2 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
4 * Copyright (c) 2011, 2015, Intel Corporation.
6 * This file is part of Portals
7 * http://sourceforge.net/projects/sandiaportals/
9 * Portals is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Portals is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
20 #define DEBUG_SUBSYSTEM S_LNET
21 #include "../../include/linux/lnet/lib-lnet.h"
23 #define LNET_NRB_TINY_MIN 512 /* min value for each CPT */
24 #define LNET_NRB_TINY (LNET_NRB_TINY_MIN * 4)
25 #define LNET_NRB_SMALL_MIN 4096 /* min value for each CPT */
26 #define LNET_NRB_SMALL (LNET_NRB_SMALL_MIN * 4)
27 #define LNET_NRB_SMALL_PAGES 1
28 #define LNET_NRB_LARGE_MIN 256 /* min value for each CPT */
29 #define LNET_NRB_LARGE (LNET_NRB_LARGE_MIN * 4)
30 #define LNET_NRB_LARGE_PAGES ((LNET_MTU + PAGE_SIZE - 1) >> \
33 static char *forwarding
= "";
34 module_param(forwarding
, charp
, 0444);
35 MODULE_PARM_DESC(forwarding
, "Explicitly enable/disable forwarding between networks");
37 static int tiny_router_buffers
;
38 module_param(tiny_router_buffers
, int, 0444);
39 MODULE_PARM_DESC(tiny_router_buffers
, "# of 0 payload messages to buffer in the router");
40 static int small_router_buffers
;
41 module_param(small_router_buffers
, int, 0444);
42 MODULE_PARM_DESC(small_router_buffers
, "# of small (1 page) messages to buffer in the router");
43 static int large_router_buffers
;
44 module_param(large_router_buffers
, int, 0444);
45 MODULE_PARM_DESC(large_router_buffers
, "# of large messages to buffer in the router");
46 static int peer_buffer_credits
;
47 module_param(peer_buffer_credits
, int, 0444);
48 MODULE_PARM_DESC(peer_buffer_credits
, "# router buffer credits per peer");
50 static int auto_down
= 1;
51 module_param(auto_down
, int, 0444);
52 MODULE_PARM_DESC(auto_down
, "Automatically mark peers down on comms error");
55 lnet_peer_buffer_credits(lnet_ni_t
*ni
)
57 /* NI option overrides LNet default */
58 if (ni
->ni_peerrtrcredits
> 0)
59 return ni
->ni_peerrtrcredits
;
60 if (peer_buffer_credits
> 0)
61 return peer_buffer_credits
;
64 * As an approximation, allow this peer the same number of router
65 * buffers as it is allowed outstanding sends
67 return ni
->ni_peertxcredits
;
71 static int lnet_router_checker(void *);
73 static int check_routers_before_use
;
74 module_param(check_routers_before_use
, int, 0444);
75 MODULE_PARM_DESC(check_routers_before_use
, "Assume routers are down and ping them before use");
77 int avoid_asym_router_failure
= 1;
78 module_param(avoid_asym_router_failure
, int, 0644);
79 MODULE_PARM_DESC(avoid_asym_router_failure
, "Avoid asymmetrical router failures (0 to disable)");
81 static int dead_router_check_interval
= 60;
82 module_param(dead_router_check_interval
, int, 0644);
83 MODULE_PARM_DESC(dead_router_check_interval
, "Seconds between dead router health checks (<= 0 to disable)");
85 static int live_router_check_interval
= 60;
86 module_param(live_router_check_interval
, int, 0644);
87 MODULE_PARM_DESC(live_router_check_interval
, "Seconds between live router health checks (<= 0 to disable)");
89 static int router_ping_timeout
= 50;
90 module_param(router_ping_timeout
, int, 0644);
91 MODULE_PARM_DESC(router_ping_timeout
, "Seconds to wait for the reply to a router health query");
94 lnet_peers_start_down(void)
96 return check_routers_before_use
;
100 lnet_notify_locked(lnet_peer_t
*lp
, int notifylnd
, int alive
,
103 if (time_before(when
, lp
->lp_timestamp
)) { /* out of date information */
104 CDEBUG(D_NET
, "Out of date\n");
108 lp
->lp_timestamp
= when
; /* update timestamp */
109 lp
->lp_ping_deadline
= 0; /* disable ping timeout */
111 if (lp
->lp_alive_count
&& /* got old news */
112 (!lp
->lp_alive
) == (!alive
)) { /* new date for old news */
113 CDEBUG(D_NET
, "Old news\n");
117 /* Flag that notification is outstanding */
119 lp
->lp_alive_count
++;
120 lp
->lp_alive
= !(!alive
); /* 1 bit! */
122 lp
->lp_notifylnd
|= notifylnd
;
124 lp
->lp_ping_feats
= LNET_PING_FEAT_INVAL
; /* reset */
126 CDEBUG(D_NET
, "set %s %d\n", libcfs_nid2str(lp
->lp_nid
), alive
);
130 lnet_ni_notify_locked(lnet_ni_t
*ni
, lnet_peer_t
*lp
)
136 * Notify only in 1 thread at any time to ensure ordered notification.
137 * NB individual events can be missed; the only guarantee is that you
138 * always get the most recent news
140 if (lp
->lp_notifying
|| !ni
)
143 lp
->lp_notifying
= 1;
145 while (lp
->lp_notify
) {
146 alive
= lp
->lp_alive
;
147 notifylnd
= lp
->lp_notifylnd
;
149 lp
->lp_notifylnd
= 0;
152 if (notifylnd
&& ni
->ni_lnd
->lnd_notify
) {
153 lnet_net_unlock(lp
->lp_cpt
);
156 * A new notification could happen now; I'll handle it
157 * when control returns to me
159 ni
->ni_lnd
->lnd_notify(ni
, lp
->lp_nid
, alive
);
161 lnet_net_lock(lp
->lp_cpt
);
165 lp
->lp_notifying
= 0;
169 lnet_rtr_addref_locked(lnet_peer_t
*lp
)
171 LASSERT(lp
->lp_refcount
> 0);
172 LASSERT(lp
->lp_rtr_refcount
>= 0);
174 /* lnet_net_lock must be exclusively locked */
175 lp
->lp_rtr_refcount
++;
176 if (lp
->lp_rtr_refcount
== 1) {
177 struct list_head
*pos
;
179 /* a simple insertion sort */
180 list_for_each_prev(pos
, &the_lnet
.ln_routers
) {
181 lnet_peer_t
*rtr
= list_entry(pos
, lnet_peer_t
,
184 if (rtr
->lp_nid
< lp
->lp_nid
)
188 list_add(&lp
->lp_rtr_list
, pos
);
189 /* addref for the_lnet.ln_routers */
190 lnet_peer_addref_locked(lp
);
191 the_lnet
.ln_routers_version
++;
196 lnet_rtr_decref_locked(lnet_peer_t
*lp
)
198 LASSERT(lp
->lp_refcount
> 0);
199 LASSERT(lp
->lp_rtr_refcount
> 0);
201 /* lnet_net_lock must be exclusively locked */
202 lp
->lp_rtr_refcount
--;
203 if (!lp
->lp_rtr_refcount
) {
204 LASSERT(list_empty(&lp
->lp_routes
));
207 list_add(&lp
->lp_rcd
->rcd_list
,
208 &the_lnet
.ln_rcd_deathrow
);
212 list_del(&lp
->lp_rtr_list
);
213 /* decref for the_lnet.ln_routers */
214 lnet_peer_decref_locked(lp
);
215 the_lnet
.ln_routers_version
++;
220 lnet_find_net_locked(__u32 net
)
222 lnet_remotenet_t
*rnet
;
223 struct list_head
*tmp
;
224 struct list_head
*rn_list
;
226 LASSERT(!the_lnet
.ln_shutdown
);
228 rn_list
= lnet_net2rnethash(net
);
229 list_for_each(tmp
, rn_list
) {
230 rnet
= list_entry(tmp
, lnet_remotenet_t
, lrn_list
);
232 if (rnet
->lrn_net
== net
)
238 static void lnet_shuffle_seed(void)
241 __u32 lnd_type
, seed
[2];
242 struct timespec64 ts
;
244 struct list_head
*tmp
;
249 cfs_get_random_bytes(seed
, sizeof(seed
));
252 * Nodes with small feet have little entropy
253 * the NID for this node gives the most entropy in the low bits
255 list_for_each(tmp
, &the_lnet
.ln_nis
) {
256 ni
= list_entry(tmp
, lnet_ni_t
, ni_list
);
257 lnd_type
= LNET_NETTYP(LNET_NIDNET(ni
->ni_nid
));
259 if (lnd_type
!= LOLND
)
260 seed
[0] ^= (LNET_NIDADDR(ni
->ni_nid
) | lnd_type
);
264 cfs_srand(ts
.tv_sec
^ seed
[0], ts
.tv_nsec
^ seed
[1]);
268 /* NB expects LNET_LOCK held */
270 lnet_add_route_to_rnet(lnet_remotenet_t
*rnet
, lnet_route_t
*route
)
272 unsigned int len
= 0;
273 unsigned int offset
= 0;
278 list_for_each(e
, &rnet
->lrn_routes
) {
282 /* len+1 positions to add a new entry, also prevents division by 0 */
283 offset
= cfs_rand() % (len
+ 1);
284 list_for_each(e
, &rnet
->lrn_routes
) {
289 list_add(&route
->lr_list
, e
);
290 list_add(&route
->lr_gwlist
, &route
->lr_gateway
->lp_routes
);
292 the_lnet
.ln_remote_nets_version
++;
293 lnet_rtr_addref_locked(route
->lr_gateway
);
297 lnet_add_route(__u32 net
, __u32 hops
, lnet_nid_t gateway
,
298 unsigned int priority
)
301 lnet_remotenet_t
*rnet
;
302 lnet_remotenet_t
*rnet2
;
308 CDEBUG(D_NET
, "Add route: net %s hops %d priority %u gw %s\n",
309 libcfs_net2str(net
), hops
, priority
, libcfs_nid2str(gateway
));
311 if (gateway
== LNET_NID_ANY
||
312 LNET_NETTYP(LNET_NIDNET(gateway
)) == LOLND
||
313 net
== LNET_NIDNET(LNET_NID_ANY
) ||
314 LNET_NETTYP(net
) == LOLND
||
315 LNET_NIDNET(gateway
) == net
||
316 (hops
!= LNET_UNDEFINED_HOPS
&& (hops
< 1 || hops
> 255)))
319 if (lnet_islocalnet(net
)) /* it's a local network */
322 /* Assume net, route, all new */
323 LIBCFS_ALLOC(route
, sizeof(*route
));
324 LIBCFS_ALLOC(rnet
, sizeof(*rnet
));
325 if (!route
|| !rnet
) {
326 CERROR("Out of memory creating route %s %d %s\n",
327 libcfs_net2str(net
), hops
, libcfs_nid2str(gateway
));
329 LIBCFS_FREE(route
, sizeof(*route
));
331 LIBCFS_FREE(rnet
, sizeof(*rnet
));
335 INIT_LIST_HEAD(&rnet
->lrn_routes
);
337 route
->lr_hops
= hops
;
339 route
->lr_priority
= priority
;
341 lnet_net_lock(LNET_LOCK_EX
);
343 rc
= lnet_nid2peer_locked(&route
->lr_gateway
, gateway
, LNET_LOCK_EX
);
345 lnet_net_unlock(LNET_LOCK_EX
);
347 LIBCFS_FREE(route
, sizeof(*route
));
348 LIBCFS_FREE(rnet
, sizeof(*rnet
));
350 if (rc
== -EHOSTUNREACH
) /* gateway is not on a local net */
351 return rc
; /* ignore the route entry */
352 CERROR("Error %d creating route %s %d %s\n", rc
,
353 libcfs_net2str(net
), hops
,
354 libcfs_nid2str(gateway
));
358 LASSERT(!the_lnet
.ln_shutdown
);
360 rnet2
= lnet_find_net_locked(net
);
363 list_add_tail(&rnet
->lrn_list
, lnet_net2rnethash(net
));
367 /* Search for a duplicate route (it's a NOOP if it is) */
369 list_for_each(e
, &rnet2
->lrn_routes
) {
370 lnet_route_t
*route2
= list_entry(e
, lnet_route_t
, lr_list
);
372 if (route2
->lr_gateway
== route
->lr_gateway
) {
377 /* our lookups must be true */
378 LASSERT(route2
->lr_gateway
->lp_nid
!= gateway
);
382 lnet_peer_addref_locked(route
->lr_gateway
); /* +1 for notify */
383 lnet_add_route_to_rnet(rnet2
, route
);
385 ni
= route
->lr_gateway
->lp_ni
;
386 lnet_net_unlock(LNET_LOCK_EX
);
388 /* XXX Assume alive */
389 if (ni
->ni_lnd
->lnd_notify
)
390 ni
->ni_lnd
->lnd_notify(ni
, gateway
, 1);
392 lnet_net_lock(LNET_LOCK_EX
);
395 /* -1 for notify or !add_route */
396 lnet_peer_decref_locked(route
->lr_gateway
);
397 lnet_net_unlock(LNET_LOCK_EX
);
402 LIBCFS_FREE(route
, sizeof(*route
));
406 LIBCFS_FREE(rnet
, sizeof(*rnet
));
408 /* indicate to startup the router checker if configured */
409 wake_up(&the_lnet
.ln_rc_waitq
);
415 lnet_check_routes(void)
417 lnet_remotenet_t
*rnet
;
419 lnet_route_t
*route2
;
420 struct list_head
*e1
;
421 struct list_head
*e2
;
423 struct list_head
*rn_list
;
426 cpt
= lnet_net_lock_current();
428 for (i
= 0; i
< LNET_REMOTE_NETS_HASH_SIZE
; i
++) {
429 rn_list
= &the_lnet
.ln_remote_nets_hash
[i
];
430 list_for_each(e1
, rn_list
) {
431 rnet
= list_entry(e1
, lnet_remotenet_t
, lrn_list
);
434 list_for_each(e2
, &rnet
->lrn_routes
) {
439 route
= list_entry(e2
, lnet_route_t
, lr_list
);
446 if (route
->lr_gateway
->lp_ni
==
447 route2
->lr_gateway
->lp_ni
)
450 nid1
= route
->lr_gateway
->lp_nid
;
451 nid2
= route2
->lr_gateway
->lp_nid
;
454 lnet_net_unlock(cpt
);
456 CERROR("Routes to %s via %s and %s not supported\n",
458 libcfs_nid2str(nid1
),
459 libcfs_nid2str(nid2
));
465 lnet_net_unlock(cpt
);
470 lnet_del_route(__u32 net
, lnet_nid_t gw_nid
)
472 struct lnet_peer
*gateway
;
473 lnet_remotenet_t
*rnet
;
475 struct list_head
*e1
;
476 struct list_head
*e2
;
478 struct list_head
*rn_list
;
481 CDEBUG(D_NET
, "Del route: net %s : gw %s\n",
482 libcfs_net2str(net
), libcfs_nid2str(gw_nid
));
485 * NB Caller may specify either all routes via the given gateway
486 * or a specific route entry actual NIDs)
488 lnet_net_lock(LNET_LOCK_EX
);
489 if (net
== LNET_NIDNET(LNET_NID_ANY
))
490 rn_list
= &the_lnet
.ln_remote_nets_hash
[0];
492 rn_list
= lnet_net2rnethash(net
);
495 list_for_each(e1
, rn_list
) {
496 rnet
= list_entry(e1
, lnet_remotenet_t
, lrn_list
);
498 if (!(net
== LNET_NIDNET(LNET_NID_ANY
) ||
499 net
== rnet
->lrn_net
))
502 list_for_each(e2
, &rnet
->lrn_routes
) {
503 route
= list_entry(e2
, lnet_route_t
, lr_list
);
505 gateway
= route
->lr_gateway
;
506 if (!(gw_nid
== LNET_NID_ANY
||
507 gw_nid
== gateway
->lp_nid
))
510 list_del(&route
->lr_list
);
511 list_del(&route
->lr_gwlist
);
512 the_lnet
.ln_remote_nets_version
++;
514 if (list_empty(&rnet
->lrn_routes
))
515 list_del(&rnet
->lrn_list
);
519 lnet_rtr_decref_locked(gateway
);
520 lnet_peer_decref_locked(gateway
);
522 lnet_net_unlock(LNET_LOCK_EX
);
524 LIBCFS_FREE(route
, sizeof(*route
));
527 LIBCFS_FREE(rnet
, sizeof(*rnet
));
530 lnet_net_lock(LNET_LOCK_EX
);
535 if (net
== LNET_NIDNET(LNET_NID_ANY
) &&
536 ++idx
< LNET_REMOTE_NETS_HASH_SIZE
) {
537 rn_list
= &the_lnet
.ln_remote_nets_hash
[idx
];
540 lnet_net_unlock(LNET_LOCK_EX
);
546 lnet_destroy_routes(void)
548 lnet_del_route(LNET_NIDNET(LNET_NID_ANY
), LNET_NID_ANY
);
551 int lnet_get_rtr_pool_cfg(int idx
, struct lnet_ioctl_pool_cfg
*pool_cfg
)
553 int i
, rc
= -ENOENT
, j
;
555 if (!the_lnet
.ln_rtrpools
)
558 for (i
= 0; i
< LNET_NRBPOOLS
; i
++) {
559 lnet_rtrbufpool_t
*rbp
;
561 lnet_net_lock(LNET_LOCK_EX
);
562 cfs_percpt_for_each(rbp
, j
, the_lnet
.ln_rtrpools
) {
566 pool_cfg
->pl_pools
[i
].pl_npages
= rbp
[i
].rbp_npages
;
567 pool_cfg
->pl_pools
[i
].pl_nbuffers
= rbp
[i
].rbp_nbuffers
;
568 pool_cfg
->pl_pools
[i
].pl_credits
= rbp
[i
].rbp_credits
;
569 pool_cfg
->pl_pools
[i
].pl_mincredits
= rbp
[i
].rbp_mincredits
;
573 lnet_net_unlock(LNET_LOCK_EX
);
576 lnet_net_lock(LNET_LOCK_EX
);
577 pool_cfg
->pl_routing
= the_lnet
.ln_routing
;
578 lnet_net_unlock(LNET_LOCK_EX
);
584 lnet_get_route(int idx
, __u32
*net
, __u32
*hops
,
585 lnet_nid_t
*gateway
, __u32
*alive
, __u32
*priority
)
587 struct list_head
*e1
;
588 struct list_head
*e2
;
589 lnet_remotenet_t
*rnet
;
593 struct list_head
*rn_list
;
595 cpt
= lnet_net_lock_current();
597 for (i
= 0; i
< LNET_REMOTE_NETS_HASH_SIZE
; i
++) {
598 rn_list
= &the_lnet
.ln_remote_nets_hash
[i
];
599 list_for_each(e1
, rn_list
) {
600 rnet
= list_entry(e1
, lnet_remotenet_t
, lrn_list
);
602 list_for_each(e2
, &rnet
->lrn_routes
) {
603 route
= list_entry(e2
, lnet_route_t
, lr_list
);
606 *net
= rnet
->lrn_net
;
607 *hops
= route
->lr_hops
;
608 *priority
= route
->lr_priority
;
609 *gateway
= route
->lr_gateway
->lp_nid
;
610 *alive
= lnet_is_route_alive(route
);
611 lnet_net_unlock(cpt
);
618 lnet_net_unlock(cpt
);
623 lnet_swap_pinginfo(lnet_ping_info_t
*info
)
626 lnet_ni_status_t
*stat
;
628 __swab32s(&info
->pi_magic
);
629 __swab32s(&info
->pi_features
);
630 __swab32s(&info
->pi_pid
);
631 __swab32s(&info
->pi_nnis
);
632 for (i
= 0; i
< info
->pi_nnis
&& i
< LNET_MAX_RTR_NIS
; i
++) {
633 stat
= &info
->pi_ni
[i
];
634 __swab64s(&stat
->ns_nid
);
635 __swab32s(&stat
->ns_status
);
640 * parse router-checker pinginfo, record number of down NIs for remote
641 * networks on that router.
644 lnet_parse_rc_info(lnet_rc_data_t
*rcd
)
646 lnet_ping_info_t
*info
= rcd
->rcd_pinginfo
;
647 struct lnet_peer
*gw
= rcd
->rcd_gateway
;
653 if (info
->pi_magic
== __swab32(LNET_PROTO_PING_MAGIC
))
654 lnet_swap_pinginfo(info
);
656 /* NB always racing with network! */
657 if (info
->pi_magic
!= LNET_PROTO_PING_MAGIC
) {
658 CDEBUG(D_NET
, "%s: Unexpected magic %08x\n",
659 libcfs_nid2str(gw
->lp_nid
), info
->pi_magic
);
660 gw
->lp_ping_feats
= LNET_PING_FEAT_INVAL
;
664 gw
->lp_ping_feats
= info
->pi_features
;
665 if (!(gw
->lp_ping_feats
& LNET_PING_FEAT_MASK
)) {
666 CDEBUG(D_NET
, "%s: Unexpected features 0x%x\n",
667 libcfs_nid2str(gw
->lp_nid
), gw
->lp_ping_feats
);
668 return; /* nothing I can understand */
671 if (!(gw
->lp_ping_feats
& LNET_PING_FEAT_NI_STATUS
))
672 return; /* can't carry NI status info */
674 list_for_each_entry(rte
, &gw
->lp_routes
, lr_gwlist
) {
679 if (gw
->lp_ping_feats
& LNET_PING_FEAT_RTE_DISABLED
) {
684 for (i
= 0; i
< info
->pi_nnis
&& i
< LNET_MAX_RTR_NIS
; i
++) {
685 lnet_ni_status_t
*stat
= &info
->pi_ni
[i
];
686 lnet_nid_t nid
= stat
->ns_nid
;
688 if (nid
== LNET_NID_ANY
) {
689 CDEBUG(D_NET
, "%s: unexpected LNET_NID_ANY\n",
690 libcfs_nid2str(gw
->lp_nid
));
691 gw
->lp_ping_feats
= LNET_PING_FEAT_INVAL
;
695 if (LNET_NETTYP(LNET_NIDNET(nid
)) == LOLND
)
698 if (stat
->ns_status
== LNET_NI_STATUS_DOWN
) {
703 if (stat
->ns_status
== LNET_NI_STATUS_UP
) {
704 if (LNET_NIDNET(nid
) == rte
->lr_net
) {
711 CDEBUG(D_NET
, "%s: Unexpected status 0x%x\n",
712 libcfs_nid2str(gw
->lp_nid
), stat
->ns_status
);
713 gw
->lp_ping_feats
= LNET_PING_FEAT_INVAL
;
717 if (up
) { /* ignore downed NIs if NI for dest network is up */
722 * if @down is zero and this route is single-hop, it means
723 * we can't find NI for target network
725 if (!down
&& rte
->lr_hops
== 1)
728 rte
->lr_downis
= down
;
733 lnet_router_checker_event(lnet_event_t
*event
)
735 lnet_rc_data_t
*rcd
= event
->md
.user_ptr
;
736 struct lnet_peer
*lp
;
740 if (event
->unlinked
) {
741 LNetInvalidateHandle(&rcd
->rcd_mdh
);
745 LASSERT(event
->type
== LNET_EVENT_SEND
||
746 event
->type
== LNET_EVENT_REPLY
);
748 lp
= rcd
->rcd_gateway
;
752 * NB: it's called with holding lnet_res_lock, we have a few
753 * places need to hold both locks at the same time, please take
754 * care of lock ordering
756 lnet_net_lock(lp
->lp_cpt
);
757 if (!lnet_isrouter(lp
) || lp
->lp_rcd
!= rcd
) {
758 /* ignore if no longer a router or rcd is replaced */
762 if (event
->type
== LNET_EVENT_SEND
) {
763 lp
->lp_ping_notsent
= 0;
768 /* LNET_EVENT_REPLY */
770 * A successful REPLY means the router is up. If _any_ comms
771 * to the router fail I assume it's down (this will happen if
772 * we ping alive routers to try to detect router death before
775 lnet_notify_locked(lp
, 1, !event
->status
, cfs_time_current());
778 * The router checker will wake up very shortly and do the
779 * actual notification.
780 * XXX If 'lp' stops being a router before then, it will still
781 * have the notification pending!!!
783 if (avoid_asym_router_failure
&& !event
->status
)
784 lnet_parse_rc_info(rcd
);
787 lnet_net_unlock(lp
->lp_cpt
);
791 lnet_wait_known_routerstate(void)
794 struct list_head
*entry
;
797 LASSERT(the_lnet
.ln_rc_state
== LNET_RC_STATE_RUNNING
);
800 int cpt
= lnet_net_lock_current();
803 list_for_each(entry
, &the_lnet
.ln_routers
) {
804 rtr
= list_entry(entry
, lnet_peer_t
, lp_rtr_list
);
806 if (!rtr
->lp_alive_count
) {
812 lnet_net_unlock(cpt
);
817 set_current_state(TASK_UNINTERRUPTIBLE
);
818 schedule_timeout(cfs_time_seconds(1));
823 lnet_router_ni_update_locked(lnet_peer_t
*gw
, __u32 net
)
827 if ((gw
->lp_ping_feats
& LNET_PING_FEAT_NI_STATUS
)) {
828 list_for_each_entry(rte
, &gw
->lp_routes
, lr_gwlist
) {
829 if (rte
->lr_net
== net
) {
838 lnet_update_ni_status_locked(void)
844 LASSERT(the_lnet
.ln_routing
);
846 timeout
= router_ping_timeout
+
847 max(live_router_check_interval
, dead_router_check_interval
);
849 now
= ktime_get_real_seconds();
850 list_for_each_entry(ni
, &the_lnet
.ln_nis
, ni_list
) {
851 if (ni
->ni_lnd
->lnd_type
== LOLND
)
854 if (now
< ni
->ni_last_alive
+ timeout
)
858 /* re-check with lock */
859 if (now
< ni
->ni_last_alive
+ timeout
) {
864 LASSERT(ni
->ni_status
);
866 if (ni
->ni_status
->ns_status
!= LNET_NI_STATUS_DOWN
) {
867 CDEBUG(D_NET
, "NI(%s:%d) status changed to down\n",
868 libcfs_nid2str(ni
->ni_nid
), timeout
);
870 * NB: so far, this is the only place to set
871 * NI status to "down"
873 ni
->ni_status
->ns_status
= LNET_NI_STATUS_DOWN
;
880 lnet_destroy_rc_data(lnet_rc_data_t
*rcd
)
882 LASSERT(list_empty(&rcd
->rcd_list
));
883 /* detached from network */
884 LASSERT(LNetHandleIsInvalid(rcd
->rcd_mdh
));
886 if (rcd
->rcd_gateway
) {
887 int cpt
= rcd
->rcd_gateway
->lp_cpt
;
890 lnet_peer_decref_locked(rcd
->rcd_gateway
);
891 lnet_net_unlock(cpt
);
894 if (rcd
->rcd_pinginfo
)
895 LIBCFS_FREE(rcd
->rcd_pinginfo
, LNET_PINGINFO_SIZE
);
897 LIBCFS_FREE(rcd
, sizeof(*rcd
));
900 static lnet_rc_data_t
*
901 lnet_create_rc_data_locked(lnet_peer_t
*gateway
)
903 lnet_rc_data_t
*rcd
= NULL
;
904 lnet_ping_info_t
*pi
;
908 lnet_net_unlock(gateway
->lp_cpt
);
910 LIBCFS_ALLOC(rcd
, sizeof(*rcd
));
914 LNetInvalidateHandle(&rcd
->rcd_mdh
);
915 INIT_LIST_HEAD(&rcd
->rcd_list
);
917 LIBCFS_ALLOC(pi
, LNET_PINGINFO_SIZE
);
921 for (i
= 0; i
< LNET_MAX_RTR_NIS
; i
++) {
922 pi
->pi_ni
[i
].ns_nid
= LNET_NID_ANY
;
923 pi
->pi_ni
[i
].ns_status
= LNET_NI_STATUS_INVALID
;
925 rcd
->rcd_pinginfo
= pi
;
927 LASSERT(!LNetHandleIsInvalid(the_lnet
.ln_rc_eqh
));
928 rc
= LNetMDBind((lnet_md_t
){.start
= pi
,
930 .length
= LNET_PINGINFO_SIZE
,
931 .threshold
= LNET_MD_THRESH_INF
,
932 .options
= LNET_MD_TRUNCATE
,
933 .eq_handle
= the_lnet
.ln_rc_eqh
},
937 CERROR("Can't bind MD: %d\n", rc
);
942 lnet_net_lock(gateway
->lp_cpt
);
943 /* router table changed or someone has created rcd for this gateway */
944 if (!lnet_isrouter(gateway
) || gateway
->lp_rcd
) {
945 lnet_net_unlock(gateway
->lp_cpt
);
949 lnet_peer_addref_locked(gateway
);
950 rcd
->rcd_gateway
= gateway
;
951 gateway
->lp_rcd
= rcd
;
952 gateway
->lp_ping_notsent
= 0;
958 if (!LNetHandleIsInvalid(rcd
->rcd_mdh
)) {
959 rc
= LNetMDUnlink(rcd
->rcd_mdh
);
962 lnet_destroy_rc_data(rcd
);
965 lnet_net_lock(gateway
->lp_cpt
);
966 return gateway
->lp_rcd
;
970 lnet_router_check_interval(lnet_peer_t
*rtr
)
974 secs
= rtr
->lp_alive
? live_router_check_interval
:
975 dead_router_check_interval
;
983 lnet_ping_router_locked(lnet_peer_t
*rtr
)
985 lnet_rc_data_t
*rcd
= NULL
;
986 unsigned long now
= cfs_time_current();
989 lnet_peer_addref_locked(rtr
);
991 if (rtr
->lp_ping_deadline
&& /* ping timed out? */
992 cfs_time_after(now
, rtr
->lp_ping_deadline
))
993 lnet_notify_locked(rtr
, 1, 0, now
);
995 /* Run any outstanding notifications */
996 lnet_ni_notify_locked(rtr
->lp_ni
, rtr
);
998 if (!lnet_isrouter(rtr
) ||
999 the_lnet
.ln_rc_state
!= LNET_RC_STATE_RUNNING
) {
1000 /* router table changed or router checker is shutting down */
1001 lnet_peer_decref_locked(rtr
);
1006 rtr
->lp_rcd
: lnet_create_rc_data_locked(rtr
);
1011 secs
= lnet_router_check_interval(rtr
);
1014 "rtr %s %d: deadline %lu ping_notsent %d alive %d alive_count %d lp_ping_timestamp %lu\n",
1015 libcfs_nid2str(rtr
->lp_nid
), secs
,
1016 rtr
->lp_ping_deadline
, rtr
->lp_ping_notsent
,
1017 rtr
->lp_alive
, rtr
->lp_alive_count
, rtr
->lp_ping_timestamp
);
1019 if (secs
&& !rtr
->lp_ping_notsent
&&
1020 cfs_time_after(now
, cfs_time_add(rtr
->lp_ping_timestamp
,
1021 cfs_time_seconds(secs
)))) {
1023 lnet_process_id_t id
;
1024 lnet_handle_md_t mdh
;
1026 id
.nid
= rtr
->lp_nid
;
1027 id
.pid
= LNET_PID_LUSTRE
;
1028 CDEBUG(D_NET
, "Check: %s\n", libcfs_id2str(id
));
1030 rtr
->lp_ping_notsent
= 1;
1031 rtr
->lp_ping_timestamp
= now
;
1035 if (!rtr
->lp_ping_deadline
) {
1036 rtr
->lp_ping_deadline
=
1037 cfs_time_shift(router_ping_timeout
);
1040 lnet_net_unlock(rtr
->lp_cpt
);
1042 rc
= LNetGet(LNET_NID_ANY
, mdh
, id
, LNET_RESERVED_PORTAL
,
1043 LNET_PROTO_PING_MATCHBITS
, 0);
1045 lnet_net_lock(rtr
->lp_cpt
);
1047 rtr
->lp_ping_notsent
= 0; /* no event pending */
1050 lnet_peer_decref_locked(rtr
);
1054 lnet_router_checker_start(void)
1056 struct task_struct
*task
;
1060 LASSERT(the_lnet
.ln_rc_state
== LNET_RC_STATE_SHUTDOWN
);
1062 if (check_routers_before_use
&&
1063 dead_router_check_interval
<= 0) {
1064 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be set if 'check_routers_before_use' is set\n");
1068 sema_init(&the_lnet
.ln_rc_signal
, 0);
1070 rc
= LNetEQAlloc(0, lnet_router_checker_event
, &the_lnet
.ln_rc_eqh
);
1072 CERROR("Can't allocate EQ(%d): %d\n", eqsz
, rc
);
1076 the_lnet
.ln_rc_state
= LNET_RC_STATE_RUNNING
;
1077 task
= kthread_run(lnet_router_checker
, NULL
, "router_checker");
1080 CERROR("Can't start router checker thread: %d\n", rc
);
1081 /* block until event callback signals exit */
1082 down(&the_lnet
.ln_rc_signal
);
1083 rc
= LNetEQFree(the_lnet
.ln_rc_eqh
);
1085 the_lnet
.ln_rc_state
= LNET_RC_STATE_SHUTDOWN
;
1089 if (check_routers_before_use
) {
1091 * Note that a helpful side-effect of pinging all known routers
1092 * at startup is that it makes them drop stale connections they
1093 * may have to a previous instance of me.
1095 lnet_wait_known_routerstate();
1102 lnet_router_checker_stop(void)
1106 if (the_lnet
.ln_rc_state
== LNET_RC_STATE_SHUTDOWN
)
1109 LASSERT(the_lnet
.ln_rc_state
== LNET_RC_STATE_RUNNING
);
1110 the_lnet
.ln_rc_state
= LNET_RC_STATE_STOPPING
;
1111 /* wakeup the RC thread if it's sleeping */
1112 wake_up(&the_lnet
.ln_rc_waitq
);
1114 /* block until event callback signals exit */
1115 down(&the_lnet
.ln_rc_signal
);
1116 LASSERT(the_lnet
.ln_rc_state
== LNET_RC_STATE_SHUTDOWN
);
1118 rc
= LNetEQFree(the_lnet
.ln_rc_eqh
);
1123 lnet_prune_rc_data(int wait_unlink
)
1125 lnet_rc_data_t
*rcd
;
1126 lnet_rc_data_t
*tmp
;
1128 struct list_head head
;
1131 if (likely(the_lnet
.ln_rc_state
== LNET_RC_STATE_RUNNING
&&
1132 list_empty(&the_lnet
.ln_rcd_deathrow
) &&
1133 list_empty(&the_lnet
.ln_rcd_zombie
)))
1136 INIT_LIST_HEAD(&head
);
1138 lnet_net_lock(LNET_LOCK_EX
);
1140 if (the_lnet
.ln_rc_state
!= LNET_RC_STATE_RUNNING
) {
1141 /* router checker is stopping, prune all */
1142 list_for_each_entry(lp
, &the_lnet
.ln_routers
,
1147 LASSERT(list_empty(&lp
->lp_rcd
->rcd_list
));
1148 list_add(&lp
->lp_rcd
->rcd_list
,
1149 &the_lnet
.ln_rcd_deathrow
);
1154 /* unlink all RCDs on deathrow list */
1155 list_splice_init(&the_lnet
.ln_rcd_deathrow
, &head
);
1157 if (!list_empty(&head
)) {
1158 lnet_net_unlock(LNET_LOCK_EX
);
1160 list_for_each_entry(rcd
, &head
, rcd_list
)
1161 LNetMDUnlink(rcd
->rcd_mdh
);
1163 lnet_net_lock(LNET_LOCK_EX
);
1166 list_splice_init(&head
, &the_lnet
.ln_rcd_zombie
);
1168 /* release all zombie RCDs */
1169 while (!list_empty(&the_lnet
.ln_rcd_zombie
)) {
1170 list_for_each_entry_safe(rcd
, tmp
, &the_lnet
.ln_rcd_zombie
,
1172 if (LNetHandleIsInvalid(rcd
->rcd_mdh
))
1173 list_move(&rcd
->rcd_list
, &head
);
1176 wait_unlink
= wait_unlink
&&
1177 !list_empty(&the_lnet
.ln_rcd_zombie
);
1179 lnet_net_unlock(LNET_LOCK_EX
);
1181 while (!list_empty(&head
)) {
1182 rcd
= list_entry(head
.next
,
1183 lnet_rc_data_t
, rcd_list
);
1184 list_del_init(&rcd
->rcd_list
);
1185 lnet_destroy_rc_data(rcd
);
1192 CDEBUG(((i
& (-i
)) == i
) ? D_WARNING
: D_NET
,
1193 "Waiting for rc buffers to unlink\n");
1194 set_current_state(TASK_UNINTERRUPTIBLE
);
1195 schedule_timeout(cfs_time_seconds(1) / 4);
1197 lnet_net_lock(LNET_LOCK_EX
);
1200 lnet_net_unlock(LNET_LOCK_EX
);
1204 * This function is called to check if the RC should block indefinitely.
1205 * It's called from lnet_router_checker() as well as being passed to
1206 * wait_event_interruptible() to avoid the lost wake_up problem.
1208 * When it's called from wait_event_interruptible() it is necessary to
1209 * also not sleep if the rc state is not running to avoid a deadlock
1210 * when the system is shutting down
1213 lnet_router_checker_active(void)
1215 if (the_lnet
.ln_rc_state
!= LNET_RC_STATE_RUNNING
)
1219 * Router Checker thread needs to run when routing is enabled in
1220 * order to call lnet_update_ni_status_locked()
1222 if (the_lnet
.ln_routing
)
1225 return !list_empty(&the_lnet
.ln_routers
) &&
1226 (live_router_check_interval
> 0 ||
1227 dead_router_check_interval
> 0);
1231 lnet_router_checker(void *arg
)
1234 struct list_head
*entry
;
1236 cfs_block_allsigs();
1238 while (the_lnet
.ln_rc_state
== LNET_RC_STATE_RUNNING
) {
1243 cpt
= lnet_net_lock_current();
1245 version
= the_lnet
.ln_routers_version
;
1247 list_for_each(entry
, &the_lnet
.ln_routers
) {
1248 rtr
= list_entry(entry
, lnet_peer_t
, lp_rtr_list
);
1250 cpt2
= lnet_cpt_of_nid_locked(rtr
->lp_nid
);
1252 lnet_net_unlock(cpt
);
1255 /* the routers list has changed */
1256 if (version
!= the_lnet
.ln_routers_version
)
1260 lnet_ping_router_locked(rtr
);
1262 /* NB dropped lock */
1263 if (version
!= the_lnet
.ln_routers_version
) {
1264 /* the routers list has changed */
1269 if (the_lnet
.ln_routing
)
1270 lnet_update_ni_status_locked();
1272 lnet_net_unlock(cpt
);
1274 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1277 * Call schedule_timeout() here always adds 1 to load average
1278 * because kernel counts # active tasks as nr_running
1279 * + nr_uninterruptible.
1282 * if there are any routes then wakeup every second. If
1283 * there are no routes then sleep indefinitely until woken
1284 * up by a user adding a route
1286 if (!lnet_router_checker_active())
1287 wait_event_interruptible(the_lnet
.ln_rc_waitq
,
1288 lnet_router_checker_active());
1290 wait_event_interruptible_timeout(the_lnet
.ln_rc_waitq
,
1292 cfs_time_seconds(1));
1295 lnet_prune_rc_data(1); /* wait for UNLINK */
1297 the_lnet
.ln_rc_state
= LNET_RC_STATE_SHUTDOWN
;
1298 up(&the_lnet
.ln_rc_signal
);
1299 /* The unlink event callback will signal final completion */
1304 lnet_destroy_rtrbuf(lnet_rtrbuf_t
*rb
, int npages
)
1306 int sz
= offsetof(lnet_rtrbuf_t
, rb_kiov
[npages
]);
1308 while (--npages
>= 0)
1309 __free_page(rb
->rb_kiov
[npages
].kiov_page
);
1311 LIBCFS_FREE(rb
, sz
);
1314 static lnet_rtrbuf_t
*
1315 lnet_new_rtrbuf(lnet_rtrbufpool_t
*rbp
, int cpt
)
1317 int npages
= rbp
->rbp_npages
;
1318 int sz
= offsetof(lnet_rtrbuf_t
, rb_kiov
[npages
]);
1323 LIBCFS_CPT_ALLOC(rb
, lnet_cpt_table(), cpt
, sz
);
1329 for (i
= 0; i
< npages
; i
++) {
1330 page
= alloc_pages_node(
1331 cfs_cpt_spread_node(lnet_cpt_table(), cpt
),
1332 GFP_KERNEL
| __GFP_ZERO
, 0);
1335 __free_page(rb
->rb_kiov
[i
].kiov_page
);
1337 LIBCFS_FREE(rb
, sz
);
1341 rb
->rb_kiov
[i
].kiov_len
= PAGE_SIZE
;
1342 rb
->rb_kiov
[i
].kiov_offset
= 0;
1343 rb
->rb_kiov
[i
].kiov_page
= page
;
1350 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t
*rbp
, int cpt
)
1352 int npages
= rbp
->rbp_npages
;
1353 struct list_head tmp
;
1355 lnet_rtrbuf_t
*temp
;
1357 if (!rbp
->rbp_nbuffers
) /* not initialized or already freed */
1360 INIT_LIST_HEAD(&tmp
);
1363 lnet_drop_routed_msgs_locked(&rbp
->rbp_msgs
, cpt
);
1364 list_splice_init(&rbp
->rbp_bufs
, &tmp
);
1365 rbp
->rbp_req_nbuffers
= 0;
1366 rbp
->rbp_nbuffers
= 0;
1367 rbp
->rbp_credits
= 0;
1368 rbp
->rbp_mincredits
= 0;
1369 lnet_net_unlock(cpt
);
1371 /* Free buffers on the free list. */
1372 list_for_each_entry_safe(rb
, temp
, &tmp
, rb_list
) {
1373 list_del(&rb
->rb_list
);
1374 lnet_destroy_rtrbuf(rb
, npages
);
1379 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t
*rbp
, int nbufs
, int cpt
)
1381 struct list_head rb_list
;
1384 int num_buffers
= 0;
1386 int npages
= rbp
->rbp_npages
;
1390 * If we are called for less buffers than already in the pool, we
1391 * just lower the req_nbuffers number and excess buffers will be
1392 * thrown away as they are returned to the free list. Credits
1393 * then get adjusted as well.
1394 * If we already have enough buffers allocated to serve the
1395 * increase requested, then we can treat that the same way as we
1398 num_rb
= nbufs
- rbp
->rbp_nbuffers
;
1399 if (nbufs
<= rbp
->rbp_req_nbuffers
|| num_rb
<= 0) {
1400 rbp
->rbp_req_nbuffers
= nbufs
;
1401 lnet_net_unlock(cpt
);
1405 * store the older value of rbp_req_nbuffers and then set it to
1406 * the new request to prevent lnet_return_rx_credits_locked() from
1407 * freeing buffers that we need to keep around
1409 old_req_nbufs
= rbp
->rbp_req_nbuffers
;
1410 rbp
->rbp_req_nbuffers
= nbufs
;
1411 lnet_net_unlock(cpt
);
1413 INIT_LIST_HEAD(&rb_list
);
1416 * allocate the buffers on a local list first. If all buffers are
1417 * allocated successfully then join this list to the rbp buffer
1418 * list. If not then free all allocated buffers.
1420 while (num_rb
-- > 0) {
1421 rb
= lnet_new_rtrbuf(rbp
, cpt
);
1423 CERROR("Failed to allocate %d route bufs of %d pages\n",
1427 rbp
->rbp_req_nbuffers
= old_req_nbufs
;
1428 lnet_net_unlock(cpt
);
1433 list_add(&rb
->rb_list
, &rb_list
);
1439 list_splice_tail(&rb_list
, &rbp
->rbp_bufs
);
1440 rbp
->rbp_nbuffers
+= num_buffers
;
1441 rbp
->rbp_credits
+= num_buffers
;
1442 rbp
->rbp_mincredits
= rbp
->rbp_credits
;
1444 * We need to schedule blocked msg using the newly
1447 while (!list_empty(&rbp
->rbp_bufs
) &&
1448 !list_empty(&rbp
->rbp_msgs
))
1449 lnet_schedule_blocked_locked(rbp
);
1451 lnet_net_unlock(cpt
);
1456 while (!list_empty(&rb_list
)) {
1457 rb
= list_entry(rb_list
.next
, lnet_rtrbuf_t
, rb_list
);
1458 list_del(&rb
->rb_list
);
1459 lnet_destroy_rtrbuf(rb
, npages
);
1466 lnet_rtrpool_init(lnet_rtrbufpool_t
*rbp
, int npages
)
1468 INIT_LIST_HEAD(&rbp
->rbp_msgs
);
1469 INIT_LIST_HEAD(&rbp
->rbp_bufs
);
1471 rbp
->rbp_npages
= npages
;
1472 rbp
->rbp_credits
= 0;
1473 rbp
->rbp_mincredits
= 0;
1477 lnet_rtrpools_free(int keep_pools
)
1479 lnet_rtrbufpool_t
*rtrp
;
1482 if (!the_lnet
.ln_rtrpools
) /* uninitialized or freed */
1485 cfs_percpt_for_each(rtrp
, i
, the_lnet
.ln_rtrpools
) {
1486 lnet_rtrpool_free_bufs(&rtrp
[LNET_TINY_BUF_IDX
], i
);
1487 lnet_rtrpool_free_bufs(&rtrp
[LNET_SMALL_BUF_IDX
], i
);
1488 lnet_rtrpool_free_bufs(&rtrp
[LNET_LARGE_BUF_IDX
], i
);
1492 cfs_percpt_free(the_lnet
.ln_rtrpools
);
1493 the_lnet
.ln_rtrpools
= NULL
;
1498 lnet_nrb_tiny_calculate(void)
1500 int nrbs
= LNET_NRB_TINY
;
1502 if (tiny_router_buffers
< 0) {
1503 LCONSOLE_ERROR_MSG(0x10c,
1504 "tiny_router_buffers=%d invalid when routing enabled\n",
1505 tiny_router_buffers
);
1509 if (tiny_router_buffers
> 0)
1510 nrbs
= tiny_router_buffers
;
1512 nrbs
/= LNET_CPT_NUMBER
;
1513 return max(nrbs
, LNET_NRB_TINY_MIN
);
1517 lnet_nrb_small_calculate(void)
1519 int nrbs
= LNET_NRB_SMALL
;
1521 if (small_router_buffers
< 0) {
1522 LCONSOLE_ERROR_MSG(0x10c,
1523 "small_router_buffers=%d invalid when routing enabled\n",
1524 small_router_buffers
);
1528 if (small_router_buffers
> 0)
1529 nrbs
= small_router_buffers
;
1531 nrbs
/= LNET_CPT_NUMBER
;
1532 return max(nrbs
, LNET_NRB_SMALL_MIN
);
1536 lnet_nrb_large_calculate(void)
1538 int nrbs
= LNET_NRB_LARGE
;
1540 if (large_router_buffers
< 0) {
1541 LCONSOLE_ERROR_MSG(0x10c,
1542 "large_router_buffers=%d invalid when routing enabled\n",
1543 large_router_buffers
);
1547 if (large_router_buffers
> 0)
1548 nrbs
= large_router_buffers
;
1550 nrbs
/= LNET_CPT_NUMBER
;
1551 return max(nrbs
, LNET_NRB_LARGE_MIN
);
1555 lnet_rtrpools_alloc(int im_a_router
)
1557 lnet_rtrbufpool_t
*rtrp
;
1564 if (!strcmp(forwarding
, "")) {
1565 /* not set either way */
1568 } else if (!strcmp(forwarding
, "disabled")) {
1569 /* explicitly disabled */
1571 } else if (!strcmp(forwarding
, "enabled")) {
1572 /* explicitly enabled */
1574 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either 'enabled' or 'disabled'\n");
1578 nrb_tiny
= lnet_nrb_tiny_calculate();
1582 nrb_small
= lnet_nrb_small_calculate();
1586 nrb_large
= lnet_nrb_large_calculate();
1590 the_lnet
.ln_rtrpools
= cfs_percpt_alloc(lnet_cpt_table(),
1592 sizeof(lnet_rtrbufpool_t
));
1593 if (!the_lnet
.ln_rtrpools
) {
1594 LCONSOLE_ERROR_MSG(0x10c,
1595 "Failed to initialize router buffe pool\n");
1599 cfs_percpt_for_each(rtrp
, i
, the_lnet
.ln_rtrpools
) {
1600 lnet_rtrpool_init(&rtrp
[LNET_TINY_BUF_IDX
], 0);
1601 rc
= lnet_rtrpool_adjust_bufs(&rtrp
[LNET_TINY_BUF_IDX
],
1606 lnet_rtrpool_init(&rtrp
[LNET_SMALL_BUF_IDX
],
1607 LNET_NRB_SMALL_PAGES
);
1608 rc
= lnet_rtrpool_adjust_bufs(&rtrp
[LNET_SMALL_BUF_IDX
],
1613 lnet_rtrpool_init(&rtrp
[LNET_LARGE_BUF_IDX
],
1614 LNET_NRB_LARGE_PAGES
);
1615 rc
= lnet_rtrpool_adjust_bufs(&rtrp
[LNET_LARGE_BUF_IDX
],
1621 lnet_net_lock(LNET_LOCK_EX
);
1622 the_lnet
.ln_routing
= 1;
1623 lnet_net_unlock(LNET_LOCK_EX
);
1628 lnet_rtrpools_free(0);
1633 lnet_rtrpools_adjust_helper(int tiny
, int small
, int large
)
1638 lnet_rtrbufpool_t
*rtrp
;
1641 * If the provided values for each buffer pool are different than the
1642 * configured values, we need to take action.
1645 tiny_router_buffers
= tiny
;
1646 nrb
= lnet_nrb_tiny_calculate();
1647 cfs_percpt_for_each(rtrp
, i
, the_lnet
.ln_rtrpools
) {
1648 rc
= lnet_rtrpool_adjust_bufs(&rtrp
[LNET_TINY_BUF_IDX
],
1655 small_router_buffers
= small
;
1656 nrb
= lnet_nrb_small_calculate();
1657 cfs_percpt_for_each(rtrp
, i
, the_lnet
.ln_rtrpools
) {
1658 rc
= lnet_rtrpool_adjust_bufs(&rtrp
[LNET_SMALL_BUF_IDX
],
1665 large_router_buffers
= large
;
1666 nrb
= lnet_nrb_large_calculate();
1667 cfs_percpt_for_each(rtrp
, i
, the_lnet
.ln_rtrpools
) {
1668 rc
= lnet_rtrpool_adjust_bufs(&rtrp
[LNET_LARGE_BUF_IDX
],
1679 lnet_rtrpools_adjust(int tiny
, int small
, int large
)
1682 * this function doesn't revert the changes if adding new buffers
1683 * failed. It's up to the user space caller to revert the
1686 if (!the_lnet
.ln_routing
)
1689 return lnet_rtrpools_adjust_helper(tiny
, small
, large
);
1693 lnet_rtrpools_enable(void)
1697 if (the_lnet
.ln_routing
)
1700 if (!the_lnet
.ln_rtrpools
)
1702 * If routing is turned off, and we have never
1703 * initialized the pools before, just call the
1704 * standard buffer pool allocation routine as
1705 * if we are just configuring this for the first
1708 return lnet_rtrpools_alloc(1);
1710 rc
= lnet_rtrpools_adjust_helper(0, 0, 0);
1714 lnet_net_lock(LNET_LOCK_EX
);
1715 the_lnet
.ln_routing
= 1;
1717 the_lnet
.ln_ping_info
->pi_features
&= ~LNET_PING_FEAT_RTE_DISABLED
;
1718 lnet_net_unlock(LNET_LOCK_EX
);
1724 lnet_rtrpools_disable(void)
1726 if (!the_lnet
.ln_routing
)
1729 lnet_net_lock(LNET_LOCK_EX
);
1730 the_lnet
.ln_routing
= 0;
1731 the_lnet
.ln_ping_info
->pi_features
|= LNET_PING_FEAT_RTE_DISABLED
;
1733 tiny_router_buffers
= 0;
1734 small_router_buffers
= 0;
1735 large_router_buffers
= 0;
1736 lnet_net_unlock(LNET_LOCK_EX
);
1737 lnet_rtrpools_free(1);
1741 lnet_notify(lnet_ni_t
*ni
, lnet_nid_t nid
, int alive
, unsigned long when
)
1743 struct lnet_peer
*lp
= NULL
;
1744 unsigned long now
= cfs_time_current();
1745 int cpt
= lnet_cpt_of_nid(nid
);
1747 LASSERT(!in_interrupt());
1749 CDEBUG(D_NET
, "%s notifying %s: %s\n",
1750 !ni
? "userspace" : libcfs_nid2str(ni
->ni_nid
),
1751 libcfs_nid2str(nid
),
1752 alive
? "up" : "down");
1755 LNET_NIDNET(ni
->ni_nid
) != LNET_NIDNET(nid
)) {
1756 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1757 libcfs_nid2str(nid
), alive
? "birth" : "death",
1758 libcfs_nid2str(ni
->ni_nid
));
1762 /* can't do predictions... */
1763 if (cfs_time_after(when
, now
)) {
1764 CWARN("Ignoring prediction from %s of %s %s %ld seconds in the future\n",
1765 !ni
? "userspace" : libcfs_nid2str(ni
->ni_nid
),
1766 libcfs_nid2str(nid
), alive
? "up" : "down",
1767 cfs_duration_sec(cfs_time_sub(when
, now
)));
1771 if (ni
&& !alive
&& /* LND telling me she's down */
1772 !auto_down
) { /* auto-down disabled */
1773 CDEBUG(D_NET
, "Auto-down disabled\n");
1779 if (the_lnet
.ln_shutdown
) {
1780 lnet_net_unlock(cpt
);
1784 lp
= lnet_find_peer_locked(the_lnet
.ln_peer_tables
[cpt
], nid
);
1787 lnet_net_unlock(cpt
);
1788 CDEBUG(D_NET
, "%s not found\n", libcfs_nid2str(nid
));
1793 * We can't fully trust LND on reporting exact peer last_alive
1794 * if he notifies us about dead peer. For example ksocklnd can
1795 * call us with when == _time_when_the_node_was_booted_ if
1796 * no connections were successfully established
1798 if (ni
&& !alive
&& when
< lp
->lp_last_alive
)
1799 when
= lp
->lp_last_alive
;
1801 lnet_notify_locked(lp
, !ni
, alive
, when
);
1804 lnet_ni_notify_locked(ni
, lp
);
1806 lnet_peer_decref_locked(lp
);
1808 lnet_net_unlock(cpt
);
1811 EXPORT_SYMBOL(lnet_notify
);