80d578ad00d30a133df0ee9c90aef3f3a513365f
[dragonfly.git] / sbin / hammer2 / msg_lnk.c
1 /*
2  * Copyright (c) 2012 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@dragonflybsd.org>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 /*
35  * LNK_SPAN PROTOCOL SUPPORT FUNCTIONS
36  *
37  * This code supports the LNK_SPAN protocol.  Essentially all PFS's
38  * clients and services rendezvous with the userland hammer2 service and
39  * open LNK_SPAN transactions using a message header linkid of 0,
40  * registering any PFS's they have connectivity to with us.
41  *
42  * --
43  *
44  * Each registration maintains its own open LNK_SPAN message transaction.
45  * The SPANs are collected, aggregated, and retransmitted over available
46  * connections through the maintainance of additional LNK_SPAN message
47  * transactions on each link.
48  *
49  * The msgid for each active LNK_SPAN transaction we receive allows us to
50  * send a message to the target PFS (which might be one of many belonging
51  * to the same cluster), by specifying that msgid as the linkid in any
52  * message we send to the target PFS.
53  *
54  * Similarly the msgid we allocate for any LNK_SPAN transaction we transmit
55  * (and remember we will maintain multiple open LNK_SPAN transactions on
56  * each connection representing the topology span, so every node sees every
57  * other node as a separate open transaction).  So, similarly the msgid for
58  * these active transactions which we initiated can be used by the other
59  * end to route messages through us to another node, ultimately winding up
60  * at the identified hammer2 PFS.  We have to adjust the spanid in the message
61  * header at each hop to be representative of the outgoing LNK_SPAN we
62  * are forwarding the message through.
63  *
64  * --
65  *
66  * If we were to retransmit every LNK_SPAN transaction we receive it would
67  * create a huge mess, so we have to aggregate all received LNK_SPAN
68  * transactions, sort them by the fsid (the cluster) and sub-sort them by
69  * the pfs_fsid (individual nodes in the cluster), and only retransmit
70  * (create outgoing transactions) for a subset of the nearest distance-hops
71  * for each individual node.
72  *
73  * The higher level protocols can then issue transactions to the nodes making
74  * up a cluster to perform all actions required.
75  *
76  * --
77  *
78  * Since this is a large topology and a spanning tree protocol, links can
79  * go up and down all the time.  Any time a link goes down its transaction
80  * is closed.  The transaction has to be closed on both ends before we can
81  * delete (and potentially reuse) the related spanid.  The LNK_SPAN being
82  * closed may have been propagated out to other connections and those related
83  * LNK_SPANs are also closed.  Ultimately all routes via the lost LNK_SPAN
84  * go away, ultimately reaching all sources and all targets.
85  *
86  * Any messages in-transit using a route that goes away will be thrown away.
87  * Open transactions are only tracked at the two end-points.  When a link
88  * failure propagates to an end-point the related open transactions lose
89  * their spanid and are automatically aborted.
90  *
91  * It is important to note that internal route nodes cannot just associate
92  * a lost LNK_SPAN transaction with another route to the same destination.
93  * Message transactions MUST be serialized and MUST be ordered.  All messages
94  * for a transaction must run over the same route.  So if the route used by
95  * an active transaction is lost, the related messages will be fully aborted
96  * and the higher protocol levels will retry as appropriate.
97  *
98  * It is also important to note that several paths to the same PFS can be
99  * propagated along the same link, which allows concurrency and even
100  * redundancy over several network interfaces or via different routes through
101  * the topology.  Any given transaction will use only a single route but busy
102  * servers will often have hundreds of transactions active simultaniously,
103  * so having multiple active paths through the network topology for A<->B
104  * will improve performance.
105  *
106  * --
107  *
108  * Most protocols consolidate operations rather than simply relaying them.
109  * This is particularly true of LEAF protocols (such as strict HAMMER2
110  * clients), of which there can be millions connecting into the cluster at
111  * various points.  The SPAN protocol is not used for these LEAF elements.
112  *
113  * Instead the primary service they connect to implements a proxy for the
114  * client protocols so the core topology only has to propagate a couple of
115  * LNK_SPANs and not millions.  LNK_SPANs are meant to be used only for
116  * core master nodes and satellite slaves and cache nodes.
117  */
118
119 #include "hammer2.h"
120
121 /*
122  * RED-BLACK TREE DEFINITIONS
123  *
124  * We need to track:
125  *
126  * (1) shared fsid's (a cluster).
127  * (2) unique fsid's (a node in a cluster) <--- LNK_SPAN transactions.
128  *
129  * We need to aggegate all active LNK_SPANs, aggregate, and create our own
130  * outgoing LNK_SPAN transactions on each of our connections representing
131  * the aggregated state.
132  *
133  * h2span_connect       - list of iocom connections who wish to receive SPAN
134  *                        propagation from other connections.  Might contain
135  *                        a filter string.  Only iocom's with an open
136  *                        LNK_CONN transactions are applicable for SPAN
137  *                        propagation.
138  *
139  * h2span_relay         - List of links relayed (via SPAN).  Essentially
140  *                        each relay structure represents a LNK_SPAN
141  *                        transaction that we initiated, verses h2span_link
142  *                        which is a LNK_SPAN transaction that we received.
143  *
144  * --
145  *
146  * h2span_cluster       - Organizes the shared fsid's.  One structure for
147  *                        each cluster.
148  *
149  * h2span_node          - Organizes the nodes in a cluster.  One structure
150  *                        for each unique {cluster,node}, aka {fsid, pfs_fsid}.
151  *
152  * h2span_link          - Organizes all incoming and outgoing LNK_SPAN message
153  *                        transactions related to a node.
154  *
155  *                        One h2span_link structure for each incoming LNK_SPAN
156  *                        transaction.  Links selected for propagation back
157  *                        out are also where the outgoing LNK_SPAN messages
158  *                        are indexed into (so we can propagate changes).
159  *
160  *                        The h2span_link's use a red-black tree to sort the
161  *                        distance hop metric for the incoming LNK_SPAN.  We
162  *                        then select the top N for outgoing.  When the
163  *                        topology changes the top N may also change and cause
164  *                        new outgoing LNK_SPAN transactions to be opened
165  *                        and less desireable ones to be closed, causing
166  *                        transactional aborts within the message flow in
167  *                        the process.
168  *
169  * Also note            - All outgoing LNK_SPAN message transactions are also
170  *                        entered into a red-black tree for use by the routing
171  *                        function.  This is handled by msg.c in the state
172  *                        code, not here.
173  */
174
175 struct h2span_link;
176 struct h2span_relay;
177 TAILQ_HEAD(h2span_connect_queue, h2span_connect);
178 TAILQ_HEAD(h2span_relay_queue, h2span_relay);
179
180 RB_HEAD(h2span_cluster_tree, h2span_cluster);
181 RB_HEAD(h2span_node_tree, h2span_node);
182 RB_HEAD(h2span_link_tree, h2span_link);
183 RB_HEAD(h2span_relay_tree, h2span_relay);
184
185 /*
186  * Received LNK_CONN transaction enables SPAN protocol over connection.
187  * (may contain filter).
188  */
189 struct h2span_connect {
190         TAILQ_ENTRY(h2span_connect) entry;
191         struct h2span_relay_tree tree;
192         hammer2_state_t *state;
193 };
194
195 /*
196  * All received LNK_SPANs are organized by cluster (pfs_clid),
197  * node (pfs_fsid), and link (received LNK_SPAN transaction).
198  */
199 struct h2span_cluster {
200         RB_ENTRY(h2span_cluster) rbnode;
201         struct h2span_node_tree tree;
202         uuid_t  pfs_clid;               /* shared fsid */
203 };
204
205 struct h2span_node {
206         RB_ENTRY(h2span_node) rbnode;
207         struct h2span_link_tree tree;
208         struct h2span_cluster *cls;
209         uuid_t  pfs_fsid;               /* unique fsid */
210 };
211
212 struct h2span_link {
213         RB_ENTRY(h2span_link) rbnode;
214         hammer2_state_t *state;         /* state<->link */
215         struct h2span_node *node;       /* related node */
216         int32_t dist;
217         struct h2span_relay_queue relayq; /* relay out */
218 };
219
220 /*
221  * Any LNK_SPAN transactions we receive which are relayed out other
222  * connections utilize this structure to track the LNK_SPAN transaction
223  * we initiate on the other connections, if selected for relay.
224  *
225  * In many respects this is the core of the protocol... actually figuring
226  * out what LNK_SPANs to relay.  The spanid used for relaying is the
227  * address of the 'state' structure, which is why h2span_relay has to
228  * be entered into a RB-TREE based at h2span_connect (so we can look
229  * up the spanid to validate it).
230  */
231 struct h2span_relay {
232         RB_ENTRY(h2span_relay) rbnode;  /* from h2span_connect */
233         TAILQ_ENTRY(h2span_relay) entry; /* from link */
234         struct h2span_connect *conn;
235         hammer2_state_t *state;         /* transmitted LNK_SPAN */
236         struct h2span_link *link;       /* received LNK_SPAN */
237 };
238
239
240 typedef struct h2span_connect h2span_connect_t;
241 typedef struct h2span_cluster h2span_cluster_t;
242 typedef struct h2span_node h2span_node_t;
243 typedef struct h2span_link h2span_link_t;
244 typedef struct h2span_relay h2span_relay_t;
245
246 static
247 int
248 h2span_cluster_cmp(h2span_cluster_t *cls1, h2span_cluster_t *cls2)
249 {
250         return(uuid_compare(&cls1->pfs_clid, &cls2->pfs_clid, NULL));
251 }
252
253 static
254 int
255 h2span_node_cmp(h2span_node_t *node1, h2span_node_t *node2)
256 {
257         return(uuid_compare(&node1->pfs_fsid, &node2->pfs_fsid, NULL));
258 }
259
260 static
261 int
262 h2span_link_cmp(h2span_link_t *link1, h2span_link_t *link2)
263 {
264         if (link1->dist < link2->dist)
265                 return(-1);
266         if (link1->dist > link2->dist)
267                 return(1);
268         if ((intptr_t)link1 < (intptr_t)link2)
269                 return(-1);
270         if ((intptr_t)link1 > (intptr_t)link2)
271                 return(1);
272         return(0);
273 }
274
275 /*
276  * Relay entries are sorted by node, subsorted by distance and link
277  * address (so we can match up the conn->tree relay topology with
278  * a node's link topology).
279  */
280 static
281 int
282 h2span_relay_cmp(h2span_relay_t *relay1, h2span_relay_t *relay2)
283 {
284         if ((intptr_t)relay1->link->node < (intptr_t)relay2->link->node)
285                 return(-1);
286         if ((intptr_t)relay1->link->node > (intptr_t)relay2->link->node)
287                 return(1);
288         if ((intptr_t)relay1->link->dist < (intptr_t)relay2->link->dist)
289                 return(-1);
290         if ((intptr_t)relay1->link->dist > (intptr_t)relay2->link->dist)
291                 return(1);
292         if ((intptr_t)relay1->link < (intptr_t)relay2->link)
293                 return(-1);
294         if ((intptr_t)relay1->link > (intptr_t)relay2->link)
295                 return(1);
296         return(0);
297 }
298
299 RB_PROTOTYPE_STATIC(h2span_cluster_tree, h2span_cluster,
300              rbnode, h2span_cluster_cmp);
301 RB_PROTOTYPE_STATIC(h2span_node_tree, h2span_node,
302              rbnode, h2span_node_cmp);
303 RB_PROTOTYPE_STATIC(h2span_link_tree, h2span_link,
304              rbnode, h2span_link_cmp);
305 RB_PROTOTYPE_STATIC(h2span_relay_tree, h2span_relay,
306              rbnode, h2span_relay_cmp);
307
308 RB_GENERATE_STATIC(h2span_cluster_tree, h2span_cluster,
309              rbnode, h2span_cluster_cmp);
310 RB_GENERATE_STATIC(h2span_node_tree, h2span_node,
311              rbnode, h2span_node_cmp);
312 RB_GENERATE_STATIC(h2span_link_tree, h2span_link,
313              rbnode, h2span_link_cmp);
314 RB_GENERATE_STATIC(h2span_relay_tree, h2span_relay,
315              rbnode, h2span_relay_cmp);
316
317 /*
318  * Global mutex protects cluster_tree lookups.
319  */
320 static pthread_mutex_t cluster_mtx;
321 static struct h2span_cluster_tree cluster_tree = RB_INITIALIZER(cluster_tree);
322 static struct h2span_connect_queue connq = TAILQ_HEAD_INITIALIZER(connq);
323
324 static void hammer2_lnk_span(hammer2_state_t *state, hammer2_msg_t *msg);
325 static void hammer2_lnk_conn(hammer2_state_t *state, hammer2_msg_t *msg);
326 static void hammer2_lnk_relay(hammer2_state_t *state, hammer2_msg_t *msg);
327 static void hammer2_relay_scan(h2span_connect_t *conn, h2span_node_t *node);
328 static void hammer2_relay_delete(h2span_relay_t *relay);
329
330 /*
331  * Receive a HAMMER2_MSG_PROTO_LNK message.  This only called for
332  * one-way and opening-transactions since state->func will be assigned
333  * in all other cases.
334  */
335 void
336 hammer2_msg_lnk(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
337 {
338         switch(msg->any.head.cmd & HAMMER2_MSGF_BASECMDMASK) {
339         case HAMMER2_LNK_CONN:
340                 hammer2_lnk_conn(msg->state, msg);
341                 break;
342         case HAMMER2_LNK_SPAN:
343                 hammer2_lnk_span(msg->state, msg);
344                 break;
345         default:
346                 fprintf(stderr,
347                         "MSG_PROTO_LNK: Unknown msg %08x\n", msg->any.head.cmd);
348                 hammer2_msg_reply(iocom, msg, HAMMER2_MSG_ERR_UNKNOWN);
349                 /* state invalid after reply */
350                 break;
351         }
352 }
353
354 void
355 hammer2_lnk_conn(hammer2_state_t *state, hammer2_msg_t *msg)
356 {
357         h2span_connect_t *conn;
358         h2span_relay_t *relay;
359         char *alloc = NULL;
360
361         pthread_mutex_lock(&cluster_mtx);
362
363         /*
364          * On transaction start we allocate a new h2span_connect and
365          * acknowledge the request, leaving the transaction open.
366          * We then relay priority-selected SPANs.
367          */
368         if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
369                 state->func = hammer2_lnk_conn;
370
371                 fprintf(stderr, "LNK_CONN(%016jx): %s/%s\n",
372                         (intmax_t)msg->any.head.msgid,
373                         hammer2_uuid_to_str(&msg->any.lnk_conn.pfs_clid,
374                                             &alloc),
375                         msg->any.lnk_conn.label);
376                 free(alloc);
377
378                 conn = hammer2_alloc(sizeof(*conn));
379
380                 RB_INIT(&conn->tree);
381                 conn->state = state;
382                 state->any.conn = conn;
383                 TAILQ_INSERT_TAIL(&connq, conn, entry);
384
385                 hammer2_msg_result(state->iocom, msg, 0);
386
387                 /*
388                  * Span-synchronize all nodes with the new connection
389                  */
390                 hammer2_relay_scan(conn, NULL);
391         }
392
393         /*
394          * On transaction terminate we clean out our h2span_connect
395          * and acknowledge the request, closing the transaction.
396          */
397         if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
398                 fprintf(stderr, "LNK_CONN: Terminated\n");
399                 conn = state->any.conn;
400                 assert(conn);
401
402                 /*
403                  * Clean out all relays.  This requires terminating each
404                  * relay transaction.
405                  */
406                 while ((relay = RB_ROOT(&conn->tree)) != NULL) {
407                         hammer2_relay_delete(relay);
408                 }
409
410                 /*
411                  * Clean out conn
412                  */
413                 conn->state = NULL;
414                 msg->state->any.conn = NULL;
415                 TAILQ_REMOVE(&connq, conn, entry);
416                 hammer2_free(conn);
417
418                 hammer2_msg_reply(state->iocom, msg, 0);
419                 /* state invalid after reply */
420         }
421         pthread_mutex_unlock(&cluster_mtx);
422 }
423
424 void
425 hammer2_lnk_span(hammer2_state_t *state, hammer2_msg_t *msg)
426 {
427         h2span_cluster_t dummy_cls;
428         h2span_node_t dummy_node;
429         h2span_cluster_t *cls;
430         h2span_node_t *node;
431         h2span_link_t *slink;
432         h2span_relay_t *relay;
433         char *alloc = NULL;
434
435         pthread_mutex_lock(&cluster_mtx);
436
437         /*
438          * On transaction start we initialize the tracking infrastructure
439          */
440         if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
441                 state->func = hammer2_lnk_span;
442
443                 fprintf(stderr, "LNK_SPAN: %s/%s\n",
444                         hammer2_uuid_to_str(&msg->any.lnk_span.pfs_clid,
445                                             &alloc),
446                         msg->any.lnk_span.label);
447                 free(alloc);
448
449                 /*
450                  * Find the cluster
451                  */
452                 dummy_cls.pfs_clid = msg->any.lnk_span.pfs_clid;
453                 cls = RB_FIND(h2span_cluster_tree, &cluster_tree, &dummy_cls);
454                 if (cls == NULL) {
455                         cls = hammer2_alloc(sizeof(*cls));
456                         cls->pfs_clid = msg->any.lnk_span.pfs_clid;
457                         RB_INIT(&cls->tree);
458                         RB_INSERT(h2span_cluster_tree, &cluster_tree, cls);
459                 }
460
461                 /*
462                  * Find the node
463                  */
464                 dummy_node.pfs_fsid = msg->any.lnk_span.pfs_fsid;
465                 node = RB_FIND(h2span_node_tree, &cls->tree, &dummy_node);
466                 if (node == NULL) {
467                         node = hammer2_alloc(sizeof(*node));
468                         node->pfs_fsid = msg->any.lnk_span.pfs_fsid;
469                         node->cls = cls;
470                         RB_INIT(&node->tree);
471                         RB_INSERT(h2span_node_tree, &cls->tree, node);
472                 }
473
474                 /*
475                  * Create the link
476                  */
477                 assert(state->any.link == NULL);
478                 slink = hammer2_alloc(sizeof(*slink));
479                 TAILQ_INIT(&slink->relayq);
480                 slink->node = node;
481                 slink->dist = msg->any.lnk_span.dist;
482                 slink->state = state;
483                 state->any.link = slink;
484                 RB_INSERT(h2span_link_tree, &node->tree, slink);
485
486                 hammer2_relay_scan(NULL, node);
487         }
488
489         /*
490          * On transaction terminate we remove the tracking infrastructure.
491          */
492         if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
493                 slink = state->any.link;
494                 assert(slink != NULL);
495                 node = slink->node;
496                 cls = node->cls;
497
498                 /*
499                  * Clean out all relays.  This requires terminating each
500                  * relay transaction.
501                  */
502                 while ((relay = TAILQ_FIRST(&slink->relayq)) != NULL) {
503                         hammer2_relay_delete(relay);
504                 }
505
506                 /*
507                  * Clean out the topology
508                  */
509                 RB_REMOVE(h2span_link_tree, &node->tree, slink);
510                 if (RB_EMPTY(&node->tree)) {
511                         RB_REMOVE(h2span_node_tree, &cls->tree, node);
512                         if (RB_EMPTY(&cls->tree)) {
513                                 RB_REMOVE(h2span_cluster_tree,
514                                           &cluster_tree, cls);
515                                 hammer2_free(cls);
516                         }
517                         node->cls = NULL;
518                         hammer2_free(node);
519                         node = NULL;
520                 }
521                 state->any.link = NULL;
522                 slink->state = NULL;
523                 slink->node = NULL;
524                 hammer2_free(slink);
525
526                 /*
527                  * We have to terminate the transaction
528                  */
529                 hammer2_state_reply(state, 0);
530                 /* state invalid after reply */
531
532                 /*
533                  * If the node still exists issue any required updates.  If
534                  * it doesn't then all related relays have already been
535                  * removed and there's nothing left to do.
536                  */
537                 if (node)
538                         hammer2_relay_scan(NULL, node);
539         }
540
541         pthread_mutex_unlock(&cluster_mtx);
542 }
543
544 /*
545  * Messages received on relay SPANs.  These are open transactions so it is
546  * in fact possible for the other end to close the transaction.
547  *
548  * XXX MPRACE on state structure
549  */
550 static void
551 hammer2_lnk_relay(hammer2_state_t *state, hammer2_msg_t *msg)
552 {
553         h2span_relay_t *relay;
554
555         if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
556                 pthread_mutex_lock(&cluster_mtx);
557                 if ((relay = state->any.relay) != NULL) {
558                         hammer2_relay_delete(relay);
559                 } else {
560                         hammer2_state_reply(state, 0);
561                 }
562                 pthread_mutex_unlock(&cluster_mtx);
563         }
564 }
565
566 /*
567  * Update relay transactions for SPANs.
568  *
569  * Called with cluster_mtx held.
570  */
571 static void hammer2_relay_scan_specific(h2span_node_t *node,
572                                         h2span_connect_t *conn);
573
574 static void
575 hammer2_relay_scan(h2span_connect_t *conn, h2span_node_t *node)
576 {
577         h2span_cluster_t *cls;
578
579         if (node) {
580                 /*
581                  * Iterate specific node
582                  */
583                 TAILQ_FOREACH(conn, &connq, entry)
584                         hammer2_relay_scan_specific(node, conn);
585         } else {
586                 /*
587                  * Full iteration.
588                  *
589                  * Iterate cluster ids, nodes, and either a specific connection
590                  * or all connections.
591                  */
592                 RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
593                         /*
594                          * Iterate node ids
595                          */
596                         RB_FOREACH(node, h2span_node_tree, &cls->tree) {
597                                 /*
598                                  * Synchronize the node's link (received SPANs)
599                                  * with each connection's relays.
600                                  */
601                                 if (conn) {
602                                         hammer2_relay_scan_specific(node, conn);
603                                 } else {
604                                         TAILQ_FOREACH(conn, &connq, entry) {
605                                             hammer2_relay_scan_specific(node,
606                                                                         conn);
607                                         }
608                                         assert(conn == NULL);
609                                 }
610                         }
611                 }
612         }
613 }
614
615 /*
616  * Update the relay'd SPANs for this (node, conn).
617  *
618  * Iterate links and adjust relays to match.  We only propagate the top link
619  * for now (XXX we want to propagate the top two).
620  *
621  * The hammer2_relay_scan_cmp() function locates the first relay element
622  * for any given node.  The relay elements will be sub-sorted by dist.
623  */
624 struct relay_scan_info {
625         h2span_node_t *node;
626         h2span_relay_t *relay;
627 };
628
629 static int
630 hammer2_relay_scan_cmp(h2span_relay_t *relay, void *arg)
631 {
632         struct relay_scan_info *info = arg;
633
634         if ((intptr_t)relay->link->node < (intptr_t)info->node)
635                 return(-1);
636         if ((intptr_t)relay->link->node > (intptr_t)info->node)
637                 return(1);
638         return(0);
639 }
640
641 static int
642 hammer2_relay_scan_callback(h2span_relay_t *relay, void *arg)
643 {
644         struct relay_scan_info *info = arg;
645
646         info->relay = relay;
647         return(-1);
648 }
649
650 static void
651 hammer2_relay_scan_specific(h2span_node_t *node, h2span_connect_t *conn)
652 {
653         struct relay_scan_info info;
654         h2span_relay_t *relay;
655         h2span_relay_t *next_relay;
656         h2span_link_t *slink;
657         int count = 2;
658
659         info.node = node;
660         info.relay = NULL;
661
662         /*
663          * Locate the first related relay for the connection.  relay will
664          * be NULL if there were none.
665          */
666         RB_SCAN(h2span_relay_tree, &conn->tree,
667                 hammer2_relay_scan_cmp, hammer2_relay_scan_callback, &info);
668         relay = info.relay;
669
670         fprintf(stderr, "relay scan for connection %p\n", conn);
671
672         /*
673          * Iterate the node's links (received SPANs) in distance order,
674          * lowest (best) dist first.
675          */
676         RB_FOREACH(slink, h2span_link_tree, &node->tree) {
677                 /*
678                  * PROPAGATE THE BEST RELAYS BY TRANSMITTING SPANs.
679                  *
680                  * Check for match against current best relay.
681                  *
682                  * A match failure means that the current best relay is not
683                  * as good as the link, create a new relay for the link.
684                  *
685                  * (If some prior better link was removed it would have also
686                  *  removed the relay, so the relay can only match exactly or
687                  *  be worst).
688                  */
689                 info.relay = relay;
690                 if (relay == NULL || relay->link != slink) {
691                         hammer2_msg_t *msg;
692
693                         assert(relay == NULL ||
694                                relay->link->dist <= slink->dist);
695                         relay = hammer2_alloc(sizeof(*relay));
696                         relay->conn = conn;
697                         relay->link = slink;
698
699                         msg = hammer2_msg_alloc(conn->state->iocom, 0,
700                                                 HAMMER2_LNK_SPAN |
701                                                 HAMMER2_MSGF_CREATE);
702                         msg->any.lnk_span = slink->state->msg->any.lnk_span;
703                         ++msg->any.lnk_span.dist; /* XXX add weighting */
704
705                         hammer2_msg_write(conn->state->iocom, msg,
706                                           hammer2_lnk_relay, relay,
707                                           &relay->state);
708                         fprintf(stderr, "RELAY SPAN ON CLS=%p NODE=%p FD %d state %p\n",
709                                 node->cls, node,
710                                 conn->state->iocom->sock_fd, relay->state);
711
712                         RB_INSERT(h2span_relay_tree, &conn->tree, relay);
713                         TAILQ_INSERT_TAIL(&slink->relayq, relay, entry);
714                 }
715
716                 /*
717                  * Iterate, figure out the next relay.
718                  */
719                 relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
720                 if (--count == 0) {
721                         break;
722                         continue;
723                 }
724         }
725
726         /*
727          * Any remaining relay's belonging to this connection which match
728          * the node are in excess of the current aggregate spanning state
729          * and should be removed.
730          */
731         while (relay && relay->link->node == node) {
732                 next_relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
733                 hammer2_relay_delete(relay);
734                 relay = next_relay;
735         }
736 }
737
738 static
739 void
740 hammer2_relay_delete(h2span_relay_t *relay)
741 {
742         fprintf(stderr, "RELAY DELETE ON CLS=%p NODE=%p FD %d STATE %p\n",
743                 relay->link->node->cls, relay->link->node,
744                 relay->conn->state->iocom->sock_fd, relay->state);
745         fprintf(stderr, "RELAY TX %08x RX %08x\n", relay->state->txcmd, relay->state->rxcmd);
746
747         RB_REMOVE(h2span_relay_tree, &relay->conn->tree, relay);
748         TAILQ_REMOVE(&relay->link->relayq, relay, entry);
749
750         if (relay->state) {
751                 relay->state->any.relay = NULL;
752                 hammer2_state_reply(relay->state, 0);
753                 /* state invalid after reply */
754                 relay->state = NULL;
755         }
756         relay->conn = NULL;
757         relay->link = NULL;
758         hammer2_free(relay);
759 }