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 (dev->operstate != IF_OPER_UNKNOWN))) ||
116 (lnet_get_link_status(dev) == 0))
122 static struct ksock_conn_cb *
123 ksocknal_create_conn_cb(struct sockaddr *addr)
125 struct ksock_conn_cb *conn_cb;
127 LIBCFS_ALLOC(conn_cb, sizeof(*conn_cb));
131 refcount_set(&conn_cb->ksnr_refcount, 1);
132 conn_cb->ksnr_peer = NULL;
133 conn_cb->ksnr_retry_interval = 0; /* OK to connect at any time */
134 rpc_copy_addr((struct sockaddr *)&conn_cb->ksnr_addr, addr);
135 rpc_set_port((struct sockaddr *)&conn_cb->ksnr_addr,
137 conn_cb->ksnr_scheduled = 0;
138 conn_cb->ksnr_connecting = 0;
139 conn_cb->ksnr_connected = 0;
140 conn_cb->ksnr_deleted = 0;
141 conn_cb->ksnr_conn_count = 0;
142 conn_cb->ksnr_ctrl_conn_count = 0;
143 conn_cb->ksnr_blki_conn_count = 0;
144 conn_cb->ksnr_blko_conn_count = 0;
145 conn_cb->ksnr_max_conns = 0;
146 conn_cb->ksnr_busy_retry_count = 0;
152 ksocknal_destroy_conn_cb(struct ksock_conn_cb *conn_cb)
154 LASSERT(refcount_read(&conn_cb->ksnr_refcount) == 0);
156 if (conn_cb->ksnr_peer)
157 ksocknal_peer_decref(conn_cb->ksnr_peer);
159 LIBCFS_FREE(conn_cb, sizeof(*conn_cb));
162 static struct ksock_peer_ni *
163 ksocknal_create_peer(struct lnet_ni *ni, struct lnet_processid *id)
165 int cpt = lnet_nid2cpt(&id->nid, ni);
166 struct ksock_net *net = ni->ni_data;
167 struct ksock_peer_ni *peer_ni;
169 LASSERT(!LNET_NID_IS_ANY(&id->nid));
170 LASSERT(id->pid != LNET_PID_ANY);
171 LASSERT(!in_interrupt());
173 if (!atomic_inc_unless_negative(&net->ksnn_npeers)) {
174 CERROR("Can't create peer_ni: network shutdown\n");
175 return ERR_PTR(-ESHUTDOWN);
178 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
180 atomic_dec(&net->ksnn_npeers);
181 return ERR_PTR(-ENOMEM);
184 peer_ni->ksnp_ni = ni;
185 peer_ni->ksnp_id = *id;
186 refcount_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
187 peer_ni->ksnp_closing = 0;
188 peer_ni->ksnp_accepting = 0;
189 peer_ni->ksnp_proto = NULL;
190 peer_ni->ksnp_last_alive = 0;
191 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
192 peer_ni->ksnp_conn_cb = NULL;
194 INIT_LIST_HEAD(&peer_ni->ksnp_conns);
195 INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
196 INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
197 spin_lock_init(&peer_ni->ksnp_lock);
203 ksocknal_destroy_peer(struct ksock_peer_ni *peer_ni)
205 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
207 CDEBUG (D_NET, "peer_ni %s %p deleted\n",
208 libcfs_idstr(&peer_ni->ksnp_id), peer_ni);
210 LASSERT(refcount_read(&peer_ni->ksnp_refcount) == 0);
211 LASSERT(peer_ni->ksnp_accepting == 0);
212 LASSERT(list_empty(&peer_ni->ksnp_conns));
213 LASSERT(peer_ni->ksnp_conn_cb == NULL);
214 LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
215 LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
217 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
219 /* NB a peer_ni's connections and conn_cb keep a reference on their
220 * peer_ni until they are destroyed, so we can be assured that _all_
221 * state to do with this peer_ni has been cleaned up when its refcount
224 if (atomic_dec_and_test(&net->ksnn_npeers))
225 wake_up_var(&net->ksnn_npeers);
228 struct ksock_peer_ni *
229 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_processid *id)
231 struct ksock_peer_ni *peer_ni;
232 unsigned long hash = nidhash(&id->nid);
234 hash_for_each_possible(ksocknal_data.ksnd_peers, peer_ni,
236 LASSERT(!peer_ni->ksnp_closing);
238 if (peer_ni->ksnp_ni != ni)
241 if (!nid_same(&peer_ni->ksnp_id.nid, &id->nid) ||
242 peer_ni->ksnp_id.pid != id->pid)
245 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
246 peer_ni, libcfs_idstr(id),
247 refcount_read(&peer_ni->ksnp_refcount));
253 struct ksock_peer_ni *
254 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_processid *id)
256 struct ksock_peer_ni *peer_ni;
258 read_lock(&ksocknal_data.ksnd_global_lock);
259 peer_ni = ksocknal_find_peer_locked(ni, id);
260 if (peer_ni != NULL) /* +1 ref for caller? */
261 ksocknal_peer_addref(peer_ni);
262 read_unlock(&ksocknal_data.ksnd_global_lock);
268 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
270 LASSERT(list_empty(&peer_ni->ksnp_conns));
271 LASSERT(peer_ni->ksnp_conn_cb == NULL);
272 LASSERT(!peer_ni->ksnp_closing);
273 peer_ni->ksnp_closing = 1;
274 hlist_del(&peer_ni->ksnp_list);
275 /* lose peerlist's ref */
276 ksocknal_peer_decref(peer_ni);
281 ksocknal_dump_peer_debug_info(struct ksock_peer_ni *peer_ni)
283 struct ksock_conn *conn;
284 struct list_head *ctmp;
285 struct list_head *txtmp;
289 list_for_each(ctmp, &peer_ni->ksnp_conns) {
290 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
292 if (!list_empty(&conn->ksnc_tx_queue))
293 list_for_each(txtmp, &conn->ksnc_tx_queue) txcount++;
295 CDEBUG(D_CONSOLE, "Conn %d [type, closing, crefcnt, srefcnt]: %d, %d, %d, %d\n",
299 refcount_read(&conn->ksnc_conn_refcount),
300 refcount_read(&conn->ksnc_sock_refcount));
301 CDEBUG(D_CONSOLE, "Conn %d rx [scheduled, ready, state]: %d, %d, %d\n",
303 conn->ksnc_rx_scheduled,
305 conn->ksnc_rx_state);
306 CDEBUG(D_CONSOLE, "Conn %d tx [txqcnt, scheduled, last_post, ready, deadline]: %d, %d, %lld, %d, %lld\n",
309 conn->ksnc_tx_scheduled,
310 conn->ksnc_tx_last_post,
312 conn->ksnc_rx_deadline);
314 if (conn->ksnc_scheduler)
315 CDEBUG(D_CONSOLE, "Conn %d sched [nconns, cpt]: %d, %d\n",
317 conn->ksnc_scheduler->kss_nconns,
318 conn->ksnc_scheduler->kss_cpt);
326 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
327 struct lnet_processid *id, __u32 *myip, __u32 *peer_ip,
328 int *port, int *conn_count, int *share_count)
330 struct ksock_peer_ni *peer_ni;
331 struct ksock_conn_cb *conn_cb;
334 struct ksock_net *net;
336 read_lock(&ksocknal_data.ksnd_global_lock);
338 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
340 if (peer_ni->ksnp_ni != ni)
345 *id = peer_ni->ksnp_id;
346 conn_cb = peer_ni->ksnp_conn_cb;
347 if (conn_cb == NULL) {
355 ksocknal_dump_peer_debug_info(peer_ni);
357 if (conn_cb->ksnr_addr.ss_family == AF_INET) {
358 struct sockaddr_in *sa =
359 (void *)&conn_cb->ksnr_addr;
361 rc = choose_ipv4_src(myip,
362 net->ksnn_interface.ksni_index,
363 ntohl(sa->sin_addr.s_addr),
365 *peer_ip = ntohl(sa->sin_addr.s_addr);
366 *port = ntohs(sa->sin_port);
370 *peer_ip = 0xFFFFFFFF;
374 *conn_count = conn_cb->ksnr_conn_count;
379 read_unlock(&ksocknal_data.ksnd_global_lock);
384 ksocknal_get_conns_per_peer(struct ksock_peer_ni *peer_ni)
386 struct lnet_ni *ni = peer_ni->ksnp_ni;
387 struct lnet_ioctl_config_socklnd_tunables *tunables;
391 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
393 return tunables->lnd_conns_per_peer;
397 ksocknal_incr_conn_count(struct ksock_conn_cb *conn_cb,
400 conn_cb->ksnr_conn_count++;
402 /* check if all connections of the given type got created */
404 case SOCKLND_CONN_CONTROL:
405 conn_cb->ksnr_ctrl_conn_count++;
406 /* there's a single control connection per peer,
407 * two in case of loopback
409 conn_cb->ksnr_connected |= BIT(type);
411 case SOCKLND_CONN_BULK_IN:
412 conn_cb->ksnr_blki_conn_count++;
413 if (conn_cb->ksnr_blki_conn_count >= conn_cb->ksnr_max_conns)
414 conn_cb->ksnr_connected |= BIT(type);
416 case SOCKLND_CONN_BULK_OUT:
417 conn_cb->ksnr_blko_conn_count++;
418 if (conn_cb->ksnr_blko_conn_count >= conn_cb->ksnr_max_conns)
419 conn_cb->ksnr_connected |= BIT(type);
421 case SOCKLND_CONN_ANY:
422 if (conn_cb->ksnr_conn_count >= conn_cb->ksnr_max_conns)
423 conn_cb->ksnr_connected |= BIT(type);
430 CDEBUG(D_NET, "Add conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
431 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
436 ksocknal_decr_conn_count(struct ksock_conn_cb *conn_cb,
439 conn_cb->ksnr_conn_count--;
441 /* check if all connections of the given type got created */
443 case SOCKLND_CONN_CONTROL:
444 conn_cb->ksnr_ctrl_conn_count--;
445 /* there's a single control connection per peer,
446 * two in case of loopback
448 if (conn_cb->ksnr_ctrl_conn_count == 0)
449 conn_cb->ksnr_connected &= ~BIT(type);
451 case SOCKLND_CONN_BULK_IN:
452 conn_cb->ksnr_blki_conn_count--;
453 if (conn_cb->ksnr_blki_conn_count == 0)
454 conn_cb->ksnr_connected &= ~BIT(type);
456 case SOCKLND_CONN_BULK_OUT:
457 conn_cb->ksnr_blko_conn_count--;
458 if (conn_cb->ksnr_blko_conn_count == 0)
459 conn_cb->ksnr_connected &= ~BIT(type);
461 case SOCKLND_CONN_ANY:
462 if (conn_cb->ksnr_conn_count == 0)
463 conn_cb->ksnr_connected &= ~BIT(type);
470 CDEBUG(D_NET, "Del conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
471 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
475 ksocknal_associate_cb_conn_locked(struct ksock_conn_cb *conn_cb,
476 struct ksock_conn *conn)
478 int type = conn->ksnc_type;
480 conn->ksnc_conn_cb = conn_cb;
481 ksocknal_conn_cb_addref(conn_cb);
482 ksocknal_incr_conn_count(conn_cb, type);
484 /* Successful connection => further attempts can
485 * proceed immediately
487 conn_cb->ksnr_retry_interval = 0;
491 ksocknal_add_conn_cb_locked(struct ksock_peer_ni *peer_ni,
492 struct ksock_conn_cb *conn_cb)
494 struct ksock_conn *conn;
495 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
497 LASSERT(!peer_ni->ksnp_closing);
498 LASSERT(!conn_cb->ksnr_peer);
499 LASSERT(!conn_cb->ksnr_scheduled);
500 LASSERT(!conn_cb->ksnr_connecting);
501 LASSERT(conn_cb->ksnr_connected == 0);
503 conn_cb->ksnr_peer = peer_ni;
504 ksocknal_peer_addref(peer_ni);
506 /* peer_ni's route list takes over my ref on 'route' */
507 peer_ni->ksnp_conn_cb = conn_cb;
508 net->ksnn_interface.ksni_nroutes++;
510 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
511 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
512 (struct sockaddr *)&conn_cb->ksnr_addr))
514 CDEBUG(D_NET, "call ksocknal_associate_cb_conn_locked\n");
515 ksocknal_associate_cb_conn_locked(conn_cb, conn);
516 /* keep going (typed conns) */
521 ksocknal_del_conn_cb_locked(struct ksock_conn_cb *conn_cb)
523 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
524 struct ksock_conn *conn;
525 struct ksock_conn *cnxt;
526 struct ksock_net *net;
528 LASSERT(!conn_cb->ksnr_deleted);
530 /* Close associated conns */
531 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
532 if (conn->ksnc_conn_cb != conn_cb)
535 ksocknal_close_conn_locked(conn, 0);
538 net = (struct ksock_net *)(peer_ni->ksnp_ni->ni_data);
539 net->ksnn_interface.ksni_nroutes--;
540 LASSERT(net->ksnn_interface.ksni_nroutes >= 0);
542 conn_cb->ksnr_deleted = 1;
543 ksocknal_conn_cb_decref(conn_cb); /* drop peer_ni's ref */
544 peer_ni->ksnp_conn_cb = NULL;
546 if (list_empty(&peer_ni->ksnp_conns)) {
547 /* I've just removed the last route to a peer_ni with no active
550 ksocknal_unlink_peer_locked(peer_ni);
555 ksocknal_get_conn_count_by_type(struct ksock_conn_cb *conn_cb,
558 unsigned int count = 0;
561 case SOCKLND_CONN_CONTROL:
562 count = conn_cb->ksnr_ctrl_conn_count;
564 case SOCKLND_CONN_BULK_IN:
565 count = conn_cb->ksnr_blki_conn_count;
567 case SOCKLND_CONN_BULK_OUT:
568 count = conn_cb->ksnr_blko_conn_count;
570 case SOCKLND_CONN_ANY:
571 count = conn_cb->ksnr_conn_count;
582 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_processid *id,
583 struct sockaddr *addr)
585 struct ksock_peer_ni *peer_ni;
586 struct ksock_peer_ni *peer2;
587 struct ksock_conn_cb *conn_cb;
589 if (LNET_NID_IS_ANY(&id->nid) ||
590 id->pid == LNET_PID_ANY)
593 /* Have a brand new peer_ni ready... */
594 peer_ni = ksocknal_create_peer(ni, id);
596 return PTR_ERR(peer_ni);
598 conn_cb = ksocknal_create_conn_cb(addr);
600 ksocknal_peer_decref(peer_ni);
604 write_lock_bh(&ksocknal_data.ksnd_global_lock);
606 /* always called with a ref on ni, so shutdown can't have started */
607 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
610 peer2 = ksocknal_find_peer_locked(ni, id);
612 ksocknal_peer_decref(peer_ni);
615 /* peer_ni table takes my ref on peer_ni */
616 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
620 if (peer_ni->ksnp_conn_cb) {
621 ksocknal_conn_cb_decref(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);
629 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
632 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
638 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni)
640 struct ksock_conn *conn;
641 struct ksock_conn *cnxt;
642 struct ksock_conn_cb *conn_cb;
644 LASSERT(!peer_ni->ksnp_closing);
646 /* Extra ref prevents peer_ni disappearing until I'm done with it */
647 ksocknal_peer_addref(peer_ni);
648 conn_cb = peer_ni->ksnp_conn_cb;
650 ksocknal_del_conn_cb_locked(conn_cb);
652 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
654 ksocknal_close_conn_locked(conn, 0);
656 ksocknal_peer_decref(peer_ni);
657 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
661 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_processid *id)
664 struct hlist_node *pnxt;
665 struct ksock_peer_ni *peer_ni;
671 write_lock_bh(&ksocknal_data.ksnd_global_lock);
673 if (id && !LNET_NID_IS_ANY(&id->nid)) {
674 lo = hash_min(nidhash(&id->nid),
675 HASH_BITS(ksocknal_data.ksnd_peers));
679 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
682 for (i = lo; i <= hi; i++) {
683 hlist_for_each_entry_safe(peer_ni, pnxt,
684 &ksocknal_data.ksnd_peers[i],
686 if (peer_ni->ksnp_ni != ni)
689 if (!((!id || LNET_NID_IS_ANY(&id->nid) ||
690 nid_same(&peer_ni->ksnp_id.nid, &id->nid)) &&
691 (!id || id->pid == LNET_PID_ANY ||
692 peer_ni->ksnp_id.pid == id->pid)))
695 ksocknal_peer_addref(peer_ni); /* a ref for me... */
697 ksocknal_del_peer_locked(peer_ni);
699 if (peer_ni->ksnp_closing &&
700 !list_empty(&peer_ni->ksnp_tx_queue)) {
701 LASSERT(list_empty(&peer_ni->ksnp_conns));
702 LASSERT(peer_ni->ksnp_conn_cb == NULL);
704 list_splice_init(&peer_ni->ksnp_tx_queue,
708 ksocknal_peer_decref(peer_ni); /* ...till here */
710 rc = 0; /* matched! */
714 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
716 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
721 static struct ksock_conn *
722 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
724 struct ksock_peer_ni *peer_ni;
725 struct ksock_conn *conn;
728 read_lock(&ksocknal_data.ksnd_global_lock);
730 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
731 LASSERT(!peer_ni->ksnp_closing);
733 if (peer_ni->ksnp_ni != ni)
736 list_for_each_entry(conn, &peer_ni->ksnp_conns,
741 ksocknal_conn_addref(conn);
742 read_unlock(&ksocknal_data.ksnd_global_lock);
747 read_unlock(&ksocknal_data.ksnd_global_lock);
751 static struct ksock_sched *
752 ksocknal_choose_scheduler_locked(unsigned int cpt)
754 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
757 if (sched->kss_nthreads == 0) {
758 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
759 if (sched->kss_nthreads > 0) {
760 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
761 cpt, sched->kss_cpt);
772 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
774 struct ksock_connreq *cr;
776 struct sockaddr_storage peer;
778 rc = lnet_sock_getaddr(sock, true, &peer);
780 CERROR("Can't determine new connection's address\n");
784 LIBCFS_ALLOC(cr, sizeof(*cr));
786 LCONSOLE_ERROR_MSG(0x12f,
787 "Dropping connection request from %pISc: memory exhausted\n",
794 cr->ksncr_sock = sock;
796 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
798 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
799 wake_up(&ksocknal_data.ksnd_connd_waitq);
801 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
805 static const struct ln_key_list ksocknal_tunables_keys = {
806 .lkl_maxattr = LNET_NET_SOCKLND_TUNABLES_ATTR_MAX,
808 [LNET_NET_SOCKLND_TUNABLES_ATTR_CONNS_PER_PEER] = {
809 .lkp_value = "conns_per_peer",
810 .lkp_data_type = NLA_U16
812 [LNET_NET_SOCKLND_TUNABLES_ATTR_LND_TIMEOUT] = {
813 .lkp_value = "timeout",
814 .lkp_data_type = NLA_U32
820 ksocknal_nl_get(int cmd, struct sk_buff *msg, int type, void *data)
822 struct lnet_lnd_tunables *tun;
823 struct lnet_ni *ni = data;
828 if (cmd != LNET_CMD_NETS || type != LNET_NET_LOCAL_NI_ATTR_LND_TUNABLES)
831 tun = &ni->ni_lnd_tunables;
832 nla_put_u16(msg, LNET_NET_SOCKLND_TUNABLES_ATTR_CONNS_PER_PEER,
833 tun->lnd_tun_u.lnd_sock.lnd_conns_per_peer);
834 nla_put_u32(msg, LNET_NET_SOCKLND_TUNABLES_ATTR_LND_TIMEOUT,
841 ksocknal_nl_set_default(int cmd, int type, void *data)
843 struct lnet_lnd_tunables *tunables = data;
844 struct lnet_ioctl_config_socklnd_tunables *lt;
845 struct lnet_ioctl_config_socklnd_tunables *df;
847 lt = &tunables->lnd_tun_u.lnd_sock;
848 df = &ksock_default_tunables;
850 case LNET_NET_SOCKLND_TUNABLES_ATTR_CONNS_PER_PEER:
851 lt->lnd_conns_per_peer = df->lnd_conns_per_peer;
853 case LNET_NET_SOCKLND_TUNABLES_ATTR_LND_TIMEOUT:
854 lt->lnd_timeout = df->lnd_timeout;
862 ksocknal_nl_set(int cmd, struct nlattr *attr, int type, void *data)
864 struct lnet_lnd_tunables *tunables = data;
868 if (cmd != LNET_CMD_NETS)
872 ksocknal_nl_set_default(cmd, type, data);
876 if (nla_type(attr) != LN_SCALAR_ATTR_INT_VALUE)
880 case LNET_NET_SOCKLND_TUNABLES_ATTR_CONNS_PER_PEER:
881 /* value values are 1 to 127. Zero mean calculate the value */
882 num = nla_get_s64(attr);
883 if (num > -1 && num < 128)
884 tunables->lnd_tun_u.lnd_sock.lnd_conns_per_peer = num;
888 case LNET_NET_SOCKLND_TUNABLES_ATTR_LND_TIMEOUT:
889 num = nla_get_s64(attr);
890 tunables->lnd_tun_u.lnd_sock.lnd_timeout = num;
900 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
903 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
904 return conn_cb->ksnr_connecting;
909 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
910 struct socket *sock, int type)
912 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
914 struct lnet_processid peerid;
916 struct ksock_conn *conn;
917 struct ksock_conn *conn2;
918 struct ksock_peer_ni *peer_ni = NULL;
919 struct ksock_peer_ni *peer2;
920 struct ksock_sched *sched;
921 struct ksock_hello_msg *hello;
924 struct ksock_tx *txtmp;
931 active = (conn_cb != NULL);
933 LASSERT(active == (type != SOCKLND_CONN_NONE));
935 LIBCFS_ALLOC(conn, sizeof(*conn));
941 conn->ksnc_peer = NULL;
942 conn->ksnc_conn_cb = NULL;
943 conn->ksnc_sock = sock;
944 /* 2 ref, 1 for conn, another extra ref prevents socket
945 * being closed before establishment of connection */
946 refcount_set(&conn->ksnc_sock_refcount, 2);
947 conn->ksnc_type = type;
948 ksocknal_lib_save_callback(sock, conn);
949 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
951 conn->ksnc_rx_ready = 0;
952 conn->ksnc_rx_scheduled = 0;
954 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
955 conn->ksnc_tx_ready = 0;
956 conn->ksnc_tx_scheduled = 0;
957 conn->ksnc_tx_carrier = NULL;
958 atomic_set (&conn->ksnc_tx_nob, 0);
960 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
961 kshm_ips[LNET_INTERFACES_NUM]));
967 /* stash conn's local and remote addrs */
968 rc = ksocknal_lib_get_conn_addrs(conn);
972 /* Find out/confirm peer_ni's NID and connection type and get the
973 * vector of interfaces she's willing to let me connect to.
974 * Passive connections use the listener timeout since the peer_ni sends
978 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
980 peer_ni = conn_cb->ksnr_peer;
981 LASSERT(ni == peer_ni->ksnp_ni);
983 /* Active connection sends HELLO eagerly */
984 hello->kshm_nips = 0;
985 peerid = peer_ni->ksnp_id;
987 write_lock_bh(global_lock);
988 conn->ksnc_proto = peer_ni->ksnp_proto;
989 write_unlock_bh(global_lock);
991 if (conn->ksnc_proto == NULL) {
992 if (psa->sin_family == AF_INET6)
993 conn->ksnc_proto = &ksocknal_protocol_v4x;
994 else if (psa->sin_family == AF_INET)
995 conn->ksnc_proto = &ksocknal_protocol_v3x;
996 #if SOCKNAL_VERSION_DEBUG
997 if (*ksocknal_tunables.ksnd_protocol == 2)
998 conn->ksnc_proto = &ksocknal_protocol_v2x;
999 else if (*ksocknal_tunables.ksnd_protocol == 1)
1000 conn->ksnc_proto = &ksocknal_protocol_v1x;
1003 if (!conn->ksnc_proto) {
1008 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1012 peerid.nid = LNET_ANY_NID;
1013 peerid.pid = LNET_PID_ANY;
1015 /* Passive, get protocol from peer_ni */
1016 conn->ksnc_proto = NULL;
1019 rc = ksocknal_recv_hello(ni, conn, hello, &peerid, &incarnation);
1023 LASSERT(rc == 0 || active);
1024 LASSERT(conn->ksnc_proto != NULL);
1025 LASSERT(!LNET_NID_IS_ANY(&peerid.nid));
1027 cpt = lnet_nid2cpt(&peerid.nid, ni);
1030 ksocknal_peer_addref(peer_ni);
1031 write_lock_bh(global_lock);
1033 peer_ni = ksocknal_create_peer(ni, &peerid);
1034 if (IS_ERR(peer_ni)) {
1035 rc = PTR_ERR(peer_ni);
1039 write_lock_bh(global_lock);
1041 /* called with a ref on ni, so shutdown can't have started */
1042 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
1044 peer2 = ksocknal_find_peer_locked(ni, &peerid);
1045 if (peer2 == NULL) {
1046 /* NB this puts an "empty" peer_ni in the peer_ni
1047 * table (which takes my ref) */
1048 hash_add(ksocknal_data.ksnd_peers,
1049 &peer_ni->ksnp_list, nidhash(&peerid.nid));
1051 ksocknal_peer_decref(peer_ni);
1056 ksocknal_peer_addref(peer_ni);
1057 peer_ni->ksnp_accepting++;
1059 /* Am I already connecting to this guy? Resolve in
1060 * favour of higher NID...
1062 if (memcmp(&peerid.nid, &ni->ni_nid, sizeof(peerid.nid)) < 0 &&
1063 ksocknal_connecting(peer_ni->ksnp_conn_cb,
1064 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
1066 warn = "connection race resolution";
1071 if (peer_ni->ksnp_closing ||
1072 (active && conn_cb->ksnr_deleted)) {
1073 /* peer_ni/conn_cb got closed under me */
1075 warn = "peer_ni/conn_cb removed";
1079 if (peer_ni->ksnp_proto == NULL) {
1080 /* Never connected before.
1081 * NB recv_hello may have returned EPROTO to signal my peer_ni
1082 * wants a different protocol than the one I asked for.
1084 LASSERT(list_empty(&peer_ni->ksnp_conns));
1086 peer_ni->ksnp_proto = conn->ksnc_proto;
1087 peer_ni->ksnp_incarnation = incarnation;
1090 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1091 peer_ni->ksnp_incarnation != incarnation) {
1092 /* peer_ni rebooted or I've got the wrong protocol version */
1093 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
1095 peer_ni->ksnp_proto = NULL;
1097 warn = peer_ni->ksnp_incarnation != incarnation ?
1098 "peer_ni rebooted" :
1099 "wrong proto version";
1109 warn = "lost conn race";
1112 warn = "retry with different protocol version";
1116 /* Refuse to duplicate an existing connection, unless this is a
1117 * loopback connection */
1118 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1119 (struct sockaddr *)&conn->ksnc_myaddr)) {
1120 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1122 (struct sockaddr *)&conn2->ksnc_peeraddr,
1123 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1125 (struct sockaddr *)&conn2->ksnc_myaddr,
1126 (struct sockaddr *)&conn->ksnc_myaddr) ||
1127 conn2->ksnc_type != conn->ksnc_type)
1131 /* If max conns per type is not registered in conn_cb
1132 * as ksnr_max_conns, use ni's conns_per_peer
1134 if ((peer_ni->ksnp_conn_cb &&
1135 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1136 (!peer_ni->ksnp_conn_cb &&
1137 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1140 /* Reply on a passive connection attempt so the peer_ni
1141 * realises we're connected.
1151 /* If the connection created by this route didn't bind to the IP
1152 * address the route connected to, the connection/route matching
1153 * code below probably isn't going to work.
1156 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1157 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1158 CERROR("Route %s %pISc connected to %pISc\n",
1159 libcfs_idstr(&peer_ni->ksnp_id),
1160 &conn_cb->ksnr_addr,
1161 &conn->ksnc_peeraddr);
1164 /* Search for a conn_cb corresponding to the new connection and
1165 * create an association. This allows incoming connections created
1166 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1167 * continually create duplicate conn_cbs.
1169 conn_cb = peer_ni->ksnp_conn_cb;
1171 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1172 (struct sockaddr *)&conn_cb->ksnr_addr))
1173 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1175 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1176 peer_ni->ksnp_last_alive = ktime_get_seconds();
1177 peer_ni->ksnp_send_keepalive = 0;
1178 peer_ni->ksnp_error = 0;
1180 sched = ksocknal_choose_scheduler_locked(cpt);
1182 CERROR("no schedulers available. node is unhealthy\n");
1186 * The cpt might have changed if we ended up selecting a non cpt
1187 * native scheduler. So use the scheduler's cpt instead.
1189 cpt = sched->kss_cpt;
1190 sched->kss_nconns++;
1191 conn->ksnc_scheduler = sched;
1193 conn->ksnc_tx_last_post = ktime_get_seconds();
1194 /* Set the deadline for the outgoing HELLO to drain */
1195 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1196 conn->ksnc_tx_deadline = ktime_get_seconds() +
1198 smp_mb(); /* order with adding to peer_ni's conn list */
1200 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1201 ksocknal_conn_addref(conn);
1203 ksocknal_new_packet(conn, 0);
1205 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1207 /* Take packets blocking for this connection. */
1208 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1209 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1213 list_del(&tx->tx_list);
1214 ksocknal_queue_tx_locked(tx, conn);
1217 write_unlock_bh(global_lock);
1218 /* We've now got a new connection. Any errors from here on are just
1219 * like "normal" comms errors and we close the connection normally.
1220 * NB (a) we still have to send the reply HELLO for passive
1222 * (b) normal I/O on the conn is blocked until I setup and call the
1226 CDEBUG(D_NET, "New conn %s p %d.x %pISc -> %pIScp"
1227 " incarnation:%lld sched[%d]\n",
1228 libcfs_idstr(&peerid), conn->ksnc_proto->pro_version,
1229 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1233 hello->kshm_nips = 0;
1234 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1237 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1238 kshm_ips[LNET_INTERFACES_NUM]));
1240 /* setup the socket AFTER I've received hello (it disables
1241 * SO_LINGER). I might call back to the acceptor who may want
1242 * to send a protocol version response and then close the
1243 * socket; this ensures the socket only tears down after the
1244 * response has been sent.
1247 rc = ksocknal_lib_setup_sock(sock);
1249 write_lock_bh(global_lock);
1251 /* NB my callbacks block while I hold ksnd_global_lock */
1252 ksocknal_lib_set_callback(sock, conn);
1255 peer_ni->ksnp_accepting--;
1257 write_unlock_bh(global_lock);
1260 write_lock_bh(global_lock);
1261 if (!conn->ksnc_closing) {
1262 /* could be closed by another thread */
1263 ksocknal_close_conn_locked(conn, rc);
1265 write_unlock_bh(global_lock);
1266 } else if (ksocknal_connsock_addref(conn) == 0) {
1267 /* Allow I/O to proceed. */
1268 ksocknal_read_callback(conn);
1269 ksocknal_write_callback(conn);
1270 ksocknal_connsock_decref(conn);
1273 ksocknal_connsock_decref(conn);
1274 ksocknal_conn_decref(conn);
1279 if (!peer_ni->ksnp_closing &&
1280 list_empty(&peer_ni->ksnp_conns) &&
1281 peer_ni->ksnp_conn_cb == NULL) {
1282 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1283 ksocknal_unlink_peer_locked(peer_ni);
1286 write_unlock_bh(global_lock);
1290 CERROR("Not creating conn %s type %d: %s\n",
1291 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1293 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1294 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1299 /* Request retry by replying with CONN_NONE
1300 * ksnc_proto has been set already
1302 conn->ksnc_type = SOCKLND_CONN_NONE;
1303 hello->kshm_nips = 0;
1304 ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1307 write_lock_bh(global_lock);
1308 peer_ni->ksnp_accepting--;
1309 write_unlock_bh(global_lock);
1313 * If we get here without an error code, just use -EALREADY.
1314 * Depending on how we got here, the error may be positive
1315 * or negative. Normalize the value for ksocknal_txlist_done().
1317 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1318 ksocknal_txlist_done(ni, &zombies, rc2);
1319 ksocknal_peer_decref(peer_ni);
1323 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1324 kshm_ips[LNET_INTERFACES_NUM]));
1326 LIBCFS_FREE(conn, sizeof(*conn));
1335 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1337 /* This just does the immmediate housekeeping, and queues the
1338 * connection for the reaper to terminate.
1339 * Caller holds ksnd_global_lock exclusively in irq context */
1340 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1341 struct ksock_conn_cb *conn_cb;
1342 struct ksock_conn *conn2;
1344 int duplicate_count = 0;
1346 LASSERT(peer_ni->ksnp_error == 0);
1347 LASSERT(!conn->ksnc_closing);
1348 conn->ksnc_closing = 1;
1350 /* ksnd_deathrow_conns takes over peer_ni's ref */
1351 list_del(&conn->ksnc_list);
1353 conn_cb = conn->ksnc_conn_cb;
1354 if (conn_cb != NULL) {
1355 /* dissociate conn from cb... */
1356 LASSERT(!conn_cb->ksnr_deleted);
1358 conn_count = ksocknal_get_conn_count_by_type(conn_cb,
1360 /* connected bit is set only if all connections
1361 * of the given type got created
1363 if (conn_count == conn_cb->ksnr_max_conns)
1364 LASSERT((conn_cb->ksnr_connected &
1365 BIT(conn->ksnc_type)) != 0);
1367 if (conn_count == 1) {
1368 list_for_each_entry(conn2, &peer_ni->ksnp_conns,
1370 if (conn2->ksnc_conn_cb == conn_cb &&
1371 conn2->ksnc_type == conn->ksnc_type)
1372 duplicate_count += 1;
1374 if (duplicate_count > 0)
1375 CERROR("Found %d duplicate conns type %d\n",
1379 ksocknal_decr_conn_count(conn_cb, conn->ksnc_type);
1381 conn->ksnc_conn_cb = NULL;
1383 /* drop conn's ref on conn_cb */
1384 ksocknal_conn_cb_decref(conn_cb);
1387 if (list_empty(&peer_ni->ksnp_conns)) {
1388 /* No more connections to this peer_ni */
1390 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1391 struct ksock_tx *tx;
1393 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1395 /* throw them to the last connection...,
1396 * these TXs will be send to /dev/null by scheduler */
1397 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1399 ksocknal_tx_prep(conn, tx);
1401 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1402 list_splice_init(&peer_ni->ksnp_tx_queue,
1403 &conn->ksnc_tx_queue);
1404 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1407 /* renegotiate protocol version */
1408 peer_ni->ksnp_proto = NULL;
1409 /* stash last conn close reason */
1410 peer_ni->ksnp_error = error;
1412 if (peer_ni->ksnp_conn_cb == NULL) {
1413 /* I've just closed last conn belonging to a
1414 * peer_ni with no connections to it
1416 ksocknal_unlink_peer_locked(peer_ni);
1420 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1422 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1423 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1425 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1429 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1431 bool notify = false;
1432 time64_t last_alive = 0;
1434 /* There has been a connection failure or comms error; but I'll only
1435 * tell LNET I think the peer_ni is dead if it's to another kernel and
1436 * there are no connections or connection attempts in existence. */
1438 read_lock(&ksocknal_data.ksnd_global_lock);
1440 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1441 list_empty(&peer_ni->ksnp_conns) &&
1442 peer_ni->ksnp_accepting == 0 &&
1443 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1445 last_alive = peer_ni->ksnp_last_alive;
1448 read_unlock(&ksocknal_data.ksnd_global_lock);
1451 lnet_notify(peer_ni->ksnp_ni,
1452 &peer_ni->ksnp_id.nid,
1453 false, false, last_alive);
1457 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1459 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1460 struct ksock_tx *tx;
1461 struct ksock_tx *tmp;
1464 /* NB safe to finalize TXs because closing of socket will
1465 * abort all buffered data */
1466 LASSERT(conn->ksnc_sock == NULL);
1468 spin_lock(&peer_ni->ksnp_lock);
1470 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1472 if (tx->tx_conn != conn)
1475 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1477 tx->tx_msg.ksm_zc_cookies[0] = 0;
1478 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1479 list_move(&tx->tx_zc_list, &zlist);
1482 spin_unlock(&peer_ni->ksnp_lock);
1484 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1485 tx_zc_list)) != NULL) {
1486 list_del(&tx->tx_zc_list);
1487 ksocknal_tx_decref(tx);
1492 ksocknal_terminate_conn(struct ksock_conn *conn)
1494 /* This gets called by the reaper (guaranteed thread context) to
1495 * disengage the socket from its callbacks and close it.
1496 * ksnc_refcount will eventually hit zero, and then the reaper will
1499 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1500 struct ksock_sched *sched = conn->ksnc_scheduler;
1501 bool failed = false;
1503 LASSERT(conn->ksnc_closing);
1505 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1506 spin_lock_bh(&sched->kss_lock);
1508 /* a closing conn is always ready to tx */
1509 conn->ksnc_tx_ready = 1;
1511 if (!conn->ksnc_tx_scheduled &&
1512 !list_empty(&conn->ksnc_tx_queue)) {
1513 list_add_tail(&conn->ksnc_tx_list,
1514 &sched->kss_tx_conns);
1515 conn->ksnc_tx_scheduled = 1;
1516 /* extra ref for scheduler */
1517 ksocknal_conn_addref(conn);
1519 wake_up(&sched->kss_waitq);
1522 spin_unlock_bh(&sched->kss_lock);
1524 /* serialise with callbacks */
1525 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1527 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1529 /* OK, so this conn may not be completely disengaged from its
1530 * scheduler yet, but it _has_ committed to terminate...
1532 conn->ksnc_scheduler->kss_nconns--;
1534 if (peer_ni->ksnp_error != 0) {
1535 /* peer_ni's last conn closed in error */
1536 LASSERT(list_empty(&peer_ni->ksnp_conns));
1538 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1541 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1544 ksocknal_peer_failed(peer_ni);
1546 /* The socket is closed on the final put; either here, or in
1547 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1548 * when the connection was established, this will close the socket
1549 * immediately, aborting anything buffered in it. Any hung
1550 * zero-copy transmits will therefore complete in finite time.
1552 ksocknal_connsock_decref(conn);
1556 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1558 /* Queue the conn for the reaper to destroy */
1559 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1560 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1562 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1563 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1565 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1569 ksocknal_destroy_conn(struct ksock_conn *conn)
1573 /* Final coup-de-grace of the reaper */
1574 CDEBUG(D_NET, "connection %p\n", conn);
1576 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1577 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1578 LASSERT(conn->ksnc_sock == NULL);
1579 LASSERT(conn->ksnc_conn_cb == NULL);
1580 LASSERT(!conn->ksnc_tx_scheduled);
1581 LASSERT(!conn->ksnc_rx_scheduled);
1582 LASSERT(list_empty(&conn->ksnc_tx_queue));
1584 /* complete current receive if any */
1585 switch (conn->ksnc_rx_state) {
1586 case SOCKNAL_RX_LNET_PAYLOAD:
1587 last_rcv = conn->ksnc_rx_deadline -
1589 CERROR("Completing partial receive from %s[%d], ip %pIScp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1590 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1592 &conn->ksnc_peeraddr,
1593 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1594 ktime_get_seconds() - last_rcv);
1595 if (conn->ksnc_lnet_msg)
1596 conn->ksnc_lnet_msg->msg_health_status =
1597 LNET_MSG_STATUS_REMOTE_ERROR;
1598 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1600 case SOCKNAL_RX_LNET_HEADER:
1601 if (conn->ksnc_rx_started)
1602 CERROR("Incomplete receive of lnet header from %s, ip %pIScp, with error, protocol: %d.x.\n",
1603 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1604 &conn->ksnc_peeraddr,
1605 conn->ksnc_proto->pro_version);
1607 case SOCKNAL_RX_KSM_HEADER:
1608 if (conn->ksnc_rx_started)
1609 CERROR("Incomplete receive of ksock message from %s, ip %pIScp, with error, protocol: %d.x.\n",
1610 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1611 &conn->ksnc_peeraddr,
1612 conn->ksnc_proto->pro_version);
1614 case SOCKNAL_RX_SLOP:
1615 if (conn->ksnc_rx_started)
1616 CERROR("Incomplete receive of slops from %s, ip %pIScp, with error\n",
1617 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1618 &conn->ksnc_peeraddr);
1625 ksocknal_peer_decref(conn->ksnc_peer);
1627 LIBCFS_FREE(conn, sizeof(*conn));
1631 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1632 struct sockaddr *addr, int why)
1634 struct ksock_conn *conn;
1635 struct ksock_conn *cnxt;
1638 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1641 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1643 ksocknal_close_conn_locked(conn, why);
1651 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1653 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1656 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1658 count = ksocknal_close_peer_conns_locked(
1659 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1661 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1667 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1669 struct ksock_peer_ni *peer_ni;
1670 struct hlist_node *pnxt;
1675 struct sockaddr_in sa = {.sin_family = AF_INET};
1677 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1679 if (!LNET_NID_IS_ANY(&id->nid)) {
1680 lo = hash_min(nidhash(&id->nid),
1681 HASH_BITS(ksocknal_data.ksnd_peers));
1685 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1688 sa.sin_addr.s_addr = htonl(ipaddr);
1689 for (i = lo; i <= hi; i++) {
1690 hlist_for_each_entry_safe(peer_ni, pnxt,
1691 &ksocknal_data.ksnd_peers[i],
1694 if (!((LNET_NID_IS_ANY(&id->nid) ||
1695 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1696 (id->pid == LNET_PID_ANY ||
1697 id->pid == peer_ni->ksnp_id.pid)))
1700 count += ksocknal_close_peer_conns_locked(
1702 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1706 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1708 /* wildcards always succeed */
1709 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1713 return (count == 0 ? -ENOENT : 0);
1717 ksocknal_notify_gw_down(struct lnet_nid *gw_nid)
1719 /* The router is telling me she's been notified of a change in
1722 struct lnet_processid id = {
1723 .pid = LNET_PID_ANY,
1727 CDEBUG(D_NET, "gw %s down\n", libcfs_nidstr(gw_nid));
1729 /* If the gateway crashed, close all open connections... */
1730 ksocknal_close_matching_conns(&id, 0);
1733 /* We can only establish new connections
1734 * if we have autroutes, and these connect on demand.
1739 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1743 struct ksock_conn *conn;
1745 for (index = 0; ; index++) {
1746 read_lock(&ksocknal_data.ksnd_global_lock);
1751 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1753 ksocknal_conn_addref(conn);
1758 read_unlock(&ksocknal_data.ksnd_global_lock);
1763 ksocknal_lib_push_conn (conn);
1764 ksocknal_conn_decref(conn);
1769 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1776 if (!LNET_NID_IS_ANY(&id->nid)) {
1777 lo = hash_min(nidhash(&id->nid),
1778 HASH_BITS(ksocknal_data.ksnd_peers));
1782 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1785 for (bkt = lo; bkt <= hi; bkt++) {
1786 int peer_off; /* searching offset in peer_ni hash table */
1788 for (peer_off = 0; ; peer_off++) {
1789 struct ksock_peer_ni *peer_ni;
1792 read_lock(&ksocknal_data.ksnd_global_lock);
1793 hlist_for_each_entry(peer_ni,
1794 &ksocknal_data.ksnd_peers[bkt],
1796 if (!((LNET_NID_IS_ANY(&id->nid) ||
1798 &peer_ni->ksnp_id.nid)) &&
1799 (id->pid == LNET_PID_ANY ||
1800 id->pid == peer_ni->ksnp_id.pid)))
1803 if (i++ == peer_off) {
1804 ksocknal_peer_addref(peer_ni);
1808 read_unlock(&ksocknal_data.ksnd_global_lock);
1810 if (i <= peer_off) /* no match */
1814 ksocknal_push_peer(peer_ni);
1815 ksocknal_peer_decref(peer_ni);
1822 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1824 struct lnet_processid id = {};
1825 struct libcfs_ioctl_data *data = arg;
1829 case IOC_LIBCFS_GET_INTERFACE: {
1830 struct ksock_net *net = ni->ni_data;
1831 struct ksock_interface *iface;
1832 struct sockaddr_in *sa;
1834 read_lock(&ksocknal_data.ksnd_global_lock);
1836 if (data->ioc_count >= 1) {
1840 iface = &net->ksnn_interface;
1842 sa = (void *)&iface->ksni_addr;
1843 if (sa->sin_family == AF_INET) {
1844 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1845 data->ioc_u32[1] = iface->ksni_netmask;
1847 data->ioc_u32[0] = 0xFFFFFFFF;
1848 data->ioc_u32[1] = 0;
1850 data->ioc_u32[2] = iface->ksni_npeers;
1851 data->ioc_u32[3] = iface->ksni_nroutes;
1854 read_unlock(&ksocknal_data.ksnd_global_lock);
1858 case IOC_LIBCFS_GET_PEER: {
1863 int share_count = 0;
1865 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1866 &id, &myip, &ip, &port,
1867 &conn_count, &share_count);
1871 if (!nid_is_nid4(&id.nid))
1873 data->ioc_nid = lnet_nid_to_nid4(&id.nid);
1874 data->ioc_count = share_count;
1875 data->ioc_u32[0] = ip;
1876 data->ioc_u32[1] = port;
1877 data->ioc_u32[2] = myip;
1878 data->ioc_u32[3] = conn_count;
1879 data->ioc_u32[4] = id.pid;
1883 case IOC_LIBCFS_ADD_PEER: {
1884 struct sockaddr_in sa = {.sin_family = AF_INET};
1886 id.pid = LNET_PID_LUSTRE;
1887 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1888 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1889 sa.sin_port = htons(data->ioc_u32[1]);
1890 return ksocknal_add_peer(ni, &id, (struct sockaddr *)&sa);
1892 case IOC_LIBCFS_DEL_PEER:
1893 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1894 id.pid = LNET_PID_ANY;
1895 return ksocknal_del_peer(ni, &id);
1897 case IOC_LIBCFS_GET_CONN: {
1901 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1902 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1903 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1908 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1910 data->ioc_count = txmem;
1911 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1912 data->ioc_flags = nagle;
1913 if (psa->sin_family == AF_INET)
1914 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1916 data->ioc_u32[0] = 0xFFFFFFFF;
1917 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1918 &conn->ksnc_peeraddr);
1919 if (mysa->sin_family == AF_INET)
1920 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1922 data->ioc_u32[2] = 0xFFFFFFFF;
1923 data->ioc_u32[3] = conn->ksnc_type;
1924 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1925 data->ioc_u32[5] = rxmem;
1926 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1927 ksocknal_conn_decref(conn);
1931 case IOC_LIBCFS_CLOSE_CONNECTION:
1932 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1933 id.pid = LNET_PID_ANY;
1934 return ksocknal_close_matching_conns(&id,
1937 case IOC_LIBCFS_REGISTER_MYNID:
1938 /* Ignore if this is a noop */
1939 if (nid_is_nid4(&ni->ni_nid) &&
1940 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1943 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1944 libcfs_nid2str(data->ioc_nid),
1945 libcfs_nidstr(&ni->ni_nid));
1948 case IOC_LIBCFS_PUSH_CONNECTION:
1949 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1950 id.pid = LNET_PID_ANY;
1951 return ksocknal_push(ni, &id);
1960 ksocknal_free_buffers (void)
1962 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1964 if (ksocknal_data.ksnd_schedulers != NULL)
1965 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1967 spin_lock(&ksocknal_data.ksnd_tx_lock);
1969 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1971 struct ksock_tx *tx;
1973 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1974 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1976 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1977 tx_list)) != NULL) {
1978 list_del(&tx->tx_list);
1979 LIBCFS_FREE(tx, tx->tx_desc_size);
1982 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1987 ksocknal_handle_link_state_change(struct net_device *dev,
1988 unsigned char operstate)
1990 struct lnet_ni *ni = NULL;
1991 struct ksock_net *net;
1992 struct ksock_net *cnxt;
1994 unsigned char link_down;
1995 struct in_device *in_dev;
1996 bool found_ip = false;
1997 struct ksock_interface *ksi = NULL;
1998 struct sockaddr_in *sa;
1999 __u32 ni_state_before;
2000 bool update_ping_buf = false;
2002 DECLARE_CONST_IN_IFADDR(ifa);
2004 link_down = !((operstate == IF_OPER_UP) || (operstate == IF_OPER_UNKNOWN));
2005 ifindex = dev->ifindex;
2007 if (!ksocknal_data.ksnd_nnets)
2010 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
2013 ksi = &net->ksnn_interface;
2014 sa = (void *)&ksi->ksni_addr;
2017 if (strcmp(ksi->ksni_name, dev->name))
2020 if (ksi->ksni_index == -1) {
2021 if (dev->reg_state != NETREG_REGISTERED)
2023 /* A registration just happened: save the new index for
2025 ksi->ksni_index = ifindex;
2029 if (ksi->ksni_index != ifindex)
2032 if (dev->reg_state == NETREG_UNREGISTERING) {
2033 /* Device is being unregistered, we need to clear the
2034 * index, it can change when device will be back */
2035 ksi->ksni_index = -1;
2041 in_dev = __in_dev_get_rtnl(dev);
2043 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
2045 ni_state_before = lnet_set_link_fatal_state(ni, 1);
2048 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
2049 if (sa->sin_addr.s_addr == ifa->ifa_local)
2055 CDEBUG(D_NET, "Interface %s has no matching ip\n",
2057 ni_state_before = lnet_set_link_fatal_state(ni, 1);
2062 ni_state_before = lnet_set_link_fatal_state(ni, 1);
2064 state = (lnet_get_link_status(dev) == 0);
2065 ni_state_before = lnet_set_link_fatal_state(ni,
2069 if (!update_ping_buf &&
2070 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
2071 (atomic_read(&ni->ni_fatal_error_on) != ni_state_before))
2072 update_ping_buf = true;
2075 if (update_ping_buf)
2076 lnet_mark_ping_buffer_for_update();
2083 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
2085 struct lnet_ni *ni = NULL;
2086 struct ksock_net *net;
2087 struct ksock_net *cnxt;
2088 struct net_device *event_netdev = ifa->ifa_dev->dev;
2090 struct ksock_interface *ksi = NULL;
2091 struct sockaddr_in *sa;
2092 __u32 ni_state_before;
2093 bool update_ping_buf = false;
2096 if (!ksocknal_data.ksnd_nnets)
2099 ifindex = event_netdev->ifindex;
2101 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
2104 ksi = &net->ksnn_interface;
2105 sa = (void *)&ksi->ksni_addr;
2107 if (ksi->ksni_index != ifindex ||
2108 strcmp(ksi->ksni_name, event_netdev->name))
2111 if (sa->sin_addr.s_addr == ifa->ifa_local) {
2113 link_down = (event == NETDEV_DOWN);
2114 ni_state_before = lnet_set_link_fatal_state(ni,
2117 if (!update_ping_buf &&
2118 (ni->ni_state == LNET_NI_STATE_ACTIVE) &&
2119 ((event == NETDEV_DOWN) != ni_state_before))
2120 update_ping_buf = true;
2124 if (update_ping_buf)
2125 lnet_mark_ping_buffer_for_update();
2130 /************************************
2131 * Net device notifier event handler
2132 ************************************/
2133 static int ksocknal_device_event(struct notifier_block *unused,
2134 unsigned long event, void *ptr)
2136 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2137 unsigned char operstate;
2139 operstate = dev->operstate;
2141 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2142 event, dev->name, dev->ifindex, operstate);
2148 case NETDEV_REGISTER:
2149 case NETDEV_UNREGISTER:
2150 ksocknal_handle_link_state_change(dev, operstate);
2157 /************************************
2158 * Inetaddr notifier event handler
2159 ************************************/
2160 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2161 unsigned long event, void *ptr)
2163 struct in_ifaddr *ifa = ptr;
2165 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2166 event, &ifa->ifa_address, &ifa->ifa_mask);
2172 ksocknal_handle_inetaddr_change(ifa, event);
2179 static struct notifier_block ksocknal_dev_notifier_block = {
2180 .notifier_call = ksocknal_device_event,
2183 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2184 .notifier_call = ksocknal_inetaddr_event,
2188 ksocknal_base_shutdown(void)
2190 struct ksock_sched *sched;
2191 struct ksock_peer_ni *peer_ni;
2194 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2195 libcfs_kmem_read());
2196 LASSERT (ksocknal_data.ksnd_nnets == 0);
2198 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2199 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2200 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2203 switch (ksocknal_data.ksnd_init) {
2208 case SOCKNAL_INIT_ALL:
2209 case SOCKNAL_INIT_DATA:
2210 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2213 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2214 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2215 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2216 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2217 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2219 if (ksocknal_data.ksnd_schedulers != NULL) {
2220 cfs_percpt_for_each(sched, i,
2221 ksocknal_data.ksnd_schedulers) {
2223 LASSERT(list_empty(&sched->kss_tx_conns));
2224 LASSERT(list_empty(&sched->kss_rx_conns));
2225 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2226 LASSERT(sched->kss_nconns == 0);
2230 /* flag threads to terminate; wake and wait for them to die */
2231 ksocknal_data.ksnd_shuttingdown = 1;
2232 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2233 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2235 if (ksocknal_data.ksnd_schedulers != NULL) {
2236 cfs_percpt_for_each(sched, i,
2237 ksocknal_data.ksnd_schedulers)
2238 wake_up_all(&sched->kss_waitq);
2241 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2242 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2243 "waiting for %d threads to terminate\n",
2244 atomic_read(&ksocknal_data.ksnd_nthreads));
2246 ksocknal_free_buffers();
2248 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2252 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2253 libcfs_kmem_read());
2255 module_put(THIS_MODULE);
2259 ksocknal_base_startup(void)
2261 struct ksock_sched *sched;
2265 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2266 LASSERT(ksocknal_data.ksnd_nnets == 0);
2268 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2270 hash_init(ksocknal_data.ksnd_peers);
2272 rwlock_init(&ksocknal_data.ksnd_global_lock);
2273 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2275 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2276 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2277 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2278 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2279 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2281 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2282 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2283 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2284 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2286 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2287 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2289 /* NB memset above zeros whole of ksocknal_data */
2291 /* flag lists/ptrs/locks initialised */
2292 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2293 if (!try_module_get(THIS_MODULE))
2296 /* Create a scheduler block per available CPT */
2297 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2299 if (ksocknal_data.ksnd_schedulers == NULL)
2302 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2306 * make sure not to allocate more threads than there are
2307 * cores/CPUs in teh CPT
2309 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2310 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2311 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2314 * max to half of CPUs, assume another half should be
2315 * reserved for upper layer modules
2317 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2320 sched->kss_nthreads_max = nthrs;
2323 spin_lock_init(&sched->kss_lock);
2324 INIT_LIST_HEAD(&sched->kss_rx_conns);
2325 INIT_LIST_HEAD(&sched->kss_tx_conns);
2326 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2327 init_waitqueue_head(&sched->kss_waitq);
2330 ksocknal_data.ksnd_connd_starting = 0;
2331 ksocknal_data.ksnd_connd_failed_stamp = 0;
2332 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2333 /* must have at least 2 connds to remain responsive to accepts while
2335 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2336 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2338 if (*ksocknal_tunables.ksnd_nconnds_max <
2339 *ksocknal_tunables.ksnd_nconnds) {
2340 ksocknal_tunables.ksnd_nconnds_max =
2341 ksocknal_tunables.ksnd_nconnds;
2344 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2345 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2346 ksocknal_data.ksnd_connd_starting++;
2347 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2349 rc = ksocknal_thread_start(ksocknal_connd,
2350 (void *)((uintptr_t)i),
2351 "socknal_cd%02d", i);
2353 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2354 ksocknal_data.ksnd_connd_starting--;
2355 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2356 CERROR("Can't spawn socknal connd: %d\n", rc);
2361 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2363 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2367 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2368 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2370 /* flag everything initialised */
2371 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2376 ksocknal_base_shutdown();
2381 ksocknal_debug_peerhash(struct lnet_ni *ni)
2383 struct ksock_peer_ni *peer_ni;
2386 read_lock(&ksocknal_data.ksnd_global_lock);
2388 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2389 struct ksock_conn_cb *conn_cb;
2390 struct ksock_conn *conn;
2392 if (peer_ni->ksnp_ni != ni)
2395 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2396 libcfs_idstr(&peer_ni->ksnp_id),
2397 refcount_read(&peer_ni->ksnp_refcount),
2398 peer_ni->ksnp_closing,
2399 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2400 peer_ni->ksnp_zc_next_cookie,
2401 !list_empty(&peer_ni->ksnp_tx_queue),
2402 !list_empty(&peer_ni->ksnp_zc_req_list));
2404 conn_cb = peer_ni->ksnp_conn_cb;
2406 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2407 refcount_read(&conn_cb->ksnr_refcount),
2408 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2409 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2412 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2413 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2414 refcount_read(&conn->ksnc_conn_refcount),
2415 refcount_read(&conn->ksnc_sock_refcount),
2416 conn->ksnc_type, conn->ksnc_closing);
2421 read_unlock(&ksocknal_data.ksnd_global_lock);
2426 ksocknal_shutdown(struct lnet_ni *ni)
2428 struct ksock_net *net = ni->ni_data;
2430 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2431 LASSERT(ksocknal_data.ksnd_nnets > 0);
2433 /* prevent new peers */
2434 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2436 /* Delete all peers */
2437 ksocknal_del_peer(ni, NULL);
2439 /* Wait for all peer_ni state to clean up */
2440 wait_var_event_warning(&net->ksnn_npeers,
2441 atomic_read(&net->ksnn_npeers) ==
2442 SOCKNAL_SHUTDOWN_BIAS,
2443 "waiting for %d peers to disconnect\n",
2444 ksocknal_debug_peerhash(ni) +
2445 atomic_read(&net->ksnn_npeers) -
2446 SOCKNAL_SHUTDOWN_BIAS);
2448 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2449 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2451 list_del(&net->ksnn_list);
2452 LIBCFS_FREE(net, sizeof(*net));
2454 ksocknal_data.ksnd_nnets--;
2455 if (ksocknal_data.ksnd_nnets == 0)
2456 ksocknal_base_shutdown();
2460 ksocknal_search_new_ipif(struct ksock_net *net)
2463 char *ifnam = &net->ksnn_interface.ksni_name[0];
2464 char *colon = strchr(ifnam, ':');
2466 struct ksock_net *tmp;
2471 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2472 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2473 char *colon2 = strchr(ifnam2, ':');
2478 found = strcmp(ifnam, ifnam2) == 0;
2491 ksocknal_start_schedulers(struct ksock_sched *sched)
2497 if (sched->kss_nthreads == 0) {
2498 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2499 nthrs = sched->kss_nthreads_max;
2501 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2503 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2504 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2506 nthrs = min(nthrs, sched->kss_nthreads_max);
2508 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2509 /* increase two threads if there is new interface */
2510 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2513 for (i = 0; i < nthrs; i++) {
2516 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2517 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2518 "socknal_sd%02d_%02d",
2520 (int)KSOCK_THREAD_SID(id));
2524 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2525 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2529 sched->kss_nthreads += i;
2534 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2536 int newif = ksocknal_search_new_ipif(net);
2540 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2543 for (i = 0; i < ncpts; i++) {
2544 struct ksock_sched *sched;
2545 int cpt = (cpts == NULL) ? i : cpts[i];
2547 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2548 sched = ksocknal_data.ksnd_schedulers[cpt];
2550 if (!newif && sched->kss_nthreads > 0)
2553 rc = ksocknal_start_schedulers(sched);
2561 ksocknal_startup(struct lnet_ni *ni)
2563 struct ksock_net *net;
2564 struct ksock_interface *ksi = NULL;
2565 struct lnet_inetdev *ifaces = NULL;
2569 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2570 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2571 rc = ksocknal_base_startup();
2575 LIBCFS_ALLOC(net, sizeof(*net));
2579 net->ksnn_incarnation = ktime_get_real_ns();
2582 ksocknal_tunables_setup(ni);
2584 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, true);
2588 ksi = &net->ksnn_interface;
2590 /* Interface and/or IP address is specified otherwise default to
2591 * the first Interface
2593 if_idx = lnet_inet_select(ni, ifaces, rc);
2597 if (!ni->ni_interface) {
2598 rc = lnet_ni_add_interface(ni, ifaces[if_idx].li_name);
2600 CWARN("ksocklnd failed to allocate ni_interface\n");
2603 ni->ni_dev_cpt = ifaces[if_idx].li_cpt;
2604 ksi->ksni_index = ifaces[if_idx].li_index;
2605 if (ifaces[if_idx].li_size == sizeof(struct in6_addr)) {
2606 struct sockaddr_in6 *sa;
2607 sa = (void *)&ksi->ksni_addr;
2608 memset(sa, 0, sizeof(*sa));
2609 sa->sin6_family = AF_INET6;
2610 memcpy(&sa->sin6_addr, ifaces[if_idx].li_ipv6addr,
2611 sizeof(struct in6_addr));
2612 ni->ni_nid.nid_size = sizeof(struct in6_addr) - 4;
2613 memcpy(&ni->ni_nid.nid_addr, ifaces[if_idx].li_ipv6addr,
2614 sizeof(struct in6_addr));
2616 struct sockaddr_in *sa;
2617 sa = (void *)&ksi->ksni_addr;
2618 memset(sa, 0, sizeof(*sa));
2619 sa->sin_family = AF_INET;
2620 sa->sin_addr.s_addr = ifaces[if_idx].li_ipaddr;
2621 ksi->ksni_netmask = ifaces[if_idx].li_netmask;
2622 ni->ni_nid.nid_size = 0;
2623 ni->ni_nid.nid_addr[0] = sa->sin_addr.s_addr;
2625 strlcpy(ksi->ksni_name, ifaces[if_idx].li_name, sizeof(ksi->ksni_name));
2627 /* call it before add it to ksocknal_data.ksnd_nets */
2628 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2632 if ((ksocknal_ip2index((struct sockaddr *)&ksi->ksni_addr,
2634 &dev_status) < 0) ||
2636 lnet_set_link_fatal_state(ni, 1);
2638 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2640 ksocknal_data.ksnd_nnets++;
2645 LIBCFS_FREE(net, sizeof(*net));
2647 if (ksocknal_data.ksnd_nnets == 0)
2648 ksocknal_base_shutdown();
2653 static void __exit ksocklnd_exit(void)
2655 lnet_unregister_lnd(&the_ksocklnd);
2658 static const struct lnet_lnd the_ksocklnd = {
2659 .lnd_type = SOCKLND,
2660 .lnd_startup = ksocknal_startup,
2661 .lnd_shutdown = ksocknal_shutdown,
2662 .lnd_ctl = ksocknal_ctl,
2663 .lnd_send = ksocknal_send,
2664 .lnd_recv = ksocknal_recv,
2665 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2666 .lnd_accept = ksocknal_accept,
2667 .lnd_nl_get = ksocknal_nl_get,
2668 .lnd_nl_set = ksocknal_nl_set,
2669 .lnd_keys = &ksocknal_tunables_keys,
2672 static int __init ksocklnd_init(void)
2676 /* check ksnr_connected/connecting field large enough */
2677 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2678 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2680 rc = ksocknal_tunables_init();
2684 rc = libcfs_setup();
2688 lnet_register_lnd(&the_ksocklnd);
2693 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2694 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2695 MODULE_VERSION("2.8.0");
2696 MODULE_LICENSE("GPL");
2698 module_init(ksocklnd_init);
2699 module_exit(ksocklnd_exit);