
626 (conn->ksnc_closing) ? 0 : rc);
632 ksocknal_launch_connection_locked(struct ksock_conn_cb *conn_cb)
634 /* called holding write lock on ksnd_global_lock */
636 LASSERT(!conn_cb->ksnr_scheduled);
637 LASSERT(!conn_cb->ksnr_connecting);
638 LASSERT((ksocknal_conn_cb_mask() & ~conn_cb->ksnr_connected) != 0);
640 /* scheduling conn for connd */
641 conn_cb->ksnr_scheduled = 1;
643 /* extra ref for connd */
644 ksocknal_conn_cb_addref(conn_cb);
646 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
648 list_add_tail(&conn_cb->ksnr_connd_list,
649 &ksocknal_data.ksnd_connd_routes);
650 wake_up(&ksocknal_data.ksnd_connd_waitq);
652 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
656 ksocknal_launch_all_connections_locked(struct ksock_peer_ni *peer_ni)
658 struct ksock_conn_cb *conn_cb;
660 /* called holding write lock on ksnd_global_lock */
662 /* launch any/all connections that need it */
663 conn_cb = ksocknal_find_connectable_conn_cb_locked(peer_ni);
667 ksocknal_launch_connection_locked(conn_cb);
672 ksocknal_find_conn_locked(struct ksock_peer_ni *peer_ni, struct ksock_tx *tx, int nonblk)
674 struct ksock_conn *c;
675 struct ksock_conn *conn;
676 struct ksock_conn *typed = NULL;
677 struct ksock_conn *fallback = NULL;
681 list_for_each_entry(c, &peer_ni->ksnp_conns, ksnc_list) {
682 int nob = atomic_read(&c->ksnc_tx_nob) +
683 c->ksnc_sock->sk->sk_wmem_queued;
686 LASSERT (!c->ksnc_closing);
687 LASSERT (c->ksnc_proto != NULL &&
688 c->ksnc_proto->pro_match_tx != NULL);
690 rc = c->ksnc_proto->pro_match_tx(c, tx, nonblk);
695 case SOCKNAL_MATCH_NO: /* protocol rejected the tx */
698 case SOCKNAL_MATCH_YES: /* typed connection */
699 if (typed == NULL || tnob > nob ||
700 (tnob == nob && *ksocknal_tunables.ksnd_round_robin &&
701 typed->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
707 case SOCKNAL_MATCH_MAY: /* fallback connection */
708 if (fallback == NULL || fnob > nob ||
709 (fnob == nob && *ksocknal_tunables.ksnd_round_robin &&
710 fallback->ksnc_tx_last_post > c->ksnc_tx_last_post)) {
718 /* prefer the typed selection */
719 conn = (typed != NULL) ? typed : fallback;
722 conn->ksnc_tx_last_post = ktime_get_seconds();
728 ksocknal_tx_prep(struct ksock_conn *conn, struct ksock_tx *tx)
730 conn->ksnc_proto->pro_pack(tx);
732 atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
733 ksocknal_conn_addref(conn); /* +1 ref for tx */
738 ksocknal_queue_tx_locked(struct ksock_tx *tx, struct ksock_conn *conn)
740 struct ksock_sched *sched = conn->ksnc_scheduler;
741 struct ksock_msg *msg = &tx->tx_msg;
742 struct ksock_tx *ztx = NULL;
745 /* called holding global lock (read or irq-write) and caller may
746 * not have dropped this lock between finding conn and calling me,
747 * so we don't need the {get,put}connsock dance to deref
749 LASSERT(!conn->ksnc_closing);
751 CDEBUG(D_NET, "Sending to %s ip %pISp\n",
752 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
753 &conn->ksnc_peeraddr);
755 ksocknal_tx_prep(conn, tx);
757 /* Ensure the frags we've been given EXACTLY match the number of
758 * bytes we want to send. Many TCP/IP stacks disregard any total
759 * size parameters passed to them and just look at the frags.
761 * We always expect at least 1 mapped fragment containing the
762 * complete ksocknal message header.
764 LASSERT(lnet_iov_nob(tx->tx_niov, &tx->tx_hdr) +
765 lnet_kiov_nob(tx->tx_nkiov, tx->tx_kiov) ==
766 (unsigned int)tx->tx_nob);
767 LASSERT(tx->tx_niov >= 1);
768 LASSERT(tx->tx_resid == tx->tx_nob);
770 CDEBUG(D_NET, "Packet %p type %d, nob %d niov %d nkiov %d\n",
771 tx, tx->tx_lnetmsg ? tx->tx_lnetmsg->msg_type : KSOCK_MSG_NOOP,
772 tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
774 bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
775 spin_lock_bh(&sched->kss_lock);
777 if (list_empty(&conn->ksnc_tx_queue) && bufnob == 0) {
778 /* First packet starts the timeout */
779 conn->ksnc_tx_deadline = ktime_get_seconds() +
781 if (conn->ksnc_tx_bufnob > 0) /* something got ACKed */
782 conn->ksnc_peer->ksnp_last_alive = ktime_get_seconds();
783 conn->ksnc_tx_bufnob = 0;
784 smp_mb(); /* order with adding to tx_queue */
787 if (msg->ksm_type == KSOCK_MSG_NOOP) {
788 /* The packet is noop ZC ACK, try to piggyback the ack_cookie
789 * on a normal packet so I don't need to send it */
790 LASSERT (msg->ksm_zc_cookies[1] != 0);
791 LASSERT (conn->ksnc_proto->pro_queue_tx_zcack != NULL);
793 if (conn->ksnc_proto->pro_queue_tx_zcack(conn, tx, 0))
794 ztx = tx; /* ZC ACK piggybacked on ztx release tx later */
797 /* It's a normal packet - can it piggback a noop zc-ack that
798 * has been queued already? */
799 LASSERT (msg->ksm_zc_cookies[1] == 0);
800 LASSERT (conn->ksnc_proto->pro_queue_tx_msg != NULL);
802 ztx = conn->ksnc_proto->pro_queue_tx_msg(conn, tx);
803 /* ztx will be released later */
807 atomic_sub (ztx->tx_nob, &conn->ksnc_tx_nob);
808 list_add_tail(&ztx->tx_list, &sched->kss_zombie_noop_txs);
811 if (conn->ksnc_tx_ready && /* able to send */
812 !conn->ksnc_tx_scheduled) { /* not scheduled to send */
813 /* +1 ref for scheduler */
814 ksocknal_conn_addref(conn);
815 list_add_tail(&conn->ksnc_tx_list,
816 &sched->kss_tx_conns);
817 conn->ksnc_tx_scheduled = 1;
818 wake_up(&sched->kss_waitq);
821 spin_unlock_bh(&sched->kss_lock);
825 struct ksock_conn_cb *
826 ksocknal_find_connectable_conn_cb_locked(struct ksock_peer_ni *peer_ni)
828 time64_t now = ktime_get_seconds();
829 struct ksock_conn_cb *conn_cb;
831 conn_cb = peer_ni->ksnp_conn_cb;
835 LASSERT(!conn_cb->ksnr_connecting || conn_cb->ksnr_scheduled);
837 if (conn_cb->ksnr_scheduled) /* connections being established */
840 /* all conn types connected ? */
841 if ((ksocknal_conn_cb_mask() & ~conn_cb->ksnr_connected) == 0)
844 if (!(conn_cb->ksnr_retry_interval == 0 || /* first attempt */
845 now >= conn_cb->ksnr_timeout)) {
847 "Too soon to retry route %pIS (cnted %d, interval %lld, %lld secs later)\n",
849 conn_cb->ksnr_connected,
850 conn_cb->ksnr_retry_interval,
851 conn_cb->ksnr_timeout - now);
858 struct ksock_conn_cb *
859 ksocknal_find_connecting_conn_cb_locked(struct ksock_peer_ni *peer_ni)
861 struct ksock_conn_cb *conn_cb;
863 conn_cb = peer_ni->ksnp_conn_cb;
867 LASSERT(!conn_cb->ksnr_connecting || conn_cb->ksnr_scheduled);
869 return conn_cb->ksnr_scheduled ? conn_cb : NULL;
873 ksocknal_launch_packet(struct lnet_ni *ni, struct ksock_tx *tx,
874 struct lnet_processid *id)
876 struct ksock_peer_ni *peer_ni;
877 struct ksock_conn *conn;
878 struct sockaddr_storage sa;
883 LASSERT(tx->tx_conn == NULL);
885 g_lock = &ksocknal_data.ksnd_global_lock;
887 for (retry = 0;; retry = 1) {
889 peer_ni = ksocknal_find_peer_locked(ni, id);
890 if (peer_ni != NULL) {
891 if (ksocknal_find_connectable_conn_cb_locked(peer_ni) == NULL) {
892 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
894 /* I've got nothing that need to be
895 * connecting and I do have an actual
898 ksocknal_queue_tx_locked(tx, conn);
905 /* I'll need a write lock... */
908 write_lock_bh(g_lock);
910 peer_ni = ksocknal_find_peer_locked(ni, id);
914 write_unlock_bh(g_lock);
916 if ((id->pid & LNET_PID_USERFLAG) != 0) {
917 CERROR("Refusing to create a connection to userspace process %s\n",
919 return -EHOSTUNREACH;
923 CERROR("Can't find peer_ni %s\n", libcfs_idstr(id));
924 return -EHOSTUNREACH;
927 memset(&sa, 0, sizeof(sa));
928 switch (NID_ADDR_BYTES(&id->nid)) {
929 struct sockaddr_in *sin;
930 struct sockaddr_in6 *sin6;
933 sin->sin_family = AF_INET;
934 sin->sin_addr.s_addr = id->nid.nid_addr[0];
935 sin->sin_port = htons(lnet_acceptor_port());
939 sin6->sin6_family = AF_INET6;
940 memcpy(&sin6->sin6_addr, id->nid.nid_addr,
941 sizeof(sin6->sin6_addr));
942 sin6->sin6_port = htons(lnet_acceptor_port());
945 rc = ksocknal_add_peer(ni, id, (struct sockaddr *)&sa);
947 CERROR("Can't add peer_ni %s: %d\n",
948 libcfs_idstr(id), rc);
953 ksocknal_launch_all_connections_locked(peer_ni);
955 conn = ksocknal_find_conn_locked(peer_ni, tx, tx->tx_nonblk);
957 /* Connection exists; queue message on it */
958 ksocknal_queue_tx_locked (tx, conn);
959 write_unlock_bh(g_lock);
963 if (peer_ni->ksnp_accepting > 0 ||
964 ksocknal_find_connecting_conn_cb_locked(peer_ni) != NULL) {
965 /* the message is going to be pinned to the peer_ni */
966 tx->tx_deadline = ktime_get_seconds() +
969 /* Queue the message until a connection is established */
970 list_add_tail(&tx->tx_list, &peer_ni->ksnp_tx_queue);
971 write_unlock_bh(g_lock);
975 write_unlock_bh(g_lock);
977 /* NB Routes may be ignored if connections to them failed recently */
978 CNETERR("No usable routes to %s\n", libcfs_idstr(id));
979 tx->tx_hstatus = LNET_MSG_STATUS_REMOTE_ERROR;
980 return (-EHOSTUNREACH);
984 ksocknal_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg)
986 /* '1' for consistency with code that checks !mpflag to restore */
987 unsigned int mpflag = 1;
988 int type = lntmsg->msg_type;
989 struct lnet_processid *target = &lntmsg->msg_target;
990 unsigned int payload_niov = lntmsg->msg_niov;
991 struct bio_vec *payload_kiov = lntmsg->msg_kiov;
992 unsigned int payload_offset = lntmsg->msg_offset;
993 unsigned int payload_nob = lntmsg->msg_len;
998 /* NB 'private' is different depending on what we're sending.
1002 CDEBUG(D_NET, "sending %u bytes in %d frags to %s\n",
1003 payload_nob, payload_niov, libcfs_idstr(target));
1005 LASSERT (payload_nob == 0 || payload_niov > 0);
1006 LASSERT (payload_niov <= LNET_MAX_IOV);
1007 LASSERT (!in_interrupt ());
1009 desc_size = offsetof(struct ksock_tx,
1010 tx_payload[payload_niov]);
1012 if (lntmsg->msg_vmflush)
1013 mpflag = memalloc_noreclaim_save();
1015 tx = ksocknal_alloc_tx(KSOCK_MSG_LNET, desc_size);
1017 CERROR("Can't allocate tx desc type %d size %d\n",
1019 if (lntmsg->msg_vmflush)
1020 memalloc_noreclaim_restore(mpflag);
1024 tx->tx_conn = NULL; /* set when assigned a conn */
1025 tx->tx_lnetmsg = lntmsg;
1028 tx->tx_kiov = tx->tx_payload;
1029 tx->tx_nkiov = lnet_extract_kiov(payload_niov, tx->tx_kiov,
1030 payload_niov, payload_kiov,
1031 payload_offset, payload_nob);
1033 if (payload_nob >= *ksocknal_tunables.ksnd_zc_min_payload)
1034 tx->tx_zc_capable = 1;
1036 tx->tx_msg.ksm_csum = 0;
1037 tx->tx_msg.ksm_type = KSOCK_MSG_LNET;
1038 tx->tx_msg.ksm_zc_cookies[0] = 0;
1039 tx->tx_msg.ksm_zc_cookies[1] = 0;
1041 /* The first fragment will be set later in pro_pack */
1042 rc = ksocknal_launch_packet(ni, tx, target);
1044 * We can't test lntsmg->msg_vmflush again as lntmsg may
1048 memalloc_noreclaim_restore(mpflag);
1053 lntmsg->msg_health_status = tx->tx_hstatus;
1054 ksocknal_free_tx(tx);
1059 ksocknal_thread_fini (void)
1061 if (atomic_dec_and_test(&ksocknal_data.ksnd_nthreads))
1062 wake_up_var(&ksocknal_data.ksnd_nthreads);
1066 ksocknal_new_packet(struct ksock_conn *conn, int nob_to_skip)
1068 static char ksocknal_slop_buffer[4096];
1073 LASSERT(conn->ksnc_proto != NULL);
1075 if ((*ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0) {
1076 /* Remind the socket to ack eagerly... */
1077 ksocknal_lib_eager_ack(conn);
1080 if (nob_to_skip == 0) { /* right at next packet boundary now */
1081 conn->ksnc_rx_started = 0;
1082 smp_mb(); /* racing with timeout thread */
1084 switch (conn->ksnc_proto->pro_version) {
1085 case KSOCK_PROTO_V2:
1086 case KSOCK_PROTO_V3:
1087 conn->ksnc_rx_state = SOCKNAL_RX_KSM_HEADER;
1088 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1089 conn->ksnc_rx_iov[0].iov_base = (char *)&conn->ksnc_msg;
1091 conn->ksnc_rx_nob_wanted = sizeof(struct ksock_msg_hdr);
1092 conn->ksnc_rx_nob_left = sizeof(struct ksock_msg_hdr);
1093 conn->ksnc_rx_iov[0].iov_len =
1094 sizeof(struct ksock_msg_hdr);
1097 case KSOCK_PROTO_V1:
1098 /* Receiving bare struct lnet_hdr_nid4 */
1099 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1100 conn->ksnc_rx_nob_wanted = sizeof(struct lnet_hdr_nid4);
1101 conn->ksnc_rx_nob_left = sizeof(struct lnet_hdr_nid4);
1103 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1104 conn->ksnc_rx_iov[0].iov_base =
1105 (void *)&conn->ksnc_msg.ksm_u.lnetmsg_nid4;
1106 conn->ksnc_rx_iov[0].iov_len =
1107 sizeof(struct lnet_hdr_nid4);
1113 conn->ksnc_rx_niov = 1;
1115 conn->ksnc_rx_kiov = NULL;
1116 conn->ksnc_rx_nkiov = 0;
1117 conn->ksnc_rx_csum = ~0;
1121 /* Set up to skip as much as possible now. If there's more left
1122 * (ran out of iov entries) we'll get called again */
1124 conn->ksnc_rx_state = SOCKNAL_RX_SLOP;
1125 conn->ksnc_rx_nob_left = nob_to_skip;
1126 conn->ksnc_rx_iov = (struct kvec *)&conn->ksnc_rx_iov_space;
1131 nob = min_t(int, nob_to_skip, sizeof(ksocknal_slop_buffer));
1133 conn->ksnc_rx_iov[niov].iov_base = ksocknal_slop_buffer;
1134 conn->ksnc_rx_iov[niov].iov_len = nob;
1139 } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
1140 niov < sizeof(conn->ksnc_rx_iov_space) / sizeof(struct kvec));
1142 conn->ksnc_rx_niov = niov;
1143 conn->ksnc_rx_kiov = NULL;
1144 conn->ksnc_rx_nkiov = 0;
1145 conn->ksnc_rx_nob_wanted = skipped;
1150 ksocknal_process_receive(struct ksock_conn *conn,
1151 struct page **rx_scratch_pgs,
1152 struct kvec *scratch_iov)
1154 struct _lnet_hdr_nid4 *lhdr;
1155 struct lnet_processid *id;
1156 struct lnet_hdr hdr;
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_processid *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_idstr(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_idstr(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_csum != 0 && /* has checksum */
1209 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1210 /* NOOP Checksum error */
1211 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1212 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1213 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1214 ksocknal_new_packet(conn, 0);
1215 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1219 if (conn->ksnc_msg.ksm_zc_cookies[1] != 0) {
1222 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1224 if (conn->ksnc_msg.ksm_type == KSOCK_MSG_NOOP)
1225 cookie = conn->ksnc_msg.ksm_zc_cookies[0];
1227 rc = conn->ksnc_proto->pro_handle_zcack(
1228 conn, cookie, conn->ksnc_msg.ksm_zc_cookies[1]);
1231 CERROR("%s: Unknown ZC-ACK cookie: %llu, %llu\n",
1232 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1234 conn->ksnc_msg.ksm_zc_cookies[1]);
1235 ksocknal_new_packet(conn, 0);
1236 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1241 switch (conn->ksnc_msg.ksm_type) {
1242 case KSOCK_MSG_NOOP:
1243 ksocknal_new_packet(conn, 0);
1244 return 0; /* NOOP is done and just return */
1246 case KSOCK_MSG_LNET:
1248 conn->ksnc_rx_state = SOCKNAL_RX_LNET_HEADER;
1249 conn->ksnc_rx_nob_wanted = sizeof(struct lnet_hdr_nid4);
1250 conn->ksnc_rx_nob_left = sizeof(struct lnet_hdr_nid4);
1252 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1253 conn->ksnc_rx_iov[0].iov_base =
1254 (void *)&conn->ksnc_msg.ksm_u.lnetmsg_nid4;
1255 conn->ksnc_rx_iov[0].iov_len =
1256 sizeof(struct lnet_hdr_nid4);
1258 conn->ksnc_rx_niov = 1;
1259 conn->ksnc_rx_kiov = NULL;
1260 conn->ksnc_rx_nkiov = 0;
1262 goto again; /* read lnet header now */
1265 CERROR("%s: Unknown message type: %x\n",
1266 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1267 conn->ksnc_msg.ksm_type);
1268 ksocknal_new_packet(conn, 0);
1269 ksocknal_close_conn_and_siblings(conn, -EPROTO);
1273 case SOCKNAL_RX_LNET_HEADER:
1274 /* unpack message header */
1275 conn->ksnc_proto->pro_unpack(&conn->ksnc_msg, &hdr);
1277 if ((conn->ksnc_peer->ksnp_id.pid & LNET_PID_USERFLAG) != 0) {
1278 /* Userspace peer_ni */
1279 id = &conn->ksnc_peer->ksnp_id;
1281 /* Substitute process ID assigned at connection time */
1282 hdr.src_pid = id->pid;
1283 hdr.src_nid = id->nid;
1286 conn->ksnc_rx_state = SOCKNAL_RX_PARSE;
1287 ksocknal_conn_addref(conn); /* ++ref while parsing */
1290 rc = lnet_parse(conn->ksnc_peer->ksnp_ni,
1292 &conn->ksnc_peer->ksnp_id.nid,
1295 /* I just received garbage: give up on this conn */
1296 ksocknal_new_packet(conn, 0);
1297 ksocknal_close_conn_and_siblings(conn, rc);
1298 ksocknal_conn_decref(conn);
1302 /* I'm racing with ksocknal_recv() */
1303 LASSERT(conn->ksnc_rx_state == SOCKNAL_RX_PARSE ||
1304 conn->ksnc_rx_state == SOCKNAL_RX_LNET_PAYLOAD);
1306 if (conn->ksnc_rx_state != SOCKNAL_RX_LNET_PAYLOAD)
1309 /* ksocknal_recv() got called */
1312 case SOCKNAL_RX_LNET_PAYLOAD:
1313 /* payload all received */
1316 if (conn->ksnc_rx_nob_left == 0 && /* not truncating */
1317 conn->ksnc_msg.ksm_csum != 0 && /* has checksum */
1318 conn->ksnc_msg.ksm_csum != conn->ksnc_rx_csum) {
1319 CERROR("%s: Checksum error, wire:0x%08X data:0x%08X\n",
1320 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1321 conn->ksnc_msg.ksm_csum, conn->ksnc_rx_csum);
1325 if (rc == 0 && conn->ksnc_msg.ksm_zc_cookies[0] != 0) {
1326 LASSERT(conn->ksnc_proto != &ksocknal_protocol_v1x);
1328 lhdr = (void *)&conn->ksnc_msg.ksm_u.lnetmsg_nid4;
1329 id = &conn->ksnc_peer->ksnp_id;
1331 rc = conn->ksnc_proto->pro_handle_zcreq(
1333 conn->ksnc_msg.ksm_zc_cookies[0],
1334 *ksocknal_tunables.ksnd_nonblk_zcack ||
1335 le64_to_cpu(lhdr->src_nid) !=
1336 lnet_nid_to_nid4(&id->nid));
1339 if (rc && conn->ksnc_lnet_msg)
1340 conn->ksnc_lnet_msg->msg_health_status =
1341 LNET_MSG_STATUS_REMOTE_ERROR;
1342 lnet_finalize(conn->ksnc_lnet_msg, rc);
1345 ksocknal_new_packet(conn, 0);
1346 ksocknal_close_conn_and_siblings(conn, rc);
1351 case SOCKNAL_RX_SLOP:
1352 /* starting new packet? */
1353 if (ksocknal_new_packet(conn, conn->ksnc_rx_nob_left))
1354 return 0; /* come back later */
1355 goto again; /* try to finish reading slop now */
1363 return (-EINVAL); /* keep gcc happy */
1367 ksocknal_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
1368 int delayed, unsigned int niov,
1369 struct bio_vec *kiov, unsigned int offset, unsigned int mlen,
1372 struct ksock_conn *conn = private;
1373 struct ksock_sched *sched = conn->ksnc_scheduler;
1375 LASSERT (mlen <= rlen);
1376 LASSERT (niov <= LNET_MAX_IOV);
1378 conn->ksnc_lnet_msg = msg;
1379 conn->ksnc_rx_nob_wanted = mlen;
1380 conn->ksnc_rx_nob_left = rlen;
1383 conn->ksnc_rx_nkiov = 0;
1384 conn->ksnc_rx_kiov = NULL;
1385 conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
1386 conn->ksnc_rx_niov = 0;
1388 conn->ksnc_rx_niov = 0;
1389 conn->ksnc_rx_iov = NULL;
1390 conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
1391 conn->ksnc_rx_nkiov =
1392 lnet_extract_kiov(LNET_MAX_IOV, conn->ksnc_rx_kiov,
1393 niov, kiov, offset, mlen);
1397 lnet_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
1398 lnet_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
1400 LASSERT (conn->ksnc_rx_scheduled);
1402 spin_lock_bh(&sched->kss_lock);
1404 switch (conn->ksnc_rx_state) {
1405 case SOCKNAL_RX_PARSE_WAIT:
1406 list_add_tail(&conn->ksnc_rx_list, &sched->kss_rx_conns);
1407 wake_up(&sched->kss_waitq);
1408 LASSERT(conn->ksnc_rx_ready);
1411 case SOCKNAL_RX_PARSE:
1412 /* scheduler hasn't noticed I'm parsing yet */
1416 conn->ksnc_rx_state = SOCKNAL_RX_LNET_PAYLOAD;
1418 spin_unlock_bh(&sched->kss_lock);
1419 ksocknal_conn_decref(conn);
1424 ksocknal_sched_cansleep(struct ksock_sched *sched)
1428 spin_lock_bh(&sched->kss_lock);
1430 rc = (!ksocknal_data.ksnd_shuttingdown &&
1431 list_empty(&sched->kss_rx_conns) &&
1432 list_empty(&sched->kss_tx_conns));
1434 spin_unlock_bh(&sched->kss_lock);
1438 int ksocknal_scheduler(void *arg)
1440 struct ksock_sched *sched;
1441 struct ksock_conn *conn;
1442 struct ksock_tx *tx;
1444 long id = (long)arg;
1445 struct page **rx_scratch_pgs;
1446 struct kvec *scratch_iov;
1448 sched = ksocknal_data.ksnd_schedulers[KSOCK_THREAD_CPT(id)];
1450 LIBCFS_CPT_ALLOC(rx_scratch_pgs, lnet_cpt_table(), sched->kss_cpt,
1451 sizeof(*rx_scratch_pgs) * LNET_MAX_IOV);
1452 if (!rx_scratch_pgs) {
1453 CERROR("Unable to allocate scratch pages\n");
1457 LIBCFS_CPT_ALLOC(scratch_iov, lnet_cpt_table(), sched->kss_cpt,
1458 sizeof(*scratch_iov) * LNET_MAX_IOV);
1460 CERROR("Unable to allocate scratch iov\n");
1464 rc = cfs_cpt_bind(lnet_cpt_table(), sched->kss_cpt);
1466 CWARN("Can't set CPU partition affinity to %d: %d\n",
1467 sched->kss_cpt, rc);
1470 spin_lock_bh(&sched->kss_lock);
1472 while (!ksocknal_data.ksnd_shuttingdown) {
1473 bool did_something = false;
1475 /* Ensure I progress everything semi-fairly */
1476 conn = list_first_entry_or_null(&sched->kss_rx_conns,
1480 list_del(&conn->ksnc_rx_list);
1482 LASSERT(conn->ksnc_rx_scheduled);
1483 LASSERT(conn->ksnc_rx_ready);
1485 /* clear rx_ready in case receive isn't complete.
1486 * Do it BEFORE we call process_recv, since
1487 * data_ready can set it any time after we release
1489 conn->ksnc_rx_ready = 0;
1490 spin_unlock_bh(&sched->kss_lock);
1492 rc = ksocknal_process_receive(conn, rx_scratch_pgs,
1495 spin_lock_bh(&sched->kss_lock);
1497 /* I'm the only one that can clear this flag */
1498 LASSERT(conn->ksnc_rx_scheduled);
1500 /* Did process_receive get everything it wanted? */
1502 conn->ksnc_rx_ready = 1;
1504 if (conn->ksnc_rx_state == SOCKNAL_RX_PARSE) {
1505 /* Conn blocked waiting for ksocknal_recv()
1506 * I change its state (under lock) to signal
1507 * it can be rescheduled */
1508 conn->ksnc_rx_state = SOCKNAL_RX_PARSE_WAIT;
1509 } else if (conn->ksnc_rx_ready) {
1510 /* reschedule for rx */
1511 list_add_tail(&conn->ksnc_rx_list,
1512 &sched->kss_rx_conns);
1514 conn->ksnc_rx_scheduled = 0;
1516 ksocknal_conn_decref(conn);
1519 did_something = true;
1522 if (!list_empty(&sched->kss_tx_conns)) {
1525 list_splice_init(&sched->kss_zombie_noop_txs, &zlist);
1527 conn = list_first_entry(&sched->kss_tx_conns,
1530 list_del(&conn->ksnc_tx_list);
1532 LASSERT(conn->ksnc_tx_scheduled);
1533 LASSERT(conn->ksnc_tx_ready);
1534 LASSERT(!list_empty(&conn->ksnc_tx_queue));
1536 tx = list_first_entry(&conn->ksnc_tx_queue,
1537 struct ksock_tx, tx_list);
1539 if (conn->ksnc_tx_carrier == tx)
1540 ksocknal_next_tx_carrier(conn);
1542 /* dequeue now so empty list => more to send */
1543 list_del(&tx->tx_list);
1545 /* Clear tx_ready in case send isn't complete. Do
1546 * it BEFORE we call process_transmit, since
1547 * write_space can set it any time after we release
1549 conn->ksnc_tx_ready = 0;
1550 spin_unlock_bh(&sched->kss_lock);
1552 if (!list_empty(&zlist)) {
1553 /* free zombie noop txs, it's fast because
1554 * noop txs are just put in freelist */
1555 ksocknal_txlist_done(NULL, &zlist, 0);
1558 rc = ksocknal_process_transmit(conn, tx, scratch_iov);
1560 if (rc == -ENOMEM || rc == -EAGAIN) {
1561 /* Incomplete send: replace tx on HEAD of tx_queue */
1562 spin_lock_bh(&sched->kss_lock);
1563 list_add(&tx->tx_list,
1564 &conn->ksnc_tx_queue);
1566 /* Complete send; tx -ref */
1567 ksocknal_tx_decref(tx);
1569 spin_lock_bh(&sched->kss_lock);
1570 /* assume space for more */
1571 conn->ksnc_tx_ready = 1;
1574 if (rc == -ENOMEM) {
1575 /* Do nothing; after a short timeout, this
1576 * conn will be reposted on kss_tx_conns. */
1577 } else if (conn->ksnc_tx_ready &&
1578 !list_empty(&conn->ksnc_tx_queue)) {
1579 /* reschedule for tx */
1580 list_add_tail(&conn->ksnc_tx_list,
1581 &sched->kss_tx_conns);
1583 conn->ksnc_tx_scheduled = 0;
1585 ksocknal_conn_decref(conn);
1588 did_something = true;
1590 if (!did_something || /* nothing to do */
1591 need_resched()) { /* hogging CPU? */
1592 spin_unlock_bh(&sched->kss_lock);
1594 if (!did_something) { /* wait for something to do */
1595 rc = wait_event_interruptible_exclusive(
1597 !ksocknal_sched_cansleep(sched));
1603 spin_lock_bh(&sched->kss_lock);
1607 spin_unlock_bh(&sched->kss_lock);
1608 CFS_FREE_PTR_ARRAY(rx_scratch_pgs, LNET_MAX_IOV);
1609 CFS_FREE_PTR_ARRAY(scratch_iov, LNET_MAX_IOV);
1610 ksocknal_thread_fini();
1615 * Add connection to kss_rx_conns of scheduler
1616 * and wakeup the scheduler.
1618 void ksocknal_read_callback(struct ksock_conn *conn)
1620 struct ksock_sched *sched;
1622 sched = conn->ksnc_scheduler;
1624 spin_lock_bh(&sched->kss_lock);
1626 conn->ksnc_rx_ready = 1;
1628 if (!conn->ksnc_rx_scheduled) { /* not being progressed */
1629 list_add_tail(&conn->ksnc_rx_list,
1630 &sched->kss_rx_conns);
1631 conn->ksnc_rx_scheduled = 1;
1632 /* extra ref for scheduler */
1633 ksocknal_conn_addref(conn);
1635 wake_up (&sched->kss_waitq);
1637 spin_unlock_bh(&sched->kss_lock);
1641 * Add connection to kss_tx_conns of scheduler
1642 * and wakeup the scheduler.
1644 void ksocknal_write_callback(struct ksock_conn *conn)
1646 struct ksock_sched *sched;
1648 sched = conn->ksnc_scheduler;
1650 spin_lock_bh(&sched->kss_lock);
1652 conn->ksnc_tx_ready = 1;
1654 if (!conn->ksnc_tx_scheduled && /* not being progressed */
1655 !list_empty(&conn->ksnc_tx_queue)) { /* packets to send */
1656 list_add_tail(&conn->ksnc_tx_list, &sched->kss_tx_conns);
1657 conn->ksnc_tx_scheduled = 1;
1658 /* extra ref for scheduler */
1659 ksocknal_conn_addref(conn);
1661 wake_up(&sched->kss_waitq);
1664 spin_unlock_bh(&sched->kss_lock);
1667 static const struct ksock_proto *
1668 ksocknal_parse_proto_version(struct ksock_hello_msg *hello)
1672 if (hello->kshm_magic == LNET_PROTO_MAGIC)
1673 version = hello->kshm_version;
1674 else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
1675 version = __swab32(hello->kshm_version);
1678 #if SOCKNAL_VERSION_DEBUG
1679 if (*ksocknal_tunables.ksnd_protocol == 1)
1682 if (*ksocknal_tunables.ksnd_protocol == 2 &&
1683 version == KSOCK_PROTO_V3)
1686 if (version == KSOCK_PROTO_V2)
1687 return &ksocknal_protocol_v2x;
1689 if (version == KSOCK_PROTO_V3)
1690 return &ksocknal_protocol_v3x;
1695 if (hello->kshm_magic == le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1696 struct lnet_magicversion *hmv;
1698 BUILD_BUG_ON(sizeof(struct lnet_magicversion) !=
1699 offsetof(struct ksock_hello_msg, kshm_src_nid));
1701 hmv = (struct lnet_magicversion *)hello;
1703 if (hmv->version_major == cpu_to_le16 (KSOCK_PROTO_V1_MAJOR) &&
1704 hmv->version_minor == cpu_to_le16 (KSOCK_PROTO_V1_MINOR))
1705 return &ksocknal_protocol_v1x;
1712 ksocknal_send_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1713 struct lnet_nid *peer_nid, struct ksock_hello_msg *hello)
1715 /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
1716 struct ksock_net *net = (struct ksock_net *)ni->ni_data;
1718 LASSERT(hello->kshm_nips <= LNET_INTERFACES_NUM);
1720 /* rely on caller to hold a ref on socket so it wouldn't disappear */
1721 LASSERT(conn->ksnc_proto != NULL);
1723 hello->kshm_src_nid = ni->ni_nid;
1724 hello->kshm_dst_nid = *peer_nid;
1725 hello->kshm_src_pid = the_lnet.ln_pid;
1727 hello->kshm_src_incarnation = net->ksnn_incarnation;
1728 hello->kshm_ctype = conn->ksnc_type;
1730 return conn->ksnc_proto->pro_send_hello(conn, hello);
1734 ksocknal_invert_type(int type)
1737 case SOCKLND_CONN_ANY:
1738 case SOCKLND_CONN_CONTROL:
1740 case SOCKLND_CONN_BULK_IN:
1741 return SOCKLND_CONN_BULK_OUT;
1742 case SOCKLND_CONN_BULK_OUT:
1743 return SOCKLND_CONN_BULK_IN;
1745 return (SOCKLND_CONN_NONE);
1750 ksocknal_recv_hello(struct lnet_ni *ni, struct ksock_conn *conn,
1751 struct ksock_hello_msg *hello,
1752 struct lnet_processid *peerid,
1755 /* Return < 0 fatal error
1757 * EALREADY lost connection race
1758 * EPROTO protocol version mismatch
1760 struct socket *sock = conn->ksnc_sock;
1761 int active = (conn->ksnc_proto != NULL);
1765 const struct ksock_proto *proto;
1766 struct lnet_processid recv_id;
1768 /* socket type set on active connections - not set on passive */
1769 LASSERT(!active == !(conn->ksnc_type != SOCKLND_CONN_NONE));
1771 timeout = active ? ksocknal_timeout() :
1772 lnet_acceptor_timeout();
1774 rc = lnet_sock_read(sock, &hello->kshm_magic,
1775 sizeof(hello->kshm_magic), timeout);
1777 CERROR("Error %d reading HELLO from %pIS\n",
1778 rc, &conn->ksnc_peeraddr);
1783 if (hello->kshm_magic != LNET_PROTO_MAGIC &&
1784 hello->kshm_magic != __swab32(LNET_PROTO_MAGIC) &&
1785 hello->kshm_magic != le32_to_cpu(LNET_PROTO_TCP_MAGIC)) {
1786 /* Unexpected magic! */
1787 CERROR("Bad magic(1) %#08x (%#08x expected) from %pIS\n",
1788 __cpu_to_le32 (hello->kshm_magic),
1789 LNET_PROTO_TCP_MAGIC, &conn->ksnc_peeraddr);
1793 rc = lnet_sock_read(sock, &hello->kshm_version,
1794 sizeof(hello->kshm_version), timeout);
1796 CERROR("Error %d reading HELLO from %pIS\n",
1797 rc, &conn->ksnc_peeraddr);
1802 proto = ksocknal_parse_proto_version(hello);
1803 if (proto == NULL) {
1805 /* unknown protocol from peer_ni,
1806 * tell peer_ni my protocol.
1808 conn->ksnc_proto = &ksocknal_protocol_v3x;
1809 #if SOCKNAL_VERSION_DEBUG
1810 if (*ksocknal_tunables.ksnd_protocol == 2)
1811 conn->ksnc_proto = &ksocknal_protocol_v2x;
1812 else if (*ksocknal_tunables.ksnd_protocol == 1)
1813 conn->ksnc_proto = &ksocknal_protocol_v1x;
1815 hello->kshm_nips = 0;
1816 ksocknal_send_hello(ni, conn, &ni->ni_nid,
1820 CERROR("Unknown protocol version (%d.x expected) from %pIS\n",
1821 conn->ksnc_proto->pro_version, &conn->ksnc_peeraddr);
1826 proto_match = (conn->ksnc_proto == proto);
1827 conn->ksnc_proto = proto;
1829 /* receive the rest of hello message anyway */
1830 rc = conn->ksnc_proto->pro_recv_hello(conn, hello, timeout);
1832 CERROR("Error %d reading or checking hello from from %pIS\n",
1833 rc, &conn->ksnc_peeraddr);
1838 *incarnation = hello->kshm_src_incarnation;
1840 if (LNET_NID_IS_ANY(&hello->kshm_src_nid)) {
1841 CERROR("Expecting a HELLO hdr with a NID, but got LNET_NID_ANY from %pIS\n",
1842 &conn->ksnc_peeraddr);
1847 rpc_get_port((struct sockaddr *)&conn->ksnc_peeraddr) >
1848 LNET_ACCEPTOR_MAX_RESERVED_PORT) {
1849 /* Userspace NAL assigns peer_ni process ID from socket */
1850 recv_id.pid = rpc_get_port((struct sockaddr *)
1851 &conn->ksnc_peeraddr) |
1853 LASSERT(conn->ksnc_peeraddr.ss_family == AF_INET);
1854 memset(&recv_id.nid, 0, sizeof(recv_id.nid));
1855 recv_id.nid.nid_type = ni->ni_nid.nid_type;
1856 recv_id.nid.nid_num = ni->ni_nid.nid_num;
1857 recv_id.nid.nid_addr[0] =
1858 ((struct sockaddr_in *)
1859 &conn->ksnc_peeraddr)->sin_addr.s_addr;
1861 recv_id.nid = hello->kshm_src_nid;
1862 recv_id.pid = hello->kshm_src_pid;
1868 /* peer_ni determines type */
1869 conn->ksnc_type = ksocknal_invert_type(hello->kshm_ctype);
1870 if (conn->ksnc_type == SOCKLND_CONN_NONE) {
1871 CERROR("Unexpected type %d from %s ip %pIS\n",
1872 hello->kshm_ctype, libcfs_idstr(peerid),
1873 &conn->ksnc_peeraddr);
1879 if (peerid->pid != recv_id.pid ||
1880 !nid_same(&peerid->nid, &recv_id.nid)) {
1881 LCONSOLE_ERROR_MSG(0x130,
1882 "Connected successfully to %s on host %pIS, but they claimed they were %s; please check your Lustre configuration.\n",
1883 libcfs_idstr(peerid),
1884 &conn->ksnc_peeraddr,
1885 libcfs_idstr(&recv_id));
1889 if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
1890 /* Possible protocol mismatch or I lost the connection race */
1891 return proto_match ? EALREADY : EPROTO;
1894 if (ksocknal_invert_type(hello->kshm_ctype) != conn->ksnc_type) {
1895 CERROR("Mismatched types: me %d, %s ip %pIS %d\n",
1896 conn->ksnc_type, libcfs_idstr(peerid),
1897 &conn->ksnc_peeraddr,
1905 ksocknal_connect(struct ksock_conn_cb *conn_cb)
1908 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
1911 struct socket *sock;
1913 bool retry_later = false;
1916 deadline = ktime_get_seconds() + ksocknal_timeout();
1918 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1920 LASSERT(conn_cb->ksnr_scheduled);
1921 LASSERT(!conn_cb->ksnr_connecting);
1923 conn_cb->ksnr_connecting = 1;
1926 wanted = ksocknal_conn_cb_mask() & ~conn_cb->ksnr_connected;
1928 /* stop connecting if peer_ni/cb got closed under me, or
1929 * conn cb got connected while queued
1931 if (peer_ni->ksnp_closing || conn_cb->ksnr_deleted ||
1933 retry_later = false;
1937 /* reschedule if peer_ni is connecting to me */
1938 if (peer_ni->ksnp_accepting > 0) {
1940 "peer_ni %s(%d) already connecting to me, retry later.\n",
1941 libcfs_nidstr(&peer_ni->ksnp_id.nid),
1942 peer_ni->ksnp_accepting);
1946 if (retry_later) /* needs reschedule */
1949 if ((wanted & BIT(SOCKLND_CONN_ANY)) != 0) {
1950 type = SOCKLND_CONN_ANY;
1951 } else if ((wanted & BIT(SOCKLND_CONN_CONTROL)) != 0) {
1952 type = SOCKLND_CONN_CONTROL;
1953 } else if ((wanted & BIT(SOCKLND_CONN_BULK_IN)) != 0 &&
1954 conn_cb->ksnr_blki_conn_count <= conn_cb->ksnr_blko_conn_count) {
1955 type = SOCKLND_CONN_BULK_IN;
1957 LASSERT ((wanted & BIT(SOCKLND_CONN_BULK_OUT)) != 0);
1958 type = SOCKLND_CONN_BULK_OUT;
1961 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1963 if (ktime_get_seconds() >= deadline) {
1965 lnet_connect_console_error(
1966 rc, &peer_ni->ksnp_id.nid,
1967 (struct sockaddr *)&conn_cb->ksnr_addr);
1971 sock = lnet_connect(&peer_ni->ksnp_id.nid,
1972 conn_cb->ksnr_myiface,
1973 (struct sockaddr *)&conn_cb->ksnr_addr,
1974 peer_ni->ksnp_ni->ni_net_ns);
1980 rc = ksocknal_create_conn(peer_ni->ksnp_ni, conn_cb, sock,
1983 lnet_connect_console_error(
1984 rc, &peer_ni->ksnp_id.nid,
1985 (struct sockaddr *)&conn_cb->ksnr_addr);
1989 /* A +ve RC means I have to retry because I lost the connection
1990 * race or I have to renegotiate protocol version
1992 retry_later = (rc != 0);
1994 CDEBUG(D_NET, "peer_ni %s: conn race, retry later.\n",
1995 libcfs_nidstr(&peer_ni->ksnp_id.nid));
1997 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2000 conn_cb->ksnr_scheduled = 0;
2001 conn_cb->ksnr_connecting = 0;
2004 /* re-queue for attention; this frees me up to handle
2005 * the peer_ni's incoming connection request
2008 if (rc == EALREADY ||
2009 (rc == 0 && peer_ni->ksnp_accepting > 0)) {
2010 /* We want to introduce a delay before next
2011 * attempt to connect if we lost conn race, but
2012 * the race is resolved quickly usually, so
2013 * min_reconnectms should be good heuristic
2015 conn_cb->ksnr_retry_interval =
2016 *ksocknal_tunables.ksnd_min_reconnectms / 1000;
2017 conn_cb->ksnr_timeout = ktime_get_seconds() +
2018 conn_cb->ksnr_retry_interval;
2021 ksocknal_launch_connection_locked(conn_cb);
2024 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2028 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2030 conn_cb->ksnr_scheduled = 0;
2031 conn_cb->ksnr_connecting = 0;
2033 /* This is a retry rather than a new connection */
2034 conn_cb->ksnr_retry_interval *= 2;
2035 conn_cb->ksnr_retry_interval =
2036 max_t(time64_t, conn_cb->ksnr_retry_interval,
2037 *ksocknal_tunables.ksnd_min_reconnectms / 1000);
2038 conn_cb->ksnr_retry_interval =
2039 min_t(time64_t, conn_cb->ksnr_retry_interval,
2040 *ksocknal_tunables.ksnd_max_reconnectms / 1000);
2042 LASSERT(conn_cb->ksnr_retry_interval);
2043 conn_cb->ksnr_timeout = ktime_get_seconds() +
2044 conn_cb->ksnr_retry_interval;
2046 if (!list_empty(&peer_ni->ksnp_tx_queue) &&
2047 peer_ni->ksnp_accepting == 0 &&
2048 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
2049 struct ksock_conn *conn;
2051 /* ksnp_tx_queue is queued on a conn on successful
2052 * connection for V1.x and V2.x
2054 conn = list_first_entry_or_null(&peer_ni->ksnp_conns,
2055 struct ksock_conn, ksnc_list);
2057 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x ||
2058 conn->ksnc_proto == &ksocknal_protocol_v4x);
2060 /* take all the blocked packets while I've got the lock and
2063 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
2066 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2068 ksocknal_peer_failed(peer_ni);
2069 ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc);
2074 * check whether we need to create more connds.
2075 * It will try to create new thread if it's necessary, @timeout can
2076 * be updated if failed to create, so caller wouldn't keep try while
2077 * running out of resource.
2080 ksocknal_connd_check_start(time64_t sec, long *timeout)
2083 int total = ksocknal_data.ksnd_connd_starting +
2084 ksocknal_data.ksnd_connd_running;
2086 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2087 /* still in initializing */
2091 if (total >= *ksocknal_tunables.ksnd_nconnds_max ||
2092 total > ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV) {
2093 /* can't create more connd, or still have enough
2094 * threads to handle more connecting */
2098 if (list_empty(&ksocknal_data.ksnd_connd_routes)) {
2099 /* no pending connecting request */
2103 if (sec - ksocknal_data.ksnd_connd_failed_stamp <= 1) {
2104 /* may run out of resource, retry later */
2105 *timeout = cfs_time_seconds(1);
2109 if (ksocknal_data.ksnd_connd_starting > 0) {
2110 /* serialize starting to avoid flood */
2114 ksocknal_data.ksnd_connd_starting_stamp = sec;
2115 ksocknal_data.ksnd_connd_starting++;
2116 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2118 /* NB: total is the next id */
2119 rc = ksocknal_thread_start(ksocknal_connd, NULL,
2120 "socknal_cd%02d", total);
2122 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2127 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2128 ksocknal_data.ksnd_connd_starting--;
2129 ksocknal_data.ksnd_connd_failed_stamp = ktime_get_real_seconds();
2135 * check whether current thread can exit, it will return 1 if there are too
2136 * many threads and no creating in past 120 seconds.
2137 * Also, this function may update @timeout to make caller come back
2138 * again to recheck these conditions.
2141 ksocknal_connd_check_stop(time64_t sec, long *timeout)
2145 if (unlikely(ksocknal_data.ksnd_init < SOCKNAL_INIT_ALL)) {
2146 /* still in initializing */
2150 if (ksocknal_data.ksnd_connd_starting > 0) {
2151 /* in progress of starting new thread */
2155 if (ksocknal_data.ksnd_connd_running <=
2156 *ksocknal_tunables.ksnd_nconnds) { /* can't shrink */
2160 /* created thread in past 120 seconds? */
2161 val = (int)(ksocknal_data.ksnd_connd_starting_stamp +
2162 SOCKNAL_CONND_TIMEOUT - sec);
2164 *timeout = (val > 0) ? cfs_time_seconds(val) :
2165 cfs_time_seconds(SOCKNAL_CONND_TIMEOUT);
2169 /* no creating in past 120 seconds */
2171 return ksocknal_data.ksnd_connd_running >
2172 ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV;
2175 /* Go through connd_cbs queue looking for a conn_cb that we can process
2176 * right now, @timeout_p can be updated if we need to come back later */
2177 static struct ksock_conn_cb *
2178 ksocknal_connd_get_conn_cb_locked(signed long *timeout_p)
2180 time64_t now = ktime_get_seconds();
2181 time64_t conn_timeout;
2182 struct ksock_conn_cb *conn_cb;
2184 /* connd_routes can contain both pending and ordinary routes */
2185 list_for_each_entry(conn_cb, &ksocknal_data.ksnd_connd_routes,
2188 conn_timeout = conn_cb->ksnr_timeout;
2190 if (conn_cb->ksnr_retry_interval == 0 ||
2191 now >= conn_timeout)
2194 if (*timeout_p == MAX_SCHEDULE_TIMEOUT ||
2195 *timeout_p > cfs_time_seconds(conn_timeout - now))
2196 *timeout_p = cfs_time_seconds(conn_timeout - now);
2203 ksocknal_connd(void *arg)
2205 spinlock_t *connd_lock = &ksocknal_data.ksnd_connd_lock;
2206 struct ksock_connreq *cr;
2207 wait_queue_entry_t wait;
2212 spin_lock_bh(connd_lock);
2214 LASSERT(ksocknal_data.ksnd_connd_starting > 0);
2215 ksocknal_data.ksnd_connd_starting--;
2216 ksocknal_data.ksnd_connd_running++;
2218 while (!ksocknal_data.ksnd_shuttingdown) {
2219 struct ksock_conn_cb *conn_cb = NULL;
2220 time64_t sec = ktime_get_real_seconds();
2221 long timeout = MAX_SCHEDULE_TIMEOUT;
2222 bool dropped_lock = false;
2224 if (ksocknal_connd_check_stop(sec, &timeout)) {
2225 /* wakeup another one to check stop */
2226 wake_up(&ksocknal_data.ksnd_connd_waitq);
2230 if (ksocknal_connd_check_start(sec, &timeout)) {
2231 /* created new thread */
2232 dropped_lock = true;
2235 cr = list_first_entry_or_null(&ksocknal_data.ksnd_connd_connreqs,
2236 struct ksock_connreq, ksncr_list);
2238 /* Connection accepted by the listener */
2239 list_del(&cr->ksncr_list);
2240 spin_unlock_bh(connd_lock);
2241 dropped_lock = true;
2243 ksocknal_create_conn(cr->ksncr_ni, NULL,
2244 cr->ksncr_sock, SOCKLND_CONN_NONE);
2245 lnet_ni_decref(cr->ksncr_ni);
2246 LIBCFS_FREE(cr, sizeof(*cr));
2248 spin_lock_bh(connd_lock);
2251 /* Only handle an outgoing connection request if there
2252 * is a thread left to handle incoming connections and
2255 if (ksocknal_data.ksnd_connd_connecting + SOCKNAL_CONND_RESV <
2256 ksocknal_data.ksnd_connd_running)
2257 conn_cb = ksocknal_connd_get_conn_cb_locked(&timeout);
2260 list_del(&conn_cb->ksnr_connd_list);
2261 ksocknal_data.ksnd_connd_connecting++;
2262 spin_unlock_bh(connd_lock);
2263 dropped_lock = true;
2265 if (ksocknal_connect(conn_cb)) {
2266 /* consecutive retry */
2267 if (cons_retry++ > SOCKNAL_INSANITY_RECONN) {
2268 CWARN("massive consecutive re-connecting to %pIS\n",
2269 &conn_cb->ksnr_addr);
2276 ksocknal_conn_cb_decref(conn_cb);
2278 spin_lock_bh(connd_lock);
2279 ksocknal_data.ksnd_connd_connecting--;
2283 if (!need_resched())
2285 spin_unlock_bh(connd_lock);
2287 spin_lock_bh(connd_lock);
2291 /* Nothing to do for 'timeout' */
2292 set_current_state(TASK_INTERRUPTIBLE);
2293 add_wait_queue_exclusive(&ksocknal_data.ksnd_connd_waitq,
2295 spin_unlock_bh(connd_lock);
2297 schedule_timeout(timeout);
2299 remove_wait_queue(&ksocknal_data.ksnd_connd_waitq, &wait);
2300 spin_lock_bh(connd_lock);
2302 ksocknal_data.ksnd_connd_running--;
2303 spin_unlock_bh(connd_lock);
2305 ksocknal_thread_fini();
2309 static struct ksock_conn *
2310 ksocknal_find_timed_out_conn(struct ksock_peer_ni *peer_ni)
2312 /* We're called with a shared lock on ksnd_global_lock */
2313 struct ksock_conn *conn;
2314 struct ksock_tx *tx;
2315 struct ksock_sched *sched;
2317 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2320 /* Don't need the {get,put}connsock dance to deref ksnc_sock */
2321 LASSERT (!conn->ksnc_closing);
2322 sched = conn->ksnc_scheduler;
2324 error = conn->ksnc_sock->sk->sk_err;
2326 ksocknal_conn_addref(conn);
2330 CNETERR("A connection with %s (%pISp) was reset; it may have rebooted.\n",
2331 libcfs_idstr(&peer_ni->ksnp_id),
2332 &conn->ksnc_peeraddr);
2335 CNETERR("A connection with %s (%pISp) timed out; the network or node may be down.\n",
2336 libcfs_idstr(&peer_ni->ksnp_id),
2337 &conn->ksnc_peeraddr);
2340 CNETERR("An unexpected network error %d occurred with %s (%pISp\n",
2342 libcfs_idstr(&peer_ni->ksnp_id),
2343 &conn->ksnc_peeraddr);
2350 if (conn->ksnc_rx_started &&
2351 ktime_get_seconds() >= conn->ksnc_rx_deadline) {
2352 /* Timed out incomplete incoming message */
2353 ksocknal_conn_addref(conn);
2354 CNETERR("Timeout receiving from %s (%pISp), state %d wanted %d left %d\n",
2355 libcfs_idstr(&peer_ni->ksnp_id),
2356 &conn->ksnc_peeraddr,
2357 conn->ksnc_rx_state,
2358 conn->ksnc_rx_nob_wanted,
2359 conn->ksnc_rx_nob_left);
2363 spin_lock_bh(&sched->kss_lock);
2364 if ((!list_empty(&conn->ksnc_tx_queue) ||
2365 conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
2366 ktime_get_seconds() >= conn->ksnc_tx_deadline) {
2367 /* Timed out messages queued for sending or
2368 * buffered in the socket's send buffer
2370 ksocknal_conn_addref(conn);
2371 list_for_each_entry(tx, &conn->ksnc_tx_queue,
2374 LNET_MSG_STATUS_LOCAL_TIMEOUT;
2375 CNETERR("Timeout sending data to %s (%pISp) the network or that node may be down.\n",
2376 libcfs_idstr(&peer_ni->ksnp_id),
2377 &conn->ksnc_peeraddr);
2378 spin_unlock_bh(&sched->kss_lock);
2381 spin_unlock_bh(&sched->kss_lock);
2388 ksocknal_flush_stale_txs(struct ksock_peer_ni *peer_ni)
2390 struct ksock_tx *tx;
2391 LIST_HEAD(stale_txs);
2393 write_lock_bh(&ksocknal_data.ksnd_global_lock);
2395 while ((tx = list_first_entry_or_null(&peer_ni->ksnp_tx_queue,
2397 tx_list)) != NULL) {
2398 if (ktime_get_seconds() < tx->tx_deadline)
2401 tx->tx_hstatus = LNET_MSG_STATUS_LOCAL_TIMEOUT;
2403 list_move_tail(&tx->tx_list, &stale_txs);
2406 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2408 ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT);
2412 ksocknal_send_keepalive_locked(struct ksock_peer_ni *peer_ni)
2413 __must_hold(&ksocknal_data.ksnd_global_lock)
2415 struct ksock_sched *sched;
2416 struct ksock_conn *conn;
2417 struct ksock_tx *tx;
2419 /* last_alive will be updated by create_conn */
2420 if (list_empty(&peer_ni->ksnp_conns))
2423 if (peer_ni->ksnp_proto != &ksocknal_protocol_v3x &&
2424 peer_ni->ksnp_proto != &ksocknal_protocol_v4x)
2427 if (*ksocknal_tunables.ksnd_keepalive <= 0 ||
2428 ktime_get_seconds() < peer_ni->ksnp_last_alive +
2429 *ksocknal_tunables.ksnd_keepalive)
2432 if (ktime_get_seconds() < peer_ni->ksnp_send_keepalive)
2435 /* retry 10 secs later, so we wouldn't put pressure
2436 * on this peer_ni if we failed to send keepalive this time */
2437 peer_ni->ksnp_send_keepalive = ktime_get_seconds() + 10;
2439 conn = ksocknal_find_conn_locked(peer_ni, NULL, 1);
2441 sched = conn->ksnc_scheduler;
2443 spin_lock_bh(&sched->kss_lock);
2444 if (!list_empty(&conn->ksnc_tx_queue)) {
2445 spin_unlock_bh(&sched->kss_lock);
2446 /* there is an queued ACK, don't need keepalive */
2450 spin_unlock_bh(&sched->kss_lock);
2453 read_unlock(&ksocknal_data.ksnd_global_lock);
2455 /* cookie = 1 is reserved for keepalive PING */
2456 tx = ksocknal_alloc_tx_noop(1, 1);
2458 read_lock(&ksocknal_data.ksnd_global_lock);
2462 if (ksocknal_launch_packet(peer_ni->ksnp_ni, tx, &peer_ni->ksnp_id)
2464 read_lock(&ksocknal_data.ksnd_global_lock);
2468 ksocknal_free_tx(tx);
2469 read_lock(&ksocknal_data.ksnd_global_lock);
2476 ksocknal_check_peer_timeouts(int idx)
2478 struct hlist_head *peers = &ksocknal_data.ksnd_peers[idx];
2479 struct ksock_peer_ni *peer_ni;
2480 struct ksock_conn *conn;
2481 struct ksock_tx *tx;
2484 /* NB. We expect to have a look at all the peers and not find any
2485 * connections to time out, so we just use a shared lock while we
2488 read_lock(&ksocknal_data.ksnd_global_lock);
2490 hlist_for_each_entry(peer_ni, peers, ksnp_list) {
2491 struct ksock_tx *tx_stale;
2492 time64_t deadline = 0;
2496 if (ksocknal_send_keepalive_locked(peer_ni) != 0) {
2497 read_unlock(&ksocknal_data.ksnd_global_lock);
2501 conn = ksocknal_find_timed_out_conn(peer_ni);
2504 read_unlock(&ksocknal_data.ksnd_global_lock);
2506 ksocknal_close_conn_and_siblings(conn, -ETIMEDOUT);
2508 /* NB we won't find this one again, but we can't
2509 * just proceed with the next peer_ni, since we dropped
2510 * ksnd_global_lock and it might be dead already!
2512 ksocknal_conn_decref(conn);
2516 /* we can't process stale txs right here because we're
2517 * holding only shared lock
2519 tx = list_first_entry_or_null(&peer_ni->ksnp_tx_queue,
2520 struct ksock_tx, tx_list);
2521 if (tx && ktime_get_seconds() >= tx->tx_deadline) {
2522 ksocknal_peer_addref(peer_ni);
2523 read_unlock(&ksocknal_data.ksnd_global_lock);
2525 ksocknal_flush_stale_txs(peer_ni);
2527 ksocknal_peer_decref(peer_ni);
2531 if (list_empty(&peer_ni->ksnp_zc_req_list))
2535 spin_lock(&peer_ni->ksnp_lock);
2536 list_for_each_entry(tx, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
2537 if (ktime_get_seconds() < tx->tx_deadline)
2539 /* ignore the TX if connection is being closed */
2540 if (tx->tx_conn->ksnc_closing)
2543 if (tx_stale == NULL)
2547 if (tx_stale == NULL) {
2548 spin_unlock(&peer_ni->ksnp_lock);
2552 deadline = tx_stale->tx_deadline;
2553 resid = tx_stale->tx_resid;
2554 conn = tx_stale->tx_conn;
2555 ksocknal_conn_addref(conn);
2557 spin_unlock(&peer_ni->ksnp_lock);
2558 read_unlock(&ksocknal_data.ksnd_global_lock);
2560 CERROR("Total %d stale ZC_REQs for peer_ni %s detected; the "
2561 "oldest(%p) timed out %lld secs ago, "
2562 "resid: %d, wmem: %d\n",
2563 n, libcfs_nidstr(&peer_ni->ksnp_id.nid), tx_stale,
2564 ktime_get_seconds() - deadline,
2565 resid, conn->ksnc_sock->sk->sk_wmem_queued);
2567 ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
2568 ksocknal_conn_decref(conn);
2572 read_unlock(&ksocknal_data.ksnd_global_lock);
2575 int ksocknal_reaper(void *arg)
2577 wait_queue_entry_t wait;
2578 struct ksock_conn *conn;
2579 struct ksock_sched *sched;
2580 LIST_HEAD(enomem_conns);
2585 time64_t deadline = ktime_get_seconds();
2589 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2591 while (!ksocknal_data.ksnd_shuttingdown) {
2592 conn = list_first_entry_or_null(&ksocknal_data.ksnd_deathrow_conns,
2593 struct ksock_conn, ksnc_list);
2595 list_del(&conn->ksnc_list);
2597 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2599 ksocknal_terminate_conn(conn);
2600 ksocknal_conn_decref(conn);
2602 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2606 conn = list_first_entry_or_null(&ksocknal_data.ksnd_zombie_conns,
2607 struct ksock_conn, ksnc_list);
2609 list_del(&conn->ksnc_list);
2611 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2613 ksocknal_destroy_conn(conn);
2615 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2619 list_splice_init(&ksocknal_data.ksnd_enomem_conns,
2622 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2624 /* reschedule all the connections that stalled with ENOMEM... */
2626 while ((conn = list_first_entry_or_null(&enomem_conns,
2628 ksnc_tx_list)) != NULL) {
2629 list_del(&conn->ksnc_tx_list);
2631 sched = conn->ksnc_scheduler;
2633 spin_lock_bh(&sched->kss_lock);
2635 LASSERT(conn->ksnc_tx_scheduled);
2636 conn->ksnc_tx_ready = 1;
2637 list_add_tail(&conn->ksnc_tx_list,
2638 &sched->kss_tx_conns);
2639 wake_up(&sched->kss_waitq);
2641 spin_unlock_bh(&sched->kss_lock);
2645 /* careful with the jiffy wrap... */
2646 while ((timeout = deadline - ktime_get_seconds()) <= 0) {
2649 int chunk = HASH_SIZE(ksocknal_data.ksnd_peers);
2650 unsigned int lnd_timeout;
2652 /* Time to check for timeouts on a few more peers: I
2653 * do checks every 'p' seconds on a proportion of the
2654 * peer_ni table and I need to check every connection
2655 * 'n' times within a timeout interval, to ensure I
2656 * detect a timeout on any connection within (n+1)/n
2657 * times the timeout interval.
2660 lnd_timeout = ksocknal_timeout();
2661 if (lnd_timeout > n * p)
2662 chunk = (chunk * n * p) / lnd_timeout;
2666 for (i = 0; i < chunk; i++) {
2667 ksocknal_check_peer_timeouts(peer_index);
2668 peer_index = (peer_index + 1) %
2669 HASH_SIZE(ksocknal_data.ksnd_peers);
2675 if (nenomem_conns != 0) {
2676 /* Reduce my timeout if I rescheduled ENOMEM conns.
2677 * This also prevents me getting woken immediately
2678 * if any go back on my enomem list. */
2679 timeout = SOCKNAL_ENOMEM_RETRY;
2681 ksocknal_data.ksnd_reaper_waketime = ktime_get_seconds() +
2684 set_current_state(TASK_INTERRUPTIBLE);
2685 add_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2687 if (!ksocknal_data.ksnd_shuttingdown &&
2688 list_empty(&ksocknal_data.ksnd_deathrow_conns) &&
2689 list_empty(&ksocknal_data.ksnd_zombie_conns))
2690 schedule_timeout(cfs_time_seconds(timeout));
2692 set_current_state(TASK_RUNNING);
2693 remove_wait_queue(&ksocknal_data.ksnd_reaper_waitq, &wait);
2695 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
2698 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
2700 ksocknal_thread_fini();