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, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/socklnd/socklnd.c
34 * Author: Zach Brown <zab@zabbo.net>
35 * Author: Peter J. Braam <braam@clusterfs.com>
36 * Author: Phil Schwan <phil@clusterfs.com>
37 * Author: Eric Barton <eric@bartonsoftware.com>
42 static lnd_t the_ksocklnd
;
43 struct ksock_nal_data ksocknal_data
;
45 static struct ksock_interface
*
46 ksocknal_ip2iface(lnet_ni_t
*ni
, __u32 ip
)
48 struct ksock_net
*net
= ni
->ni_data
;
50 struct ksock_interface
*iface
;
52 for (i
= 0; i
< net
->ksnn_ninterfaces
; i
++) {
53 LASSERT(i
< LNET_MAX_INTERFACES
);
54 iface
= &net
->ksnn_interfaces
[i
];
56 if (iface
->ksni_ipaddr
== ip
)
63 static struct ksock_route
*
64 ksocknal_create_route(__u32 ipaddr
, int port
)
66 struct ksock_route
*route
;
68 LIBCFS_ALLOC(route
, sizeof(*route
));
72 atomic_set(&route
->ksnr_refcount
, 1);
73 route
->ksnr_peer
= NULL
;
74 route
->ksnr_retry_interval
= 0; /* OK to connect at any time */
75 route
->ksnr_ipaddr
= ipaddr
;
76 route
->ksnr_port
= port
;
77 route
->ksnr_scheduled
= 0;
78 route
->ksnr_connecting
= 0;
79 route
->ksnr_connected
= 0;
80 route
->ksnr_deleted
= 0;
81 route
->ksnr_conn_count
= 0;
82 route
->ksnr_share_count
= 0;
88 ksocknal_destroy_route(struct ksock_route
*route
)
90 LASSERT(!atomic_read(&route
->ksnr_refcount
));
93 ksocknal_peer_decref(route
->ksnr_peer
);
95 LIBCFS_FREE(route
, sizeof(*route
));
99 ksocknal_create_peer(struct ksock_peer
**peerp
, lnet_ni_t
*ni
, lnet_process_id_t id
)
101 int cpt
= lnet_cpt_of_nid(id
.nid
);
102 struct ksock_net
*net
= ni
->ni_data
;
103 struct ksock_peer
*peer
;
105 LASSERT(id
.nid
!= LNET_NID_ANY
);
106 LASSERT(id
.pid
!= LNET_PID_ANY
);
107 LASSERT(!in_interrupt());
109 LIBCFS_CPT_ALLOC(peer
, lnet_cpt_table(), cpt
, sizeof(*peer
));
115 atomic_set(&peer
->ksnp_refcount
, 1); /* 1 ref for caller */
116 peer
->ksnp_closing
= 0;
117 peer
->ksnp_accepting
= 0;
118 peer
->ksnp_proto
= NULL
;
119 peer
->ksnp_last_alive
= 0;
120 peer
->ksnp_zc_next_cookie
= SOCKNAL_KEEPALIVE_PING
+ 1;
122 INIT_LIST_HEAD(&peer
->ksnp_conns
);
123 INIT_LIST_HEAD(&peer
->ksnp_routes
);
124 INIT_LIST_HEAD(&peer
->ksnp_tx_queue
);
125 INIT_LIST_HEAD(&peer
->ksnp_zc_req_list
);
126 spin_lock_init(&peer
->ksnp_lock
);
128 spin_lock_bh(&net
->ksnn_lock
);
130 if (net
->ksnn_shutdown
) {
131 spin_unlock_bh(&net
->ksnn_lock
);
133 LIBCFS_FREE(peer
, sizeof(*peer
));
134 CERROR("Can't create peer: network shutdown\n");
140 spin_unlock_bh(&net
->ksnn_lock
);
147 ksocknal_destroy_peer(struct ksock_peer
*peer
)
149 struct ksock_net
*net
= peer
->ksnp_ni
->ni_data
;
151 CDEBUG(D_NET
, "peer %s %p deleted\n",
152 libcfs_id2str(peer
->ksnp_id
), peer
);
154 LASSERT(!atomic_read(&peer
->ksnp_refcount
));
155 LASSERT(!peer
->ksnp_accepting
);
156 LASSERT(list_empty(&peer
->ksnp_conns
));
157 LASSERT(list_empty(&peer
->ksnp_routes
));
158 LASSERT(list_empty(&peer
->ksnp_tx_queue
));
159 LASSERT(list_empty(&peer
->ksnp_zc_req_list
));
161 LIBCFS_FREE(peer
, sizeof(*peer
));
164 * NB a peer's connections and routes keep a reference on their peer
165 * until they are destroyed, so we can be assured that _all_ state to
166 * do with this peer has been cleaned up when its refcount drops to
169 spin_lock_bh(&net
->ksnn_lock
);
171 spin_unlock_bh(&net
->ksnn_lock
);
175 ksocknal_find_peer_locked(lnet_ni_t
*ni
, lnet_process_id_t id
)
177 struct list_head
*peer_list
= ksocknal_nid2peerlist(id
.nid
);
178 struct list_head
*tmp
;
179 struct ksock_peer
*peer
;
181 list_for_each(tmp
, peer_list
) {
182 peer
= list_entry(tmp
, struct ksock_peer
, ksnp_list
);
184 LASSERT(!peer
->ksnp_closing
);
186 if (peer
->ksnp_ni
!= ni
)
189 if (peer
->ksnp_id
.nid
!= id
.nid
||
190 peer
->ksnp_id
.pid
!= id
.pid
)
193 CDEBUG(D_NET
, "got peer [%p] -> %s (%d)\n",
194 peer
, libcfs_id2str(id
),
195 atomic_read(&peer
->ksnp_refcount
));
202 ksocknal_find_peer(lnet_ni_t
*ni
, lnet_process_id_t id
)
204 struct ksock_peer
*peer
;
206 read_lock(&ksocknal_data
.ksnd_global_lock
);
207 peer
= ksocknal_find_peer_locked(ni
, id
);
208 if (peer
) /* +1 ref for caller? */
209 ksocknal_peer_addref(peer
);
210 read_unlock(&ksocknal_data
.ksnd_global_lock
);
216 ksocknal_unlink_peer_locked(struct ksock_peer
*peer
)
220 struct ksock_interface
*iface
;
222 for (i
= 0; i
< peer
->ksnp_n_passive_ips
; i
++) {
223 LASSERT(i
< LNET_MAX_INTERFACES
);
224 ip
= peer
->ksnp_passive_ips
[i
];
226 iface
= ksocknal_ip2iface(peer
->ksnp_ni
, ip
);
228 * All IPs in peer->ksnp_passive_ips[] come from the
229 * interface list, therefore the call must succeed.
233 CDEBUG(D_NET
, "peer=%p iface=%p ksni_nroutes=%d\n",
234 peer
, iface
, iface
->ksni_nroutes
);
235 iface
->ksni_npeers
--;
238 LASSERT(list_empty(&peer
->ksnp_conns
));
239 LASSERT(list_empty(&peer
->ksnp_routes
));
240 LASSERT(!peer
->ksnp_closing
);
241 peer
->ksnp_closing
= 1;
242 list_del(&peer
->ksnp_list
);
243 /* lose peerlist's ref */
244 ksocknal_peer_decref(peer
);
248 ksocknal_get_peer_info(lnet_ni_t
*ni
, int index
,
249 lnet_process_id_t
*id
, __u32
*myip
, __u32
*peer_ip
,
250 int *port
, int *conn_count
, int *share_count
)
252 struct ksock_peer
*peer
;
253 struct list_head
*ptmp
;
254 struct ksock_route
*route
;
255 struct list_head
*rtmp
;
260 read_lock(&ksocknal_data
.ksnd_global_lock
);
262 for (i
= 0; i
< ksocknal_data
.ksnd_peer_hash_size
; i
++) {
263 list_for_each(ptmp
, &ksocknal_data
.ksnd_peers
[i
]) {
264 peer
= list_entry(ptmp
, struct ksock_peer
, ksnp_list
);
266 if (peer
->ksnp_ni
!= ni
)
269 if (!peer
->ksnp_n_passive_ips
&&
270 list_empty(&peer
->ksnp_routes
)) {
284 for (j
= 0; j
< peer
->ksnp_n_passive_ips
; j
++) {
289 *myip
= peer
->ksnp_passive_ips
[j
];
298 list_for_each(rtmp
, &peer
->ksnp_routes
) {
302 route
= list_entry(rtmp
, struct ksock_route
,
306 *myip
= route
->ksnr_myipaddr
;
307 *peer_ip
= route
->ksnr_ipaddr
;
308 *port
= route
->ksnr_port
;
309 *conn_count
= route
->ksnr_conn_count
;
310 *share_count
= route
->ksnr_share_count
;
317 read_unlock(&ksocknal_data
.ksnd_global_lock
);
322 ksocknal_associate_route_conn_locked(struct ksock_route
*route
, struct ksock_conn
*conn
)
324 struct ksock_peer
*peer
= route
->ksnr_peer
;
325 int type
= conn
->ksnc_type
;
326 struct ksock_interface
*iface
;
328 conn
->ksnc_route
= route
;
329 ksocknal_route_addref(route
);
331 if (route
->ksnr_myipaddr
!= conn
->ksnc_myipaddr
) {
332 if (!route
->ksnr_myipaddr
) {
333 /* route wasn't bound locally yet (the initial route) */
334 CDEBUG(D_NET
, "Binding %s %pI4h to %pI4h\n",
335 libcfs_id2str(peer
->ksnp_id
),
337 &conn
->ksnc_myipaddr
);
339 CDEBUG(D_NET
, "Rebinding %s %pI4h from %pI4h to %pI4h\n",
340 libcfs_id2str(peer
->ksnp_id
),
342 &route
->ksnr_myipaddr
,
343 &conn
->ksnc_myipaddr
);
345 iface
= ksocknal_ip2iface(route
->ksnr_peer
->ksnp_ni
,
346 route
->ksnr_myipaddr
);
348 iface
->ksni_nroutes
--;
350 route
->ksnr_myipaddr
= conn
->ksnc_myipaddr
;
351 iface
= ksocknal_ip2iface(route
->ksnr_peer
->ksnp_ni
,
352 route
->ksnr_myipaddr
);
354 iface
->ksni_nroutes
++;
357 route
->ksnr_connected
|= (1 << type
);
358 route
->ksnr_conn_count
++;
361 * Successful connection => further attempts can
362 * proceed immediately
364 route
->ksnr_retry_interval
= 0;
368 ksocknal_add_route_locked(struct ksock_peer
*peer
, struct ksock_route
*route
)
370 struct list_head
*tmp
;
371 struct ksock_conn
*conn
;
372 struct ksock_route
*route2
;
374 LASSERT(!peer
->ksnp_closing
);
375 LASSERT(!route
->ksnr_peer
);
376 LASSERT(!route
->ksnr_scheduled
);
377 LASSERT(!route
->ksnr_connecting
);
378 LASSERT(!route
->ksnr_connected
);
380 /* LASSERT(unique) */
381 list_for_each(tmp
, &peer
->ksnp_routes
) {
382 route2
= list_entry(tmp
, struct ksock_route
, ksnr_list
);
384 if (route2
->ksnr_ipaddr
== route
->ksnr_ipaddr
) {
385 CERROR("Duplicate route %s %pI4h\n",
386 libcfs_id2str(peer
->ksnp_id
),
387 &route
->ksnr_ipaddr
);
392 route
->ksnr_peer
= peer
;
393 ksocknal_peer_addref(peer
);
394 /* peer's routelist takes over my ref on 'route' */
395 list_add_tail(&route
->ksnr_list
, &peer
->ksnp_routes
);
397 list_for_each(tmp
, &peer
->ksnp_conns
) {
398 conn
= list_entry(tmp
, struct ksock_conn
, ksnc_list
);
400 if (conn
->ksnc_ipaddr
!= route
->ksnr_ipaddr
)
403 ksocknal_associate_route_conn_locked(route
, conn
);
404 /* keep going (typed routes) */
409 ksocknal_del_route_locked(struct ksock_route
*route
)
411 struct ksock_peer
*peer
= route
->ksnr_peer
;
412 struct ksock_interface
*iface
;
413 struct ksock_conn
*conn
;
414 struct list_head
*ctmp
;
415 struct list_head
*cnxt
;
417 LASSERT(!route
->ksnr_deleted
);
419 /* Close associated conns */
420 list_for_each_safe(ctmp
, cnxt
, &peer
->ksnp_conns
) {
421 conn
= list_entry(ctmp
, struct ksock_conn
, ksnc_list
);
423 if (conn
->ksnc_route
!= route
)
426 ksocknal_close_conn_locked(conn
, 0);
429 if (route
->ksnr_myipaddr
) {
430 iface
= ksocknal_ip2iface(route
->ksnr_peer
->ksnp_ni
,
431 route
->ksnr_myipaddr
);
433 iface
->ksni_nroutes
--;
436 route
->ksnr_deleted
= 1;
437 list_del(&route
->ksnr_list
);
438 ksocknal_route_decref(route
); /* drop peer's ref */
440 if (list_empty(&peer
->ksnp_routes
) &&
441 list_empty(&peer
->ksnp_conns
)) {
443 * I've just removed the last route to a peer with no active
446 ksocknal_unlink_peer_locked(peer
);
451 ksocknal_add_peer(lnet_ni_t
*ni
, lnet_process_id_t id
, __u32 ipaddr
, int port
)
453 struct list_head
*tmp
;
454 struct ksock_peer
*peer
;
455 struct ksock_peer
*peer2
;
456 struct ksock_route
*route
;
457 struct ksock_route
*route2
;
460 if (id
.nid
== LNET_NID_ANY
||
461 id
.pid
== LNET_PID_ANY
)
464 /* Have a brand new peer ready... */
465 rc
= ksocknal_create_peer(&peer
, ni
, id
);
469 route
= ksocknal_create_route(ipaddr
, port
);
471 ksocknal_peer_decref(peer
);
475 write_lock_bh(&ksocknal_data
.ksnd_global_lock
);
477 /* always called with a ref on ni, so shutdown can't have started */
478 LASSERT(!((struct ksock_net
*)ni
->ni_data
)->ksnn_shutdown
);
480 peer2
= ksocknal_find_peer_locked(ni
, id
);
482 ksocknal_peer_decref(peer
);
485 /* peer table takes my ref on peer */
486 list_add_tail(&peer
->ksnp_list
,
487 ksocknal_nid2peerlist(id
.nid
));
491 list_for_each(tmp
, &peer
->ksnp_routes
) {
492 route2
= list_entry(tmp
, struct ksock_route
, ksnr_list
);
494 if (route2
->ksnr_ipaddr
== ipaddr
)
500 ksocknal_add_route_locked(peer
, route
);
501 route
->ksnr_share_count
++;
503 ksocknal_route_decref(route
);
504 route2
->ksnr_share_count
++;
507 write_unlock_bh(&ksocknal_data
.ksnd_global_lock
);
513 ksocknal_del_peer_locked(struct ksock_peer
*peer
, __u32 ip
)
515 struct ksock_conn
*conn
;
516 struct ksock_route
*route
;
517 struct list_head
*tmp
;
518 struct list_head
*nxt
;
521 LASSERT(!peer
->ksnp_closing
);
523 /* Extra ref prevents peer disappearing until I'm done with it */
524 ksocknal_peer_addref(peer
);
526 list_for_each_safe(tmp
, nxt
, &peer
->ksnp_routes
) {
527 route
= list_entry(tmp
, struct ksock_route
, ksnr_list
);
530 if (!(!ip
|| route
->ksnr_ipaddr
== ip
))
533 route
->ksnr_share_count
= 0;
534 /* This deletes associated conns too */
535 ksocknal_del_route_locked(route
);
539 list_for_each_safe(tmp
, nxt
, &peer
->ksnp_routes
) {
540 route
= list_entry(tmp
, struct ksock_route
, ksnr_list
);
541 nshared
+= route
->ksnr_share_count
;
546 * remove everything else if there are no explicit entries
549 list_for_each_safe(tmp
, nxt
, &peer
->ksnp_routes
) {
550 route
= list_entry(tmp
, struct ksock_route
, ksnr_list
);
552 /* we should only be removing auto-entries */
553 LASSERT(!route
->ksnr_share_count
);
554 ksocknal_del_route_locked(route
);
557 list_for_each_safe(tmp
, nxt
, &peer
->ksnp_conns
) {
558 conn
= list_entry(tmp
, struct ksock_conn
, ksnc_list
);
560 ksocknal_close_conn_locked(conn
, 0);
564 ksocknal_peer_decref(peer
);
565 /* NB peer unlinks itself when last conn/route is removed */
569 ksocknal_del_peer(lnet_ni_t
*ni
, lnet_process_id_t id
, __u32 ip
)
572 struct list_head
*ptmp
;
573 struct list_head
*pnxt
;
574 struct ksock_peer
*peer
;
580 write_lock_bh(&ksocknal_data
.ksnd_global_lock
);
582 if (id
.nid
!= LNET_NID_ANY
) {
583 lo
= (int)(ksocknal_nid2peerlist(id
.nid
) - ksocknal_data
.ksnd_peers
);
584 hi
= (int)(ksocknal_nid2peerlist(id
.nid
) - ksocknal_data
.ksnd_peers
);
587 hi
= ksocknal_data
.ksnd_peer_hash_size
- 1;
590 for (i
= lo
; i
<= hi
; i
++) {
591 list_for_each_safe(ptmp
, pnxt
, &ksocknal_data
.ksnd_peers
[i
]) {
592 peer
= list_entry(ptmp
, struct ksock_peer
, ksnp_list
);
594 if (peer
->ksnp_ni
!= ni
)
597 if (!((id
.nid
== LNET_NID_ANY
|| peer
->ksnp_id
.nid
== id
.nid
) &&
598 (id
.pid
== LNET_PID_ANY
|| peer
->ksnp_id
.pid
== id
.pid
)))
601 ksocknal_peer_addref(peer
); /* a ref for me... */
603 ksocknal_del_peer_locked(peer
, ip
);
605 if (peer
->ksnp_closing
&&
606 !list_empty(&peer
->ksnp_tx_queue
)) {
607 LASSERT(list_empty(&peer
->ksnp_conns
));
608 LASSERT(list_empty(&peer
->ksnp_routes
));
610 list_splice_init(&peer
->ksnp_tx_queue
,
614 ksocknal_peer_decref(peer
); /* ...till here */
616 rc
= 0; /* matched! */
620 write_unlock_bh(&ksocknal_data
.ksnd_global_lock
);
622 ksocknal_txlist_done(ni
, &zombies
, 1);
627 static struct ksock_conn
*
628 ksocknal_get_conn_by_idx(lnet_ni_t
*ni
, int index
)
630 struct ksock_peer
*peer
;
631 struct list_head
*ptmp
;
632 struct ksock_conn
*conn
;
633 struct list_head
*ctmp
;
636 read_lock(&ksocknal_data
.ksnd_global_lock
);
638 for (i
= 0; i
< ksocknal_data
.ksnd_peer_hash_size
; i
++) {
639 list_for_each(ptmp
, &ksocknal_data
.ksnd_peers
[i
]) {
640 peer
= list_entry(ptmp
, struct ksock_peer
, ksnp_list
);
642 LASSERT(!peer
->ksnp_closing
);
644 if (peer
->ksnp_ni
!= ni
)
647 list_for_each(ctmp
, &peer
->ksnp_conns
) {
651 conn
= list_entry(ctmp
, struct ksock_conn
,
653 ksocknal_conn_addref(conn
);
654 read_unlock(&ksocknal_data
.ksnd_global_lock
);
660 read_unlock(&ksocknal_data
.ksnd_global_lock
);
664 static struct ksock_sched
*
665 ksocknal_choose_scheduler_locked(unsigned int cpt
)
667 struct ksock_sched_info
*info
= ksocknal_data
.ksnd_sched_info
[cpt
];
668 struct ksock_sched
*sched
;
671 LASSERT(info
->ksi_nthreads
> 0);
673 sched
= &info
->ksi_scheds
[0];
675 * NB: it's safe so far, but info->ksi_nthreads could be changed
676 * at runtime when we have dynamic LNet configuration, then we
677 * need to take care of this.
679 for (i
= 1; i
< info
->ksi_nthreads
; i
++) {
680 if (sched
->kss_nconns
> info
->ksi_scheds
[i
].kss_nconns
)
681 sched
= &info
->ksi_scheds
[i
];
688 ksocknal_local_ipvec(lnet_ni_t
*ni
, __u32
*ipaddrs
)
690 struct ksock_net
*net
= ni
->ni_data
;
694 read_lock(&ksocknal_data
.ksnd_global_lock
);
696 nip
= net
->ksnn_ninterfaces
;
697 LASSERT(nip
<= LNET_MAX_INTERFACES
);
700 * Only offer interfaces for additional connections if I have
704 read_unlock(&ksocknal_data
.ksnd_global_lock
);
708 for (i
= 0; i
< nip
; i
++) {
709 ipaddrs
[i
] = net
->ksnn_interfaces
[i
].ksni_ipaddr
;
713 read_unlock(&ksocknal_data
.ksnd_global_lock
);
718 ksocknal_match_peerip(struct ksock_interface
*iface
, __u32
*ips
, int nips
)
720 int best_netmatch
= 0;
727 for (i
= 0; i
< nips
; i
++) {
731 this_xor
= ips
[i
] ^ iface
->ksni_ipaddr
;
732 this_netmatch
= !(this_xor
& iface
->ksni_netmask
) ? 1 : 0;
735 best_netmatch
< this_netmatch
||
736 (best_netmatch
== this_netmatch
&&
737 best_xor
> this_xor
)))
741 best_netmatch
= this_netmatch
;
750 ksocknal_select_ips(struct ksock_peer
*peer
, __u32
*peerips
, int n_peerips
)
752 rwlock_t
*global_lock
= &ksocknal_data
.ksnd_global_lock
;
753 struct ksock_net
*net
= peer
->ksnp_ni
->ni_data
;
754 struct ksock_interface
*iface
;
755 struct ksock_interface
*best_iface
;
767 * CAVEAT EMPTOR: We do all our interface matching with an
768 * exclusive hold of global lock at IRQ priority. We're only
769 * expecting to be dealing with small numbers of interfaces, so the
770 * O(n**3)-ness shouldn't matter
773 * Also note that I'm not going to return more than n_peerips
774 * interfaces, even if I have more myself
776 write_lock_bh(global_lock
);
778 LASSERT(n_peerips
<= LNET_MAX_INTERFACES
);
779 LASSERT(net
->ksnn_ninterfaces
<= LNET_MAX_INTERFACES
);
782 * Only match interfaces for additional connections
783 * if I have > 1 interface
785 n_ips
= (net
->ksnn_ninterfaces
< 2) ? 0 :
786 min(n_peerips
, net
->ksnn_ninterfaces
);
788 for (i
= 0; peer
->ksnp_n_passive_ips
< n_ips
; i
++) {
789 /* ^ yes really... */
792 * If we have any new interfaces, first tick off all the
793 * peer IPs that match old interfaces, then choose new
794 * interfaces to match the remaining peer IPS.
795 * We don't forget interfaces we've stopped using; we might
796 * start using them again...
798 if (i
< peer
->ksnp_n_passive_ips
) {
800 ip
= peer
->ksnp_passive_ips
[i
];
801 best_iface
= ksocknal_ip2iface(peer
->ksnp_ni
, ip
);
803 /* peer passive ips are kept up to date */
806 /* choose a new interface */
807 LASSERT(i
== peer
->ksnp_n_passive_ips
);
813 for (j
= 0; j
< net
->ksnn_ninterfaces
; j
++) {
814 iface
= &net
->ksnn_interfaces
[j
];
815 ip
= iface
->ksni_ipaddr
;
817 for (k
= 0; k
< peer
->ksnp_n_passive_ips
; k
++)
818 if (peer
->ksnp_passive_ips
[k
] == ip
)
821 if (k
< peer
->ksnp_n_passive_ips
) /* using it already */
824 k
= ksocknal_match_peerip(iface
, peerips
, n_peerips
);
825 xor = ip
^ peerips
[k
];
826 this_netmatch
= !(xor & iface
->ksni_netmask
) ? 1 : 0;
829 best_netmatch
< this_netmatch
||
830 (best_netmatch
== this_netmatch
&&
831 best_npeers
> iface
->ksni_npeers
)))
835 best_netmatch
= this_netmatch
;
836 best_npeers
= iface
->ksni_npeers
;
841 best_iface
->ksni_npeers
++;
842 ip
= best_iface
->ksni_ipaddr
;
843 peer
->ksnp_passive_ips
[i
] = ip
;
844 peer
->ksnp_n_passive_ips
= i
+ 1;
847 /* mark the best matching peer IP used */
848 j
= ksocknal_match_peerip(best_iface
, peerips
, n_peerips
);
852 /* Overwrite input peer IP addresses */
853 memcpy(peerips
, peer
->ksnp_passive_ips
, n_ips
* sizeof(*peerips
));
855 write_unlock_bh(global_lock
);
861 ksocknal_create_routes(struct ksock_peer
*peer
, int port
,
862 __u32
*peer_ipaddrs
, int npeer_ipaddrs
)
864 struct ksock_route
*newroute
= NULL
;
865 rwlock_t
*global_lock
= &ksocknal_data
.ksnd_global_lock
;
866 lnet_ni_t
*ni
= peer
->ksnp_ni
;
867 struct ksock_net
*net
= ni
->ni_data
;
868 struct list_head
*rtmp
;
869 struct ksock_route
*route
;
870 struct ksock_interface
*iface
;
871 struct ksock_interface
*best_iface
;
879 * CAVEAT EMPTOR: We do all our interface matching with an
880 * exclusive hold of global lock at IRQ priority. We're only
881 * expecting to be dealing with small numbers of interfaces, so the
882 * O(n**3)-ness here shouldn't matter
884 write_lock_bh(global_lock
);
886 if (net
->ksnn_ninterfaces
< 2) {
888 * Only create additional connections
889 * if I have > 1 interface
891 write_unlock_bh(global_lock
);
895 LASSERT(npeer_ipaddrs
<= LNET_MAX_INTERFACES
);
897 for (i
= 0; i
< npeer_ipaddrs
; i
++) {
899 newroute
->ksnr_ipaddr
= peer_ipaddrs
[i
];
901 write_unlock_bh(global_lock
);
903 newroute
= ksocknal_create_route(peer_ipaddrs
[i
], port
);
907 write_lock_bh(global_lock
);
910 if (peer
->ksnp_closing
) {
911 /* peer got closed under me */
915 /* Already got a route? */
917 list_for_each(rtmp
, &peer
->ksnp_routes
) {
918 route
= list_entry(rtmp
, struct ksock_route
, ksnr_list
);
920 if (route
->ksnr_ipaddr
== newroute
->ksnr_ipaddr
)
932 LASSERT(net
->ksnn_ninterfaces
<= LNET_MAX_INTERFACES
);
934 /* Select interface to connect from */
935 for (j
= 0; j
< net
->ksnn_ninterfaces
; j
++) {
936 iface
= &net
->ksnn_interfaces
[j
];
938 /* Using this interface already? */
939 list_for_each(rtmp
, &peer
->ksnp_routes
) {
940 route
= list_entry(rtmp
, struct ksock_route
,
943 if (route
->ksnr_myipaddr
== iface
->ksni_ipaddr
)
951 this_netmatch
= (!((iface
->ksni_ipaddr
^
952 newroute
->ksnr_ipaddr
) &
953 iface
->ksni_netmask
)) ? 1 : 0;
956 best_netmatch
< this_netmatch
||
957 (best_netmatch
== this_netmatch
&&
958 best_nroutes
> iface
->ksni_nroutes
)))
962 best_netmatch
= this_netmatch
;
963 best_nroutes
= iface
->ksni_nroutes
;
969 newroute
->ksnr_myipaddr
= best_iface
->ksni_ipaddr
;
970 best_iface
->ksni_nroutes
++;
972 ksocknal_add_route_locked(peer
, newroute
);
976 write_unlock_bh(global_lock
);
978 ksocknal_route_decref(newroute
);
982 ksocknal_accept(lnet_ni_t
*ni
, struct socket
*sock
)
984 struct ksock_connreq
*cr
;
989 rc
= lnet_sock_getaddr(sock
, 1, &peer_ip
, &peer_port
);
990 LASSERT(!rc
); /* we succeeded before */
992 LIBCFS_ALLOC(cr
, sizeof(*cr
));
994 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from %pI4h: memory exhausted\n",
1001 cr
->ksncr_sock
= sock
;
1003 spin_lock_bh(&ksocknal_data
.ksnd_connd_lock
);
1005 list_add_tail(&cr
->ksncr_list
, &ksocknal_data
.ksnd_connd_connreqs
);
1006 wake_up(&ksocknal_data
.ksnd_connd_waitq
);
1008 spin_unlock_bh(&ksocknal_data
.ksnd_connd_lock
);
1013 ksocknal_connecting(struct ksock_peer
*peer
, __u32 ipaddr
)
1015 struct ksock_route
*route
;
1017 list_for_each_entry(route
, &peer
->ksnp_routes
, ksnr_list
) {
1018 if (route
->ksnr_ipaddr
== ipaddr
)
1019 return route
->ksnr_connecting
;
1025 ksocknal_create_conn(lnet_ni_t
*ni
, struct ksock_route
*route
,
1026 struct socket
*sock
, int type
)
1028 rwlock_t
*global_lock
= &ksocknal_data
.ksnd_global_lock
;
1030 lnet_process_id_t peerid
;
1031 struct list_head
*tmp
;
1033 struct ksock_conn
*conn
;
1034 struct ksock_conn
*conn2
;
1035 struct ksock_peer
*peer
= NULL
;
1036 struct ksock_peer
*peer2
;
1037 struct ksock_sched
*sched
;
1038 ksock_hello_msg_t
*hello
;
1040 struct ksock_tx
*tx
;
1041 struct ksock_tx
*txtmp
;
1048 LASSERT(active
== (type
!= SOCKLND_CONN_NONE
));
1050 LIBCFS_ALLOC(conn
, sizeof(*conn
));
1056 conn
->ksnc_peer
= NULL
;
1057 conn
->ksnc_route
= NULL
;
1058 conn
->ksnc_sock
= sock
;
1060 * 2 ref, 1 for conn, another extra ref prevents socket
1061 * being closed before establishment of connection
1063 atomic_set(&conn
->ksnc_sock_refcount
, 2);
1064 conn
->ksnc_type
= type
;
1065 ksocknal_lib_save_callback(sock
, conn
);
1066 atomic_set(&conn
->ksnc_conn_refcount
, 1); /* 1 ref for me */
1068 conn
->ksnc_rx_ready
= 0;
1069 conn
->ksnc_rx_scheduled
= 0;
1071 INIT_LIST_HEAD(&conn
->ksnc_tx_queue
);
1072 conn
->ksnc_tx_ready
= 0;
1073 conn
->ksnc_tx_scheduled
= 0;
1074 conn
->ksnc_tx_carrier
= NULL
;
1075 atomic_set(&conn
->ksnc_tx_nob
, 0);
1077 LIBCFS_ALLOC(hello
, offsetof(ksock_hello_msg_t
,
1078 kshm_ips
[LNET_MAX_INTERFACES
]));
1084 /* stash conn's local and remote addrs */
1085 rc
= ksocknal_lib_get_conn_addrs(conn
);
1090 * Find out/confirm peer's NID and connection type and get the
1091 * vector of interfaces she's willing to let me connect to.
1092 * Passive connections use the listener timeout since the peer sends
1096 peer
= route
->ksnr_peer
;
1097 LASSERT(ni
== peer
->ksnp_ni
);
1099 /* Active connection sends HELLO eagerly */
1100 hello
->kshm_nips
= ksocknal_local_ipvec(ni
, hello
->kshm_ips
);
1101 peerid
= peer
->ksnp_id
;
1103 write_lock_bh(global_lock
);
1104 conn
->ksnc_proto
= peer
->ksnp_proto
;
1105 write_unlock_bh(global_lock
);
1107 if (!conn
->ksnc_proto
) {
1108 conn
->ksnc_proto
= &ksocknal_protocol_v3x
;
1109 #if SOCKNAL_VERSION_DEBUG
1110 if (*ksocknal_tunables
.ksnd_protocol
== 2)
1111 conn
->ksnc_proto
= &ksocknal_protocol_v2x
;
1112 else if (*ksocknal_tunables
.ksnd_protocol
== 1)
1113 conn
->ksnc_proto
= &ksocknal_protocol_v1x
;
1117 rc
= ksocknal_send_hello(ni
, conn
, peerid
.nid
, hello
);
1121 peerid
.nid
= LNET_NID_ANY
;
1122 peerid
.pid
= LNET_PID_ANY
;
1124 /* Passive, get protocol from peer */
1125 conn
->ksnc_proto
= NULL
;
1128 rc
= ksocknal_recv_hello(ni
, conn
, hello
, &peerid
, &incarnation
);
1132 LASSERT(!rc
|| active
);
1133 LASSERT(conn
->ksnc_proto
);
1134 LASSERT(peerid
.nid
!= LNET_NID_ANY
);
1136 cpt
= lnet_cpt_of_nid(peerid
.nid
);
1139 ksocknal_peer_addref(peer
);
1140 write_lock_bh(global_lock
);
1142 rc
= ksocknal_create_peer(&peer
, ni
, peerid
);
1146 write_lock_bh(global_lock
);
1148 /* called with a ref on ni, so shutdown can't have started */
1149 LASSERT(!((struct ksock_net
*)ni
->ni_data
)->ksnn_shutdown
);
1151 peer2
= ksocknal_find_peer_locked(ni
, peerid
);
1154 * NB this puts an "empty" peer in the peer
1155 * table (which takes my ref)
1157 list_add_tail(&peer
->ksnp_list
,
1158 ksocknal_nid2peerlist(peerid
.nid
));
1160 ksocknal_peer_decref(peer
);
1165 ksocknal_peer_addref(peer
);
1166 peer
->ksnp_accepting
++;
1169 * Am I already connecting to this guy? Resolve in
1170 * favour of higher NID...
1172 if (peerid
.nid
< ni
->ni_nid
&&
1173 ksocknal_connecting(peer
, conn
->ksnc_ipaddr
)) {
1175 warn
= "connection race resolution";
1180 if (peer
->ksnp_closing
||
1181 (active
&& route
->ksnr_deleted
)) {
1182 /* peer/route got closed under me */
1184 warn
= "peer/route removed";
1188 if (!peer
->ksnp_proto
) {
1190 * Never connected before.
1191 * NB recv_hello may have returned EPROTO to signal my peer
1192 * wants a different protocol than the one I asked for.
1194 LASSERT(list_empty(&peer
->ksnp_conns
));
1196 peer
->ksnp_proto
= conn
->ksnc_proto
;
1197 peer
->ksnp_incarnation
= incarnation
;
1200 if (peer
->ksnp_proto
!= conn
->ksnc_proto
||
1201 peer
->ksnp_incarnation
!= incarnation
) {
1202 /* Peer rebooted or I've got the wrong protocol version */
1203 ksocknal_close_peer_conns_locked(peer
, 0, 0);
1205 peer
->ksnp_proto
= NULL
;
1207 warn
= peer
->ksnp_incarnation
!= incarnation
?
1209 "wrong proto version";
1219 warn
= "lost conn race";
1222 warn
= "retry with different protocol version";
1227 * Refuse to duplicate an existing connection, unless this is a
1228 * loopback connection
1230 if (conn
->ksnc_ipaddr
!= conn
->ksnc_myipaddr
) {
1231 list_for_each(tmp
, &peer
->ksnp_conns
) {
1232 conn2
= list_entry(tmp
, struct ksock_conn
, ksnc_list
);
1234 if (conn2
->ksnc_ipaddr
!= conn
->ksnc_ipaddr
||
1235 conn2
->ksnc_myipaddr
!= conn
->ksnc_myipaddr
||
1236 conn2
->ksnc_type
!= conn
->ksnc_type
)
1240 * Reply on a passive connection attempt so the peer
1241 * realises we're connected.
1253 * If the connection created by this route didn't bind to the IP
1254 * address the route connected to, the connection/route matching
1255 * code below probably isn't going to work.
1258 route
->ksnr_ipaddr
!= conn
->ksnc_ipaddr
) {
1259 CERROR("Route %s %pI4h connected to %pI4h\n",
1260 libcfs_id2str(peer
->ksnp_id
),
1261 &route
->ksnr_ipaddr
,
1262 &conn
->ksnc_ipaddr
);
1266 * Search for a route corresponding to the new connection and
1267 * create an association. This allows incoming connections created
1268 * by routes in my peer to match my own route entries so I don't
1269 * continually create duplicate routes.
1271 list_for_each(tmp
, &peer
->ksnp_routes
) {
1272 route
= list_entry(tmp
, struct ksock_route
, ksnr_list
);
1274 if (route
->ksnr_ipaddr
!= conn
->ksnc_ipaddr
)
1277 ksocknal_associate_route_conn_locked(route
, conn
);
1281 conn
->ksnc_peer
= peer
; /* conn takes my ref on peer */
1282 peer
->ksnp_last_alive
= cfs_time_current();
1283 peer
->ksnp_send_keepalive
= 0;
1284 peer
->ksnp_error
= 0;
1286 sched
= ksocknal_choose_scheduler_locked(cpt
);
1287 sched
->kss_nconns
++;
1288 conn
->ksnc_scheduler
= sched
;
1290 conn
->ksnc_tx_last_post
= cfs_time_current();
1291 /* Set the deadline for the outgoing HELLO to drain */
1292 conn
->ksnc_tx_bufnob
= sock
->sk
->sk_wmem_queued
;
1293 conn
->ksnc_tx_deadline
= cfs_time_shift(*ksocknal_tunables
.ksnd_timeout
);
1294 mb(); /* order with adding to peer's conn list */
1296 list_add(&conn
->ksnc_list
, &peer
->ksnp_conns
);
1297 ksocknal_conn_addref(conn
);
1299 ksocknal_new_packet(conn
, 0);
1301 conn
->ksnc_zc_capable
= ksocknal_lib_zc_capable(conn
);
1303 /* Take packets blocking for this connection. */
1304 list_for_each_entry_safe(tx
, txtmp
, &peer
->ksnp_tx_queue
, tx_list
) {
1305 if (conn
->ksnc_proto
->pro_match_tx(conn
, tx
, tx
->tx_nonblk
) == SOCKNAL_MATCH_NO
)
1308 list_del(&tx
->tx_list
);
1309 ksocknal_queue_tx_locked(tx
, conn
);
1312 write_unlock_bh(global_lock
);
1315 * We've now got a new connection. Any errors from here on are just
1316 * like "normal" comms errors and we close the connection normally.
1317 * NB (a) we still have to send the reply HELLO for passive
1319 * (b) normal I/O on the conn is blocked until I setup and call the
1322 CDEBUG(D_NET
, "New conn %s p %d.x %pI4h -> %pI4h/%d incarnation:%lld sched[%d:%d]\n",
1323 libcfs_id2str(peerid
), conn
->ksnc_proto
->pro_version
,
1324 &conn
->ksnc_myipaddr
, &conn
->ksnc_ipaddr
,
1325 conn
->ksnc_port
, incarnation
, cpt
,
1326 (int)(sched
- &sched
->kss_info
->ksi_scheds
[0]));
1329 /* additional routes after interface exchange? */
1330 ksocknal_create_routes(peer
, conn
->ksnc_port
,
1331 hello
->kshm_ips
, hello
->kshm_nips
);
1333 hello
->kshm_nips
= ksocknal_select_ips(peer
, hello
->kshm_ips
,
1335 rc
= ksocknal_send_hello(ni
, conn
, peerid
.nid
, hello
);
1338 LIBCFS_FREE(hello
, offsetof(ksock_hello_msg_t
,
1339 kshm_ips
[LNET_MAX_INTERFACES
]));
1342 * setup the socket AFTER I've received hello (it disables
1343 * SO_LINGER). I might call back to the acceptor who may want
1344 * to send a protocol version response and then close the
1345 * socket; this ensures the socket only tears down after the
1346 * response has been sent.
1349 rc
= ksocknal_lib_setup_sock(sock
);
1351 write_lock_bh(global_lock
);
1353 /* NB my callbacks block while I hold ksnd_global_lock */
1354 ksocknal_lib_set_callback(sock
, conn
);
1357 peer
->ksnp_accepting
--;
1359 write_unlock_bh(global_lock
);
1362 write_lock_bh(global_lock
);
1363 if (!conn
->ksnc_closing
) {
1364 /* could be closed by another thread */
1365 ksocknal_close_conn_locked(conn
, rc
);
1367 write_unlock_bh(global_lock
);
1368 } else if (!ksocknal_connsock_addref(conn
)) {
1369 /* Allow I/O to proceed. */
1370 ksocknal_read_callback(conn
);
1371 ksocknal_write_callback(conn
);
1372 ksocknal_connsock_decref(conn
);
1375 ksocknal_connsock_decref(conn
);
1376 ksocknal_conn_decref(conn
);
1380 if (!peer
->ksnp_closing
&&
1381 list_empty(&peer
->ksnp_conns
) &&
1382 list_empty(&peer
->ksnp_routes
)) {
1383 list_add(&zombies
, &peer
->ksnp_tx_queue
);
1384 list_del_init(&peer
->ksnp_tx_queue
);
1385 ksocknal_unlink_peer_locked(peer
);
1388 write_unlock_bh(global_lock
);
1392 CERROR("Not creating conn %s type %d: %s\n",
1393 libcfs_id2str(peerid
), conn
->ksnc_type
, warn
);
1395 CDEBUG(D_NET
, "Not creating conn %s type %d: %s\n",
1396 libcfs_id2str(peerid
), conn
->ksnc_type
, warn
);
1402 * Request retry by replying with CONN_NONE
1403 * ksnc_proto has been set already
1405 conn
->ksnc_type
= SOCKLND_CONN_NONE
;
1406 hello
->kshm_nips
= 0;
1407 ksocknal_send_hello(ni
, conn
, peerid
.nid
, hello
);
1410 write_lock_bh(global_lock
);
1411 peer
->ksnp_accepting
--;
1412 write_unlock_bh(global_lock
);
1415 ksocknal_txlist_done(ni
, &zombies
, 1);
1416 ksocknal_peer_decref(peer
);
1420 LIBCFS_FREE(hello
, offsetof(ksock_hello_msg_t
,
1421 kshm_ips
[LNET_MAX_INTERFACES
]));
1423 LIBCFS_FREE(conn
, sizeof(*conn
));
1431 ksocknal_close_conn_locked(struct ksock_conn
*conn
, int error
)
1434 * This just does the immmediate housekeeping, and queues the
1435 * connection for the reaper to terminate.
1436 * Caller holds ksnd_global_lock exclusively in irq context
1438 struct ksock_peer
*peer
= conn
->ksnc_peer
;
1439 struct ksock_route
*route
;
1440 struct ksock_conn
*conn2
;
1441 struct list_head
*tmp
;
1443 LASSERT(!peer
->ksnp_error
);
1444 LASSERT(!conn
->ksnc_closing
);
1445 conn
->ksnc_closing
= 1;
1447 /* ksnd_deathrow_conns takes over peer's ref */
1448 list_del(&conn
->ksnc_list
);
1450 route
= conn
->ksnc_route
;
1452 /* dissociate conn from route... */
1453 LASSERT(!route
->ksnr_deleted
);
1454 LASSERT(route
->ksnr_connected
& (1 << conn
->ksnc_type
));
1457 list_for_each(tmp
, &peer
->ksnp_conns
) {
1458 conn2
= list_entry(tmp
, struct ksock_conn
, ksnc_list
);
1460 if (conn2
->ksnc_route
== route
&&
1461 conn2
->ksnc_type
== conn
->ksnc_type
)
1467 route
->ksnr_connected
&= ~(1 << conn
->ksnc_type
);
1469 conn
->ksnc_route
= NULL
;
1471 ksocknal_route_decref(route
); /* drop conn's ref on route */
1474 if (list_empty(&peer
->ksnp_conns
)) {
1475 /* No more connections to this peer */
1477 if (!list_empty(&peer
->ksnp_tx_queue
)) {
1478 struct ksock_tx
*tx
;
1480 LASSERT(conn
->ksnc_proto
== &ksocknal_protocol_v3x
);
1483 * throw them to the last connection...,
1484 * these TXs will be send to /dev/null by scheduler
1486 list_for_each_entry(tx
, &peer
->ksnp_tx_queue
,
1488 ksocknal_tx_prep(conn
, tx
);
1490 spin_lock_bh(&conn
->ksnc_scheduler
->kss_lock
);
1491 list_splice_init(&peer
->ksnp_tx_queue
,
1492 &conn
->ksnc_tx_queue
);
1493 spin_unlock_bh(&conn
->ksnc_scheduler
->kss_lock
);
1496 peer
->ksnp_proto
= NULL
; /* renegotiate protocol version */
1497 peer
->ksnp_error
= error
; /* stash last conn close reason */
1499 if (list_empty(&peer
->ksnp_routes
)) {
1501 * I've just closed last conn belonging to a
1502 * peer with no routes to it
1504 ksocknal_unlink_peer_locked(peer
);
1508 spin_lock_bh(&ksocknal_data
.ksnd_reaper_lock
);
1510 list_add_tail(&conn
->ksnc_list
,
1511 &ksocknal_data
.ksnd_deathrow_conns
);
1512 wake_up(&ksocknal_data
.ksnd_reaper_waitq
);
1514 spin_unlock_bh(&ksocknal_data
.ksnd_reaper_lock
);
1518 ksocknal_peer_failed(struct ksock_peer
*peer
)
1521 unsigned long last_alive
= 0;
1524 * There has been a connection failure or comms error; but I'll only
1525 * tell LNET I think the peer is dead if it's to another kernel and
1526 * there are no connections or connection attempts in existence.
1528 read_lock(&ksocknal_data
.ksnd_global_lock
);
1530 if (!(peer
->ksnp_id
.pid
& LNET_PID_USERFLAG
) &&
1531 list_empty(&peer
->ksnp_conns
) &&
1532 !peer
->ksnp_accepting
&&
1533 !ksocknal_find_connecting_route_locked(peer
)) {
1535 last_alive
= peer
->ksnp_last_alive
;
1538 read_unlock(&ksocknal_data
.ksnd_global_lock
);
1541 lnet_notify(peer
->ksnp_ni
, peer
->ksnp_id
.nid
, 0,
1546 ksocknal_finalize_zcreq(struct ksock_conn
*conn
)
1548 struct ksock_peer
*peer
= conn
->ksnc_peer
;
1549 struct ksock_tx
*tx
;
1550 struct ksock_tx
*temp
;
1551 struct ksock_tx
*tmp
;
1555 * NB safe to finalize TXs because closing of socket will
1556 * abort all buffered data
1558 LASSERT(!conn
->ksnc_sock
);
1560 spin_lock(&peer
->ksnp_lock
);
1562 list_for_each_entry_safe(tx
, tmp
, &peer
->ksnp_zc_req_list
, tx_zc_list
) {
1563 if (tx
->tx_conn
!= conn
)
1566 LASSERT(tx
->tx_msg
.ksm_zc_cookies
[0]);
1568 tx
->tx_msg
.ksm_zc_cookies
[0] = 0;
1569 tx
->tx_zc_aborted
= 1; /* mark it as not-acked */
1570 list_del(&tx
->tx_zc_list
);
1571 list_add(&tx
->tx_zc_list
, &zlist
);
1574 spin_unlock(&peer
->ksnp_lock
);
1576 list_for_each_entry_safe(tx
, temp
, &zlist
, tx_zc_list
) {
1577 list_del(&tx
->tx_zc_list
);
1578 ksocknal_tx_decref(tx
);
1583 ksocknal_terminate_conn(struct ksock_conn
*conn
)
1586 * This gets called by the reaper (guaranteed thread context) to
1587 * disengage the socket from its callbacks and close it.
1588 * ksnc_refcount will eventually hit zero, and then the reaper will
1591 struct ksock_peer
*peer
= conn
->ksnc_peer
;
1592 struct ksock_sched
*sched
= conn
->ksnc_scheduler
;
1595 LASSERT(conn
->ksnc_closing
);
1597 /* wake up the scheduler to "send" all remaining packets to /dev/null */
1598 spin_lock_bh(&sched
->kss_lock
);
1600 /* a closing conn is always ready to tx */
1601 conn
->ksnc_tx_ready
= 1;
1603 if (!conn
->ksnc_tx_scheduled
&&
1604 !list_empty(&conn
->ksnc_tx_queue
)) {
1605 list_add_tail(&conn
->ksnc_tx_list
,
1606 &sched
->kss_tx_conns
);
1607 conn
->ksnc_tx_scheduled
= 1;
1608 /* extra ref for scheduler */
1609 ksocknal_conn_addref(conn
);
1611 wake_up(&sched
->kss_waitq
);
1614 spin_unlock_bh(&sched
->kss_lock
);
1616 /* serialise with callbacks */
1617 write_lock_bh(&ksocknal_data
.ksnd_global_lock
);
1619 ksocknal_lib_reset_callback(conn
->ksnc_sock
, conn
);
1622 * OK, so this conn may not be completely disengaged from its
1623 * scheduler yet, but it _has_ committed to terminate...
1625 conn
->ksnc_scheduler
->kss_nconns
--;
1627 if (peer
->ksnp_error
) {
1628 /* peer's last conn closed in error */
1629 LASSERT(list_empty(&peer
->ksnp_conns
));
1631 peer
->ksnp_error
= 0; /* avoid multiple notifications */
1634 write_unlock_bh(&ksocknal_data
.ksnd_global_lock
);
1637 ksocknal_peer_failed(peer
);
1640 * The socket is closed on the final put; either here, or in
1641 * ksocknal_{send,recv}msg(). Since we set up the linger2 option
1642 * when the connection was established, this will close the socket
1643 * immediately, aborting anything buffered in it. Any hung
1644 * zero-copy transmits will therefore complete in finite time.
1646 ksocknal_connsock_decref(conn
);
1650 ksocknal_queue_zombie_conn(struct ksock_conn
*conn
)
1652 /* Queue the conn for the reaper to destroy */
1654 LASSERT(!atomic_read(&conn
->ksnc_conn_refcount
));
1655 spin_lock_bh(&ksocknal_data
.ksnd_reaper_lock
);
1657 list_add_tail(&conn
->ksnc_list
, &ksocknal_data
.ksnd_zombie_conns
);
1658 wake_up(&ksocknal_data
.ksnd_reaper_waitq
);
1660 spin_unlock_bh(&ksocknal_data
.ksnd_reaper_lock
);
1664 ksocknal_destroy_conn(struct ksock_conn
*conn
)
1666 unsigned long last_rcv
;
1668 /* Final coup-de-grace of the reaper */
1669 CDEBUG(D_NET
, "connection %p\n", conn
);
1671 LASSERT(!atomic_read(&conn
->ksnc_conn_refcount
));
1672 LASSERT(!atomic_read(&conn
->ksnc_sock_refcount
));
1673 LASSERT(!conn
->ksnc_sock
);
1674 LASSERT(!conn
->ksnc_route
);
1675 LASSERT(!conn
->ksnc_tx_scheduled
);
1676 LASSERT(!conn
->ksnc_rx_scheduled
);
1677 LASSERT(list_empty(&conn
->ksnc_tx_queue
));
1679 /* complete current receive if any */
1680 switch (conn
->ksnc_rx_state
) {
1681 case SOCKNAL_RX_LNET_PAYLOAD
:
1682 last_rcv
= conn
->ksnc_rx_deadline
-
1683 cfs_time_seconds(*ksocknal_tunables
.ksnd_timeout
);
1684 CERROR("Completing partial receive from %s[%d], ip %pI4h:%d, with error, wanted: %d, left: %d, last alive is %ld secs ago\n",
1685 libcfs_id2str(conn
->ksnc_peer
->ksnp_id
), conn
->ksnc_type
,
1686 &conn
->ksnc_ipaddr
, conn
->ksnc_port
,
1687 conn
->ksnc_rx_nob_wanted
, conn
->ksnc_rx_nob_left
,
1688 cfs_duration_sec(cfs_time_sub(cfs_time_current(),
1690 lnet_finalize(conn
->ksnc_peer
->ksnp_ni
,
1691 conn
->ksnc_cookie
, -EIO
);
1693 case SOCKNAL_RX_LNET_HEADER
:
1694 if (conn
->ksnc_rx_started
)
1695 CERROR("Incomplete receive of lnet header from %s, ip %pI4h:%d, with error, protocol: %d.x.\n",
1696 libcfs_id2str(conn
->ksnc_peer
->ksnp_id
),
1697 &conn
->ksnc_ipaddr
, conn
->ksnc_port
,
1698 conn
->ksnc_proto
->pro_version
);
1700 case SOCKNAL_RX_KSM_HEADER
:
1701 if (conn
->ksnc_rx_started
)
1702 CERROR("Incomplete receive of ksock message from %s, ip %pI4h:%d, with error, protocol: %d.x.\n",
1703 libcfs_id2str(conn
->ksnc_peer
->ksnp_id
),
1704 &conn
->ksnc_ipaddr
, conn
->ksnc_port
,
1705 conn
->ksnc_proto
->pro_version
);
1707 case SOCKNAL_RX_SLOP
:
1708 if (conn
->ksnc_rx_started
)
1709 CERROR("Incomplete receive of slops from %s, ip %pI4h:%d, with error\n",
1710 libcfs_id2str(conn
->ksnc_peer
->ksnp_id
),
1711 &conn
->ksnc_ipaddr
, conn
->ksnc_port
);
1718 ksocknal_peer_decref(conn
->ksnc_peer
);
1720 LIBCFS_FREE(conn
, sizeof(*conn
));
1724 ksocknal_close_peer_conns_locked(struct ksock_peer
*peer
, __u32 ipaddr
, int why
)
1726 struct ksock_conn
*conn
;
1727 struct list_head
*ctmp
;
1728 struct list_head
*cnxt
;
1731 list_for_each_safe(ctmp
, cnxt
, &peer
->ksnp_conns
) {
1732 conn
= list_entry(ctmp
, struct ksock_conn
, ksnc_list
);
1734 if (!ipaddr
|| conn
->ksnc_ipaddr
== ipaddr
) {
1736 ksocknal_close_conn_locked(conn
, why
);
1744 ksocknal_close_conn_and_siblings(struct ksock_conn
*conn
, int why
)
1746 struct ksock_peer
*peer
= conn
->ksnc_peer
;
1747 __u32 ipaddr
= conn
->ksnc_ipaddr
;
1750 write_lock_bh(&ksocknal_data
.ksnd_global_lock
);
1752 count
= ksocknal_close_peer_conns_locked(peer
, ipaddr
, why
);
1754 write_unlock_bh(&ksocknal_data
.ksnd_global_lock
);
1760 ksocknal_close_matching_conns(lnet_process_id_t id
, __u32 ipaddr
)
1762 struct ksock_peer
*peer
;
1763 struct list_head
*ptmp
;
1764 struct list_head
*pnxt
;
1770 write_lock_bh(&ksocknal_data
.ksnd_global_lock
);
1772 if (id
.nid
!= LNET_NID_ANY
) {
1773 lo
= (int)(ksocknal_nid2peerlist(id
.nid
) - ksocknal_data
.ksnd_peers
);
1774 hi
= (int)(ksocknal_nid2peerlist(id
.nid
) - ksocknal_data
.ksnd_peers
);
1777 hi
= ksocknal_data
.ksnd_peer_hash_size
- 1;
1780 for (i
= lo
; i
<= hi
; i
++) {
1781 list_for_each_safe(ptmp
, pnxt
,
1782 &ksocknal_data
.ksnd_peers
[i
]) {
1783 peer
= list_entry(ptmp
, struct ksock_peer
, ksnp_list
);
1785 if (!((id
.nid
== LNET_NID_ANY
|| id
.nid
== peer
->ksnp_id
.nid
) &&
1786 (id
.pid
== LNET_PID_ANY
|| id
.pid
== peer
->ksnp_id
.pid
)))
1789 count
+= ksocknal_close_peer_conns_locked(peer
, ipaddr
, 0);
1793 write_unlock_bh(&ksocknal_data
.ksnd_global_lock
);
1795 /* wildcards always succeed */
1796 if (id
.nid
== LNET_NID_ANY
|| id
.pid
== LNET_PID_ANY
|| !ipaddr
)
1806 ksocknal_notify(lnet_ni_t
*ni
, lnet_nid_t gw_nid
, int alive
)
1809 * The router is telling me she's been notified of a change in
1812 lnet_process_id_t id
= {0};
1815 id
.pid
= LNET_PID_ANY
;
1817 CDEBUG(D_NET
, "gw %s %s\n", libcfs_nid2str(gw_nid
),
1818 alive
? "up" : "down");
1821 /* If the gateway crashed, close all open connections... */
1822 ksocknal_close_matching_conns(id
, 0);
1827 * ...otherwise do nothing. We can only establish new connections
1828 * if we have autroutes, and these connect on demand.
1833 ksocknal_query(lnet_ni_t
*ni
, lnet_nid_t nid
, unsigned long *when
)
1836 unsigned long last_alive
= 0;
1837 unsigned long now
= cfs_time_current();
1838 struct ksock_peer
*peer
= NULL
;
1839 rwlock_t
*glock
= &ksocknal_data
.ksnd_global_lock
;
1840 lnet_process_id_t id
= {
1842 .pid
= LNET_PID_LUSTRE
,
1847 peer
= ksocknal_find_peer_locked(ni
, id
);
1849 struct list_head
*tmp
;
1850 struct ksock_conn
*conn
;
1853 list_for_each(tmp
, &peer
->ksnp_conns
) {
1854 conn
= list_entry(tmp
, struct ksock_conn
, ksnc_list
);
1855 bufnob
= conn
->ksnc_sock
->sk
->sk_wmem_queued
;
1857 if (bufnob
< conn
->ksnc_tx_bufnob
) {
1858 /* something got ACKed */
1859 conn
->ksnc_tx_deadline
=
1860 cfs_time_shift(*ksocknal_tunables
.ksnd_timeout
);
1861 peer
->ksnp_last_alive
= now
;
1862 conn
->ksnc_tx_bufnob
= bufnob
;
1866 last_alive
= peer
->ksnp_last_alive
;
1867 if (!ksocknal_find_connectable_route_locked(peer
))
1876 CDEBUG(D_NET
, "Peer %s %p, alive %ld secs ago, connect %d\n",
1877 libcfs_nid2str(nid
), peer
,
1878 last_alive
? cfs_duration_sec(now
- last_alive
) : -1,
1884 ksocknal_add_peer(ni
, id
, LNET_NIDADDR(nid
), lnet_acceptor_port());
1886 write_lock_bh(glock
);
1888 peer
= ksocknal_find_peer_locked(ni
, id
);
1890 ksocknal_launch_all_connections_locked(peer
);
1892 write_unlock_bh(glock
);
1896 ksocknal_push_peer(struct ksock_peer
*peer
)
1900 struct list_head
*tmp
;
1901 struct ksock_conn
*conn
;
1903 for (index
= 0; ; index
++) {
1904 read_lock(&ksocknal_data
.ksnd_global_lock
);
1909 list_for_each(tmp
, &peer
->ksnp_conns
) {
1911 conn
= list_entry(tmp
, struct ksock_conn
,
1913 ksocknal_conn_addref(conn
);
1918 read_unlock(&ksocknal_data
.ksnd_global_lock
);
1923 ksocknal_lib_push_conn(conn
);
1924 ksocknal_conn_decref(conn
);
1928 static int ksocknal_push(lnet_ni_t
*ni
, lnet_process_id_t id
)
1930 struct list_head
*start
;
1931 struct list_head
*end
;
1932 struct list_head
*tmp
;
1934 unsigned int hsize
= ksocknal_data
.ksnd_peer_hash_size
;
1936 if (id
.nid
== LNET_NID_ANY
) {
1937 start
= &ksocknal_data
.ksnd_peers
[0];
1938 end
= &ksocknal_data
.ksnd_peers
[hsize
- 1];
1940 start
= ksocknal_nid2peerlist(id
.nid
);
1941 end
= ksocknal_nid2peerlist(id
.nid
);
1944 for (tmp
= start
; tmp
<= end
; tmp
++) {
1945 int peer_off
; /* searching offset in peer hash table */
1947 for (peer_off
= 0; ; peer_off
++) {
1948 struct ksock_peer
*peer
;
1951 read_lock(&ksocknal_data
.ksnd_global_lock
);
1952 list_for_each_entry(peer
, tmp
, ksnp_list
) {
1953 if (!((id
.nid
== LNET_NID_ANY
||
1954 id
.nid
== peer
->ksnp_id
.nid
) &&
1955 (id
.pid
== LNET_PID_ANY
||
1956 id
.pid
== peer
->ksnp_id
.pid
)))
1959 if (i
++ == peer_off
) {
1960 ksocknal_peer_addref(peer
);
1964 read_unlock(&ksocknal_data
.ksnd_global_lock
);
1966 if (!i
) /* no match */
1970 ksocknal_push_peer(peer
);
1971 ksocknal_peer_decref(peer
);
1978 ksocknal_add_interface(lnet_ni_t
*ni
, __u32 ipaddress
, __u32 netmask
)
1980 struct ksock_net
*net
= ni
->ni_data
;
1981 struct ksock_interface
*iface
;
1985 struct list_head
*ptmp
;
1986 struct ksock_peer
*peer
;
1987 struct list_head
*rtmp
;
1988 struct ksock_route
*route
;
1990 if (!ipaddress
|| !netmask
)
1993 write_lock_bh(&ksocknal_data
.ksnd_global_lock
);
1995 iface
= ksocknal_ip2iface(ni
, ipaddress
);
1997 /* silently ignore dups */
1999 } else if (net
->ksnn_ninterfaces
== LNET_MAX_INTERFACES
) {
2002 iface
= &net
->ksnn_interfaces
[net
->ksnn_ninterfaces
++];
2004 iface
->ksni_ipaddr
= ipaddress
;
2005 iface
->ksni_netmask
= netmask
;
2006 iface
->ksni_nroutes
= 0;
2007 iface
->ksni_npeers
= 0;
2009 for (i
= 0; i
< ksocknal_data
.ksnd_peer_hash_size
; i
++) {
2010 list_for_each(ptmp
, &ksocknal_data
.ksnd_peers
[i
]) {
2011 peer
= list_entry(ptmp
, struct ksock_peer
,
2014 for (j
= 0; j
< peer
->ksnp_n_passive_ips
; j
++)
2015 if (peer
->ksnp_passive_ips
[j
] == ipaddress
)
2016 iface
->ksni_npeers
++;
2018 list_for_each(rtmp
, &peer
->ksnp_routes
) {
2019 route
= list_entry(rtmp
, struct ksock_route
,
2022 if (route
->ksnr_myipaddr
== ipaddress
)
2023 iface
->ksni_nroutes
++;
2029 /* NB only new connections will pay attention to the new interface! */
2032 write_unlock_bh(&ksocknal_data
.ksnd_global_lock
);
2038 ksocknal_peer_del_interface_locked(struct ksock_peer
*peer
, __u32 ipaddr
)
2040 struct list_head
*tmp
;
2041 struct list_head
*nxt
;
2042 struct ksock_route
*route
;
2043 struct ksock_conn
*conn
;
2047 for (i
= 0; i
< peer
->ksnp_n_passive_ips
; i
++)
2048 if (peer
->ksnp_passive_ips
[i
] == ipaddr
) {
2049 for (j
= i
+ 1; j
< peer
->ksnp_n_passive_ips
; j
++)
2050 peer
->ksnp_passive_ips
[j
- 1] =
2051 peer
->ksnp_passive_ips
[j
];
2052 peer
->ksnp_n_passive_ips
--;
2056 list_for_each_safe(tmp
, nxt
, &peer
->ksnp_routes
) {
2057 route
= list_entry(tmp
, struct ksock_route
, ksnr_list
);
2059 if (route
->ksnr_myipaddr
!= ipaddr
)
2062 if (route
->ksnr_share_count
) {
2063 /* Manually created; keep, but unbind */
2064 route
->ksnr_myipaddr
= 0;
2066 ksocknal_del_route_locked(route
);
2070 list_for_each_safe(tmp
, nxt
, &peer
->ksnp_conns
) {
2071 conn
= list_entry(tmp
, struct ksock_conn
, ksnc_list
);
2073 if (conn
->ksnc_myipaddr
== ipaddr
)
2074 ksocknal_close_conn_locked(conn
, 0);
2079 ksocknal_del_interface(lnet_ni_t
*ni
, __u32 ipaddress
)
2081 struct ksock_net
*net
= ni
->ni_data
;
2083 struct list_head
*tmp
;
2084 struct list_head
*nxt
;
2085 struct ksock_peer
*peer
;
2090 write_lock_bh(&ksocknal_data
.ksnd_global_lock
);
2092 for (i
= 0; i
< net
->ksnn_ninterfaces
; i
++) {
2093 this_ip
= net
->ksnn_interfaces
[i
].ksni_ipaddr
;
2095 if (!(!ipaddress
|| ipaddress
== this_ip
))
2100 for (j
= i
+ 1; j
< net
->ksnn_ninterfaces
; j
++)
2101 net
->ksnn_interfaces
[j
- 1] =
2102 net
->ksnn_interfaces
[j
];
2104 net
->ksnn_ninterfaces
--;
2106 for (j
= 0; j
< ksocknal_data
.ksnd_peer_hash_size
; j
++) {
2107 list_for_each_safe(tmp
, nxt
,
2108 &ksocknal_data
.ksnd_peers
[j
]) {
2109 peer
= list_entry(tmp
, struct ksock_peer
, ksnp_list
);
2111 if (peer
->ksnp_ni
!= ni
)
2114 ksocknal_peer_del_interface_locked(peer
, this_ip
);
2119 write_unlock_bh(&ksocknal_data
.ksnd_global_lock
);
2125 ksocknal_ctl(lnet_ni_t
*ni
, unsigned int cmd
, void *arg
)
2127 lnet_process_id_t id
= {0};
2128 struct libcfs_ioctl_data
*data
= arg
;
2132 case IOC_LIBCFS_GET_INTERFACE
: {
2133 struct ksock_net
*net
= ni
->ni_data
;
2134 struct ksock_interface
*iface
;
2136 read_lock(&ksocknal_data
.ksnd_global_lock
);
2138 if (data
->ioc_count
>= (__u32
)net
->ksnn_ninterfaces
) {
2142 iface
= &net
->ksnn_interfaces
[data
->ioc_count
];
2144 data
->ioc_u32
[0] = iface
->ksni_ipaddr
;
2145 data
->ioc_u32
[1] = iface
->ksni_netmask
;
2146 data
->ioc_u32
[2] = iface
->ksni_npeers
;
2147 data
->ioc_u32
[3] = iface
->ksni_nroutes
;
2150 read_unlock(&ksocknal_data
.ksnd_global_lock
);
2154 case IOC_LIBCFS_ADD_INTERFACE
:
2155 return ksocknal_add_interface(ni
,
2156 data
->ioc_u32
[0], /* IP address */
2157 data
->ioc_u32
[1]); /* net mask */
2159 case IOC_LIBCFS_DEL_INTERFACE
:
2160 return ksocknal_del_interface(ni
,
2161 data
->ioc_u32
[0]); /* IP address */
2163 case IOC_LIBCFS_GET_PEER
: {
2168 int share_count
= 0;
2170 rc
= ksocknal_get_peer_info(ni
, data
->ioc_count
,
2171 &id
, &myip
, &ip
, &port
,
2172 &conn_count
, &share_count
);
2176 data
->ioc_nid
= id
.nid
;
2177 data
->ioc_count
= share_count
;
2178 data
->ioc_u32
[0] = ip
;
2179 data
->ioc_u32
[1] = port
;
2180 data
->ioc_u32
[2] = myip
;
2181 data
->ioc_u32
[3] = conn_count
;
2182 data
->ioc_u32
[4] = id
.pid
;
2186 case IOC_LIBCFS_ADD_PEER
:
2187 id
.nid
= data
->ioc_nid
;
2188 id
.pid
= LNET_PID_LUSTRE
;
2189 return ksocknal_add_peer(ni
, id
,
2190 data
->ioc_u32
[0], /* IP */
2191 data
->ioc_u32
[1]); /* port */
2193 case IOC_LIBCFS_DEL_PEER
:
2194 id
.nid
= data
->ioc_nid
;
2195 id
.pid
= LNET_PID_ANY
;
2196 return ksocknal_del_peer(ni
, id
,
2197 data
->ioc_u32
[0]); /* IP */
2199 case IOC_LIBCFS_GET_CONN
: {
2203 struct ksock_conn
*conn
= ksocknal_get_conn_by_idx(ni
, data
->ioc_count
);
2208 ksocknal_lib_get_conn_tunables(conn
, &txmem
, &rxmem
, &nagle
);
2210 data
->ioc_count
= txmem
;
2211 data
->ioc_nid
= conn
->ksnc_peer
->ksnp_id
.nid
;
2212 data
->ioc_flags
= nagle
;
2213 data
->ioc_u32
[0] = conn
->ksnc_ipaddr
;
2214 data
->ioc_u32
[1] = conn
->ksnc_port
;
2215 data
->ioc_u32
[2] = conn
->ksnc_myipaddr
;
2216 data
->ioc_u32
[3] = conn
->ksnc_type
;
2217 data
->ioc_u32
[4] = conn
->ksnc_scheduler
->kss_info
->ksi_cpt
;
2218 data
->ioc_u32
[5] = rxmem
;
2219 data
->ioc_u32
[6] = conn
->ksnc_peer
->ksnp_id
.pid
;
2220 ksocknal_conn_decref(conn
);
2224 case IOC_LIBCFS_CLOSE_CONNECTION
:
2225 id
.nid
= data
->ioc_nid
;
2226 id
.pid
= LNET_PID_ANY
;
2227 return ksocknal_close_matching_conns(id
,
2230 case IOC_LIBCFS_REGISTER_MYNID
:
2231 /* Ignore if this is a noop */
2232 if (data
->ioc_nid
== ni
->ni_nid
)
2235 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2236 libcfs_nid2str(data
->ioc_nid
),
2237 libcfs_nid2str(ni
->ni_nid
));
2240 case IOC_LIBCFS_PUSH_CONNECTION
:
2241 id
.nid
= data
->ioc_nid
;
2242 id
.pid
= LNET_PID_ANY
;
2243 return ksocknal_push(ni
, id
);
2252 ksocknal_free_buffers(void)
2254 LASSERT(!atomic_read(&ksocknal_data
.ksnd_nactive_txs
));
2256 if (ksocknal_data
.ksnd_sched_info
) {
2257 struct ksock_sched_info
*info
;
2260 cfs_percpt_for_each(info
, i
, ksocknal_data
.ksnd_sched_info
) {
2261 if (info
->ksi_scheds
) {
2262 LIBCFS_FREE(info
->ksi_scheds
,
2263 info
->ksi_nthreads_max
*
2264 sizeof(info
->ksi_scheds
[0]));
2267 cfs_percpt_free(ksocknal_data
.ksnd_sched_info
);
2270 LIBCFS_FREE(ksocknal_data
.ksnd_peers
,
2271 sizeof(struct list_head
) *
2272 ksocknal_data
.ksnd_peer_hash_size
);
2274 spin_lock(&ksocknal_data
.ksnd_tx_lock
);
2276 if (!list_empty(&ksocknal_data
.ksnd_idle_noop_txs
)) {
2277 struct list_head zlist
;
2278 struct ksock_tx
*tx
;
2279 struct ksock_tx
*temp
;
2281 list_add(&zlist
, &ksocknal_data
.ksnd_idle_noop_txs
);
2282 list_del_init(&ksocknal_data
.ksnd_idle_noop_txs
);
2283 spin_unlock(&ksocknal_data
.ksnd_tx_lock
);
2285 list_for_each_entry_safe(tx
, temp
, &zlist
, tx_list
) {
2286 list_del(&tx
->tx_list
);
2287 LIBCFS_FREE(tx
, tx
->tx_desc_size
);
2290 spin_unlock(&ksocknal_data
.ksnd_tx_lock
);
2295 ksocknal_base_shutdown(void)
2297 struct ksock_sched_info
*info
;
2298 struct ksock_sched
*sched
;
2302 LASSERT(!ksocknal_data
.ksnd_nnets
);
2304 switch (ksocknal_data
.ksnd_init
) {
2308 case SOCKNAL_INIT_ALL
:
2309 case SOCKNAL_INIT_DATA
:
2310 LASSERT(ksocknal_data
.ksnd_peers
);
2311 for (i
= 0; i
< ksocknal_data
.ksnd_peer_hash_size
; i
++)
2312 LASSERT(list_empty(&ksocknal_data
.ksnd_peers
[i
]));
2314 LASSERT(list_empty(&ksocknal_data
.ksnd_nets
));
2315 LASSERT(list_empty(&ksocknal_data
.ksnd_enomem_conns
));
2316 LASSERT(list_empty(&ksocknal_data
.ksnd_zombie_conns
));
2317 LASSERT(list_empty(&ksocknal_data
.ksnd_connd_connreqs
));
2318 LASSERT(list_empty(&ksocknal_data
.ksnd_connd_routes
));
2320 if (ksocknal_data
.ksnd_sched_info
) {
2321 cfs_percpt_for_each(info
, i
,
2322 ksocknal_data
.ksnd_sched_info
) {
2323 if (!info
->ksi_scheds
)
2326 for (j
= 0; j
< info
->ksi_nthreads_max
; j
++) {
2327 sched
= &info
->ksi_scheds
[j
];
2329 &sched
->kss_tx_conns
));
2331 &sched
->kss_rx_conns
));
2333 &sched
->kss_zombie_noop_txs
));
2334 LASSERT(!sched
->kss_nconns
);
2339 /* flag threads to terminate; wake and wait for them to die */
2340 ksocknal_data
.ksnd_shuttingdown
= 1;
2341 wake_up_all(&ksocknal_data
.ksnd_connd_waitq
);
2342 wake_up_all(&ksocknal_data
.ksnd_reaper_waitq
);
2344 if (ksocknal_data
.ksnd_sched_info
) {
2345 cfs_percpt_for_each(info
, i
,
2346 ksocknal_data
.ksnd_sched_info
) {
2347 if (!info
->ksi_scheds
)
2350 for (j
= 0; j
< info
->ksi_nthreads_max
; j
++) {
2351 sched
= &info
->ksi_scheds
[j
];
2352 wake_up_all(&sched
->kss_waitq
);
2358 read_lock(&ksocknal_data
.ksnd_global_lock
);
2359 while (ksocknal_data
.ksnd_nthreads
) {
2361 CDEBUG(((i
& (-i
)) == i
) ? D_WARNING
: D_NET
, /* power of 2? */
2362 "waiting for %d threads to terminate\n",
2363 ksocknal_data
.ksnd_nthreads
);
2364 read_unlock(&ksocknal_data
.ksnd_global_lock
);
2365 set_current_state(TASK_UNINTERRUPTIBLE
);
2366 schedule_timeout(cfs_time_seconds(1));
2367 read_lock(&ksocknal_data
.ksnd_global_lock
);
2369 read_unlock(&ksocknal_data
.ksnd_global_lock
);
2371 ksocknal_free_buffers();
2373 ksocknal_data
.ksnd_init
= SOCKNAL_INIT_NOTHING
;
2377 module_put(THIS_MODULE
);
2381 ksocknal_new_incarnation(void)
2383 /* The incarnation number is the time this module loaded and it
2384 * identifies this particular instance of the socknal.
2386 return ktime_get_ns();
2390 ksocknal_base_startup(void)
2392 struct ksock_sched_info
*info
;
2396 LASSERT(ksocknal_data
.ksnd_init
== SOCKNAL_INIT_NOTHING
);
2397 LASSERT(!ksocknal_data
.ksnd_nnets
);
2399 memset(&ksocknal_data
, 0, sizeof(ksocknal_data
)); /* zero pointers */
2401 ksocknal_data
.ksnd_peer_hash_size
= SOCKNAL_PEER_HASH_SIZE
;
2402 LIBCFS_ALLOC(ksocknal_data
.ksnd_peers
,
2403 sizeof(struct list_head
) *
2404 ksocknal_data
.ksnd_peer_hash_size
);
2405 if (!ksocknal_data
.ksnd_peers
)
2408 for (i
= 0; i
< ksocknal_data
.ksnd_peer_hash_size
; i
++)
2409 INIT_LIST_HEAD(&ksocknal_data
.ksnd_peers
[i
]);
2411 rwlock_init(&ksocknal_data
.ksnd_global_lock
);
2412 INIT_LIST_HEAD(&ksocknal_data
.ksnd_nets
);
2414 spin_lock_init(&ksocknal_data
.ksnd_reaper_lock
);
2415 INIT_LIST_HEAD(&ksocknal_data
.ksnd_enomem_conns
);
2416 INIT_LIST_HEAD(&ksocknal_data
.ksnd_zombie_conns
);
2417 INIT_LIST_HEAD(&ksocknal_data
.ksnd_deathrow_conns
);
2418 init_waitqueue_head(&ksocknal_data
.ksnd_reaper_waitq
);
2420 spin_lock_init(&ksocknal_data
.ksnd_connd_lock
);
2421 INIT_LIST_HEAD(&ksocknal_data
.ksnd_connd_connreqs
);
2422 INIT_LIST_HEAD(&ksocknal_data
.ksnd_connd_routes
);
2423 init_waitqueue_head(&ksocknal_data
.ksnd_connd_waitq
);
2425 spin_lock_init(&ksocknal_data
.ksnd_tx_lock
);
2426 INIT_LIST_HEAD(&ksocknal_data
.ksnd_idle_noop_txs
);
2428 /* NB memset above zeros whole of ksocknal_data */
2430 /* flag lists/ptrs/locks initialised */
2431 ksocknal_data
.ksnd_init
= SOCKNAL_INIT_DATA
;
2432 try_module_get(THIS_MODULE
);
2434 ksocknal_data
.ksnd_sched_info
= cfs_percpt_alloc(lnet_cpt_table(),
2436 if (!ksocknal_data
.ksnd_sched_info
)
2439 cfs_percpt_for_each(info
, i
, ksocknal_data
.ksnd_sched_info
) {
2440 struct ksock_sched
*sched
;
2443 nthrs
= cfs_cpt_weight(lnet_cpt_table(), i
);
2444 if (*ksocknal_tunables
.ksnd_nscheds
> 0) {
2445 nthrs
= min(nthrs
, *ksocknal_tunables
.ksnd_nscheds
);
2448 * max to half of CPUs, assume another half should be
2449 * reserved for upper layer modules
2451 nthrs
= min(max(SOCKNAL_NSCHEDS
, nthrs
>> 1), nthrs
);
2454 info
->ksi_nthreads_max
= nthrs
;
2457 LIBCFS_CPT_ALLOC(info
->ksi_scheds
, lnet_cpt_table(), i
,
2458 info
->ksi_nthreads_max
* sizeof(*sched
));
2459 if (!info
->ksi_scheds
)
2462 for (; nthrs
> 0; nthrs
--) {
2463 sched
= &info
->ksi_scheds
[nthrs
- 1];
2465 sched
->kss_info
= info
;
2466 spin_lock_init(&sched
->kss_lock
);
2467 INIT_LIST_HEAD(&sched
->kss_rx_conns
);
2468 INIT_LIST_HEAD(&sched
->kss_tx_conns
);
2469 INIT_LIST_HEAD(&sched
->kss_zombie_noop_txs
);
2470 init_waitqueue_head(&sched
->kss_waitq
);
2474 ksocknal_data
.ksnd_connd_starting
= 0;
2475 ksocknal_data
.ksnd_connd_failed_stamp
= 0;
2476 ksocknal_data
.ksnd_connd_starting_stamp
= ktime_get_real_seconds();
2478 * must have at least 2 connds to remain responsive to accepts while
2481 if (*ksocknal_tunables
.ksnd_nconnds
< SOCKNAL_CONND_RESV
+ 1)
2482 *ksocknal_tunables
.ksnd_nconnds
= SOCKNAL_CONND_RESV
+ 1;
2484 if (*ksocknal_tunables
.ksnd_nconnds_max
<
2485 *ksocknal_tunables
.ksnd_nconnds
) {
2486 ksocknal_tunables
.ksnd_nconnds_max
=
2487 ksocknal_tunables
.ksnd_nconnds
;
2490 for (i
= 0; i
< *ksocknal_tunables
.ksnd_nconnds
; i
++) {
2493 spin_lock_bh(&ksocknal_data
.ksnd_connd_lock
);
2494 ksocknal_data
.ksnd_connd_starting
++;
2495 spin_unlock_bh(&ksocknal_data
.ksnd_connd_lock
);
2497 snprintf(name
, sizeof(name
), "socknal_cd%02d", i
);
2498 rc
= ksocknal_thread_start(ksocknal_connd
,
2499 (void *)((ulong_ptr_t
)i
), name
);
2501 spin_lock_bh(&ksocknal_data
.ksnd_connd_lock
);
2502 ksocknal_data
.ksnd_connd_starting
--;
2503 spin_unlock_bh(&ksocknal_data
.ksnd_connd_lock
);
2504 CERROR("Can't spawn socknal connd: %d\n", rc
);
2509 rc
= ksocknal_thread_start(ksocknal_reaper
, NULL
, "socknal_reaper");
2511 CERROR("Can't spawn socknal reaper: %d\n", rc
);
2515 /* flag everything initialised */
2516 ksocknal_data
.ksnd_init
= SOCKNAL_INIT_ALL
;
2521 ksocknal_base_shutdown();
2526 ksocknal_debug_peerhash(lnet_ni_t
*ni
)
2528 struct ksock_peer
*peer
= NULL
;
2529 struct list_head
*tmp
;
2532 read_lock(&ksocknal_data
.ksnd_global_lock
);
2534 for (i
= 0; i
< ksocknal_data
.ksnd_peer_hash_size
; i
++) {
2535 list_for_each(tmp
, &ksocknal_data
.ksnd_peers
[i
]) {
2536 peer
= list_entry(tmp
, struct ksock_peer
, ksnp_list
);
2538 if (peer
->ksnp_ni
== ni
)
2546 struct ksock_route
*route
;
2547 struct ksock_conn
*conn
;
2549 CWARN("Active peer on shutdown: %s, ref %d, scnt %d, closing %d, accepting %d, err %d, zcookie %llu, txq %d, zc_req %d\n",
2550 libcfs_id2str(peer
->ksnp_id
),
2551 atomic_read(&peer
->ksnp_refcount
),
2552 peer
->ksnp_sharecount
, peer
->ksnp_closing
,
2553 peer
->ksnp_accepting
, peer
->ksnp_error
,
2554 peer
->ksnp_zc_next_cookie
,
2555 !list_empty(&peer
->ksnp_tx_queue
),
2556 !list_empty(&peer
->ksnp_zc_req_list
));
2558 list_for_each(tmp
, &peer
->ksnp_routes
) {
2559 route
= list_entry(tmp
, struct ksock_route
, ksnr_list
);
2560 CWARN("Route: ref %d, schd %d, conn %d, cnted %d, del %d\n",
2561 atomic_read(&route
->ksnr_refcount
),
2562 route
->ksnr_scheduled
, route
->ksnr_connecting
,
2563 route
->ksnr_connected
, route
->ksnr_deleted
);
2566 list_for_each(tmp
, &peer
->ksnp_conns
) {
2567 conn
= list_entry(tmp
, struct ksock_conn
, ksnc_list
);
2568 CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2569 atomic_read(&conn
->ksnc_conn_refcount
),
2570 atomic_read(&conn
->ksnc_sock_refcount
),
2571 conn
->ksnc_type
, conn
->ksnc_closing
);
2575 read_unlock(&ksocknal_data
.ksnd_global_lock
);
2579 ksocknal_shutdown(lnet_ni_t
*ni
)
2581 struct ksock_net
*net
= ni
->ni_data
;
2583 lnet_process_id_t anyid
= {0};
2585 anyid
.nid
= LNET_NID_ANY
;
2586 anyid
.pid
= LNET_PID_ANY
;
2588 LASSERT(ksocknal_data
.ksnd_init
== SOCKNAL_INIT_ALL
);
2589 LASSERT(ksocknal_data
.ksnd_nnets
> 0);
2591 spin_lock_bh(&net
->ksnn_lock
);
2592 net
->ksnn_shutdown
= 1; /* prevent new peers */
2593 spin_unlock_bh(&net
->ksnn_lock
);
2595 /* Delete all peers */
2596 ksocknal_del_peer(ni
, anyid
, 0);
2598 /* Wait for all peer state to clean up */
2600 spin_lock_bh(&net
->ksnn_lock
);
2601 while (net
->ksnn_npeers
) {
2602 spin_unlock_bh(&net
->ksnn_lock
);
2605 CDEBUG(((i
& (-i
)) == i
) ? D_WARNING
: D_NET
, /* power of 2? */
2606 "waiting for %d peers to disconnect\n",
2608 set_current_state(TASK_UNINTERRUPTIBLE
);
2609 schedule_timeout(cfs_time_seconds(1));
2611 ksocknal_debug_peerhash(ni
);
2613 spin_lock_bh(&net
->ksnn_lock
);
2615 spin_unlock_bh(&net
->ksnn_lock
);
2617 for (i
= 0; i
< net
->ksnn_ninterfaces
; i
++) {
2618 LASSERT(!net
->ksnn_interfaces
[i
].ksni_npeers
);
2619 LASSERT(!net
->ksnn_interfaces
[i
].ksni_nroutes
);
2622 list_del(&net
->ksnn_list
);
2623 LIBCFS_FREE(net
, sizeof(*net
));
2625 ksocknal_data
.ksnd_nnets
--;
2626 if (!ksocknal_data
.ksnd_nnets
)
2627 ksocknal_base_shutdown();
2631 ksocknal_enumerate_interfaces(struct ksock_net
*net
)
2639 n
= lnet_ipif_enumerate(&names
);
2641 CERROR("Can't enumerate interfaces: %d\n", n
);
2645 for (i
= j
= 0; i
< n
; i
++) {
2650 if (!strcmp(names
[i
], "lo")) /* skip the loopback IF */
2653 rc
= lnet_ipif_query(names
[i
], &up
, &ip
, &mask
);
2655 CWARN("Can't get interface %s info: %d\n",
2661 CWARN("Ignoring interface %s (down)\n",
2666 if (j
== LNET_MAX_INTERFACES
) {
2667 CWARN("Ignoring interface %s (too many interfaces)\n",
2672 net
->ksnn_interfaces
[j
].ksni_ipaddr
= ip
;
2673 net
->ksnn_interfaces
[j
].ksni_netmask
= mask
;
2674 strlcpy(net
->ksnn_interfaces
[j
].ksni_name
,
2675 names
[i
], sizeof(net
->ksnn_interfaces
[j
].ksni_name
));
2679 lnet_ipif_free_enumeration(names
, n
);
2682 CERROR("Can't find any usable interfaces\n");
2688 ksocknal_search_new_ipif(struct ksock_net
*net
)
2693 for (i
= 0; i
< net
->ksnn_ninterfaces
; i
++) {
2694 char *ifnam
= &net
->ksnn_interfaces
[i
].ksni_name
[0];
2695 char *colon
= strchr(ifnam
, ':');
2697 struct ksock_net
*tmp
;
2700 if (colon
) /* ignore alias device */
2703 list_for_each_entry(tmp
, &ksocknal_data
.ksnd_nets
, ksnn_list
) {
2704 for (j
= 0; !found
&& j
< tmp
->ksnn_ninterfaces
; j
++) {
2706 &tmp
->ksnn_interfaces
[j
].ksni_name
[0];
2707 char *colon2
= strchr(ifnam2
, ':');
2712 found
= !strcmp(ifnam
, ifnam2
);
2729 ksocknal_start_schedulers(struct ksock_sched_info
*info
)
2735 if (!info
->ksi_nthreads
) {
2736 if (*ksocknal_tunables
.ksnd_nscheds
> 0) {
2737 nthrs
= info
->ksi_nthreads_max
;
2739 nthrs
= cfs_cpt_weight(lnet_cpt_table(),
2741 nthrs
= min(max(SOCKNAL_NSCHEDS
, nthrs
>> 1), nthrs
);
2742 nthrs
= min(SOCKNAL_NSCHEDS_HIGH
, nthrs
);
2744 nthrs
= min(nthrs
, info
->ksi_nthreads_max
);
2746 LASSERT(info
->ksi_nthreads
<= info
->ksi_nthreads_max
);
2747 /* increase two threads if there is new interface */
2748 nthrs
= min(2, info
->ksi_nthreads_max
- info
->ksi_nthreads
);
2751 for (i
= 0; i
< nthrs
; i
++) {
2754 struct ksock_sched
*sched
;
2756 id
= KSOCK_THREAD_ID(info
->ksi_cpt
, info
->ksi_nthreads
+ i
);
2757 sched
= &info
->ksi_scheds
[KSOCK_THREAD_SID(id
)];
2758 snprintf(name
, sizeof(name
), "socknal_sd%02d_%02d",
2759 info
->ksi_cpt
, (int)(sched
- &info
->ksi_scheds
[0]));
2761 rc
= ksocknal_thread_start(ksocknal_scheduler
,
2766 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2767 info
->ksi_cpt
, info
->ksi_nthreads
+ i
, rc
);
2771 info
->ksi_nthreads
+= i
;
2776 ksocknal_net_start_threads(struct ksock_net
*net
, __u32
*cpts
, int ncpts
)
2778 int newif
= ksocknal_search_new_ipif(net
);
2782 LASSERT(ncpts
> 0 && ncpts
<= cfs_cpt_number(lnet_cpt_table()));
2784 for (i
= 0; i
< ncpts
; i
++) {
2785 struct ksock_sched_info
*info
;
2786 int cpt
= !cpts
? i
: cpts
[i
];
2788 LASSERT(cpt
< cfs_cpt_number(lnet_cpt_table()));
2789 info
= ksocknal_data
.ksnd_sched_info
[cpt
];
2791 if (!newif
&& info
->ksi_nthreads
> 0)
2794 rc
= ksocknal_start_schedulers(info
);
2802 ksocknal_startup(lnet_ni_t
*ni
)
2804 struct ksock_net
*net
;
2808 LASSERT(ni
->ni_lnd
== &the_ksocklnd
);
2810 if (ksocknal_data
.ksnd_init
== SOCKNAL_INIT_NOTHING
) {
2811 rc
= ksocknal_base_startup();
2816 LIBCFS_ALLOC(net
, sizeof(*net
));
2820 spin_lock_init(&net
->ksnn_lock
);
2821 net
->ksnn_incarnation
= ksocknal_new_incarnation();
2823 ni
->ni_peertimeout
= *ksocknal_tunables
.ksnd_peertimeout
;
2824 ni
->ni_maxtxcredits
= *ksocknal_tunables
.ksnd_credits
;
2825 ni
->ni_peertxcredits
= *ksocknal_tunables
.ksnd_peertxcredits
;
2826 ni
->ni_peerrtrcredits
= *ksocknal_tunables
.ksnd_peerrtrcredits
;
2828 if (!ni
->ni_interfaces
[0]) {
2829 rc
= ksocknal_enumerate_interfaces(net
);
2833 net
->ksnn_ninterfaces
= 1;
2835 for (i
= 0; i
< LNET_MAX_INTERFACES
; i
++) {
2838 if (!ni
->ni_interfaces
[i
])
2841 rc
= lnet_ipif_query(ni
->ni_interfaces
[i
], &up
,
2842 &net
->ksnn_interfaces
[i
].ksni_ipaddr
,
2843 &net
->ksnn_interfaces
[i
].ksni_netmask
);
2846 CERROR("Can't get interface %s info: %d\n",
2847 ni
->ni_interfaces
[i
], rc
);
2852 CERROR("Interface %s is down\n",
2853 ni
->ni_interfaces
[i
]);
2857 strlcpy(net
->ksnn_interfaces
[i
].ksni_name
,
2858 ni
->ni_interfaces
[i
],
2859 sizeof(net
->ksnn_interfaces
[i
].ksni_name
));
2861 net
->ksnn_ninterfaces
= i
;
2864 /* call it before add it to ksocknal_data.ksnd_nets */
2865 rc
= ksocknal_net_start_threads(net
, ni
->ni_cpts
, ni
->ni_ncpts
);
2869 ni
->ni_nid
= LNET_MKNID(LNET_NIDNET(ni
->ni_nid
),
2870 net
->ksnn_interfaces
[0].ksni_ipaddr
);
2871 list_add(&net
->ksnn_list
, &ksocknal_data
.ksnd_nets
);
2873 ksocknal_data
.ksnd_nnets
++;
2878 LIBCFS_FREE(net
, sizeof(*net
));
2880 if (!ksocknal_data
.ksnd_nnets
)
2881 ksocknal_base_shutdown();
2886 static void __exit
ksocklnd_exit(void)
2888 lnet_unregister_lnd(&the_ksocklnd
);
2891 static int __init
ksocklnd_init(void)
2895 /* check ksnr_connected/connecting field large enough */
2896 CLASSERT(SOCKLND_CONN_NTYPES
<= 4);
2897 CLASSERT(SOCKLND_CONN_ACK
== SOCKLND_CONN_BULK_IN
);
2899 /* initialize the_ksocklnd */
2900 the_ksocklnd
.lnd_type
= SOCKLND
;
2901 the_ksocklnd
.lnd_startup
= ksocknal_startup
;
2902 the_ksocklnd
.lnd_shutdown
= ksocknal_shutdown
;
2903 the_ksocklnd
.lnd_ctl
= ksocknal_ctl
;
2904 the_ksocklnd
.lnd_send
= ksocknal_send
;
2905 the_ksocklnd
.lnd_recv
= ksocknal_recv
;
2906 the_ksocklnd
.lnd_notify
= ksocknal_notify
;
2907 the_ksocklnd
.lnd_query
= ksocknal_query
;
2908 the_ksocklnd
.lnd_accept
= ksocknal_accept
;
2910 rc
= ksocknal_tunables_init();
2914 lnet_register_lnd(&the_ksocklnd
);
2919 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2920 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2921 MODULE_VERSION("2.7.0");
2922 MODULE_LICENSE("GPL");
2924 module_init(ksocklnd_init
);
2925 module_exit(ksocklnd_exit
);