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>
42 #include <linux/sunrpc/addr.h>
43 #include <net/addrconf.h>
45 static const struct lnet_lnd the_ksocklnd;
46 struct ksock_nal_data ksocknal_data;
48 static struct ksock_interface *
49 ksocknal_index2iface(struct lnet_ni *ni, int index)
51 struct ksock_net *net = ni->ni_data;
52 struct ksock_interface *iface;
54 iface = &net->ksnn_interface;
56 if (iface->ksni_index == index)
62 static int ksocknal_ip2index(struct sockaddr *addr, struct lnet_ni *ni)
64 struct net_device *dev;
66 DECLARE_CONST_IN_IFADDR(ifa);
68 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
72 for_each_netdev(ni->ni_net_ns, dev) {
73 int flags = dev_get_flags(dev);
74 struct in_device *in_dev;
76 if (flags & IFF_LOOPBACK) /* skip the loopback IF */
79 if (!(flags & IFF_UP))
82 switch (addr->sa_family) {
84 in_dev = __in_dev_get_rcu(dev);
88 in_dev_for_each_ifa_rcu(ifa, in_dev) {
90 ((struct sockaddr_in *)addr)->sin_addr.s_addr)
95 #if IS_ENABLED(CONFIG_IPV6)
97 struct inet6_dev *in6_dev;
98 const struct inet6_ifaddr *ifa6;
99 struct sockaddr_in6 *addr6 = (struct sockaddr_in6*)addr;
101 in6_dev = __in6_dev_get(dev);
105 list_for_each_entry_rcu(ifa6, &in6_dev->addr_list, if_list) {
106 if (ipv6_addr_cmp(&ifa6->addr,
107 &addr6->sin6_addr) == 0)
112 #endif /* IS_ENABLED(CONFIG_IPV6) */
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_myiface = -1;
138 conn_cb->ksnr_scheduled = 0;
139 conn_cb->ksnr_connecting = 0;
140 conn_cb->ksnr_connected = 0;
141 conn_cb->ksnr_deleted = 0;
142 conn_cb->ksnr_conn_count = 0;
143 conn_cb->ksnr_ctrl_conn_count = 0;
144 conn_cb->ksnr_blki_conn_count = 0;
145 conn_cb->ksnr_blko_conn_count = 0;
146 conn_cb->ksnr_max_conns = 0;
147 conn_cb->ksnr_busy_retry_count = 0;
153 ksocknal_destroy_conn_cb(struct ksock_conn_cb *conn_cb)
155 LASSERT(refcount_read(&conn_cb->ksnr_refcount) == 0);
157 if (conn_cb->ksnr_peer)
158 ksocknal_peer_decref(conn_cb->ksnr_peer);
160 LIBCFS_FREE(conn_cb, sizeof(*conn_cb));
163 static struct ksock_peer_ni *
164 ksocknal_create_peer(struct lnet_ni *ni, struct lnet_processid *id)
166 int cpt = lnet_nid2cpt(&id->nid, ni);
167 struct ksock_net *net = ni->ni_data;
168 struct ksock_peer_ni *peer_ni;
170 LASSERT(!LNET_NID_IS_ANY(&id->nid));
171 LASSERT(id->pid != LNET_PID_ANY);
172 LASSERT(!in_interrupt());
174 if (!atomic_inc_unless_negative(&net->ksnn_npeers)) {
175 CERROR("Can't create peer_ni: network shutdown\n");
176 return ERR_PTR(-ESHUTDOWN);
179 LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
181 atomic_dec(&net->ksnn_npeers);
182 return ERR_PTR(-ENOMEM);
185 peer_ni->ksnp_ni = ni;
186 peer_ni->ksnp_id = *id;
187 refcount_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
188 peer_ni->ksnp_closing = 0;
189 peer_ni->ksnp_accepting = 0;
190 peer_ni->ksnp_proto = NULL;
191 peer_ni->ksnp_last_alive = 0;
192 peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
193 peer_ni->ksnp_conn_cb = NULL;
195 INIT_LIST_HEAD(&peer_ni->ksnp_conns);
196 INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
197 INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
198 spin_lock_init(&peer_ni->ksnp_lock);
204 ksocknal_destroy_peer(struct ksock_peer_ni *peer_ni)
206 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
208 CDEBUG (D_NET, "peer_ni %s %p deleted\n",
209 libcfs_idstr(&peer_ni->ksnp_id), peer_ni);
211 LASSERT(refcount_read(&peer_ni->ksnp_refcount) == 0);
212 LASSERT(peer_ni->ksnp_accepting == 0);
213 LASSERT(list_empty(&peer_ni->ksnp_conns));
214 LASSERT(peer_ni->ksnp_conn_cb == NULL);
215 LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
216 LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
218 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
220 /* NB a peer_ni's connections and conn_cb keep a reference on their
221 * peer_ni until they are destroyed, so we can be assured that _all_
222 * state to do with this peer_ni has been cleaned up when its refcount
225 if (atomic_dec_and_test(&net->ksnn_npeers))
226 wake_up_var(&net->ksnn_npeers);
229 struct ksock_peer_ni *
230 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_processid *id)
232 struct ksock_peer_ni *peer_ni;
233 unsigned long hash = nidhash(&id->nid);
235 hash_for_each_possible(ksocknal_data.ksnd_peers, peer_ni,
237 LASSERT(!peer_ni->ksnp_closing);
239 if (peer_ni->ksnp_ni != ni)
242 if (!nid_same(&peer_ni->ksnp_id.nid, &id->nid) ||
243 peer_ni->ksnp_id.pid != id->pid)
246 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
247 peer_ni, libcfs_idstr(id),
248 refcount_read(&peer_ni->ksnp_refcount));
254 struct ksock_peer_ni *
255 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_processid *id)
257 struct ksock_peer_ni *peer_ni;
259 read_lock(&ksocknal_data.ksnd_global_lock);
260 peer_ni = ksocknal_find_peer_locked(ni, id);
261 if (peer_ni != NULL) /* +1 ref for caller? */
262 ksocknal_peer_addref(peer_ni);
263 read_unlock(&ksocknal_data.ksnd_global_lock);
269 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
271 LASSERT(list_empty(&peer_ni->ksnp_conns));
272 LASSERT(peer_ni->ksnp_conn_cb == NULL);
273 LASSERT(!peer_ni->ksnp_closing);
274 peer_ni->ksnp_closing = 1;
275 hlist_del(&peer_ni->ksnp_list);
276 /* lose peerlist's ref */
277 ksocknal_peer_decref(peer_ni);
282 ksocknal_dump_peer_debug_info(struct ksock_peer_ni *peer_ni)
284 struct ksock_conn *conn;
285 struct list_head *ctmp;
286 struct list_head *txtmp;
290 list_for_each(ctmp, &peer_ni->ksnp_conns) {
291 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
293 if (!list_empty(&conn->ksnc_tx_queue))
294 list_for_each(txtmp, &conn->ksnc_tx_queue) txcount++;
296 CDEBUG(D_CONSOLE, "Conn %d [type, closing, crefcnt, srefcnt]: %d, %d, %d, %d\n",
300 refcount_read(&conn->ksnc_conn_refcount),
301 refcount_read(&conn->ksnc_sock_refcount));
302 CDEBUG(D_CONSOLE, "Conn %d rx [scheduled, ready, state]: %d, %d, %d\n",
304 conn->ksnc_rx_scheduled,
306 conn->ksnc_rx_state);
307 CDEBUG(D_CONSOLE, "Conn %d tx [txqcnt, scheduled, last_post, ready, deadline]: %d, %d, %lld, %d, %lld\n",
310 conn->ksnc_tx_scheduled,
311 conn->ksnc_tx_last_post,
313 conn->ksnc_rx_deadline);
315 if (conn->ksnc_scheduler)
316 CDEBUG(D_CONSOLE, "Conn %d sched [nconns, cpt]: %d, %d\n",
318 conn->ksnc_scheduler->kss_nconns,
319 conn->ksnc_scheduler->kss_cpt);
327 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
328 struct lnet_processid *id, __u32 *myip, __u32 *peer_ip,
329 int *port, int *conn_count, int *share_count)
331 struct ksock_peer_ni *peer_ni;
332 struct ksock_conn_cb *conn_cb;
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 conn_cb->ksnr_myiface,
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_conn_count_by_type(struct ksock_conn_cb *conn_cb,
387 unsigned int count = 0;
390 case SOCKLND_CONN_CONTROL:
391 count = conn_cb->ksnr_ctrl_conn_count;
393 case SOCKLND_CONN_BULK_IN:
394 count = conn_cb->ksnr_blki_conn_count;
396 case SOCKLND_CONN_BULK_OUT:
397 count = conn_cb->ksnr_blko_conn_count;
399 case SOCKLND_CONN_ANY:
400 count = conn_cb->ksnr_conn_count;
411 ksocknal_get_conns_per_peer(struct ksock_peer_ni *peer_ni)
413 struct lnet_ni *ni = peer_ni->ksnp_ni;
414 struct lnet_ioctl_config_socklnd_tunables *tunables;
418 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_sock;
420 return tunables->lnd_conns_per_peer;
424 ksocknal_incr_conn_count(struct ksock_conn_cb *conn_cb,
427 conn_cb->ksnr_conn_count++;
429 /* check if all connections of the given type got created */
431 case SOCKLND_CONN_CONTROL:
432 conn_cb->ksnr_ctrl_conn_count++;
433 /* there's a single control connection per peer,
434 * two in case of loopback
436 conn_cb->ksnr_connected |= BIT(type);
438 case SOCKLND_CONN_BULK_IN:
439 conn_cb->ksnr_blki_conn_count++;
440 if (conn_cb->ksnr_blki_conn_count >= conn_cb->ksnr_max_conns)
441 conn_cb->ksnr_connected |= BIT(type);
443 case SOCKLND_CONN_BULK_OUT:
444 conn_cb->ksnr_blko_conn_count++;
445 if (conn_cb->ksnr_blko_conn_count >= conn_cb->ksnr_max_conns)
446 conn_cb->ksnr_connected |= BIT(type);
448 case SOCKLND_CONN_ANY:
449 if (conn_cb->ksnr_conn_count >= conn_cb->ksnr_max_conns)
450 conn_cb->ksnr_connected |= BIT(type);
457 CDEBUG(D_NET, "Add conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
458 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
463 ksocknal_decr_conn_count(struct ksock_conn_cb *conn_cb,
466 conn_cb->ksnr_conn_count--;
468 /* check if all connections of the given type got created */
470 case SOCKLND_CONN_CONTROL:
471 conn_cb->ksnr_ctrl_conn_count--;
472 /* there's a single control connection per peer,
473 * two in case of loopback
475 if (conn_cb->ksnr_ctrl_conn_count == 0)
476 conn_cb->ksnr_connected &= ~BIT(type);
478 case SOCKLND_CONN_BULK_IN:
479 conn_cb->ksnr_blki_conn_count--;
480 if (conn_cb->ksnr_blki_conn_count < conn_cb->ksnr_max_conns)
481 conn_cb->ksnr_connected &= ~BIT(type);
483 case SOCKLND_CONN_BULK_OUT:
484 conn_cb->ksnr_blko_conn_count--;
485 if (conn_cb->ksnr_blko_conn_count < conn_cb->ksnr_max_conns)
486 conn_cb->ksnr_connected &= ~BIT(type);
488 case SOCKLND_CONN_ANY:
489 if (conn_cb->ksnr_conn_count < conn_cb->ksnr_max_conns)
490 conn_cb->ksnr_connected &= ~BIT(type);
497 CDEBUG(D_NET, "Del conn type %d, ksnr_connected %x ksnr_max_conns %d\n",
498 type, conn_cb->ksnr_connected, conn_cb->ksnr_max_conns);
502 ksocknal_associate_cb_conn_locked(struct ksock_conn_cb *conn_cb,
503 struct ksock_conn *conn)
505 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
506 int type = conn->ksnc_type;
507 struct ksock_interface *iface;
510 conn_iface = ksocknal_ip2index((struct sockaddr *)&conn->ksnc_myaddr,
512 conn->ksnc_conn_cb = conn_cb;
513 ksocknal_conn_cb_addref(conn_cb);
515 if (conn_cb->ksnr_myiface != conn_iface) {
516 if (conn_cb->ksnr_myiface < 0) {
517 /* route wasn't bound locally yet (the initial route) */
518 CDEBUG(D_NET, "Binding %s %pISc to interface %d\n",
519 libcfs_idstr(&peer_ni->ksnp_id),
524 "Rebinding %s %pISc from interface %d to %d\n",
525 libcfs_idstr(&peer_ni->ksnp_id),
527 conn_cb->ksnr_myiface,
530 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
531 conn_cb->ksnr_myiface);
533 iface->ksni_nroutes--;
535 conn_cb->ksnr_myiface = conn_iface;
536 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
537 conn_cb->ksnr_myiface);
539 iface->ksni_nroutes++;
542 ksocknal_incr_conn_count(conn_cb, type);
544 /* Successful connection => further attempts can
545 * proceed immediately
547 conn_cb->ksnr_retry_interval = 0;
551 ksocknal_add_conn_cb_locked(struct ksock_peer_ni *peer_ni,
552 struct ksock_conn_cb *conn_cb)
554 struct ksock_conn *conn;
555 struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
557 LASSERT(!peer_ni->ksnp_closing);
558 LASSERT(!conn_cb->ksnr_peer);
559 LASSERT(!conn_cb->ksnr_scheduled);
560 LASSERT(!conn_cb->ksnr_connecting);
561 LASSERT(conn_cb->ksnr_connected == 0);
563 conn_cb->ksnr_peer = peer_ni;
564 ksocknal_peer_addref(peer_ni);
566 /* set the conn_cb's interface to the current net's interface */
567 conn_cb->ksnr_myiface = net->ksnn_interface.ksni_index;
568 net->ksnn_interface.ksni_nroutes++;
570 /* peer_ni's route list takes over my ref on 'route' */
571 peer_ni->ksnp_conn_cb = conn_cb;
573 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
574 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
575 (struct sockaddr *)&conn_cb->ksnr_addr))
578 ksocknal_associate_cb_conn_locked(conn_cb, conn);
579 /* keep going (typed conns) */
584 ksocknal_del_conn_cb_locked(struct ksock_conn_cb *conn_cb)
586 struct ksock_peer_ni *peer_ni = conn_cb->ksnr_peer;
587 struct ksock_interface *iface;
588 struct ksock_conn *conn;
589 struct ksock_conn *cnxt;
591 LASSERT(!conn_cb->ksnr_deleted);
593 /* Close associated conns */
594 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
595 if (conn->ksnc_conn_cb != conn_cb)
598 ksocknal_close_conn_locked(conn, 0);
601 if (conn_cb->ksnr_myiface >= 0) {
602 iface = ksocknal_index2iface(peer_ni->ksnp_ni,
603 conn_cb->ksnr_myiface);
605 iface->ksni_nroutes--;
608 conn_cb->ksnr_deleted = 1;
609 ksocknal_conn_cb_decref(conn_cb); /* drop peer_ni's ref */
610 peer_ni->ksnp_conn_cb = NULL;
612 if (list_empty(&peer_ni->ksnp_conns)) {
613 /* I've just removed the last route to a peer_ni with no active
616 ksocknal_unlink_peer_locked(peer_ni);
621 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_processid *id,
622 struct sockaddr *addr)
624 struct ksock_peer_ni *peer_ni;
625 struct ksock_peer_ni *peer2;
626 struct ksock_conn_cb *conn_cb;
628 if (LNET_NID_IS_ANY(&id->nid) ||
629 id->pid == LNET_PID_ANY)
632 /* Have a brand new peer_ni ready... */
633 peer_ni = ksocknal_create_peer(ni, id);
635 return PTR_ERR(peer_ni);
637 conn_cb = ksocknal_create_conn_cb(addr);
639 ksocknal_peer_decref(peer_ni);
643 write_lock_bh(&ksocknal_data.ksnd_global_lock);
645 /* always called with a ref on ni, so shutdown can't have started */
646 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
649 peer2 = ksocknal_find_peer_locked(ni, id);
651 ksocknal_peer_decref(peer_ni);
654 /* peer_ni table takes my ref on peer_ni */
655 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list,
659 if (peer_ni->ksnp_conn_cb) {
660 ksocknal_conn_cb_decref(conn_cb);
662 ksocknal_add_conn_cb_locked(peer_ni, conn_cb);
663 /* Remember conns_per_peer setting at the time
664 * of connection initiation. It will define the
665 * max number of conns per type for this conn_cb
668 conn_cb->ksnr_max_conns = ksocknal_get_conns_per_peer(peer_ni);
671 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
677 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni)
679 struct ksock_conn *conn;
680 struct ksock_conn *cnxt;
681 struct ksock_conn_cb *conn_cb;
683 LASSERT(!peer_ni->ksnp_closing);
685 /* Extra ref prevents peer_ni disappearing until I'm done with it */
686 ksocknal_peer_addref(peer_ni);
687 conn_cb = peer_ni->ksnp_conn_cb;
689 ksocknal_del_conn_cb_locked(conn_cb);
691 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns,
693 ksocknal_close_conn_locked(conn, 0);
695 ksocknal_peer_decref(peer_ni);
696 /* NB peer_ni unlinks itself when last conn/conn_cb is removed */
700 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_processid *id)
703 struct hlist_node *pnxt;
704 struct ksock_peer_ni *peer_ni;
710 write_lock_bh(&ksocknal_data.ksnd_global_lock);
712 if (id && !LNET_NID_IS_ANY(&id->nid)) {
713 lo = hash_min(nidhash(&id->nid),
714 HASH_BITS(ksocknal_data.ksnd_peers));
718 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
721 for (i = lo; i <= hi; i++) {
722 hlist_for_each_entry_safe(peer_ni, pnxt,
723 &ksocknal_data.ksnd_peers[i],
725 if (peer_ni->ksnp_ni != ni)
728 if (!((!id || LNET_NID_IS_ANY(&id->nid) ||
729 nid_same(&peer_ni->ksnp_id.nid, &id->nid)) &&
730 (!id || id->pid == LNET_PID_ANY ||
731 peer_ni->ksnp_id.pid == id->pid)))
734 ksocknal_peer_addref(peer_ni); /* a ref for me... */
736 ksocknal_del_peer_locked(peer_ni);
738 if (peer_ni->ksnp_closing &&
739 !list_empty(&peer_ni->ksnp_tx_queue)) {
740 LASSERT(list_empty(&peer_ni->ksnp_conns));
741 LASSERT(peer_ni->ksnp_conn_cb == NULL);
743 list_splice_init(&peer_ni->ksnp_tx_queue,
747 ksocknal_peer_decref(peer_ni); /* ...till here */
749 rc = 0; /* matched! */
753 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
755 ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
760 static struct ksock_conn *
761 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
763 struct ksock_peer_ni *peer_ni;
764 struct ksock_conn *conn;
767 read_lock(&ksocknal_data.ksnd_global_lock);
769 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
770 LASSERT(!peer_ni->ksnp_closing);
772 if (peer_ni->ksnp_ni != ni)
775 list_for_each_entry(conn, &peer_ni->ksnp_conns,
780 ksocknal_conn_addref(conn);
781 read_unlock(&ksocknal_data.ksnd_global_lock);
786 read_unlock(&ksocknal_data.ksnd_global_lock);
790 static struct ksock_sched *
791 ksocknal_choose_scheduler_locked(unsigned int cpt)
793 struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
796 if (sched->kss_nthreads == 0) {
797 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
798 if (sched->kss_nthreads > 0) {
799 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
800 cpt, sched->kss_cpt);
811 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
813 struct ksock_connreq *cr;
815 struct sockaddr_storage peer;
817 rc = lnet_sock_getaddr(sock, true, &peer);
819 CERROR("Can't determine new connection's address\n");
823 LIBCFS_ALLOC(cr, sizeof(*cr));
825 LCONSOLE_ERROR_MSG(0x12f,
826 "Dropping connection request from %pISc: memory exhausted\n",
833 cr->ksncr_sock = sock;
835 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
837 list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
838 wake_up(&ksocknal_data.ksnd_connd_waitq);
840 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
845 ksocknal_connecting(struct ksock_conn_cb *conn_cb, struct sockaddr *sa)
848 rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr, sa))
849 return conn_cb->ksnr_connecting;
854 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_conn_cb *conn_cb,
855 struct socket *sock, int type)
857 rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
859 struct lnet_processid peerid;
861 struct ksock_conn *conn;
862 struct ksock_conn *conn2;
863 struct ksock_peer_ni *peer_ni = NULL;
864 struct ksock_peer_ni *peer2;
865 struct ksock_sched *sched;
866 struct ksock_hello_msg *hello;
869 struct ksock_tx *txtmp;
876 active = (conn_cb != NULL);
878 LASSERT(active == (type != SOCKLND_CONN_NONE));
880 LIBCFS_ALLOC(conn, sizeof(*conn));
886 conn->ksnc_peer = NULL;
887 conn->ksnc_conn_cb = NULL;
888 conn->ksnc_sock = sock;
889 /* 2 ref, 1 for conn, another extra ref prevents socket
890 * being closed before establishment of connection */
891 refcount_set(&conn->ksnc_sock_refcount, 2);
892 conn->ksnc_type = type;
893 ksocknal_lib_save_callback(sock, conn);
894 refcount_set(&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
896 conn->ksnc_rx_ready = 0;
897 conn->ksnc_rx_scheduled = 0;
899 INIT_LIST_HEAD(&conn->ksnc_tx_queue);
900 conn->ksnc_tx_ready = 0;
901 conn->ksnc_tx_scheduled = 0;
902 conn->ksnc_tx_carrier = NULL;
903 atomic_set (&conn->ksnc_tx_nob, 0);
905 LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
906 kshm_ips[LNET_INTERFACES_NUM]));
912 /* stash conn's local and remote addrs */
913 rc = ksocknal_lib_get_conn_addrs(conn);
917 /* Find out/confirm peer_ni's NID and connection type and get the
918 * vector of interfaces she's willing to let me connect to.
919 * Passive connections use the listener timeout since the peer_ni sends
924 peer_ni = conn_cb->ksnr_peer;
925 LASSERT(ni == peer_ni->ksnp_ni);
927 /* Active connection sends HELLO eagerly */
928 hello->kshm_nips = 0;
929 peerid = peer_ni->ksnp_id;
931 write_lock_bh(global_lock);
932 conn->ksnc_proto = peer_ni->ksnp_proto;
933 write_unlock_bh(global_lock);
935 if (conn->ksnc_proto == NULL) {
936 conn->ksnc_proto = &ksocknal_protocol_v3x;
937 #if SOCKNAL_VERSION_DEBUG
938 if (*ksocknal_tunables.ksnd_protocol == 2)
939 conn->ksnc_proto = &ksocknal_protocol_v2x;
940 else if (*ksocknal_tunables.ksnd_protocol == 1)
941 conn->ksnc_proto = &ksocknal_protocol_v1x;
945 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
949 peerid.nid = LNET_ANY_NID;
950 peerid.pid = LNET_PID_ANY;
952 /* Passive, get protocol from peer_ni */
953 conn->ksnc_proto = NULL;
956 rc = ksocknal_recv_hello(ni, conn, hello, &peerid, &incarnation);
960 LASSERT(rc == 0 || active);
961 LASSERT(conn->ksnc_proto != NULL);
962 LASSERT(!LNET_NID_IS_ANY(&peerid.nid));
964 cpt = lnet_nid2cpt(&peerid.nid, ni);
967 ksocknal_peer_addref(peer_ni);
968 write_lock_bh(global_lock);
970 peer_ni = ksocknal_create_peer(ni, &peerid);
971 if (IS_ERR(peer_ni)) {
972 rc = PTR_ERR(peer_ni);
976 write_lock_bh(global_lock);
978 /* called with a ref on ni, so shutdown can't have started */
979 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
981 peer2 = ksocknal_find_peer_locked(ni, &peerid);
983 /* NB this puts an "empty" peer_ni in the peer_ni
984 * table (which takes my ref) */
985 hash_add(ksocknal_data.ksnd_peers,
986 &peer_ni->ksnp_list, nidhash(&peerid.nid));
988 ksocknal_peer_decref(peer_ni);
993 ksocknal_peer_addref(peer_ni);
994 peer_ni->ksnp_accepting++;
996 /* Am I already connecting to this guy? Resolve in
997 * favour of higher NID...
999 if (memcmp(&peerid.nid, &ni->ni_nid, sizeof(peerid.nid)) < 0 &&
1000 ksocknal_connecting(peer_ni->ksnp_conn_cb,
1001 ((struct sockaddr *) &conn->ksnc_peeraddr))) {
1003 warn = "connection race resolution";
1008 if (peer_ni->ksnp_closing ||
1009 (active && conn_cb->ksnr_deleted)) {
1010 /* peer_ni/conn_cb got closed under me */
1012 warn = "peer_ni/conn_cb removed";
1016 if (peer_ni->ksnp_proto == NULL) {
1017 /* Never connected before.
1018 * NB recv_hello may have returned EPROTO to signal my peer_ni
1019 * wants a different protocol than the one I asked for.
1021 LASSERT(list_empty(&peer_ni->ksnp_conns));
1023 peer_ni->ksnp_proto = conn->ksnc_proto;
1024 peer_ni->ksnp_incarnation = incarnation;
1027 if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1028 peer_ni->ksnp_incarnation != incarnation) {
1029 /* peer_ni rebooted or I've got the wrong protocol version */
1030 ksocknal_close_peer_conns_locked(peer_ni, NULL, 0);
1032 peer_ni->ksnp_proto = NULL;
1034 warn = peer_ni->ksnp_incarnation != incarnation ?
1035 "peer_ni rebooted" :
1036 "wrong proto version";
1046 warn = "lost conn race";
1049 warn = "retry with different protocol version";
1053 /* Refuse to duplicate an existing connection, unless this is a
1054 * loopback connection */
1055 if (!rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1056 (struct sockaddr *)&conn->ksnc_myaddr)) {
1057 list_for_each_entry(conn2, &peer_ni->ksnp_conns, ksnc_list) {
1059 (struct sockaddr *)&conn2->ksnc_peeraddr,
1060 (struct sockaddr *)&conn->ksnc_peeraddr) ||
1062 (struct sockaddr *)&conn2->ksnc_myaddr,
1063 (struct sockaddr *)&conn->ksnc_myaddr) ||
1064 conn2->ksnc_type != conn->ksnc_type)
1068 /* If max conns per type is not registered in conn_cb
1069 * as ksnr_max_conns, use ni's conns_per_peer
1071 if ((peer_ni->ksnp_conn_cb &&
1072 num_dup < peer_ni->ksnp_conn_cb->ksnr_max_conns) ||
1073 (!peer_ni->ksnp_conn_cb &&
1074 num_dup < ksocknal_get_conns_per_peer(peer_ni)))
1077 /* Reply on a passive connection attempt so the peer_ni
1078 * realises we're connected.
1088 /* If the connection created by this route didn't bind to the IP
1089 * address the route connected to, the connection/route matching
1090 * code below probably isn't going to work.
1093 !rpc_cmp_addr((struct sockaddr *)&conn_cb->ksnr_addr,
1094 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1095 CERROR("Route %s %pISc connected to %pISc\n",
1096 libcfs_idstr(&peer_ni->ksnp_id),
1097 &conn_cb->ksnr_addr,
1098 &conn->ksnc_peeraddr);
1101 /* Search for a conn_cb corresponding to the new connection and
1102 * create an association. This allows incoming connections created
1103 * by conn_cbs in my peer_ni to match my own conn_cb entries so I don't
1104 * continually create duplicate conn_cbs.
1106 conn_cb = peer_ni->ksnp_conn_cb;
1108 if (conn_cb && rpc_cmp_addr((struct sockaddr *)&conn->ksnc_peeraddr,
1109 (struct sockaddr *)&conn_cb->ksnr_addr))
1110 ksocknal_associate_cb_conn_locked(conn_cb, conn);
1112 conn->ksnc_peer = peer_ni; /* conn takes my ref on peer_ni */
1113 peer_ni->ksnp_last_alive = ktime_get_seconds();
1114 peer_ni->ksnp_send_keepalive = 0;
1115 peer_ni->ksnp_error = 0;
1117 sched = ksocknal_choose_scheduler_locked(cpt);
1119 CERROR("no schedulers available. node is unhealthy\n");
1123 * The cpt might have changed if we ended up selecting a non cpt
1124 * native scheduler. So use the scheduler's cpt instead.
1126 cpt = sched->kss_cpt;
1127 sched->kss_nconns++;
1128 conn->ksnc_scheduler = sched;
1130 conn->ksnc_tx_last_post = ktime_get_seconds();
1131 /* Set the deadline for the outgoing HELLO to drain */
1132 conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1133 conn->ksnc_tx_deadline = ktime_get_seconds() +
1135 smp_mb(); /* order with adding to peer_ni's conn list */
1137 list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1138 ksocknal_conn_addref(conn);
1140 ksocknal_new_packet(conn, 0);
1142 conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1144 /* Take packets blocking for this connection. */
1145 list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1146 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1150 list_del(&tx->tx_list);
1151 ksocknal_queue_tx_locked(tx, conn);
1154 write_unlock_bh(global_lock);
1155 /* We've now got a new connection. Any errors from here on are just
1156 * like "normal" comms errors and we close the connection normally.
1157 * NB (a) we still have to send the reply HELLO for passive
1159 * (b) normal I/O on the conn is blocked until I setup and call the
1163 CDEBUG(D_NET, "New conn %s p %d.x %pISc -> %pIScp"
1164 " incarnation:%lld sched[%d]\n",
1165 libcfs_idstr(&peerid), conn->ksnc_proto->pro_version,
1166 &conn->ksnc_myaddr, &conn->ksnc_peeraddr,
1170 hello->kshm_nips = 0;
1171 rc = ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1174 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1175 kshm_ips[LNET_INTERFACES_NUM]));
1177 /* setup the socket AFTER I've received hello (it disables
1178 * SO_LINGER). I might call back to the acceptor who may want
1179 * to send a protocol version response and then close the
1180 * socket; this ensures the socket only tears down after the
1181 * response has been sent.
1184 rc = ksocknal_lib_setup_sock(sock);
1186 write_lock_bh(global_lock);
1188 /* NB my callbacks block while I hold ksnd_global_lock */
1189 ksocknal_lib_set_callback(sock, conn);
1192 peer_ni->ksnp_accepting--;
1194 write_unlock_bh(global_lock);
1197 write_lock_bh(global_lock);
1198 if (!conn->ksnc_closing) {
1199 /* could be closed by another thread */
1200 ksocknal_close_conn_locked(conn, rc);
1202 write_unlock_bh(global_lock);
1203 } else if (ksocknal_connsock_addref(conn) == 0) {
1204 /* Allow I/O to proceed. */
1205 ksocknal_read_callback(conn);
1206 ksocknal_write_callback(conn);
1207 ksocknal_connsock_decref(conn);
1210 ksocknal_connsock_decref(conn);
1211 ksocknal_conn_decref(conn);
1216 if (!peer_ni->ksnp_closing &&
1217 list_empty(&peer_ni->ksnp_conns) &&
1218 peer_ni->ksnp_conn_cb == NULL) {
1219 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1220 ksocknal_unlink_peer_locked(peer_ni);
1223 write_unlock_bh(global_lock);
1227 CERROR("Not creating conn %s type %d: %s\n",
1228 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1230 CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1231 libcfs_idstr(&peerid), conn->ksnc_type, warn);
1236 /* Request retry by replying with CONN_NONE
1237 * ksnc_proto has been set already
1239 conn->ksnc_type = SOCKLND_CONN_NONE;
1240 hello->kshm_nips = 0;
1241 ksocknal_send_hello(ni, conn, &peerid.nid, hello);
1244 write_lock_bh(global_lock);
1245 peer_ni->ksnp_accepting--;
1246 write_unlock_bh(global_lock);
1250 * If we get here without an error code, just use -EALREADY.
1251 * Depending on how we got here, the error may be positive
1252 * or negative. Normalize the value for ksocknal_txlist_done().
1254 rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1255 ksocknal_txlist_done(ni, &zombies, rc2);
1256 ksocknal_peer_decref(peer_ni);
1260 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1261 kshm_ips[LNET_INTERFACES_NUM]));
1263 LIBCFS_FREE(conn, sizeof(*conn));
1272 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1274 /* This just does the immmediate housekeeping, and queues the
1275 * connection for the reaper to terminate.
1276 * Caller holds ksnd_global_lock exclusively in irq context */
1277 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1278 struct ksock_conn_cb *conn_cb;
1279 struct ksock_conn *conn2;
1281 int duplicate_count = 0;
1283 LASSERT(peer_ni->ksnp_error == 0);
1284 LASSERT(!conn->ksnc_closing);
1285 conn->ksnc_closing = 1;
1287 /* ksnd_deathrow_conns takes over peer_ni's ref */
1288 list_del(&conn->ksnc_list);
1290 conn_cb = conn->ksnc_conn_cb;
1291 if (conn_cb != NULL) {
1292 /* dissociate conn from cb... */
1293 LASSERT(!conn_cb->ksnr_deleted);
1295 conn_count = ksocknal_get_conn_count_by_type(conn_cb,
1297 /* connected bit is set only if all connections
1298 * of the given type got created
1300 if (conn_count == conn_cb->ksnr_max_conns)
1301 LASSERT((conn_cb->ksnr_connected &
1302 BIT(conn->ksnc_type)) != 0);
1304 if (conn_count == 1) {
1305 list_for_each_entry(conn2, &peer_ni->ksnp_conns,
1307 if (conn2->ksnc_conn_cb == conn_cb &&
1308 conn2->ksnc_type == conn->ksnc_type)
1309 duplicate_count += 1;
1311 if (duplicate_count > 0)
1312 CERROR("Found %d duplicate conns type %d\n",
1316 ksocknal_decr_conn_count(conn_cb, conn->ksnc_type);
1318 conn->ksnc_conn_cb = NULL;
1320 /* drop conn's ref on conn_cb */
1321 ksocknal_conn_cb_decref(conn_cb);
1324 if (list_empty(&peer_ni->ksnp_conns)) {
1325 /* No more connections to this peer_ni */
1327 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1328 struct ksock_tx *tx;
1330 LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1332 /* throw them to the last connection...,
1333 * these TXs will be send to /dev/null by scheduler */
1334 list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1336 ksocknal_tx_prep(conn, tx);
1338 spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1339 list_splice_init(&peer_ni->ksnp_tx_queue,
1340 &conn->ksnc_tx_queue);
1341 spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1344 /* renegotiate protocol version */
1345 peer_ni->ksnp_proto = NULL;
1346 /* stash last conn close reason */
1347 peer_ni->ksnp_error = error;
1349 if (peer_ni->ksnp_conn_cb == NULL) {
1350 /* I've just closed last conn belonging to a
1351 * peer_ni with no connections to it
1353 ksocknal_unlink_peer_locked(peer_ni);
1357 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1359 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1360 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1362 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1366 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1368 bool notify = false;
1369 time64_t last_alive = 0;
1371 /* There has been a connection failure or comms error; but I'll only
1372 * tell LNET I think the peer_ni is dead if it's to another kernel and
1373 * there are no connections or connection attempts in existence. */
1375 read_lock(&ksocknal_data.ksnd_global_lock);
1377 if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1378 list_empty(&peer_ni->ksnp_conns) &&
1379 peer_ni->ksnp_accepting == 0 &&
1380 !ksocknal_find_connecting_conn_cb_locked(peer_ni)) {
1382 last_alive = peer_ni->ksnp_last_alive;
1385 read_unlock(&ksocknal_data.ksnd_global_lock);
1388 lnet_notify(peer_ni->ksnp_ni,
1389 lnet_nid_to_nid4(&peer_ni->ksnp_id.nid),
1390 false, false, last_alive);
1394 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1396 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1397 struct ksock_tx *tx;
1398 struct ksock_tx *tmp;
1401 /* NB safe to finalize TXs because closing of socket will
1402 * abort all buffered data */
1403 LASSERT(conn->ksnc_sock == NULL);
1405 spin_lock(&peer_ni->ksnp_lock);
1407 list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list,
1409 if (tx->tx_conn != conn)
1412 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1414 tx->tx_msg.ksm_zc_cookies[0] = 0;
1415 tx->tx_zc_aborted = 1; /* mark it as not-acked */
1416 list_move(&tx->tx_zc_list, &zlist);
1419 spin_unlock(&peer_ni->ksnp_lock);
1421 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1422 tx_zc_list)) != NULL) {
1423 list_del(&tx->tx_zc_list);
1424 ksocknal_tx_decref(tx);
1429 ksocknal_terminate_conn(struct ksock_conn *conn)
1431 /* This gets called by the reaper (guaranteed thread context) to
1432 * disengage the socket from its callbacks and close it.
1433 * ksnc_refcount will eventually hit zero, and then the reaper will
1436 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1437 struct ksock_sched *sched = conn->ksnc_scheduler;
1438 bool failed = false;
1440 LASSERT(conn->ksnc_closing);
1442 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1443 spin_lock_bh(&sched->kss_lock);
1445 /* a closing conn is always ready to tx */
1446 conn->ksnc_tx_ready = 1;
1448 if (!conn->ksnc_tx_scheduled &&
1449 !list_empty(&conn->ksnc_tx_queue)) {
1450 list_add_tail(&conn->ksnc_tx_list,
1451 &sched->kss_tx_conns);
1452 conn->ksnc_tx_scheduled = 1;
1453 /* extra ref for scheduler */
1454 ksocknal_conn_addref(conn);
1456 wake_up(&sched->kss_waitq);
1459 spin_unlock_bh(&sched->kss_lock);
1461 /* serialise with callbacks */
1462 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1464 ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1466 /* OK, so this conn may not be completely disengaged from its
1467 * scheduler yet, but it _has_ committed to terminate...
1469 conn->ksnc_scheduler->kss_nconns--;
1471 if (peer_ni->ksnp_error != 0) {
1472 /* peer_ni's last conn closed in error */
1473 LASSERT(list_empty(&peer_ni->ksnp_conns));
1475 peer_ni->ksnp_error = 0; /* avoid multiple notifications */
1478 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1481 ksocknal_peer_failed(peer_ni);
1483 /* The socket is closed on the final put; either here, or in
1484 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1485 * when the connection was established, this will close the socket
1486 * immediately, aborting anything buffered in it. Any hung
1487 * zero-copy transmits will therefore complete in finite time.
1489 ksocknal_connsock_decref(conn);
1493 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1495 /* Queue the conn for the reaper to destroy */
1496 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1497 spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1499 list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1500 wake_up(&ksocknal_data.ksnd_reaper_waitq);
1502 spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1506 ksocknal_destroy_conn(struct ksock_conn *conn)
1510 /* Final coup-de-grace of the reaper */
1511 CDEBUG(D_NET, "connection %p\n", conn);
1513 LASSERT(refcount_read(&conn->ksnc_conn_refcount) == 0);
1514 LASSERT(refcount_read(&conn->ksnc_sock_refcount) == 0);
1515 LASSERT(conn->ksnc_sock == NULL);
1516 LASSERT(conn->ksnc_conn_cb == NULL);
1517 LASSERT(!conn->ksnc_tx_scheduled);
1518 LASSERT(!conn->ksnc_rx_scheduled);
1519 LASSERT(list_empty(&conn->ksnc_tx_queue));
1521 /* complete current receive if any */
1522 switch (conn->ksnc_rx_state) {
1523 case SOCKNAL_RX_LNET_PAYLOAD:
1524 last_rcv = conn->ksnc_rx_deadline -
1526 CERROR("Completing partial receive from %s[%d], ip %pIScp, with error, wanted: %d, left: %d, last alive is %lld secs ago\n",
1527 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1529 &conn->ksnc_peeraddr,
1530 conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1531 ktime_get_seconds() - last_rcv);
1532 if (conn->ksnc_lnet_msg)
1533 conn->ksnc_lnet_msg->msg_health_status =
1534 LNET_MSG_STATUS_REMOTE_ERROR;
1535 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1537 case SOCKNAL_RX_LNET_HEADER:
1538 if (conn->ksnc_rx_started)
1539 CERROR("Incomplete receive of lnet header from %s, ip %pIScp, with error, protocol: %d.x.\n",
1540 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1541 &conn->ksnc_peeraddr,
1542 conn->ksnc_proto->pro_version);
1544 case SOCKNAL_RX_KSM_HEADER:
1545 if (conn->ksnc_rx_started)
1546 CERROR("Incomplete receive of ksock message from %s, ip %pIScp, with error, protocol: %d.x.\n",
1547 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1548 &conn->ksnc_peeraddr,
1549 conn->ksnc_proto->pro_version);
1551 case SOCKNAL_RX_SLOP:
1552 if (conn->ksnc_rx_started)
1553 CERROR("Incomplete receive of slops from %s, ip %pIScp, with error\n",
1554 libcfs_idstr(&conn->ksnc_peer->ksnp_id),
1555 &conn->ksnc_peeraddr);
1562 ksocknal_peer_decref(conn->ksnc_peer);
1564 LIBCFS_FREE(conn, sizeof(*conn));
1568 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni,
1569 struct sockaddr *addr, int why)
1571 struct ksock_conn *conn;
1572 struct ksock_conn *cnxt;
1575 list_for_each_entry_safe(conn, cnxt, &peer_ni->ksnp_conns, ksnc_list) {
1578 (struct sockaddr *)&conn->ksnc_peeraddr)) {
1580 ksocknal_close_conn_locked(conn, why);
1588 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1590 struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1593 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1595 count = ksocknal_close_peer_conns_locked(
1596 peer_ni, (struct sockaddr *)&conn->ksnc_peeraddr, why);
1598 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1604 ksocknal_close_matching_conns(struct lnet_processid *id, __u32 ipaddr)
1606 struct ksock_peer_ni *peer_ni;
1607 struct hlist_node *pnxt;
1612 struct sockaddr_in sa = {.sin_family = AF_INET};
1614 write_lock_bh(&ksocknal_data.ksnd_global_lock);
1616 if (!LNET_NID_IS_ANY(&id->nid)) {
1617 lo = hash_min(nidhash(&id->nid),
1618 HASH_BITS(ksocknal_data.ksnd_peers));
1622 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1625 sa.sin_addr.s_addr = htonl(ipaddr);
1626 for (i = lo; i <= hi; i++) {
1627 hlist_for_each_entry_safe(peer_ni, pnxt,
1628 &ksocknal_data.ksnd_peers[i],
1631 if (!((LNET_NID_IS_ANY(&id->nid) ||
1632 nid_same(&id->nid, &peer_ni->ksnp_id.nid)) &&
1633 (id->pid == LNET_PID_ANY ||
1634 id->pid == peer_ni->ksnp_id.pid)))
1637 count += ksocknal_close_peer_conns_locked(
1639 ipaddr ? (struct sockaddr *)&sa : NULL, 0);
1643 write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1645 /* wildcards always succeed */
1646 if (LNET_NID_IS_ANY(&id->nid) || id->pid == LNET_PID_ANY ||
1650 return (count == 0 ? -ENOENT : 0);
1654 ksocknal_notify_gw_down(struct lnet_nid *gw_nid)
1656 /* The router is telling me she's been notified of a change in
1659 struct lnet_processid id = {
1660 .pid = LNET_PID_ANY,
1664 CDEBUG(D_NET, "gw %s down\n", libcfs_nidstr(gw_nid));
1666 /* If the gateway crashed, close all open connections... */
1667 ksocknal_close_matching_conns(&id, 0);
1670 /* We can only establish new connections
1671 * if we have autroutes, and these connect on demand.
1676 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1680 struct ksock_conn *conn;
1682 for (index = 0; ; index++) {
1683 read_lock(&ksocknal_data.ksnd_global_lock);
1688 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
1690 ksocknal_conn_addref(conn);
1695 read_unlock(&ksocknal_data.ksnd_global_lock);
1700 ksocknal_lib_push_conn (conn);
1701 ksocknal_conn_decref(conn);
1706 ksocknal_push(struct lnet_ni *ni, struct lnet_processid *id)
1713 if (!LNET_NID_IS_ANY(&id->nid)) {
1714 lo = hash_min(nidhash(&id->nid),
1715 HASH_BITS(ksocknal_data.ksnd_peers));
1719 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1722 for (bkt = lo; bkt <= hi; bkt++) {
1723 int peer_off; /* searching offset in peer_ni hash table */
1725 for (peer_off = 0; ; peer_off++) {
1726 struct ksock_peer_ni *peer_ni;
1729 read_lock(&ksocknal_data.ksnd_global_lock);
1730 hlist_for_each_entry(peer_ni,
1731 &ksocknal_data.ksnd_peers[bkt],
1733 if (!((LNET_NID_IS_ANY(&id->nid) ||
1735 &peer_ni->ksnp_id.nid)) &&
1736 (id->pid == LNET_PID_ANY ||
1737 id->pid == peer_ni->ksnp_id.pid)))
1740 if (i++ == peer_off) {
1741 ksocknal_peer_addref(peer_ni);
1745 read_unlock(&ksocknal_data.ksnd_global_lock);
1747 if (i <= peer_off) /* no match */
1751 ksocknal_push_peer(peer_ni);
1752 ksocknal_peer_decref(peer_ni);
1759 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
1761 struct lnet_processid id = {};
1762 struct libcfs_ioctl_data *data = arg;
1766 case IOC_LIBCFS_GET_INTERFACE: {
1767 struct ksock_net *net = ni->ni_data;
1768 struct ksock_interface *iface;
1769 struct sockaddr_in *sa;
1771 read_lock(&ksocknal_data.ksnd_global_lock);
1773 if (data->ioc_count >= 1) {
1777 iface = &net->ksnn_interface;
1779 sa = (void *)&iface->ksni_addr;
1780 if (sa->sin_family == AF_INET) {
1781 data->ioc_u32[0] = ntohl(sa->sin_addr.s_addr);
1782 data->ioc_u32[1] = iface->ksni_netmask;
1784 data->ioc_u32[0] = 0xFFFFFFFF;
1785 data->ioc_u32[1] = 0;
1787 data->ioc_u32[2] = iface->ksni_npeers;
1788 data->ioc_u32[3] = iface->ksni_nroutes;
1791 read_unlock(&ksocknal_data.ksnd_global_lock);
1795 case IOC_LIBCFS_GET_PEER: {
1800 int share_count = 0;
1802 rc = ksocknal_get_peer_info(ni, data->ioc_count,
1803 &id, &myip, &ip, &port,
1804 &conn_count, &share_count);
1808 if (!nid_is_nid4(&id.nid))
1810 data->ioc_nid = lnet_nid_to_nid4(&id.nid);
1811 data->ioc_count = share_count;
1812 data->ioc_u32[0] = ip;
1813 data->ioc_u32[1] = port;
1814 data->ioc_u32[2] = myip;
1815 data->ioc_u32[3] = conn_count;
1816 data->ioc_u32[4] = id.pid;
1820 case IOC_LIBCFS_ADD_PEER: {
1821 struct sockaddr_in sa = {.sin_family = AF_INET};
1823 id.pid = LNET_PID_LUSTRE;
1824 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1825 sa.sin_addr.s_addr = htonl(data->ioc_u32[0]);
1826 sa.sin_port = htons(data->ioc_u32[1]);
1827 return ksocknal_add_peer(ni, &id, (struct sockaddr *)&sa);
1829 case IOC_LIBCFS_DEL_PEER:
1830 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1831 id.pid = LNET_PID_ANY;
1832 return ksocknal_del_peer(ni, &id);
1834 case IOC_LIBCFS_GET_CONN: {
1838 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
1839 struct sockaddr_in *psa = (void *)&conn->ksnc_peeraddr;
1840 struct sockaddr_in *mysa = (void *)&conn->ksnc_myaddr;
1845 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
1847 data->ioc_count = txmem;
1848 data->ioc_nid = lnet_nid_to_nid4(&conn->ksnc_peer->ksnp_id.nid);
1849 data->ioc_flags = nagle;
1850 if (psa->sin_family == AF_INET)
1851 data->ioc_u32[0] = ntohl(psa->sin_addr.s_addr);
1853 data->ioc_u32[0] = 0xFFFFFFFF;
1854 data->ioc_u32[1] = rpc_get_port((struct sockaddr *)
1855 &conn->ksnc_peeraddr);
1856 if (mysa->sin_family == AF_INET)
1857 data->ioc_u32[2] = ntohl(mysa->sin_addr.s_addr);
1859 data->ioc_u32[2] = 0xFFFFFFFF;
1860 data->ioc_u32[3] = conn->ksnc_type;
1861 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
1862 data->ioc_u32[5] = rxmem;
1863 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
1864 ksocknal_conn_decref(conn);
1868 case IOC_LIBCFS_CLOSE_CONNECTION:
1869 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1870 id.pid = LNET_PID_ANY;
1871 return ksocknal_close_matching_conns(&id,
1874 case IOC_LIBCFS_REGISTER_MYNID:
1875 /* Ignore if this is a noop */
1876 if (nid_is_nid4(&ni->ni_nid) &&
1877 data->ioc_nid == lnet_nid_to_nid4(&ni->ni_nid))
1880 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
1881 libcfs_nid2str(data->ioc_nid),
1882 libcfs_nidstr(&ni->ni_nid));
1885 case IOC_LIBCFS_PUSH_CONNECTION:
1886 lnet_nid4_to_nid(data->ioc_nid, &id.nid);
1887 id.pid = LNET_PID_ANY;
1888 return ksocknal_push(ni, &id);
1897 ksocknal_free_buffers (void)
1899 LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
1901 if (ksocknal_data.ksnd_schedulers != NULL)
1902 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
1904 spin_lock(&ksocknal_data.ksnd_tx_lock);
1906 if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
1908 struct ksock_tx *tx;
1910 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
1911 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1913 while ((tx = list_first_entry_or_null(&zlist, struct ksock_tx,
1914 tx_list)) != NULL) {
1915 list_del(&tx->tx_list);
1916 LIBCFS_FREE(tx, tx->tx_desc_size);
1919 spin_unlock(&ksocknal_data.ksnd_tx_lock);
1923 static int ksocknal_get_link_status(struct net_device *dev)
1929 if (!netif_running(dev)) {
1931 CDEBUG(D_NET, "device not running\n");
1933 /* Some devices may not be providing link settings */
1934 else if (dev->ethtool_ops->get_link) {
1935 ret = dev->ethtool_ops->get_link(dev);
1936 CDEBUG(D_NET, "get_link returns %u\n", ret);
1943 ksocknal_handle_link_state_change(struct net_device *dev,
1944 unsigned char operstate)
1946 struct lnet_ni *ni = NULL;
1947 struct ksock_net *net;
1948 struct ksock_net *cnxt;
1950 unsigned char link_down = !(operstate == IF_OPER_UP);
1951 struct in_device *in_dev;
1952 bool found_ip = false;
1953 struct ksock_interface *ksi = NULL;
1954 struct sockaddr_in *sa;
1955 DECLARE_CONST_IN_IFADDR(ifa);
1957 ifindex = dev->ifindex;
1959 if (!ksocknal_data.ksnd_nnets)
1962 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
1965 ksi = &net->ksnn_interface;
1966 sa = (void *)&ksi->ksni_addr;
1969 if (ksi->ksni_index != ifindex ||
1970 strcmp(ksi->ksni_name, dev->name))
1975 in_dev = __in_dev_get_rtnl(dev);
1977 CDEBUG(D_NET, "Interface %s has no IPv4 status.\n",
1979 CDEBUG(D_NET, "set link fatal state to 1\n");
1980 atomic_set(&ni->ni_fatal_error_on, 1);
1983 in_dev_for_each_ifa_rtnl(ifa, in_dev) {
1984 if (sa->sin_addr.s_addr == ifa->ifa_local)
1990 CDEBUG(D_NET, "Interface %s has no matching ip\n",
1992 CDEBUG(D_NET, "set link fatal state to 1\n");
1993 atomic_set(&ni->ni_fatal_error_on, 1);
1998 CDEBUG(D_NET, "set link fatal state to 1\n");
1999 atomic_set(&ni->ni_fatal_error_on, link_down);
2001 CDEBUG(D_NET, "set link fatal state to %u\n",
2002 (ksocknal_get_link_status(dev) == 0));
2003 atomic_set(&ni->ni_fatal_error_on,
2004 (ksocknal_get_link_status(dev) == 0));
2013 ksocknal_handle_inetaddr_change(struct in_ifaddr *ifa, unsigned long event)
2016 struct ksock_net *net;
2017 struct ksock_net *cnxt;
2018 struct net_device *event_netdev = ifa->ifa_dev->dev;
2020 struct ksock_interface *ksi = NULL;
2021 struct sockaddr_in *sa;
2023 if (!ksocknal_data.ksnd_nnets)
2026 ifindex = event_netdev->ifindex;
2028 list_for_each_entry_safe(net, cnxt, &ksocknal_data.ksnd_nets,
2031 ksi = &net->ksnn_interface;
2032 sa = (void *)&ksi->ksni_addr;
2034 if (ksi->ksni_index != ifindex ||
2035 strcmp(ksi->ksni_name, event_netdev->name))
2038 if (sa->sin_addr.s_addr == ifa->ifa_local) {
2039 CDEBUG(D_NET, "set link fatal state to %u\n",
2040 (event == NETDEV_DOWN));
2042 atomic_set(&ni->ni_fatal_error_on,
2043 (event == NETDEV_DOWN));
2050 /************************************
2051 * Net device notifier event handler
2052 ************************************/
2053 static int ksocknal_device_event(struct notifier_block *unused,
2054 unsigned long event, void *ptr)
2056 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2057 unsigned char operstate;
2059 operstate = dev->operstate;
2061 CDEBUG(D_NET, "devevent: status=%ld, iface=%s ifindex %d state %u\n",
2062 event, dev->name, dev->ifindex, operstate);
2068 ksocknal_handle_link_state_change(dev, operstate);
2075 /************************************
2076 * Inetaddr notifier event handler
2077 ************************************/
2078 static int ksocknal_inetaddr_event(struct notifier_block *unused,
2079 unsigned long event, void *ptr)
2081 struct in_ifaddr *ifa = ptr;
2083 CDEBUG(D_NET, "addrevent: status %ld ip addr %pI4, netmask %pI4.\n",
2084 event, &ifa->ifa_address, &ifa->ifa_mask);
2090 ksocknal_handle_inetaddr_change(ifa, event);
2097 static struct notifier_block ksocknal_dev_notifier_block = {
2098 .notifier_call = ksocknal_device_event,
2101 static struct notifier_block ksocknal_inetaddr_notifier_block = {
2102 .notifier_call = ksocknal_inetaddr_event,
2106 ksocknal_base_shutdown(void)
2108 struct ksock_sched *sched;
2109 struct ksock_peer_ni *peer_ni;
2112 CDEBUG(D_MALLOC, "before NAL cleanup: kmem %lld\n",
2113 libcfs_kmem_read());
2114 LASSERT (ksocknal_data.ksnd_nnets == 0);
2116 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL) {
2117 unregister_netdevice_notifier(&ksocknal_dev_notifier_block);
2118 unregister_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2121 switch (ksocknal_data.ksnd_init) {
2126 case SOCKNAL_INIT_ALL:
2127 case SOCKNAL_INIT_DATA:
2128 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2131 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2132 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2133 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2134 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2135 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2137 if (ksocknal_data.ksnd_schedulers != NULL) {
2138 cfs_percpt_for_each(sched, i,
2139 ksocknal_data.ksnd_schedulers) {
2141 LASSERT(list_empty(&sched->kss_tx_conns));
2142 LASSERT(list_empty(&sched->kss_rx_conns));
2143 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2144 LASSERT(sched->kss_nconns == 0);
2148 /* flag threads to terminate; wake and wait for them to die */
2149 ksocknal_data.ksnd_shuttingdown = 1;
2150 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2151 wake_up(&ksocknal_data.ksnd_reaper_waitq);
2153 if (ksocknal_data.ksnd_schedulers != NULL) {
2154 cfs_percpt_for_each(sched, i,
2155 ksocknal_data.ksnd_schedulers)
2156 wake_up_all(&sched->kss_waitq);
2159 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2160 atomic_read(&ksocknal_data.ksnd_nthreads) == 0,
2161 "waiting for %d threads to terminate\n",
2162 atomic_read(&ksocknal_data.ksnd_nthreads));
2164 ksocknal_free_buffers();
2166 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2170 CDEBUG(D_MALLOC, "after NAL cleanup: kmem %lld\n",
2171 libcfs_kmem_read());
2173 module_put(THIS_MODULE);
2177 ksocknal_base_startup(void)
2179 struct ksock_sched *sched;
2183 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2184 LASSERT(ksocknal_data.ksnd_nnets == 0);
2186 memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2188 hash_init(ksocknal_data.ksnd_peers);
2190 rwlock_init(&ksocknal_data.ksnd_global_lock);
2191 INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2193 spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2194 INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2195 INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2196 INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2197 init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2199 spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2200 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2201 INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2202 init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2204 spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2205 INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2207 /* NB memset above zeros whole of ksocknal_data */
2209 /* flag lists/ptrs/locks initialised */
2210 ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2211 if (!try_module_get(THIS_MODULE))
2214 /* Create a scheduler block per available CPT */
2215 ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2217 if (ksocknal_data.ksnd_schedulers == NULL)
2220 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2224 * make sure not to allocate more threads than there are
2225 * cores/CPUs in teh CPT
2227 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2228 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2229 nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2232 * max to half of CPUs, assume another half should be
2233 * reserved for upper layer modules
2235 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2238 sched->kss_nthreads_max = nthrs;
2241 spin_lock_init(&sched->kss_lock);
2242 INIT_LIST_HEAD(&sched->kss_rx_conns);
2243 INIT_LIST_HEAD(&sched->kss_tx_conns);
2244 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2245 init_waitqueue_head(&sched->kss_waitq);
2248 ksocknal_data.ksnd_connd_starting = 0;
2249 ksocknal_data.ksnd_connd_failed_stamp = 0;
2250 ksocknal_data.ksnd_connd_starting_stamp = ktime_get_real_seconds();
2251 /* must have at least 2 connds to remain responsive to accepts while
2253 if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2254 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2256 if (*ksocknal_tunables.ksnd_nconnds_max <
2257 *ksocknal_tunables.ksnd_nconnds) {
2258 ksocknal_tunables.ksnd_nconnds_max =
2259 ksocknal_tunables.ksnd_nconnds;
2262 for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2263 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2264 ksocknal_data.ksnd_connd_starting++;
2265 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2267 rc = ksocknal_thread_start(ksocknal_connd,
2268 (void *)((uintptr_t)i),
2269 "socknal_cd%02d", 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);
2274 CERROR("Can't spawn socknal connd: %d\n", rc);
2279 rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2281 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2285 register_netdevice_notifier(&ksocknal_dev_notifier_block);
2286 register_inetaddr_notifier(&ksocknal_inetaddr_notifier_block);
2288 /* flag everything initialised */
2289 ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2294 ksocknal_base_shutdown();
2299 ksocknal_debug_peerhash(struct lnet_ni *ni)
2301 struct ksock_peer_ni *peer_ni;
2304 read_lock(&ksocknal_data.ksnd_global_lock);
2306 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2307 struct ksock_conn_cb *conn_cb;
2308 struct ksock_conn *conn;
2310 if (peer_ni->ksnp_ni != ni)
2313 CWARN("Active peer_ni on shutdown: %s, ref %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2314 libcfs_idstr(&peer_ni->ksnp_id),
2315 refcount_read(&peer_ni->ksnp_refcount),
2316 peer_ni->ksnp_closing,
2317 peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2318 peer_ni->ksnp_zc_next_cookie,
2319 !list_empty(&peer_ni->ksnp_tx_queue),
2320 !list_empty(&peer_ni->ksnp_zc_req_list));
2322 conn_cb = peer_ni->ksnp_conn_cb;
2324 CWARN("ConnCB: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2325 refcount_read(&conn_cb->ksnr_refcount),
2326 conn_cb->ksnr_scheduled, conn_cb->ksnr_connecting,
2327 conn_cb->ksnr_connected, conn_cb->ksnr_deleted);
2330 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2331 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2332 refcount_read(&conn->ksnc_conn_refcount),
2333 refcount_read(&conn->ksnc_sock_refcount),
2334 conn->ksnc_type, conn->ksnc_closing);
2339 read_unlock(&ksocknal_data.ksnd_global_lock);
2344 ksocknal_shutdown(struct lnet_ni *ni)
2346 struct ksock_net *net = ni->ni_data;
2348 LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2349 LASSERT(ksocknal_data.ksnd_nnets > 0);
2351 /* prevent new peers */
2352 atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2354 /* Delete all peers */
2355 ksocknal_del_peer(ni, NULL);
2357 /* Wait for all peer_ni state to clean up */
2358 wait_var_event_warning(&net->ksnn_npeers,
2359 atomic_read(&net->ksnn_npeers) ==
2360 SOCKNAL_SHUTDOWN_BIAS,
2361 "waiting for %d peers to disconnect\n",
2362 ksocknal_debug_peerhash(ni) +
2363 atomic_read(&net->ksnn_npeers) -
2364 SOCKNAL_SHUTDOWN_BIAS);
2366 LASSERT(net->ksnn_interface.ksni_npeers == 0);
2367 LASSERT(net->ksnn_interface.ksni_nroutes == 0);
2369 list_del(&net->ksnn_list);
2370 LIBCFS_FREE(net, sizeof(*net));
2372 ksocknal_data.ksnd_nnets--;
2373 if (ksocknal_data.ksnd_nnets == 0)
2374 ksocknal_base_shutdown();
2378 ksocknal_search_new_ipif(struct ksock_net *net)
2381 char *ifnam = &net->ksnn_interface.ksni_name[0];
2382 char *colon = strchr(ifnam, ':');
2384 struct ksock_net *tmp;
2389 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets, ksnn_list) {
2390 char *ifnam2 = &tmp->ksnn_interface.ksni_name[0];
2391 char *colon2 = strchr(ifnam2, ':');
2396 found = strcmp(ifnam, ifnam2) == 0;
2409 ksocknal_start_schedulers(struct ksock_sched *sched)
2415 if (sched->kss_nthreads == 0) {
2416 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2417 nthrs = sched->kss_nthreads_max;
2419 nthrs = cfs_cpt_weight(lnet_cpt_table(),
2421 nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2422 nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2424 nthrs = min(nthrs, sched->kss_nthreads_max);
2426 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2427 /* increase two threads if there is new interface */
2428 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2431 for (i = 0; i < nthrs; i++) {
2434 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2435 rc = ksocknal_thread_start(ksocknal_scheduler, (void *)id,
2436 "socknal_sd%02d_%02d",
2438 (int)KSOCK_THREAD_SID(id));
2442 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2443 sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2447 sched->kss_nthreads += i;
2452 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2454 int newif = ksocknal_search_new_ipif(net);
2458 if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2461 for (i = 0; i < ncpts; i++) {
2462 struct ksock_sched *sched;
2463 int cpt = (cpts == NULL) ? i : cpts[i];
2465 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2466 sched = ksocknal_data.ksnd_schedulers[cpt];
2468 if (!newif && sched->kss_nthreads > 0)
2471 rc = ksocknal_start_schedulers(sched);
2479 ksocknal_startup(struct lnet_ni *ni)
2481 struct ksock_net *net;
2482 struct ksock_interface *ksi = NULL;
2483 struct lnet_inetdev *ifaces = NULL;
2487 LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2488 if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2489 rc = ksocknal_base_startup();
2493 LIBCFS_ALLOC(net, sizeof(*net));
2496 net->ksnn_incarnation = ktime_get_real_ns();
2499 ksocknal_tunables_setup(ni);
2501 rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns, true);
2505 ksi = &net->ksnn_interface;
2507 /* Use the first discovered interface or look in the list */
2508 if (ni->ni_interface) {
2509 for (i = 0; i < rc; i++)
2510 if (strcmp(ifaces[i].li_name, ni->ni_interface) == 0)
2513 /* ni_interfaces doesn't contain the interface we want */
2515 CERROR("ksocklnd: failed to find interface %s\n",
2521 ni->ni_dev_cpt = ifaces[i].li_cpt;
2522 ksi->ksni_index = ifaces[i].li_index;
2523 if (ifaces[i].li_ipv6) {
2524 struct sockaddr_in6 *sa;
2525 sa = (void *)&ksi->ksni_addr;
2526 memset(sa, 0, sizeof(*sa));
2527 sa->sin6_family = AF_INET6;
2528 memcpy(&sa->sin6_addr, ifaces[i].li_ipv6addr,
2529 sizeof(struct in6_addr));
2530 ni->ni_nid.nid_size = sizeof(struct in6_addr) - 4;
2531 memcpy(&ni->ni_nid.nid_addr, ifaces[i].li_ipv6addr,
2532 sizeof(struct in6_addr));
2534 struct sockaddr_in *sa;
2535 sa = (void *)&ksi->ksni_addr;
2536 memset(sa, 0, sizeof(*sa));
2537 sa->sin_family = AF_INET;
2538 sa->sin_addr.s_addr = htonl(ifaces[i].li_ipaddr);
2539 ksi->ksni_netmask = ifaces[i].li_netmask;
2540 ni->ni_nid.nid_size = 4 - 4;
2541 ni->ni_nid.nid_addr[0] = sa->sin_addr.s_addr;
2543 strlcpy(ksi->ksni_name, ifaces[i].li_name, sizeof(ksi->ksni_name));
2545 /* call it before add it to ksocknal_data.ksnd_nets */
2546 rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2550 list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2552 ksocknal_data.ksnd_nnets++;
2557 LIBCFS_FREE(net, sizeof(*net));
2559 if (ksocknal_data.ksnd_nnets == 0)
2560 ksocknal_base_shutdown();
2565 static void __exit ksocklnd_exit(void)
2567 lnet_unregister_lnd(&the_ksocklnd);
2570 static const struct lnet_lnd the_ksocklnd = {
2571 .lnd_type = SOCKLND,
2572 .lnd_startup = ksocknal_startup,
2573 .lnd_shutdown = ksocknal_shutdown,
2574 .lnd_ctl = ksocknal_ctl,
2575 .lnd_send = ksocknal_send,
2576 .lnd_recv = ksocknal_recv,
2577 .lnd_notify_peer_down = ksocknal_notify_gw_down,
2578 .lnd_accept = ksocknal_accept,
2581 static int __init ksocklnd_init(void)
2585 /* check ksnr_connected/connecting field large enough */
2586 BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2587 BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2589 rc = ksocknal_tunables_init();
2593 lnet_register_lnd(&the_ksocklnd);
2598 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2599 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2600 MODULE_VERSION("2.8.0");
2601 MODULE_LICENSE("GPL");
2603 module_init(ksocklnd_init);
2604 module_exit(ksocklnd_exit);