4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lnet/klnds/socklnd/socklnd.c
33 * Author: Zach Brown <zab@zabbo.net>
34 * Author: Peter J. Braam <braam@clusterfs.com>
35 * Author: Phil Schwan <phil@clusterfs.com>
36 * Author: Eric Barton <eric@bartonsoftware.com>
39 #include <linux/ethtool.h>
40 #include <linux/inetdevice.h>
41 #include <linux/kernel.h>
42 #include <linux/sunrpc/addr.h>
43 #include <net/addrconf.h>
46 static const struct lnet_lnd the_ksocklnd;
47 struct ksock_nal_data ksocknal_data;
49 static int ksocknal_ip2index(struct sockaddr *addr, struct lnet_ni *ni,
52 struct net_device *dev;
54 DECLARE_CONST_IN_IFADDR(ifa);
58 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
62 for_each_netdev(ni->ni_net_ns, dev) {
63 int flags = dev_get_flags(dev);
64 struct in_device *in_dev;
66 if (flags & IFF_LOOPBACK) /* skip the loopback IF */
69 if (!(flags & IFF_UP))
72 switch (addr->sa_family) {
74 in_dev = __in_dev_get_rcu(dev);
78 in_dev_for_each_ifa_rcu(ifa, in_dev) {
80 ((struct sockaddr_in *)addr)->sin_addr.s_addr)
85 #if IS_ENABLED(CONFIG_IPV6)
87 struct inet6_dev *in6_dev;
88 const struct inet6_ifaddr *ifa6;
89 struct sockaddr_in6 *addr6 = (struct sockaddr_in6*)addr;
91 in6_dev = __in6_dev_get(dev);
95 list_for_each_entry_rcu(ifa6, &in6_dev->addr_list, if_list) {
96 if (ipv6_addr_cmp(&ifa6->addr,
97 &addr6->sin6_addr) == 0)
102 #endif /* IS_ENABLED(CONFIG_IPV6) */
113 ((dev->reg_state == NETREG_UNREGISTERING) ||
114 (dev->operstate != IF_OPER_UP)) ||
115 (lnet_get_link_status(dev) == 0))
121 static struct ksock_conn_cb *
122 ksocknal_create_conn_cb(struct sockaddr *addr)
124 struct ksock_conn_cb *conn_cb;
126 LIBCFS_ALLOC(conn_cb, sizeof(*conn_cb));
130 refcount_set(&conn_cb->ksnr_refcount, 1);
131 conn_cb->ksnr_peer = NULL;
132 conn_cb->ksnr_retry_interval = 0; /* OK to connect at any time */
133 rpc_copy_addr((struct sockaddr *)&conn_cb->ksnr_addr, addr);
134 rpc_set_port((struct sockaddr *)&conn_cb->ksnr_addr,
136 conn_cb->ksnr_scheduled = 0;
137 conn_cb->ksnr_connecting = 0;
138 conn_cb->ksnr_connected = 0;
139 conn_cb->ksnr_deleted = 0;
140 conn_cb->ksnr_conn_count = 0;
141 conn_cb->ksnr_ctrl_conn_count = 0;
142 conn_cb->ksnr_blki_conn_count = 0;
143 conn_cb->ksnr_blko_conn_count = 0;
144 conn_cb->ksnr_max_conns = 0;
145 conn_cb->ksnr_busy_retry_count = 0;
151 ksocknal_destroy_conn_cb(struct ksock_conn_cb *conn_cb)
153 LASSERT(refcount_read(&conn_cb->ksnr_refcount) == 0);
155 if (conn_cb->ksnr_peer)
156 ksocknal_peer_decref(conn_cb->ksnr_peer);
158 LIBCFS_FREE(conn_cb, sizeof(*conn_cb));
161 static struct ksock_peer_ni *
162 ksocknal_create_peer(struct lnet_ni *ni, struct lnet_processid *id)
164 int cpt = lnet_nid2cpt(&id->nid, ni);
165 struct ksock_net *net = ni->ni_data;
166 struct ksock_peer_ni *peer_ni;
168 LASSERT(!LNET_NID_IS_ANY(&id->nid));
169 LASSERT(id->pid != LNET_PID_ANY);
170 LASSERT(!in_interrupt());
172 if (!atomic_inc_unless_negative(&net->ksnn_npeers)) {
173 CERROR("Can't create peer_ni: network shutdown\n");
174 return ERR_PTR(-ESHUTDOWN);
177 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
179 atomic_dec(&net->ksnn_npeers);
180 return ERR_PTR(-ENOMEM);
183 peer_ni->ksnp_ni = ni;
184 peer_ni->ksnp_id = *id;
185 refcount_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
186 peer_ni->ksnp_closing = 0;
187 peer_ni->ksnp_accepting = 0;
188 peer_ni->ksnp_proto = NULL;
189 peer_ni->ksnp_last_alive = 0;
190 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
191 peer_ni->ksnp_conn_cb = NULL;
193 INIT_LIST_HEAD(&peer_ni->ksnp_conns);
194 INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
195 INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
196 spin_lock_init(&peer_ni->ksnp_lock);
202 ksocknal_destroy_peer(struct ksock_peer_ni *peer_ni)
204 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
206 CDEBUG (D_NET, "peer_ni %s %p deleted\n",
207 libcfs_idstr(&peer_ni->ksnp_id), peer_ni);
209 LASSERT(refcount_read(&peer_ni->ksnp_refcount) == 0);
210 LASSERT(peer_ni->ksnp_accepting == 0);
211 LASSERT(list_empty(&peer_ni->ksnp_conns));
212 LASSERT(peer_ni->ksnp_conn_cb == NULL);
213 LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
214 LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
216 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
218 /* NB a peer_ni's connections and conn_cb keep a reference on their
219 * peer_ni until they are destroyed, so we can be assured that _all_
220 * state to do with this peer_ni has been cleaned up when its refcount
223 if (atomic_dec_and_test(&net->ksnn_npeers))
224 wake_up_var(&net->ksnn_npeers);
227 struct ksock_peer_ni *
228 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_processid *id)
230 struct ksock_peer_ni *peer_ni;
231 unsigned long hash = nidhash(&id->nid);
233 hash_for_each_possible(ksocknal_data.ksnd_peers, peer_ni,
235 LASSERT(!peer_ni->ksnp_closing);
237 if (peer_ni->ksnp_ni != ni)
240 if (!nid_same(&peer_ni->ksnp_id.nid, &id->nid) ||
241 peer_ni->ksnp_id.pid != id->pid)
244 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
245 peer_ni, libcfs_idstr(id),
246 refcount_read(&peer_ni->ksnp_refcount));
252 struct ksock_peer_ni *
253 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_processid *id)
255 struct ksock_peer_ni *peer_ni;
257 read_lock(&ksocknal_data.ksnd_global_lock);
258 peer_ni = ksocknal_find_peer_locked(ni, id);
259 if (peer_ni != NULL) /* +1 ref for caller? */
260 ksocknal_peer_addref(peer_ni);
261 read_unlock(&ksocknal_data.ksnd_global_lock);
267 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
269 LASSERT(list_empty(&peer_ni->ksnp_conns));
270 LASSERT(peer_ni->ksnp_conn_cb == NULL);
271 LASSERT(!peer_ni->ksnp_closing);
272 peer_ni->ksnp_closing = 1;
273 hlist_del(&peer_ni->ksnp_list);
274 /* lose peerlist's ref */
275 ksocknal_peer_decref(peer_ni);
280 ksocknal_dump_peer_debug_info(struct ksock_peer_ni *peer_ni)
282 struct ksock_conn *conn;
283 struct list_head *ctmp;
284 struct list_head *txtmp;
288 list_for_each(ctmp, &peer_ni->ksnp_conns) {
289 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
291 if (!list_empty(&conn->ksnc_tx_queue))
292 list_for_each(txtmp, &conn->ksnc_tx_queue) txcount++;
294 CDEBUG(D_CONSOLE, "Conn %d [type, closing, crefcnt, srefcnt]: %d, %d, %d, %d\n",
298 refcount_read(&conn->ksnc_conn_refcount),
299 refcount_read(&conn->ksnc_sock_refcount));
300 CDEBUG(D_CONSOLE, "Conn %d rx [scheduled, ready, state]: %d, %d, %d\n",
302 conn->ksnc_rx_scheduled,
304 conn->ksnc_rx_state);
305 CDEBUG(D_CONSOLE, "Conn %d tx [txqcnt, scheduled, last_post, ready, deadline]: %d, %d, %lld, %d, %lld\n",
308 conn->ksnc_tx_scheduled,
309 conn->ksnc_tx_last_post,
311 conn->ksnc_rx_deadline);
313 if (conn->ksnc_scheduler)
314 CDEBUG(D_CONSOLE, "Conn %d sched [nconns, cpt]: %d, %d\n",
316 conn->ksnc_scheduler->kss_nconns,
317 conn->ksnc_scheduler->kss_cpt);
325 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
326 struct lnet_processid *id, __u32 *myip, __u32 *peer_ip,
327 int *port, int *conn_count, int *share_count)
329 struct ksock_peer_ni *peer_ni;
330 struct ksock_conn_cb *conn_cb;
333 struct ksock_net *net;
335 read_lock(&ksocknal_data.ksnd_global_lock);
337 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
339 if (peer_ni->ksnp_ni != ni)
344 *id = peer_ni->ksnp_id;
345 conn_cb = peer_ni->ksnp_conn_cb;
346 if (conn_cb == NULL) {
354 ksocknal_dump_peer_debug_info(peer_ni);
356 if (conn_cb->ksnr_addr.ss_family == AF_INET) {
357 struct sockaddr_in *sa =
358 (void *)&conn_cb->ksnr_addr;
360 rc = choose_ipv4_src(myip,
361 net->ksnn_interface.ksni_index,
362 ntohl(sa->sin_addr.s_addr),
364 *peer_ip = ntohl(sa->sin_addr.s_addr);
365 *port = ntohs(sa->sin_port);
369 *peer_ip = 0xFFFFFFFF;
373 *conn_count = conn_cb->ksnr_conn_count;
378 read_unlock(&ksocknal_data.ksnd_global_lock);
383 ksocknal_get_conn_count_by_type(struct ksock_conn_cb *conn_cb,
386 unsigned int count = 0;
389 case SOCKLND_CONN_CONTROL:
390 count = conn_cb->ksnr_ctrl_conn_count;
392 case SOCKLND_CONN_BULK_IN:
393 count = conn_cb->ksnr_blki_conn_count;
395 case SOCKLND_CONN_BULK_OUT:
396 count = conn_cb->ksnr_blko_conn_count;
398 case SOCKLND_CONN_ANY:
399 count = conn_cb->ksnr_conn_count;
410 ksocknal_get_conns_per_peer(struct ksock_peer_ni *peer_ni)
412 struct lnet_ni *ni = peer_ni->ksnp_ni;
413 struct lnet_ioctl_config_socklnd_tunables *tunables;
417 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
419 return tunables->lnd_conns_per_peer;
423 ksocknal_incr_conn_count(struct ksock_conn_cb *conn_cb,
426 conn_cb->ksnr_conn_count++;
428 /* check if all connections of the given type got created */
430 case SOCKLND_CONN_CONTROL:
431 conn_cb->ksnr_ctrl_conn_count++;
432 /* there's a single control connection per peer,
433 * two in case of loopback
435 conn_cb->ksnr_connected |= BIT(type);
437 case SOCKLND_CONN_BULK_IN:
438 conn_cb->ksnr_blki_conn_count++;
439 if (conn_cb->ksnr_blki_conn_count >= conn_cb->ksnr_max_conns)
440 conn_cb->ksnr_connected |= BIT(type);
442 case SOCKLND_CONN_BULK_OUT:
443 conn_cb->ksnr_blko_conn_count++;
444 if (conn_cb->ksnr_blko_conn_count >= conn_cb->ksnr_max_conns)
445 conn_cb->ksnr_connected |= BIT(type);
447 case SOCKLND_CONN_ANY:
448 if (conn_cb->ksnr_conn_count >= conn_cb->ksnr_max_conns)
449 conn_cb->ksnr_connected |= BIT(type);
456 CDEBUG(D_NET, "Add conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
457 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
462 ksocknal_decr_conn_count(struct ksock_conn_cb *conn_cb,
465 conn_cb->ksnr_conn_count--;
467 /* check if all connections of the given type got created */
469 case SOCKLND_CONN_CONTROL:
470 conn_cb->ksnr_ctrl_conn_count--;
471 /* there's a single control connection per peer,
472 * two in case of loopback
474 if (conn_cb->ksnr_ctrl_conn_count == 0)
475 conn_cb->ksnr_connected &= ~BIT(type);
477 case SOCKLND_CONN_BULK_IN:
478 conn_cb->ksnr_blki_conn_count--;
479 if (conn_cb->ksnr_blki_conn_count < conn_cb->ksnr_max_conns)
480 conn_cb->ksnr_connected &= ~BIT(type);
482 case SOCKLND_CONN_BULK_OUT:
483 conn_cb->ksnr_blko_conn_count--;
484 if (conn_cb->ksnr_blko_conn_count < conn_cb->ksnr_max_conns)
485 conn_cb->ksnr_connected &= ~BIT(type);
487 case SOCKLND_CONN_ANY:
488 if (conn_cb->ksnr_conn_count < conn_cb->ksnr_max_conns)
489 conn_cb->ksnr_connected &= ~BIT(type);
496 CDEBUG(D_NET, "Del conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
497 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
501 ksocknal_associate_cb_conn_locked(struct ksock_conn_cb *conn_cb,
502 struct ksock_conn *conn)
504 int type = conn->ksnc_type;
506 conn->ksnc_conn_cb = conn_cb;
507 ksocknal_conn_cb_addref(conn_cb);
508 ksocknal_incr_conn_count(conn_cb, type);
510 /* Successful connection => further attempts can
511 * proceed immediately
513 conn_cb->ksnr_retry_interval = 0;
517 ksocknal_add_conn_cb_locked(struct ksock_peer_ni *peer_ni,
518 struct ksock_conn_cb *conn_cb)
520 struct ksock_conn *conn;
521 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
523 LASSERT(!peer_ni->ksnp_closing);
524 LASSERT(!conn_cb->ksnr_peer);
525 LASSERT(!conn_cb->ksnr_scheduled);
526 LASSERT(!conn_cb->ksnr_connecting);
527 LASSERT(conn_cb->ksnr_connected == 0);
529 conn_cb->ksnr_peer = peer_ni;
530 ksocknal_peer_addref(peer_ni);
532 /* peer_ni's route list takes over my ref on 'route' */
533 peer_ni->ksnp_conn_cb = conn_cb;
534 net->ksnn_interface.ksni_nroutes++;
536 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
537 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
538 (struct sockaddr *)&conn_cb->ksnr_addr))
540 CDEBUG(D_NET, "call ksocknal_associate_cb_conn_locked\n");
541 ksocknal_associate_cb_conn_locked(conn_cb, conn);
542 /* keep going (typed conns) */
547 ksocknal_del_conn_cb_locked(struct ksock_conn_cb *conn_cb)
549 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
550 struct ksock_conn *conn;
551 struct ksock_conn *cnxt;
552 struct ksock_net *net;
554 LASSERT(!conn_cb->ksnr_deleted);
556 /* Close associated conns */
557 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
558 if (conn->ksnc_conn_cb != conn_cb)
561 ksocknal_close_conn_locked(conn, 0);
564 net = (struct ksock_net *)(peer_ni->ksnp_ni->ni_data);
565 net->ksnn_interface.ksni_nroutes--;
566 LASSERT(net->ksnn_interface.ksni_nroutes >= 0);
568 conn_cb->ksnr_deleted = 1;
569 ksocknal_conn_cb_decref(conn_cb); /* drop peer_ni's ref */
570 peer_ni->ksnp_conn_cb = NULL;
572 if (list_empty(&peer_ni->ksnp_conns)) {
573 /* I've just removed the last route to a peer_ni with no active
576 ksocknal_unlink_peer_locked(peer_ni);
581 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_processid *id,
582 struct sockaddr *addr)
584 struct ksock_peer_ni *peer_ni;
585 struct ksock_peer_ni *peer2;
586 struct ksock_conn_cb *conn_cb;
588 if (LNET_NID_IS_ANY(&id->nid) ||
589 id->pid == LNET_PID_ANY)
592 /* Have a brand new peer_ni ready... */
593 peer_ni = ksocknal_create_peer(ni, id);
595 return PTR_ERR(peer_ni);
597 conn_cb = ksocknal_create_conn_cb(addr);
599 ksocknal_peer_decref(peer_ni);
603 write_lock_bh(&ksocknal_data.ksnd_global_lock);
605 /* always called with a ref on ni, so shutdown can't have started */
606 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
609 peer2 = ksocknal_find_peer_locked(ni, id);
611 ksocknal_peer_decref(peer_ni);
614 /* peer_ni table takes my ref on peer_ni */
615 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
619 if (peer_ni->ksnp_conn_cb) {
620 ksocknal_conn_cb_decref(conn_cb);
622 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
623 /* Remember conns_per_peer setting at the time
624 * of connection initiation. It will define the
625 * max number of conns per type for this conn_cb
628 conn_cb->ksnr_max_conns = ksocknal_get_conns_per_peer(peer_ni);
631 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
637 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni)
639 struct ksock_conn *conn;
640 struct ksock_conn *cnxt;
641 struct ksock_conn_cb *conn_cb;
643 LASSERT(!peer_ni->ksnp_closing);
645 /* Extra ref prevents peer_ni disappearing until I'm done with it */
646 ksocknal_peer_addref(peer_ni);
647 conn_cb = peer_ni->ksnp_conn_cb;
649 ksocknal_del_conn_cb_locked(conn_cb);
651 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
653 ksocknal_close_conn_locked(conn, 0);
655 ksocknal_peer_decref(peer_ni);
656 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
660 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_processid *id)
663 struct hlist_node *pnxt;
664 struct ksock_peer_ni *peer_ni;
670 write_lock_bh(&ksocknal_data.ksnd_global_lock);
672 if (id && !LNET_NID_IS_ANY(&id->nid)) {
673 lo = hash_min(nidhash(&id->nid),
674 HASH_BITS(ksocknal_data.ksnd_peers));
678 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
681 for (i = lo; i <= hi; i++) {
682 hlist_for_each_entry_safe(peer_ni, pnxt,
683 &ksocknal_data.ksnd_peers[i],
685 if (peer_ni->ksnp_ni != ni)
688 if (!((!id || LNET_NID_IS_ANY(&id->nid) ||
689 nid_same(&peer_ni->ksnp_id.nid, &id->nid)) &&
690 (!id || id->pid == LNET_PID_ANY ||
691 peer_ni->ksnp_id.pid == id->pid)))
694 ksocknal_peer_addref(peer_ni); /* a ref for me... */
696 ksocknal_del_peer_locked(peer_ni);
698 if (peer_ni->ksnp_closing &&
699 !list_empty(&peer_ni->ksnp_tx_queue)) {
700 LASSERT(list_empty(&peer_ni->ksnp_conns));
701 LASSERT(peer_ni->ksnp_conn_cb == NULL);
703 list_splice_init(&peer_ni->ksnp_tx_queue,
707 ksocknal_peer_decref(peer_ni); /* ...till here */
709 rc = 0; /* matched! */
713 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
715 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
720 static struct ksock_conn *
721 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
723 struct ksock_peer_ni *peer_ni;
724 struct ksock_conn *conn;
727 read_lock(&ksocknal_data.ksnd_global_lock);
729 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
730 LASSERT(!peer_ni->ksnp_closing);
732 if (peer_ni->ksnp_ni != ni)
735 list_for_each_entry(conn, &peer_ni->ksnp_conns,
740 ksocknal_conn_addref(conn);
741 read_unlock(&ksocknal_data.ksnd_global_lock);
746 read_unlock(&ksocknal_data.ksnd_global_lock);
750 static struct ksock_sched *
751 ksocknal_choose_scheduler_locked(unsigned int cpt)
753 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
756 if (sched->kss_nthreads == 0) {
757 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
758 if (sched->kss_nthreads > 0) {
759 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
760 cpt, sched->kss_cpt);
771 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
773 struct ksock_connreq *cr;
775 struct sockaddr_storage peer;
777 rc = lnet_sock_getaddr(sock, true, &peer);
779 CERROR("Can't determine new connection's address\n");
783 LIBCFS_ALLOC(cr, sizeof(*cr));
785 LCONSOLE_ERROR_MSG(0x12f,
786 "Dropping connection request from %pISc: memory exhausted\n",
793 cr->ksncr_sock = sock;
795 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
797 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
798 wake_up(&ksocknal_data.ksnd_connd_waitq);
800 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
804 static const struct ln_key_list ksocknal_tunables_keys = {
805 .lkl_maxattr = LNET_NET_SOCKLND_TUNABLES_ATTR_MAX,
807 [LNET_NET_SOCKLND_TUNABLES_ATTR_CONNS_PER_PEER] = {
808 .lkp_value = "conns_per_peer",
809 .lkp_data_type = NLA_S32
815 ksocknal_nl_set(int cmd, struct nlattr *attr, int type, void *data)
817 struct lnet_lnd_tunables *tunables = data;
820 if (cmd != LNET_CMD_NETS)
823 if (type != LNET_NET_SOCKLND_TUNABLES_ATTR_CONNS_PER_PEER ||
824 nla_type(attr) != LN_SCALAR_ATTR_INT_VALUE)
827 num = nla_get_s64(attr);
828 clamp_t(s64, num, 1, 127);
829 tunables->lnd_tun_u.lnd_sock.lnd_conns_per_peer = num;
835 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
838 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
839 return conn_cb->ksnr_connecting;
844 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
845 struct socket *sock, int type)
847 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
849 struct lnet_processid peerid;
851 struct ksock_conn *conn;
852 struct ksock_conn *conn2;
853 struct ksock_peer_ni *peer_ni = NULL;
854 struct ksock_peer_ni *peer2;
855 struct ksock_sched *sched;
856 struct ksock_hello_msg *hello;
859 struct ksock_tx *txtmp;
866 active = (conn_cb != NULL);
868 LASSERT(active == (type != SOCKLND_CONN_NONE));
870 LIBCFS_ALLOC(conn, sizeof(*conn));
876 conn->ksnc_peer = NULL;
877 conn->ksnc_conn_cb = NULL;
878 conn->ksnc_sock = sock;
879 /* 2 ref, 1 for conn, another extra ref prevents socket
880 * being closed before establishment of connection */
881 refcount_set(&conn->ksnc_sock_refcount, 2);
882 conn->ksnc_type = type;
883 ksocknal_lib_save_callback(sock, conn);
884 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
886 conn->ksnc_rx_ready = 0;
887 conn->ksnc_rx_scheduled = 0;
889 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
890 conn->ksnc_tx_ready = 0;
891 conn->ksnc_tx_scheduled = 0;
892 conn->ksnc_tx_carrier = NULL;
893 atomic_set (&conn->ksnc_tx_nob, 0);
895 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
896 kshm_ips[LNET_INTERFACES_NUM]));
902 /* stash conn's local and remote addrs */
903 rc = ksocknal_lib_get_conn_addrs(conn);
907 /* Find out/confirm peer_ni's NID and connection type and get the
908 * vector of interfaces she's willing to let me connect to.
909 * Passive connections use the listener timeout since the peer_ni sends
914 peer_ni = conn_cb->ksnr_peer;
915 LASSERT(ni == peer_ni->ksnp_ni);
917 /* Active connection sends HELLO eagerly */
918 hello->kshm_nips = 0;
919 peerid = peer_ni->ksnp_id;
921 write_lock_bh(global_lock);
922 conn->ksnc_proto = peer_ni->ksnp_proto;
923 write_unlock_bh(global_lock);
925 if (conn->ksnc_proto == NULL) {
926 conn->ksnc_proto = &ksocknal_protocol_v3x;
927 #if SOCKNAL_VERSION_DEBUG
928 if (*ksocknal_tunables.ksnd_protocol == 2)
929 conn->ksnc_proto = &ksocknal_protocol_v2x;
930 else if (*ksocknal_tunables.ksnd_protocol == 1)
931 conn->ksnc_proto = &ksocknal_protocol_v1x;
935 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
939 peerid.nid = LNET_ANY_NID;
940 peerid.pid = LNET_PID_ANY;
942 /* Passive, get protocol from peer_ni */
943 conn->ksnc_proto = NULL;
946 rc = ksocknal_recv_hello(ni, conn, hello, &peerid, &incarnation);
950 LASSERT(rc == 0 || active);
951 LASSERT(conn->ksnc_proto != NULL);
952 LASSERT(!LNET_NID_IS_ANY(&peerid.nid));
954 cpt = lnet_nid2cpt(&peerid.nid, ni);
957 ksocknal_peer_addref(peer_ni);
958 write_lock_bh(global_lock);
960 peer_ni = ksocknal_create_peer(ni, &peerid);
961 if (IS_ERR(peer_ni)) {
962 rc = PTR_ERR(peer_ni);
966 write_lock_bh(global_lock);
968 /* called with a ref on ni, so shutdown can't have started */
969 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
971 peer2 = ksocknal_find_peer_locked(ni, &peerid);
973 /* NB this puts an "empty" peer_ni in the peer_ni
974 * table (which takes my ref) */
975 hash_add(ksocknal_data.ksnd_peers,
976 &peer_ni->ksnp_list, nidhash(&peerid.nid));
978 ksocknal_peer_decref(peer_ni);
983 ksocknal_peer_addref(peer_ni);
984 peer_ni->ksnp_accepting++;
986 /* Am I already connecting to this guy? Resolve in
987 * favour of higher NID...
989 if (memcmp(&peerid.nid, &ni->ni_nid, sizeof(peerid.nid)) < 0 &&
990 ksocknal_connecting(peer_ni->ksnp_conn_cb,
991 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
993 warn = "connection race resolution";
998 if (peer_ni->ksnp_closing ||
999 (active && conn_cb->ksnr_deleted)) {
1000 /* peer_ni/conn_cb got closed under me */
1002 warn = "peer_ni/conn_cb removed";
1006 if (peer_ni->ksnp_proto == NULL) {
1007 /* Never connected before.
1008 * NB recv_hello may have returned EPROTO to signal my peer_ni
1009 * wants a different protocol than the one I asked for.
1011 LASSERT(list_empty(&peer_ni->ksnp_conns));
1013 peer_ni->ksnp_proto = conn->ksnc_proto;
1014 peer_ni->ksnp_incarnation = incarnation;
1017 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1018 peer_ni->ksnp_incarnation != incarnation) {
1019 /* peer_ni rebooted or I've got the wrong protocol version */
1020 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
1022 peer_ni->ksnp_proto = NULL;
1024 warn = peer_ni->ksnp_incarnation != incarnation ?
1025 "peer_ni rebooted" :
1026 "wrong proto version";
1036 warn = "lost conn race";
1039 warn = "retry with different protocol version";
1043 /* Refuse to duplicate an existing connection, unless this is a
1044 * loopback connection */
1045 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1046 (struct sockaddr *)&conn->ksnc_myaddr)) {
1047 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1049 (struct sockaddr *)&conn2->ksnc_peeraddr,
1050 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1052 (struct sockaddr *)&conn2->ksnc_myaddr,
1053 (struct sockaddr *)&conn->ksnc_myaddr) ||
1054 conn2->ksnc_type != conn->ksnc_type)
1058 /* If max conns per type is not registered in conn_cb
1059 * as ksnr_max_conns, use ni's conns_per_peer
1061 if ((peer_ni->ksnp_conn_cb &&
1062 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1063 (!peer_ni->ksnp_conn_cb &&
1064 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1067 /* Reply on a passive connection attempt so the peer_ni
1068 * realises we're connected.
1078 /* If the connection created by this route didn't bind to the IP
1079 * address the route connected to, the connection/route matching
1080 * code below probably isn't going to work.
1083 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1084 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1085 CERROR("Route %s %pISc connected to %pISc\n",
1086 libcfs_idstr(&peer_ni->ksnp_id),
1087 &conn_cb->ksnr_addr,
1088 &conn->ksnc_peeraddr);
1091 /* Search for a conn_cb corresponding to the new connection and
1092 * create an association. This allows incoming connections created
1093 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1094 * continually create duplicate conn_cbs.
1096 conn_cb = peer_ni->ksnp_conn_cb;
1098 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1099 (struct sockaddr *)&conn_cb->ksnr_addr))
1100 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1102 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1103 peer_ni->ksnp_last_alive = ktime_get_seconds();
1104 peer_ni->ksnp_send_keepalive = 0;
1105 peer_ni->ksnp_error = 0;
1107 sched = ksocknal_choose_scheduler_locked(cpt);
1109 CERROR("no schedulers available. node is unhealthy\n");
1113 * The cpt might have changed if we ended up selecting a non cpt
1114 * native scheduler. So use the scheduler's cpt instead.
1116 cpt = sched->kss_cpt;
1117 sched->kss_nconns++;
1118 conn->ksnc_scheduler = sched;
1120 conn->ksnc_tx_last_post = ktime_get_seconds();
1121 /* Set the deadline for the outgoing HELLO to drain */
1122 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1123 conn->ksnc_tx_deadline = ktime_get_seconds() +
1125 smp_mb(); /* order with adding to peer_ni's conn list */
1127 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1128 ksocknal_conn_addref(conn);
1130 ksocknal_new_packet(conn, 0);
1132 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1134 /* Take packets blocking for this connection. */
1135 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1136 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1140 list_del(&tx->tx_list);
1141 ksocknal_queue_tx_locked(tx, conn);
1144 write_unlock_bh(global_lock);
1145 /* We've now got a new connection. Any errors from here on are just
1146 * like "normal" comms errors and we close the connection normally.
1147 * NB (a) we still have to send the reply HELLO for passive
1149 * (b) normal I/O on the conn is blocked until I setup and call the
1153 CDEBUG(D_NET, "New conn %s p %d.x %pISc -> %pIScp"
1154 " incarnation:%lld sched[%d]\n",
1155 libcfs_idstr(&peerid), conn->ksnc_proto->pro_version,
1156 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1160 hello->kshm_nips = 0;
1161 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1164 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1165 kshm_ips[LNET_INTERFACES_NUM]));
1167 /* setup the socket AFTER I've received hello (it disables
1168 * SO_LINGER). I might call back to the acceptor who may want
1169 * to send a protocol version response and then close the
1170 * socket; this ensures the socket only tears down after the
1171 * response has been sent.
1174 rc = ksocknal_lib_setup_sock(sock);
1176 write_lock_bh(global_lock);
1178 /* NB my callbacks block while I hold ksnd_global_lock */
1179 ksocknal_lib_set_callback(sock, conn);
1182 peer_ni->ksnp_accepting--;
1184 write_unlock_bh(global_lock);
1187 write_lock_bh(global_lock);
1188 if (!conn->ksnc_closing) {
1189 /* could be closed by another thread */
1190 ksocknal_close_conn_locked(conn, rc);
1192 write_unlock_bh(global_lock);
1193 } else if (ksocknal_connsock_addref(conn) == 0) {
1194 /* Allow I/O to proceed. */
1195 ksocknal_read_callback(conn);
1196 ksocknal_write_callback(conn);
1197 ksocknal_connsock_decref(conn);
1200 ksocknal_connsock_decref(conn);
1201 ksocknal_conn_decref(conn);
1206 if (!peer_ni->ksnp_closing &&
1207 list_empty(&peer_ni->ksnp_conns) &&
1208 peer_ni->ksnp_conn_cb == NULL) {
1209 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1210 ksocknal_unlink_peer_locked(peer_ni);
1213 write_unlock_bh(global_lock);
1217 CERROR("Not creating conn %s type %d: %s\n",
1218 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1220 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1221 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1226 /* Request retry by replying with CONN_NONE
1227 * ksnc_proto has been set already
1229 conn->ksnc_type = SOCKLND_CONN_NONE;
1230 hello->kshm_nips = 0;
1231 ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1234 write_lock_bh(global_lock);
1235 peer_ni->ksnp_accepting--;
1236 write_unlock_bh(global_lock);
1240 * If we get here without an error code, just use -EALREADY.
1241 * Depending on how we got here, the error may be positive
1242 * or negative. Normalize the value for ksocknal_txlist_done().
1244 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1245 ksocknal_txlist_done(ni, &zombies, rc2);
1246 ksocknal_peer_decref(peer_ni);
1250 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1251 kshm_ips[LNET_INTERFACES_NUM]));
1253 LIBCFS_FREE(conn, sizeof(*conn));
1262 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1264 /* This just does the immmediate housekeeping, and queues the
1265 * connection for the reaper to terminate.
1266 * Caller holds ksnd_global_lock exclusively in irq context */
1267 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1268 struct ksock_conn_cb *conn_cb;
1269 struct ksock_conn *conn2;
1271 int duplicate_count = 0;
1273 LASSERT(peer_ni->ksnp_error == 0);
1274 LASSERT(!conn->ksnc_closing);
1275 conn->ksnc_closing = 1;
1277 /* ksnd_deathrow_conns takes over peer_ni's ref */
1278 list_del(&conn->ksnc_list);
1280 conn_cb = conn->ksnc_conn_cb;
1281 if (conn_cb != NULL) {
1282 /* dissociate conn from cb... */
1283 LASSERT(!conn_cb->ksnr_deleted);
1285 conn_count = ksocknal_get_conn_count_by_type(conn_cb,
1287 /* connected bit is set only if all connections
1288 * of the given type got created
1290 if (conn_count == conn_cb->ksnr_max_conns)
1291 LASSERT((conn_cb->ksnr_connected &
1292 BIT(conn->ksnc_type)) != 0);
1294 if (conn_count == 1) {
1295 list_for_each_entry(conn2, &peer_ni->ksnp_conns,
1297 if (conn2->ksnc_conn_cb == conn_cb &&
1298 conn2->ksnc_type == conn->ksnc_type)
1299 duplicate_count += 1;
1301 if (duplicate_count > 0)
1302 CERROR("Found %d duplicate conns type %d\n",
1306 ksocknal_decr_conn_count(conn_cb, conn->ksnc_type);
1308 conn->ksnc_conn_cb = NULL;
1310 /* drop conn's ref on conn_cb */
1311 ksocknal_conn_cb_decref(conn_cb);
1314 if (list_empty(&peer_ni->ksnp_conns)) {
1315 /* No more connections to this peer_ni */
1317 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1318 struct ksock_tx *tx;
1320 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1322 /* throw them to the last connection...,
1323 * these TXs will be send to /dev/null by scheduler */
1324 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1326 ksocknal_tx_prep(conn, tx);
1328 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1329 list_splice_init(&peer_ni->ksnp_tx_queue,
1330 &conn->ksnc_tx_queue);
1331 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1334 /* renegotiate protocol version */
1335 peer_ni->ksnp_proto = NULL;
1336 /* stash last conn close reason */
1337 peer_ni->ksnp_error = error;
1339 if (peer_ni->ksnp_conn_cb == NULL) {
1340 /* I've just closed last conn belonging to a
1341 * peer_ni with no connections to it
1343 ksocknal_unlink_peer_locked(peer_ni);
1347 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1349 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1350 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1352 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1356 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1358 bool notify = false;
1359 time64_t last_alive = 0;
1361 /* There has been a connection failure or comms error; but I'll only
1362 * tell LNET I think the peer_ni is dead if it's to another kernel and
1363 * there are no connections or connection attempts in existence. */
1365 read_lock(&ksocknal_data.ksnd_global_lock);
1367 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1368 list_empty(&peer_ni->ksnp_conns) &&
1369 peer_ni->ksnp_accepting == 0 &&
1370 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1372 last_alive = peer_ni->ksnp_last_alive;
1375 read_unlock(&ksocknal_data.ksnd_global_lock);
1378 lnet_notify(peer_ni->ksnp_ni,
1379 &peer_ni->ksnp_id.nid,
1380 false, false, last_alive);
1384 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1386 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1387 struct ksock_tx *tx;
1388 struct ksock_tx *tmp;
1391 /* NB safe to finalize TXs because closing of socket will
1392 * abort all buffered data */
1393 LASSERT(conn->ksnc_sock == NULL);
1395 spin_lock(&peer_ni->ksnp_lock);
1397 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1399 if (tx->tx_conn != conn)
1402 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1404 tx->tx_msg.ksm_zc_cookies[0] = 0;
1405 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1406 list_move(&tx->tx_zc_list, &zlist);
1409 spin_unlock(&peer_ni->ksnp_lock);
1411 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1412 tx_zc_list)) != NULL) {
1413 list_del(&tx->tx_zc_list);
1414 ksocknal_tx_decref(tx);
1419 ksocknal_terminate_conn(struct ksock_conn *conn)
1421 /* This gets called by the reaper (guaranteed thread context) to
1422 * disengage the socket from its callbacks and close it.
1423 * ksnc_refcount will eventually hit zero, and then the reaper will
1426 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1427 struct ksock_sched *sched = conn->ksnc_scheduler;
1428 bool failed = false;
1430 LASSERT(conn->ksnc_closing);
1432 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1433 spin_lock_bh(&sched->kss_lock);
1435 /* a closing conn is always ready to tx */
1436 conn->ksnc_tx_ready = 1;
1438 if (!conn->ksnc_tx_scheduled &&
1439 !list_empty(&conn->ksnc_tx_queue)) {
1440 list_add_tail(&conn->ksnc_tx_list,
1441 &sched->kss_tx_conns);
1442 conn->ksnc_tx_scheduled = 1;
1443 /* extra ref for scheduler */
1444 ksocknal_conn_addref(conn);
1446 wake_up(&sched->kss_waitq);
1449 spin_unlock_bh(&sched->kss_lock);
1451 /* serialise with callbacks */
1452 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1454 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1456 /* OK, so this conn may not be completely disengaged from its
1457 * scheduler yet, but it _has_ committed to terminate...
1459 conn->ksnc_scheduler->kss_nconns--;
1461 if (peer_ni->ksnp_error != 0) {
1462 /* peer_ni's last conn closed in error */
1463 LASSERT(list_empty(&peer_ni->ksnp_conns));
1465 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1468 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1471 ksocknal_peer_failed(peer_ni);
1473 /* The socket is closed on the final put; either here, or in
1474 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1475 * when the connection was established, this will close the socket
1476 * immediately, aborting anything buffered in it. Any hung
1477 * zero-copy transmits will therefore complete in finite time.
1479 ksocknal_connsock_decref(conn);
1483 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1485 /* Queue the conn for the reaper to destroy */
1486 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1487 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1489 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1490 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1492 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1496 ksocknal_destroy_conn(struct ksock_conn *conn)
1500 /* Final coup-de-grace of the reaper */
1501 CDEBUG(D_NET, "connection %p\n", conn);
1503 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1504 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1505 LASSERT(conn->ksnc_sock == NULL);
1506 LASSERT(conn->ksnc_conn_cb == NULL);
1507 LASSERT(!conn->ksnc_tx_scheduled);
1508 LASSERT(!conn->ksnc_rx_scheduled);
1509 LASSERT(list_empty(&conn->ksnc_tx_queue));
1511 /* complete current receive if any */
1512 switch (conn->ksnc_rx_state) {
1513 case SOCKNAL_RX_LNET_PAYLOAD:
1514 last_rcv = conn->ksnc_rx_deadline -
1516 CERROR("Completing partial receive from %s[%d], ip %pIScp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1517 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1519 &conn->ksnc_peeraddr,
1520 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1521 ktime_get_seconds() - last_rcv);
1522 if (conn->ksnc_lnet_msg)
1523 conn->ksnc_lnet_msg->msg_health_status =
1524 LNET_MSG_STATUS_REMOTE_ERROR;
1525 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1527 case SOCKNAL_RX_LNET_HEADER:
1528 if (conn->ksnc_rx_started)
1529 CERROR("Incomplete receive of lnet header from %s, ip %pIScp, with error, protocol: %d.x.\n",
1530 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1531 &conn->ksnc_peeraddr,
1532 conn->ksnc_proto->pro_version);
1534 case SOCKNAL_RX_KSM_HEADER:
1535 if (conn->ksnc_rx_started)
1536 CERROR("Incomplete receive of ksock message from %s, ip %pIScp, with error, protocol: %d.x.\n",
1537 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1538 &conn->ksnc_peeraddr,
1539 conn->ksnc_proto->pro_version);
1541 case SOCKNAL_RX_SLOP:
1542 if (conn->ksnc_rx_started)
1543 CERROR("Incomplete receive of slops from %s, ip %pIScp, with error\n",
1544 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1545 &conn->ksnc_peeraddr);
1552 ksocknal_peer_decref(conn->ksnc_peer);
1554 LIBCFS_FREE(conn, sizeof(*conn));
1558 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1559 struct sockaddr *addr, int why)
1561 struct ksock_conn *conn;
1562 struct ksock_conn *cnxt;
1565 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1568 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1570 ksocknal_close_conn_locked(conn, why);
1578 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1580 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1583 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1585 count = ksocknal_close_peer_conns_locked(
1586 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1588 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1594 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1596 struct ksock_peer_ni *peer_ni;
1597 struct hlist_node *pnxt;
1602 struct sockaddr_in sa = {.sin_family = AF_INET};
1604 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1606 if (!LNET_NID_IS_ANY(&id->nid)) {
1607 lo = hash_min(nidhash(&id->nid),
1608 HASH_BITS(ksocknal_data.ksnd_peers));
1612 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1615 sa.sin_addr.s_addr = htonl(ipaddr);
1616 for (i = lo; i <= hi; i++) {
1617 hlist_for_each_entry_safe(peer_ni, pnxt,
1618 &ksocknal_data.ksnd_peers[i],
1621 if (!((LNET_NID_IS_ANY(&id->nid) ||
1622 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1623 (id->pid == LNET_PID_ANY ||
1624 id->pid == peer_ni->ksnp_id.pid)))
1627 count += ksocknal_close_peer_conns_locked(
1629 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1633 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1635 /* wildcards always succeed */
1636 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1640 return (count == 0 ? -ENOENT : 0);
1644 ksocknal_notify_gw_down(struct lnet_nid *gw_nid)
1646 /* The router is telling me she's been notified of a change in
1649 struct lnet_processid id = {
1650 .pid = LNET_PID_ANY,
1654 CDEBUG(D_NET, "gw %s down\n", libcfs_nidstr(gw_nid));
1656 /* If the gateway crashed, close all open connections... */
1657 ksocknal_close_matching_conns(&id, 0);
1660 /* We can only establish new connections
1661 * if we have autroutes, and these connect on demand.
1666 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1670 struct ksock_conn *conn;
1672 for (index = 0; ; index++) {
1673 read_lock(&ksocknal_data.ksnd_global_lock);
1678 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1680 ksocknal_conn_addref(conn);
1685 read_unlock(&ksocknal_data.ksnd_global_lock);
1690 ksocknal_lib_push_conn (conn);
1691 ksocknal_conn_decref(conn);
1696 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1703 if (!LNET_NID_IS_ANY(&id->nid)) {
1704 lo = hash_min(nidhash(&id->nid),
1705 HASH_BITS(ksocknal_data.ksnd_peers));
1709 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1712 for (bkt = lo; bkt <= hi; bkt++) {
1713 int peer_off; /* searching offset in peer_ni hash table */
1715 for (peer_off = 0; ; peer_off++) {
1716 struct ksock_peer_ni *peer_ni;
1719 read_lock(&ksocknal_data.ksnd_global_lock);
1720 hlist_for_each_entry(peer_ni,
1721 &ksocknal_data.ksnd_peers[bkt],
1723 if (!((LNET_NID_IS_ANY(&id->nid) ||
1725 &peer_ni->ksnp_id.nid)) &&
1726 (id->pid == LNET_PID_ANY ||
1727 id->pid == peer_ni->ksnp_id.pid)))
1730 if (i++ == peer_off) {
1731 ksocknal_peer_addref(peer_ni);
1735 read_unlock(&ksocknal_data.ksnd_global_lock);
1737 if (i <= peer_off) /* no match */
1741 ksocknal_push_peer(peer_ni);
1742 ksocknal_peer_decref(peer_ni);
1749 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1751 struct lnet_processid id = {};
1752 struct libcfs_ioctl_data *data = arg;
1756 case IOC_LIBCFS_GET_INTERFACE: {
1757 struct ksock_net *net = ni->ni_data;
1758 struct ksock_interface *iface;
1759 struct sockaddr_in *sa;
1761 read_lock(&ksocknal_data.ksnd_global_lock);
1763 if (data->ioc_count >= 1) {
1767 iface = &net->ksnn_interface;
1769 sa = (void *)&iface->ksni_addr;
1770 if (sa->sin_family == AF_INET) {
1771 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1772 data->ioc_u32[1] = iface->ksni_netmask;
1774 data->ioc_u32[0] = 0xFFFFFFFF;
1775 data->ioc_u32[1] = 0;
1777 data->ioc_u32[2] = iface->ksni_npeers;
1778 data->ioc_u32[3] = iface->ksni_nroutes;
1781 read_unlock(&ksocknal_data.ksnd_global_lock);
1785 case IOC_LIBCFS_GET_PEER: {
1790 int share_count = 0;
1792 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1793 &id, &myip, &ip, &port,
1794 &conn_count, &share_count);
1798 if (!nid_is_nid4(&id.nid))
1800 data->ioc_nid = lnet_nid_to_nid4(&id.nid);
1801 data->ioc_count = share_count;
1802 data->ioc_u32[0] = ip;
1803 data->ioc_u32[1] = port;
1804 data->ioc_u32[2] = myip;
1805 data->ioc_u32[3] = conn_count;
1806 data->ioc_u32[4] = id.pid;
1810 case IOC_LIBCFS_ADD_PEER: {
1811 struct sockaddr_in sa = {.sin_family = AF_INET};
1813 id.pid = LNET_PID_LUSTRE;
1814 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1815 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1816 sa.sin_port = htons(data->ioc_u32[1]);
1817 return ksocknal_add_peer(ni, &id, (struct sockaddr *)&sa);
1819 case IOC_LIBCFS_DEL_PEER:
1820 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1821 id.pid = LNET_PID_ANY;
1822 return ksocknal_del_peer(ni, &id);
1824 case IOC_LIBCFS_GET_CONN: {
1828 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1829 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1830 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1835 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1837 data->ioc_count = txmem;
1838 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1839 data->ioc_flags = nagle;
1840 if (psa->sin_family == AF_INET)
1841 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1843 data->ioc_u32[0] = 0xFFFFFFFF;
1844 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1845 &conn->ksnc_peeraddr);
1846 if (mysa->sin_family == AF_INET)
1847 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1849 data->ioc_u32[2] = 0xFFFFFFFF;
1850 data->ioc_u32[3] = conn->ksnc_type;
1851 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1852 data->ioc_u32[5] = rxmem;
1853 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1854 ksocknal_conn_decref(conn);
1858 case IOC_LIBCFS_CLOSE_CONNECTION:
1859 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1860 id.pid = LNET_PID_ANY;
1861 return ksocknal_close_matching_conns(&id,
1864 case IOC_LIBCFS_REGISTER_MYNID:
1865 /* Ignore if this is a noop */
1866 if (nid_is_nid4(&ni->ni_nid) &&
1867 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1870 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1871 libcfs_nid2str(data->ioc_nid),
1872 libcfs_nidstr(&ni->ni_nid));
1875 case IOC_LIBCFS_PUSH_CONNECTION:
1876 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1877 id.pid = LNET_PID_ANY;
1878 return ksocknal_push(ni, &id);
1887 ksocknal_free_buffers (void)
1889 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1891 if (ksocknal_data.ksnd_schedulers != NULL)
1892 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1894 spin_lock(&ksocknal_data.ksnd_tx_lock);
1896 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1898 struct ksock_tx *tx;
1900 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1901 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1903 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1904 tx_list)) != NULL) {
1905 list_del(&tx->tx_list);
1906 LIBCFS_FREE(tx, tx->tx_desc_size);
1909 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1914 ksocknal_handle_link_state_change(struct net_device *dev,
1915 unsigned char operstate)
1917 struct lnet_ni *ni = NULL;
1918 struct ksock_net *net;
1919 struct ksock_net *cnxt;
1921 unsigned char link_down = !(operstate == IF_OPER_UP);
1922 struct in_device *in_dev;
1923 bool found_ip = false;
1924 struct ksock_interface *ksi = NULL;
1925 struct sockaddr_in *sa;
1926 __u32 ni_state_before;
1927 bool update_ping_buf = false;
1929 DECLARE_CONST_IN_IFADDR(ifa);
1931 ifindex = dev->ifindex;
1933 if (!ksocknal_data.ksnd_nnets)
1936 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1939 ksi = &net->ksnn_interface;
1940 sa = (void *)&ksi->ksni_addr;
1943 if (strcmp(ksi->ksni_name, dev->name))
1946 if (ksi->ksni_index == -1) {
1947 if (dev->reg_state != NETREG_REGISTERED)
1949 /* A registration just happened: save the new index for
1951 ksi->ksni_index = ifindex;
1955 if (ksi->ksni_index != ifindex)
1958 if (dev->reg_state == NETREG_UNREGISTERING) {
1959 /* Device is being unregistered, we need to clear the
1960 * index, it can change when device will be back */
1961 ksi->ksni_index = -1;
1967 in_dev = __in_dev_get_rtnl(dev);
1969 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
1971 ni_state_before = lnet_set_link_fatal_state(ni, 1);
1974 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1975 if (sa->sin_addr.s_addr == ifa->ifa_local)
1981 CDEBUG(D_NET, "Interface %s has no matching ip\n",
1983 ni_state_before = lnet_set_link_fatal_state(ni, 1);
1988 ni_state_before = lnet_set_link_fatal_state(ni, 1);
1990 state = (lnet_get_link_status(dev) == 0);
1991 ni_state_before = lnet_set_link_fatal_state(ni,
1995 if (!update_ping_buf &&
1996 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
1997 (atomic_read(&ni->ni_fatal_error_on) != ni_state_before))
1998 update_ping_buf = true;
2001 if (update_ping_buf)
2002 lnet_mark_ping_buffer_for_update();
2009 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
2011 struct lnet_ni *ni = NULL;
2012 struct ksock_net *net;
2013 struct ksock_net *cnxt;
2014 struct net_device *event_netdev = ifa->ifa_dev->dev;
2016 struct ksock_interface *ksi = NULL;
2017 struct sockaddr_in *sa;
2018 __u32 ni_state_before;
2019 bool update_ping_buf = false;
2022 if (!ksocknal_data.ksnd_nnets)
2025 ifindex = event_netdev->ifindex;
2027 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
2030 ksi = &net->ksnn_interface;
2031 sa = (void *)&ksi->ksni_addr;
2033 if (ksi->ksni_index != ifindex ||
2034 strcmp(ksi->ksni_name, event_netdev->name))
2037 if (sa->sin_addr.s_addr == ifa->ifa_local) {
2039 link_down = (event == NETDEV_DOWN);
2040 ni_state_before = lnet_set_link_fatal_state(ni,
2043 if (!update_ping_buf &&
2044 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
2045 ((event == NETDEV_DOWN) != ni_state_before))
2046 update_ping_buf = true;
2050 if (update_ping_buf)
2051 lnet_mark_ping_buffer_for_update();
2056 /************************************
2057 * Net device notifier event handler
2058 ************************************/
2059 static int ksocknal_device_event(struct notifier_block *unused,
2060 unsigned long event, void *ptr)
2062 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2063 unsigned char operstate;
2065 operstate = dev->operstate;
2067 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2068 event, dev->name, dev->ifindex, operstate);
2074 case NETDEV_REGISTER:
2075 case NETDEV_UNREGISTER:
2076 ksocknal_handle_link_state_change(dev, operstate);
2083 /************************************
2084 * Inetaddr notifier event handler
2085 ************************************/
2086 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2087 unsigned long event, void *ptr)
2089 struct in_ifaddr *ifa = ptr;
2091 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2092 event, &ifa->ifa_address, &ifa->ifa_mask);
2098 ksocknal_handle_inetaddr_change(ifa, event);
2105 static struct notifier_block ksocknal_dev_notifier_block = {
2106 .notifier_call = ksocknal_device_event,
2109 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2110 .notifier_call = ksocknal_inetaddr_event,
2114 ksocknal_base_shutdown(void)
2116 struct ksock_sched *sched;
2117 struct ksock_peer_ni *peer_ni;
2120 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2121 libcfs_kmem_read());
2122 LASSERT (ksocknal_data.ksnd_nnets == 0);
2124 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2125 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2126 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2129 switch (ksocknal_data.ksnd_init) {
2134 case SOCKNAL_INIT_ALL:
2135 case SOCKNAL_INIT_DATA:
2136 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2139 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2140 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2141 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2142 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2143 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2145 if (ksocknal_data.ksnd_schedulers != NULL) {
2146 cfs_percpt_for_each(sched, i,
2147 ksocknal_data.ksnd_schedulers) {
2149 LASSERT(list_empty(&sched->kss_tx_conns));
2150 LASSERT(list_empty(&sched->kss_rx_conns));
2151 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2152 LASSERT(sched->kss_nconns == 0);
2156 /* flag threads to terminate; wake and wait for them to die */
2157 ksocknal_data.ksnd_shuttingdown = 1;
2158 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2159 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2161 if (ksocknal_data.ksnd_schedulers != NULL) {
2162 cfs_percpt_for_each(sched, i,
2163 ksocknal_data.ksnd_schedulers)
2164 wake_up_all(&sched->kss_waitq);
2167 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2168 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2169 "waiting for %d threads to terminate\n",
2170 atomic_read(&ksocknal_data.ksnd_nthreads));
2172 ksocknal_free_buffers();
2174 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2178 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2179 libcfs_kmem_read());
2181 module_put(THIS_MODULE);
2185 ksocknal_base_startup(void)
2187 struct ksock_sched *sched;
2191 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2192 LASSERT(ksocknal_data.ksnd_nnets == 0);
2194 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2196 hash_init(ksocknal_data.ksnd_peers);
2198 rwlock_init(&ksocknal_data.ksnd_global_lock);
2199 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2201 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2202 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2203 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2204 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2205 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2207 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2208 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2209 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2210 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2212 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2213 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2215 /* NB memset above zeros whole of ksocknal_data */
2217 /* flag lists/ptrs/locks initialised */
2218 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2219 if (!try_module_get(THIS_MODULE))
2222 /* Create a scheduler block per available CPT */
2223 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2225 if (ksocknal_data.ksnd_schedulers == NULL)
2228 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2232 * make sure not to allocate more threads than there are
2233 * cores/CPUs in teh CPT
2235 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2236 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2237 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2240 * max to half of CPUs, assume another half should be
2241 * reserved for upper layer modules
2243 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2246 sched->kss_nthreads_max = nthrs;
2249 spin_lock_init(&sched->kss_lock);
2250 INIT_LIST_HEAD(&sched->kss_rx_conns);
2251 INIT_LIST_HEAD(&sched->kss_tx_conns);
2252 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2253 init_waitqueue_head(&sched->kss_waitq);
2256 ksocknal_data.ksnd_connd_starting = 0;
2257 ksocknal_data.ksnd_connd_failed_stamp = 0;
2258 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2259 /* must have at least 2 connds to remain responsive to accepts while
2261 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2262 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2264 if (*ksocknal_tunables.ksnd_nconnds_max <
2265 *ksocknal_tunables.ksnd_nconnds) {
2266 ksocknal_tunables.ksnd_nconnds_max =
2267 ksocknal_tunables.ksnd_nconnds;
2270 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2271 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2272 ksocknal_data.ksnd_connd_starting++;
2273 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2275 rc = ksocknal_thread_start(ksocknal_connd,
2276 (void *)((uintptr_t)i),
2277 "socknal_cd%02d", i);
2279 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2280 ksocknal_data.ksnd_connd_starting--;
2281 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2282 CERROR("Can't spawn socknal connd: %d\n", rc);
2287 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2289 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2293 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2294 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2296 /* flag everything initialised */
2297 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2302 ksocknal_base_shutdown();
2307 ksocknal_debug_peerhash(struct lnet_ni *ni)
2309 struct ksock_peer_ni *peer_ni;
2312 read_lock(&ksocknal_data.ksnd_global_lock);
2314 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2315 struct ksock_conn_cb *conn_cb;
2316 struct ksock_conn *conn;
2318 if (peer_ni->ksnp_ni != ni)
2321 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2322 libcfs_idstr(&peer_ni->ksnp_id),
2323 refcount_read(&peer_ni->ksnp_refcount),
2324 peer_ni->ksnp_closing,
2325 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2326 peer_ni->ksnp_zc_next_cookie,
2327 !list_empty(&peer_ni->ksnp_tx_queue),
2328 !list_empty(&peer_ni->ksnp_zc_req_list));
2330 conn_cb = peer_ni->ksnp_conn_cb;
2332 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2333 refcount_read(&conn_cb->ksnr_refcount),
2334 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2335 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2338 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2339 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2340 refcount_read(&conn->ksnc_conn_refcount),
2341 refcount_read(&conn->ksnc_sock_refcount),
2342 conn->ksnc_type, conn->ksnc_closing);
2347 read_unlock(&ksocknal_data.ksnd_global_lock);
2352 ksocknal_shutdown(struct lnet_ni *ni)
2354 struct ksock_net *net = ni->ni_data;
2356 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2357 LASSERT(ksocknal_data.ksnd_nnets > 0);
2359 /* prevent new peers */
2360 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2362 /* Delete all peers */
2363 ksocknal_del_peer(ni, NULL);
2365 /* Wait for all peer_ni state to clean up */
2366 wait_var_event_warning(&net->ksnn_npeers,
2367 atomic_read(&net->ksnn_npeers) ==
2368 SOCKNAL_SHUTDOWN_BIAS,
2369 "waiting for %d peers to disconnect\n",
2370 ksocknal_debug_peerhash(ni) +
2371 atomic_read(&net->ksnn_npeers) -
2372 SOCKNAL_SHUTDOWN_BIAS);
2374 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2375 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2377 list_del(&net->ksnn_list);
2378 LIBCFS_FREE(net, sizeof(*net));
2380 ksocknal_data.ksnd_nnets--;
2381 if (ksocknal_data.ksnd_nnets == 0)
2382 ksocknal_base_shutdown();
2386 ksocknal_search_new_ipif(struct ksock_net *net)
2389 char *ifnam = &net->ksnn_interface.ksni_name[0];
2390 char *colon = strchr(ifnam, ':');
2392 struct ksock_net *tmp;
2397 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2398 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2399 char *colon2 = strchr(ifnam2, ':');
2404 found = strcmp(ifnam, ifnam2) == 0;
2417 ksocknal_start_schedulers(struct ksock_sched *sched)
2423 if (sched->kss_nthreads == 0) {
2424 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2425 nthrs = sched->kss_nthreads_max;
2427 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2429 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2430 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2432 nthrs = min(nthrs, sched->kss_nthreads_max);
2434 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2435 /* increase two threads if there is new interface */
2436 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2439 for (i = 0; i < nthrs; i++) {
2442 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2443 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2444 "socknal_sd%02d_%02d",
2446 (int)KSOCK_THREAD_SID(id));
2450 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2451 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2455 sched->kss_nthreads += i;
2460 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2462 int newif = ksocknal_search_new_ipif(net);
2466 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2469 for (i = 0; i < ncpts; i++) {
2470 struct ksock_sched *sched;
2471 int cpt = (cpts == NULL) ? i : cpts[i];
2473 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2474 sched = ksocknal_data.ksnd_schedulers[cpt];
2476 if (!newif && sched->kss_nthreads > 0)
2479 rc = ksocknal_start_schedulers(sched);
2487 ksocknal_startup(struct lnet_ni *ni)
2489 struct ksock_net *net;
2490 struct ksock_interface *ksi = NULL;
2491 struct lnet_inetdev *ifaces = NULL;
2495 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2496 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2497 rc = ksocknal_base_startup();
2501 LIBCFS_ALLOC(net, sizeof(*net));
2505 net->ksnn_incarnation = ktime_get_real_ns();
2508 ksocknal_tunables_setup(ni);
2510 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, true);
2514 ksi = &net->ksnn_interface;
2516 /* Interface and/or IP address is specified otherwise default to
2517 * the first Interface
2519 if_idx = lnet_inet_select(ni, ifaces, rc);
2523 if (!ni->ni_interface) {
2524 rc = lnet_ni_add_interface(ni, ifaces[if_idx].li_name);
2526 CWARN("ksocklnd failed to allocate ni_interface\n");
2529 ni->ni_dev_cpt = ifaces[if_idx].li_cpt;
2530 ksi->ksni_index = ifaces[if_idx].li_index;
2531 if (ifaces[if_idx].li_size == sizeof(struct in6_addr)) {
2532 struct sockaddr_in6 *sa;
2533 sa = (void *)&ksi->ksni_addr;
2534 memset(sa, 0, sizeof(*sa));
2535 sa->sin6_family = AF_INET6;
2536 memcpy(&sa->sin6_addr, ifaces[if_idx].li_ipv6addr,
2537 sizeof(struct in6_addr));
2538 ni->ni_nid.nid_size = sizeof(struct in6_addr) - 4;
2539 memcpy(&ni->ni_nid.nid_addr, ifaces[if_idx].li_ipv6addr,
2540 sizeof(struct in6_addr));
2542 struct sockaddr_in *sa;
2543 sa = (void *)&ksi->ksni_addr;
2544 memset(sa, 0, sizeof(*sa));
2545 sa->sin_family = AF_INET;
2546 sa->sin_addr.s_addr = ifaces[if_idx].li_ipaddr;
2547 ksi->ksni_netmask = ifaces[if_idx].li_netmask;
2548 ni->ni_nid.nid_size = 0;
2549 ni->ni_nid.nid_addr[0] = sa->sin_addr.s_addr;
2551 strlcpy(ksi->ksni_name, ifaces[if_idx].li_name, sizeof(ksi->ksni_name));
2553 /* call it before add it to ksocknal_data.ksnd_nets */
2554 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2558 if ((ksocknal_ip2index((struct sockaddr *)&ksi->ksni_addr,
2560 &dev_status) < 0) ||
2562 lnet_set_link_fatal_state(ni, 1);
2564 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2566 ksocknal_data.ksnd_nnets++;
2571 LIBCFS_FREE(net, sizeof(*net));
2573 if (ksocknal_data.ksnd_nnets == 0)
2574 ksocknal_base_shutdown();
2579 static void __exit ksocklnd_exit(void)
2581 lnet_unregister_lnd(&the_ksocklnd);
2584 static const struct lnet_lnd the_ksocklnd = {
2585 .lnd_type = SOCKLND,
2586 .lnd_startup = ksocknal_startup,
2587 .lnd_shutdown = ksocknal_shutdown,
2588 .lnd_ctl = ksocknal_ctl,
2589 .lnd_send = ksocknal_send,
2590 .lnd_recv = ksocknal_recv,
2591 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2592 .lnd_accept = ksocknal_accept,
2593 .lnd_nl_set = ksocknal_nl_set,
2594 .lnd_keys = &ksocknal_tunables_keys,
2597 static int __init ksocklnd_init(void)
2601 /* check ksnr_connected/connecting field large enough */
2602 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2603 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2605 rc = ksocknal_tunables_init();
2609 lnet_register_lnd(&the_ksocklnd);
2614 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2615 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2616 MODULE_VERSION("2.8.0");
2617 MODULE_LICENSE("GPL");
2619 module_init(ksocklnd_init);
2620 module_exit(ksocklnd_exit);