lib/libdmsg/msg_lnk.c

   1 /*
   2  * Copyright (c) 2012-2014 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 - Please see sys/dmsg.h for an
  36  * involved explanation of the protocol.
  37  */
  38
  39 #include "dmsg_local.h"
  40
  41 /*
  42  * Maximum spanning tree distance.  This has the practical effect of
  43  * stopping tail-chasing closed loops when a feeder span is lost.
  44  */
  45 #define DMSG_SPAN_MAXDIST       16
  46
  47 /*
  48  * RED-BLACK TREE DEFINITIONS
  49  *
  50  * We need to track:
  51  *
  52  * (1) shared fsid's (a cluster).
  53  * (2) unique fsid's (a node in a cluster) <--- LNK_SPAN transactions.
  54  *
  55  * We need to aggegate all active LNK_SPANs, aggregate, and create our own
  56  * outgoing LNK_SPAN transactions on each of our connections representing
  57  * the aggregated state.
  58  *
  59  * h2span_conn          - list of iocom connections who wish to receive SPAN
  60  *                        propagation from other connections.  Might contain
  61  *                        a filter string.  Only iocom's with an open
  62  *                        LNK_CONN transactions are applicable for SPAN
  63  *                        propagation.
  64  *
  65  * h2span_relay         - List of links relayed (via SPAN).  Essentially
  66  *                        each relay structure represents a LNK_SPAN
  67  *                        transaction that we initiated, verses h2span_link
  68  *                        which is a LNK_SPAN transaction that we received.
  69  *
  70  * --
  71  *
  72  * h2span_cluster       - Organizes the shared fsid's.  One structure for
  73  *                        each cluster.
  74  *
  75  * h2span_node          - Organizes the nodes in a cluster.  One structure
  76  *                        for each unique {cluster,node}, aka {fsid, pfs_fsid}.
  77  *
  78  * h2span_link          - Organizes all incoming and outgoing LNK_SPAN message
  79  *                        transactions related to a node.
  80  *
  81  *                        One h2span_link structure for each incoming LNK_SPAN
  82  *                        transaction.  Links selected for propagation back
  83  *                        out are also where the outgoing LNK_SPAN messages
  84  *                        are indexed into (so we can propagate changes).
  85  *
  86  *                        The h2span_link's use a red-black tree to sort the
  87  *                        distance hop metric for the incoming LNK_SPAN.  We
  88  *                        then select the top N for outgoing.  When the
  89  *                        topology changes the top N may also change and cause
  90  *                        new outgoing LNK_SPAN transactions to be opened
  91  *                        and less desireable ones to be closed, causing
  92  *                        transactional aborts within the message flow in
  93  *                        the process.
  94  *
  95  * Also note            - All outgoing LNK_SPAN message transactions are also
  96  *                        entered into a red-black tree for use by the routing
  97  *                        function.  This is handled by msg.c in the state
  98  *                        code, not here.
  99  */
 100
 101 struct h2span_link;
 102 struct h2span_relay;
 103 TAILQ_HEAD(h2span_conn_queue, h2span_conn);
 104 TAILQ_HEAD(h2span_relay_queue, h2span_relay);
 105
 106 RB_HEAD(h2span_cluster_tree, h2span_cluster);
 107 RB_HEAD(h2span_node_tree, h2span_node);
 108 RB_HEAD(h2span_link_tree, h2span_link);
 109 RB_HEAD(h2span_relay_tree, h2span_relay);
 110 uint32_t DMsgRNSS;
 111
 112 /*
 113  * Received LNK_CONN transaction enables SPAN protocol over connection.
 114  * (may contain filter).  Typically one for each mount and several may
 115  * share the same media.
 116  */
 117 struct h2span_conn {
 118         TAILQ_ENTRY(h2span_conn) entry;
 119         struct h2span_relay_tree tree;
 120         dmsg_state_t *state;
 121         dmsg_lnk_conn_t lnk_conn;
 122 };
 123
 124 /*
 125  * All received LNK_SPANs are organized by cluster (pfs_clid),
 126  * node (pfs_fsid), and link (received LNK_SPAN transaction).
 127  */
 128 struct h2span_cluster {
 129         RB_ENTRY(h2span_cluster) rbnode;
 130         struct h2span_node_tree tree;
 131         uuid_t  pfs_clid;               /* shared fsid */
 132         uint8_t peer_type;
 133         char    cl_label[128];          /* cluster label (typ PEER_BLOCK) */
 134         int     refs;                   /* prevents destruction */
 135 };
 136
 137 struct h2span_node {
 138         RB_ENTRY(h2span_node) rbnode;
 139         struct h2span_link_tree tree;
 140         struct h2span_cluster *cls;
 141         uint8_t pfs_type;
 142         uuid_t  pfs_fsid;               /* unique fsid */
 143         char    fs_label[128];          /* fs label (typ PEER_HAMMER2) */
 144         void    *opaque;
 145 };
 146
 147 struct h2span_link {
 148         RB_ENTRY(h2span_link) rbnode;
 149         dmsg_state_t    *state;         /* state<->link */
 150         struct h2span_node *node;       /* related node */
 151         struct h2span_relay_queue relayq; /* relay out */
 152         dmsg_lnk_span_t lnk_span;
 153 };
 154
 155 /*
 156  * Any LNK_SPAN transactions we receive which are relayed out other
 157  * connections utilize this structure to track the LNK_SPAN transactions
 158  * we initiate (relay out) on other connections.  We only relay out
 159  * LNK_SPANs on connections we have an open CONN transaction for.
 160  *
 161  * The relay structure points to the outgoing LNK_SPAN trans (out_state)
 162  * and to the incoming LNK_SPAN transaction (in_state).  The relay
 163  * structure holds refs on the related states.
 164  *
 165  * In many respects this is the core of the protocol... actually figuring
 166  * out what LNK_SPANs to relay.  The spanid used for relaying is the
 167  * address of the 'state' structure, which is why h2span_relay has to
 168  * be entered into a RB-TREE based at h2span_conn (so we can look
 169  * up the spanid to validate it).
 170  */
 171 struct h2span_relay {
 172         TAILQ_ENTRY(h2span_relay) entry;        /* from link */
 173         RB_ENTRY(h2span_relay) rbnode;          /* from h2span_conn */
 174         struct h2span_conn      *conn;          /* related CONN transaction */
 175         dmsg_state_t            *source_rt;     /* h2span_link state */
 176         dmsg_state_t            *target_rt;     /* h2span_relay state */
 177 };
 178
 179 typedef struct h2span_conn h2span_conn_t;
 180 typedef struct h2span_cluster h2span_cluster_t;
 181 typedef struct h2span_node h2span_node_t;
 182 typedef struct h2span_link h2span_link_t;
 183 typedef struct h2span_relay h2span_relay_t;
 184
 185 #define dmsg_termstr(array)     _dmsg_termstr((array), sizeof(array))
 186
 187 static h2span_relay_t *dmsg_generate_relay(h2span_conn_t *conn,
 188                                         h2span_link_t *slink);
 189 static uint32_t dmsg_rnss(void);
 190
 191 static __inline
 192 void
 193 _dmsg_termstr(char *base, size_t size)
 194 {
 195         base[size-1] = 0;
 196 }
 197
 198 /*
 199  * Cluster peer_type, uuid, AND label must match for a match
 200  */
 201 static
 202 int
 203 h2span_cluster_cmp(h2span_cluster_t *cls1, h2span_cluster_t *cls2)
 204 {
 205         int r;
 206
 207         if (cls1->peer_type < cls2->peer_type)
 208                 return(-1);
 209         if (cls1->peer_type > cls2->peer_type)
 210                 return(1);
 211         r = uuid_compare(&cls1->pfs_clid, &cls2->pfs_clid, NULL);
 212         if (r == 0)
 213                 r = strcmp(cls1->cl_label, cls2->cl_label);
 214
 215         return r;
 216 }
 217
 218 /*
 219  * Match against fs_label/pfs_fsid.  Together these two items represent a
 220  * unique node.  In most cases the primary differentiator is pfs_fsid but
 221  * we also string-match fs_label.
 222  */
 223 static
 224 int
 225 h2span_node_cmp(h2span_node_t *node1, h2span_node_t *node2)
 226 {
 227         int r;
 228
 229         r = strcmp(node1->fs_label, node2->fs_label);
 230         if (r == 0)
 231                 r = uuid_compare(&node1->pfs_fsid, &node2->pfs_fsid, NULL);
 232         return (r);
 233 }
 234
 235 /*
 236  * Sort/subsort must match h2span_relay_cmp() under any given node
 237  * to make the aggregation algorithm easier, so the best links are
 238  * in the same sorted order as the best relays.
 239  *
 240  * NOTE: We cannot use link*->state->msgid because this msgid is created
 241  *       by each remote host and thus might wind up being the same.
 242  */
 243 static
 244 int
 245 h2span_link_cmp(h2span_link_t *link1, h2span_link_t *link2)
 246 {
 247         if (link1->lnk_span.dist < link2->lnk_span.dist)
 248                 return(-1);
 249         if (link1->lnk_span.dist > link2->lnk_span.dist)
 250                 return(1);
 251         if (link1->lnk_span.rnss < link2->lnk_span.rnss)
 252                 return(-1);
 253         if (link1->lnk_span.rnss > link2->lnk_span.rnss)
 254                 return(1);
 255 #if 1
 256         if ((uintptr_t)link1->state < (uintptr_t)link2->state)
 257                 return(-1);
 258         if ((uintptr_t)link1->state > (uintptr_t)link2->state)
 259                 return(1);
 260 #else
 261         if (link1->state->msgid < link2->state->msgid)
 262                 return(-1);
 263         if (link1->state->msgid > link2->state->msgid)
 264                 return(1);
 265 #endif
 266         return(0);
 267 }
 268
 269 /*
 270  * Relay entries are sorted by node, subsorted by distance and link
 271  * address (so we can match up the conn->tree relay topology with
 272  * a node's link topology).
 273  */
 274 static
 275 int
 276 h2span_relay_cmp(h2span_relay_t *relay1, h2span_relay_t *relay2)
 277 {
 278         h2span_link_t *link1 = relay1->source_rt->any.link;
 279         h2span_link_t *link2 = relay2->source_rt->any.link;
 280
 281         if ((intptr_t)link1->node < (intptr_t)link2->node)
 282                 return(-1);
 283         if ((intptr_t)link1->node > (intptr_t)link2->node)
 284                 return(1);
 285         if (link1->lnk_span.dist < link2->lnk_span.dist)
 286                 return(-1);
 287         if (link1->lnk_span.dist > link2->lnk_span.dist)
 288                 return(1);
 289         if (link1->lnk_span.rnss < link2->lnk_span.rnss)
 290                 return(-1);
 291         if (link1->lnk_span.rnss > link2->lnk_span.rnss)
 292                 return(1);
 293 #if 1
 294         if ((uintptr_t)link1->state < (uintptr_t)link2->state)
 295                 return(-1);
 296         if ((uintptr_t)link1->state > (uintptr_t)link2->state)
 297                 return(1);
 298 #else
 299         if (link1->state->msgid < link2->state->msgid)
 300                 return(-1);
 301         if (link1->state->msgid > link2->state->msgid)
 302                 return(1);
 303 #endif
 304         return(0);
 305 }
 306
 307 RB_PROTOTYPE_STATIC(h2span_cluster_tree, h2span_cluster,
 308              rbnode, h2span_cluster_cmp);
 309 RB_PROTOTYPE_STATIC(h2span_node_tree, h2span_node,
 310              rbnode, h2span_node_cmp);
 311 RB_PROTOTYPE_STATIC(h2span_link_tree, h2span_link,
 312              rbnode, h2span_link_cmp);
 313 RB_PROTOTYPE_STATIC(h2span_relay_tree, h2span_relay,
 314              rbnode, h2span_relay_cmp);
 315
 316 RB_GENERATE_STATIC(h2span_cluster_tree, h2span_cluster,
 317              rbnode, h2span_cluster_cmp);
 318 RB_GENERATE_STATIC(h2span_node_tree, h2span_node,
 319              rbnode, h2span_node_cmp);
 320 RB_GENERATE_STATIC(h2span_link_tree, h2span_link,
 321              rbnode, h2span_link_cmp);
 322 RB_GENERATE_STATIC(h2span_relay_tree, h2span_relay,
 323              rbnode, h2span_relay_cmp);
 324
 325 /*
 326  * Global mutex protects cluster_tree lookups, connq, mediaq.
 327  */
 328 static pthread_mutex_t cluster_mtx;
 329 static struct h2span_cluster_tree cluster_tree = RB_INITIALIZER(cluster_tree);
 330 static struct h2span_conn_queue connq = TAILQ_HEAD_INITIALIZER(connq);
 331 static struct dmsg_media_queue mediaq = TAILQ_HEAD_INITIALIZER(mediaq);
 332
 333 static void dmsg_lnk_span(dmsg_msg_t *msg);
 334 static void dmsg_lnk_conn(dmsg_msg_t *msg);
 335 static void dmsg_lnk_ping(dmsg_msg_t *msg);
 336 static void dmsg_lnk_relay(dmsg_msg_t *msg);
 337 static void dmsg_relay_scan(h2span_conn_t *conn, h2span_node_t *node);
 338 static void dmsg_relay_delete(h2span_relay_t *relay);
 339
 340 void
 341 dmsg_msg_lnk_signal(dmsg_iocom_t *iocom __unused)
 342 {
 343         pthread_mutex_lock(&cluster_mtx);
 344         dmsg_relay_scan(NULL, NULL);
 345         pthread_mutex_unlock(&cluster_mtx);
 346 }
 347
 348 /*
 349  * DMSG_PROTO_LNK - Generic DMSG_PROTO_LNK.
 350  *            (incoming iocom lock not held)
 351  *
 352  * This function is typically called for one-way and opening-transactions
 353  * since state->func is assigned after that, but it will also be called
 354  * if no state->func is assigned on transaction-open.
 355  */
 356 void
 357 dmsg_msg_lnk(dmsg_msg_t *msg)
 358 {
 359         dmsg_iocom_t *iocom = msg->state->iocom;
 360
 361         switch(msg->tcmd & DMSGF_BASECMDMASK) {
 362         case DMSG_LNK_CONN:
 363                 dmsg_lnk_conn(msg);
 364                 break;
 365         case DMSG_LNK_SPAN:
 366                 dmsg_lnk_span(msg);
 367                 break;
 368         case DMSG_LNK_PING:
 369                 dmsg_lnk_ping(msg);
 370                 break;
 371         default:
 372                 iocom->usrmsg_callback(msg, 1);
 373                 /* state invalid after reply */
 374                 break;
 375         }
 376 }
 377
 378 /*
 379  * LNK_CONN - iocom identify message reception.
 380  *            (incoming iocom lock not held)
 381  *
 382  * Remote node identifies itself to us, sets up a SPAN filter, and gives us
 383  * the ok to start transmitting SPANs.
 384  */
 385 void
 386 dmsg_lnk_conn(dmsg_msg_t *msg)
 387 {
 388         dmsg_state_t *state = msg->state;
 389         dmsg_iocom_t *iocom = state->iocom;
 390         dmsg_media_t *media;
 391         h2span_conn_t *conn;
 392         h2span_relay_t *relay;
 393         char *alloc = NULL;
 394
 395         pthread_mutex_lock(&cluster_mtx);
 396
 397         dmio_printf(iocom, 3,
 398                 "dmsg_lnk_conn: msg %p cmd %08x state %p "
 399                 "txcmd %08x rxcmd %08x\n",
 400                 msg, msg->any.head.cmd, state,
 401                 state->txcmd, state->rxcmd);
 402
 403         switch(msg->any.head.cmd & DMSGF_TRANSMASK) {
 404         case DMSG_LNK_CONN | DMSGF_CREATE:
 405         case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE:
 406                 /*
 407                  * On transaction start we allocate a new h2span_conn and
 408                  * acknowledge the request, leaving the transaction open.
 409                  * We then relay priority-selected SPANs.
 410                  */
 411                 dmio_printf(iocom, 3, "LNK_CONN(%08x): %s/%s/%s\n",
 412                         (uint32_t)msg->any.head.msgid,
 413                         dmsg_uuid_to_str(&msg->any.lnk_conn.pfs_clid,
 414                                             &alloc),
 415                         msg->any.lnk_conn.cl_label,
 416                         msg->any.lnk_conn.fs_label);
 417                 free(alloc);
 418
 419                 conn = dmsg_alloc(sizeof(*conn));
 420                 assert(state->iocom->conn == NULL);
 421
 422                 RB_INIT(&conn->tree);
 423                 state->iocom->conn = conn;      /* XXX only one */
 424                 state->iocom->conn_msgid = state->msgid;
 425                 dmsg_state_hold(state);
 426                 conn->state = state;
 427                 state->func = dmsg_lnk_conn;
 428                 state->any.conn = conn;
 429                 TAILQ_INSERT_TAIL(&connq, conn, entry);
 430                 conn->lnk_conn = msg->any.lnk_conn;
 431
 432                 /*
 433                  * Set up media
 434                  */
 435                 TAILQ_FOREACH(media, &mediaq, entry) {
 436                         if (uuid_compare(&msg->any.lnk_conn.mediaid,
 437                                          &media->mediaid, NULL) == 0) {
 438                                 break;
 439                         }
 440                 }
 441                 if (media == NULL) {
 442                         media = dmsg_alloc(sizeof(*media));
 443                         media->mediaid = msg->any.lnk_conn.mediaid;
 444                         TAILQ_INSERT_TAIL(&mediaq, media, entry);
 445                 }
 446                 state->media = media;
 447                 ++media->refs;
 448
 449                 if ((msg->any.head.cmd & DMSGF_DELETE) == 0) {
 450                         iocom->usrmsg_callback(msg, 0);
 451                         dmsg_msg_result(msg, 0);
 452                         dmsg_iocom_signal(iocom);
 453                         break;
 454                 }
 455                 /* FALL THROUGH */
 456         case DMSG_LNK_CONN | DMSGF_DELETE:
 457         case DMSG_LNK_ERROR | DMSGF_DELETE:
 458                 /*
 459                  * On transaction terminate we clean out our h2span_conn
 460                  * and acknowledge the request, closing the transaction.
 461                  */
 462                 dmio_printf(iocom, 3, "%s\n", "LNK_CONN: Terminated");
 463                 conn = state->any.conn;
 464                 assert(conn);
 465
 466                 /*
 467                  * Adjust media refs
 468                  *
 469                  * Callback will clean out media config / user-opaque state
 470                  */
 471                 media = state->media;
 472                 --media->refs;
 473                 if (media->refs == 0) {
 474                         dmio_printf(iocom, 3, "%s\n", "Media shutdown");
 475                         TAILQ_REMOVE(&mediaq, media, entry);
 476                         pthread_mutex_unlock(&cluster_mtx);
 477                         iocom->usrmsg_callback(msg, 0);
 478                         pthread_mutex_lock(&cluster_mtx);
 479                         dmsg_free(media);
 480                 }
 481                 state->media = NULL;
 482
 483                 /*
 484                  * Clean out all relays.  This requires terminating each
 485                  * relay transaction.
 486                  */
 487                 while ((relay = RB_ROOT(&conn->tree)) != NULL) {
 488                         dmsg_relay_delete(relay);
 489                 }
 490
 491                 /*
 492                  * Clean out conn
 493                  */
 494                 conn->state = NULL;
 495                 msg->state->any.conn = NULL;
 496                 msg->state->iocom->conn = NULL;
 497                 TAILQ_REMOVE(&connq, conn, entry);
 498                 dmsg_free(conn);
 499
 500                 dmsg_msg_reply(msg, 0);
 501                 dmsg_state_drop(state);
 502                 /* state invalid after reply */
 503                 break;
 504         default:
 505                 iocom->usrmsg_callback(msg, 1);
 506 #if 0
 507                 if (msg->any.head.cmd & DMSGF_DELETE)
 508                         goto deleteconn;
 509                 dmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
 510 #endif
 511                 break;
 512         }
 513         pthread_mutex_unlock(&cluster_mtx);
 514 }
 515
 516 /*
 517  * LNK_SPAN - Spanning tree protocol message reception
 518  *            (incoming iocom lock not held)
 519  *
 520  * Receive a spanning tree transactional message, creating or destroying
 521  * a SPAN and propagating it to other iocoms.
 522  */
 523 void
 524 dmsg_lnk_span(dmsg_msg_t *msg)
 525 {
 526         dmsg_state_t *state = msg->state;
 527         dmsg_iocom_t *iocom = state->iocom;
 528         h2span_cluster_t dummy_cls;
 529         h2span_node_t dummy_node;
 530         h2span_cluster_t *cls;
 531         h2span_node_t *node;
 532         h2span_link_t *slink;
 533         h2span_relay_t *relay;
 534         char *alloc = NULL;
 535
 536         /*
 537          * XXX
 538          *
 539          * Ignore reply to LNK_SPAN.  The reply is expected and will commands
 540          * to flow in both directions on the open transaction.  This will also
 541          * ignore DMSGF_REPLY|DMSGF_DELETE messages.  Since we take no action
 542          * if the other end unexpectedly closes their side of the transaction,
 543          * we can ignore that too.
 544          */
 545         if (msg->any.head.cmd & DMSGF_REPLY) {
 546                 dmio_printf(iocom, 2, "%s\n",
 547                             "Ignore reply to LNK_SPAN");
 548                 return;
 549         }
 550
 551         pthread_mutex_lock(&cluster_mtx);
 552
 553         /*
 554          * On transaction start we initialize the tracking infrastructure
 555          */
 556         if (msg->any.head.cmd & DMSGF_CREATE) {
 557                 assert(state->func == NULL);
 558                 state->func = dmsg_lnk_span;
 559
 560                 dmsg_termstr(msg->any.lnk_span.cl_label);
 561                 dmsg_termstr(msg->any.lnk_span.fs_label);
 562
 563                 /*
 564                  * Find the cluster
 565                  */
 566                 dummy_cls.pfs_clid = msg->any.lnk_span.pfs_clid;
 567                 dummy_cls.peer_type = msg->any.lnk_span.peer_type;
 568                 bcopy(msg->any.lnk_span.cl_label,
 569                       dummy_cls.cl_label,
 570                       sizeof(dummy_cls.cl_label));
 571                 cls = RB_FIND(h2span_cluster_tree, &cluster_tree, &dummy_cls);
 572                 if (cls == NULL) {
 573                         cls = dmsg_alloc(sizeof(*cls));
 574                         cls->pfs_clid = msg->any.lnk_span.pfs_clid;
 575                         cls->peer_type = msg->any.lnk_span.peer_type;
 576                         bcopy(msg->any.lnk_span.cl_label,
 577                               cls->cl_label,
 578                               sizeof(cls->cl_label));
 579                         RB_INIT(&cls->tree);
 580                         RB_INSERT(h2span_cluster_tree, &cluster_tree, cls);
 581                 }
 582
 583                 /*
 584                  * Find the node
 585                  */
 586                 dummy_node.pfs_fsid = msg->any.lnk_span.pfs_fsid;
 587                 bcopy(msg->any.lnk_span.fs_label, dummy_node.fs_label,
 588                       sizeof(dummy_node.fs_label));
 589                 node = RB_FIND(h2span_node_tree, &cls->tree, &dummy_node);
 590                 if (node == NULL) {
 591                         node = dmsg_alloc(sizeof(*node));
 592                         node->pfs_fsid = msg->any.lnk_span.pfs_fsid;
 593                         node->pfs_type = msg->any.lnk_span.pfs_type;
 594                         bcopy(msg->any.lnk_span.fs_label,
 595                               node->fs_label,
 596                               sizeof(node->fs_label));
 597                         node->cls = cls;
 598                         RB_INIT(&node->tree);
 599                         RB_INSERT(h2span_node_tree, &cls->tree, node);
 600                 }
 601
 602                 /*
 603                  * Create the link
 604                  *
 605                  * NOTE: Sub-transactions on the incoming SPAN can be used
 606                  *       to talk to the originator.  We should not set-up
 607                  *       state->relay for incoming SPANs since our sub-trans
 608                  *       is running on the same interface (i.e. no actual
 609                  *       relaying need be done).
 610                  *
 611                  * NOTE: Later on when we relay the SPAN out the outgoing
 612                  *       SPAN state will be set up to relay back to this
 613                  *       state.
 614                  *
 615                  * NOTE: It is possible for SPAN targets to send one-way
 616                  *       messages to the originator but it is not possible
 617                  *       for the originator to (currently) broadcast one-way
 618                  *       messages to all of its SPAN targets.  The protocol
 619                  *       allows such a feature to be added in the future.
 620                  */
 621                 assert(state->any.link == NULL);
 622                 dmsg_state_hold(state);
 623                 slink = dmsg_alloc(sizeof(*slink));
 624                 TAILQ_INIT(&slink->relayq);
 625                 slink->node = node;
 626                 slink->state = state;
 627                 state->any.link = slink;
 628                 slink->lnk_span = msg->any.lnk_span;
 629
 630                 RB_INSERT(h2span_link_tree, &node->tree, slink);
 631
 632                 dmio_printf(iocom, 3,
 633                             "LNK_SPAN(thr %p): %p %s cl=%s fs=%s dist=%d\n",
 634                             iocom, slink,
 635                             dmsg_uuid_to_str(&msg->any.lnk_span.pfs_clid,
 636                                              &alloc),
 637                             msg->any.lnk_span.cl_label,
 638                             msg->any.lnk_span.fs_label,
 639                             msg->any.lnk_span.dist);
 640                 free(alloc);
 641 #if 0
 642                 dmsg_relay_scan(NULL, node);
 643 #endif
 644                 /*
 645                  * Ack the open, which will issue a CREATE on our side, and
 646                  * leave the transaction open.  Necessary to allow the
 647                  * transaction to be used as a virtual circuit.
 648                  */
 649                 dmsg_state_result(state, 0);
 650                 dmsg_iocom_signal(iocom);
 651         }
 652
 653         /*
 654          * On transaction terminate we remove the tracking infrastructure.
 655          */
 656         if (msg->any.head.cmd & DMSGF_DELETE) {
 657                 slink = state->any.link;
 658                 assert(slink->state == state);
 659                 assert(slink != NULL);
 660                 node = slink->node;
 661                 cls = node->cls;
 662
 663                 dmio_printf(iocom, 3,
 664                             "LNK_DELE(thr %p): %p %s cl=%s fs=%s\n",
 665                             iocom, slink,
 666                             dmsg_uuid_to_str(&cls->pfs_clid, &alloc),
 667                             cls->cl_label,
 668                             node->fs_label);
 669                 free(alloc);
 670
 671                 /*
 672                  * Clean out all relays.  This requires terminating each
 673                  * relay transaction.
 674                  */
 675                 while ((relay = TAILQ_FIRST(&slink->relayq)) != NULL) {
 676                         dmsg_relay_delete(relay);
 677                 }
 678
 679                 /*
 680                  * Clean out the topology
 681                  */
 682                 RB_REMOVE(h2span_link_tree, &node->tree, slink);
 683                 if (RB_EMPTY(&node->tree)) {
 684                         RB_REMOVE(h2span_node_tree, &cls->tree, node);
 685                         if (RB_EMPTY(&cls->tree) && cls->refs == 0) {
 686                                 RB_REMOVE(h2span_cluster_tree,
 687                                           &cluster_tree, cls);
 688                                 dmsg_free(cls);
 689                         }
 690                         node->cls = NULL;
 691                         dmsg_free(node);
 692                         node = NULL;
 693                 }
 694                 state->any.link = NULL;
 695                 slink->state = NULL;
 696                 slink->node = NULL;
 697                 dmsg_state_drop(state);
 698                 dmsg_free(slink);
 699
 700                 /*
 701                  * We have to terminate the transaction
 702                  */
 703                 dmsg_state_reply(state, 0);
 704                 /* state invalid after reply */
 705
 706                 /*
 707                  * If the node still exists issue any required updates.  If
 708                  * it doesn't then all related relays have already been
 709                  * removed and there's nothing left to do.
 710                  */
 711 #if 0
 712                 if (node)
 713                         dmsg_relay_scan(NULL, node);
 714 #endif
 715                 if (node)
 716                         dmsg_iocom_signal(iocom);
 717         }
 718
 719         pthread_mutex_unlock(&cluster_mtx);
 720 }
 721
 722 /*
 723  * Respond to a PING with a PING|REPLY, forward replies to the usermsg
 724  * callback.
 725  */
 726 static
 727 void
 728 dmsg_lnk_ping(dmsg_msg_t *msg)
 729 {
 730         dmsg_msg_t *rep;
 731
 732         if (msg->any.head.cmd & DMSGF_REPLY) {
 733                 msg->state->iocom->usrmsg_callback(msg, 1);
 734         } else {
 735                 rep = dmsg_msg_alloc(msg->state, 0,
 736                                      DMSG_LNK_PING | DMSGF_REPLY,
 737                                      NULL, NULL);
 738                 dmsg_msg_write(rep);
 739         }
 740 }
 741
 742 /*
 743  * Update relay transactions for SPANs.
 744  *
 745  * Called with cluster_mtx held.
 746  */
 747 static void dmsg_relay_scan_specific(h2span_node_t *node,
 748                                         h2span_conn_t *conn);
 749
 750 static void
 751 dmsg_relay_scan(h2span_conn_t *conn, h2span_node_t *node)
 752 {
 753         h2span_cluster_t *cls;
 754
 755         if (node) {
 756                 /*
 757                  * Iterate specific node
 758                  */
 759                 TAILQ_FOREACH(conn, &connq, entry)
 760                         dmsg_relay_scan_specific(node, conn);
 761         } else {
 762                 /*
 763                  * Full iteration.
 764                  *
 765                  * Iterate cluster ids, nodes, and either a specific connection
 766                  * or all connections.
 767                  */
 768                 RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
 769                         /*
 770                          * Iterate node ids
 771                          */
 772                         RB_FOREACH(node, h2span_node_tree, &cls->tree) {
 773                                 /*
 774                                  * Synchronize the node's link (received SPANs)
 775                                  * with each connection's relays.
 776                                  */
 777                                 if (conn) {
 778                                         dmsg_relay_scan_specific(node, conn);
 779                                 } else {
 780                                         TAILQ_FOREACH(conn, &connq, entry) {
 781                                             dmsg_relay_scan_specific(node,
 782                                                                         conn);
 783                                         }
 784                                         assert(conn == NULL);
 785                                 }
 786                         }
 787                 }
 788         }
 789 }
 790
 791 /*
 792  * Update the relay'd SPANs for this (node, conn).
 793  *
 794  * Iterate links and adjust relays to match.  We only propagate the top link
 795  * for now (XXX we want to propagate the top two).
 796  *
 797  * The dmsg_relay_scan_cmp() function locates the first relay element
 798  * for any given node.  The relay elements will be sub-sorted by dist.
 799  */
 800 struct relay_scan_info {
 801         h2span_node_t *node;
 802         h2span_relay_t *relay;
 803 };
 804
 805 static int
 806 dmsg_relay_scan_cmp(h2span_relay_t *relay, void *arg)
 807 {
 808         struct relay_scan_info *info = arg;
 809
 810         if ((intptr_t)relay->source_rt->any.link->node < (intptr_t)info->node)
 811                 return(-1);
 812         if ((intptr_t)relay->source_rt->any.link->node > (intptr_t)info->node)
 813                 return(1);
 814         return(0);
 815 }
 816
 817 static int
 818 dmsg_relay_scan_callback(h2span_relay_t *relay, void *arg)
 819 {
 820         struct relay_scan_info *info = arg;
 821
 822         info->relay = relay;
 823         return(-1);
 824 }
 825
 826 static void
 827 dmsg_relay_scan_specific(h2span_node_t *node, h2span_conn_t *conn)
 828 {
 829         struct relay_scan_info info;
 830         h2span_relay_t *relay;
 831         h2span_relay_t *next_relay;
 832         h2span_link_t *slink;
 833         dmsg_lnk_conn_t *lconn;
 834         dmsg_lnk_span_t *lspan;
 835         int count;
 836         int maxcount = 2;
 837 #ifdef REQUIRE_SYMMETRICAL
 838         uint32_t lastdist = DMSG_SPAN_MAXDIST;
 839         uint32_t lastrnss = 0;
 840 #endif
 841
 842         info.node = node;
 843         info.relay = NULL;
 844
 845         /*
 846          * Locate the first related relay for the node on this connection.
 847          * relay will be NULL if there were none.
 848          */
 849         RB_SCAN(h2span_relay_tree, &conn->tree,
 850                 dmsg_relay_scan_cmp, dmsg_relay_scan_callback, &info);
 851         relay = info.relay;
 852         info.relay = NULL;
 853         if (relay)
 854                 assert(relay->source_rt->any.link->node == node);
 855
 856         dm_printf(9, "relay scan for connection %p\n", conn);
 857
 858         /*
 859          * Iterate the node's links (received SPANs) in distance order,
 860          * lowest (best) dist first.
 861          *
 862          * PROPAGATE THE BEST LINKS OVER THE SPECIFIED CONNECTION.
 863          *
 864          * Track relays while iterating the best links and construct
 865          * missing relays when necessary.
 866          *
 867          * (If some prior better link was removed it would have also
 868          *  removed the relay, so the relay can only match exactly or
 869          *  be worse).
 870          */
 871         count = 0;
 872         RB_FOREACH(slink, h2span_link_tree, &node->tree) {
 873                 /*
 874                  * Increment count of successful relays.  This isn't
 875                  * quite accurate if we break out but nothing after
 876                  * the loop uses (count).
 877                  *
 878                  * If count exceeds the maximum number of relays we desire
 879                  * we normally want to break out.  However, in order to
 880                  * guarantee a symmetric path we have to continue if both
 881                  * (dist) and (rnss) continue to match.  Otherwise the SPAN
 882                  * propagation in the reverse direction may choose different
 883                  * routes and we will not have a symmetric path.
 884                  *
 885                  * NOTE: Spanning tree does not have to be symmetrical so
 886                  *       this code is not currently enabled.
 887                  */
 888                 if (++count >= maxcount) {
 889 #ifdef REQUIRE_SYMMETRICAL
 890                         if (lastdist != slink->lnk_span.dist ||
 891                             lastrnss != slink->lnk_span.rnss) {
 892                                 break;
 893                         }
 894 #else
 895                         break;
 896 #endif
 897                         /* go beyond the nominal maximum desired relays */
 898                 }
 899
 900                 /*
 901                  * Match, relay already in-place, get the next
 902                  * relay to match against the next slink.
 903                  */
 904                 if (relay && relay->source_rt->any.link == slink) {
 905                         relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
 906                         continue;
 907                 }
 908
 909                 /*
 910                  * We might want this SLINK, if it passes our filters.
 911                  *
 912                  * The spanning tree can cause closed loops so we have
 913                  * to limit slink->dist.
 914                  */
 915                 if (slink->lnk_span.dist > DMSG_SPAN_MAXDIST)
 916                         break;
 917
 918                 /*
 919                  * Don't bother transmitting a LNK_SPAN out the same
 920                  * connection it came in on.  Trivial optimization.
 921                  */
 922                 if (slink->state->iocom == conn->state->iocom)
 923                         break;
 924
 925                 /*
 926                  * NOTE ON FILTERS: The protocol spec allows non-requested
 927                  * SPANs to be transmitted, the other end is expected to
 928                  * leave their transactions open but otherwise ignore them.
 929                  *
 930                  * Don't bother transmitting if the remote connection
 931                  * is not accepting this SPAN's peer_type.
 932                  *
 933                  * pfs_mask is typically used so pure clients can filter
 934                  * out receiving SPANs for other pure clients.
 935                  */
 936                 lspan = &slink->lnk_span;
 937                 lconn = &conn->lnk_conn;
 938                 if (((1LLU << lspan->peer_type) & lconn->peer_mask) == 0)
 939                         break;
 940                 if (((1LLU << lspan->pfs_type) & lconn->pfs_mask) == 0)
 941                         break;
 942
 943                 /*
 944                  * Do not give pure clients visibility to other pure clients
 945                  */
 946                 if (lconn->pfs_type == DMSG_PFSTYPE_CLIENT &&
 947                     lspan->pfs_type == DMSG_PFSTYPE_CLIENT) {
 948                         break;
 949                 }
 950
 951                 /*
 952                  * Connection filter, if cluster uuid is not NULL it must
 953                  * match the span cluster uuid.  Only applies when the
 954                  * peer_type matches.
 955                  */
 956                 if (lspan->peer_type == lconn->peer_type &&
 957                     !uuid_is_nil(&lconn->pfs_clid, NULL) &&
 958                     uuid_compare(&slink->node->cls->pfs_clid,
 959                                  &lconn->pfs_clid, NULL)) {
 960                         break;
 961                 }
 962
 963                 /*
 964                  * Connection filter, if cluster label is not empty it must
 965                  * match the span cluster label.  Only applies when the
 966                  * peer_type matches.
 967                  */
 968                 if (lspan->peer_type == lconn->peer_type &&
 969                     lconn->cl_label[0] &&
 970                     strcmp(lconn->cl_label, slink->node->cls->cl_label)) {
 971                         break;
 972                 }
 973
 974                 /*
 975                  * NOTE! pfs_fsid differentiates nodes within the same cluster
 976                  *       so we obviously don't want to match those.  Similarly
 977                  *       for fs_label.
 978                  */
 979
 980                 /*
 981                  * Ok, we've accepted this SPAN for relaying.
 982                  */
 983                 assert(relay == NULL ||
 984                        relay->source_rt->any.link->node != slink->node ||
 985                        relay->source_rt->any.link->lnk_span.dist >=
 986                         slink->lnk_span.dist);
 987                 relay = dmsg_generate_relay(conn, slink);
 988 #ifdef REQUIRE_SYMMETRICAL
 989                 lastdist = slink->lnk_span.dist;
 990                 lastrnss = slink->lnk_span.rnss;
 991 #endif
 992
 993                 /*
 994                  * Match (created new relay), get the next relay to
 995                  * match against the next slink.
 996                  */
 997                 relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
 998         }
 999
1000         /*
1001          * Any remaining relay's belonging to this connection which match
1002          * the node are in excess of the current aggregate spanning state
1003          * and should be removed.
1004          */
1005         while (relay && relay->source_rt->any.link->node == node) {
1006                 next_relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
1007                 dm_printf(9, "%s\n", "RELAY DELETE FROM EXTRAS");
1008                 dmsg_relay_delete(relay);
1009                 relay = next_relay;
1010         }
1011 }
1012
1013 /*
1014  * Find the slink associated with the msgid and return its state,
1015  * so the caller can issue a transaction.
1016  */
1017 dmsg_state_t *
1018 dmsg_findspan(const char *label)
1019 {
1020         dmsg_state_t *state;
1021         h2span_cluster_t *cls;
1022         h2span_node_t *node;
1023         h2span_link_t *slink;
1024         uint64_t msgid = strtoull(label, NULL, 16);
1025
1026         pthread_mutex_lock(&cluster_mtx);
1027
1028         state = NULL;
1029         RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
1030                 RB_FOREACH(node, h2span_node_tree, &cls->tree) {
1031                         RB_FOREACH(slink, h2span_link_tree, &node->tree) {
1032                                 if (slink->state->msgid == msgid) {
1033                                         state = slink->state;
1034                                         goto done;
1035                                 }
1036                         }
1037                 }
1038         }
1039 done:
1040         pthread_mutex_unlock(&cluster_mtx);
1041
1042         dm_printf(8, "findspan: %p\n", state);
1043
1044         return state;
1045 }
1046
1047
1048 /*
1049  * Helper function to generate missing relay on target connection.
1050  *
1051  * cluster_mtx must be held
1052  */
1053 static
1054 h2span_relay_t *
1055 dmsg_generate_relay(h2span_conn_t *conn, h2span_link_t *slink)
1056 {
1057         h2span_relay_t *relay;
1058         dmsg_msg_t *msg;
1059
1060         dmsg_state_hold(slink->state);
1061         relay = dmsg_alloc(sizeof(*relay));
1062         relay->conn = conn;
1063         relay->source_rt = slink->state;
1064         /* relay->source_rt->any.link = slink; */
1065
1066         /*
1067          * NOTE: relay->target_rt->any.relay set to relay by alloc.
1068          *
1069          * NOTE: LNK_SPAN is transmitted as a top-level transaction.
1070          */
1071         msg = dmsg_msg_alloc(&conn->state->iocom->state0,
1072                              0, DMSG_LNK_SPAN | DMSGF_CREATE,
1073                              dmsg_lnk_relay, relay);
1074         dmsg_state_hold(msg->state);
1075         relay->target_rt = msg->state;
1076
1077         msg->any.lnk_span = slink->lnk_span;
1078         msg->any.lnk_span.dist = slink->lnk_span.dist + 1;
1079         msg->any.lnk_span.rnss = slink->lnk_span.rnss + dmsg_rnss();
1080
1081         RB_INSERT(h2span_relay_tree, &conn->tree, relay);
1082         TAILQ_INSERT_TAIL(&slink->relayq, relay, entry);
1083
1084         /*
1085          * Seed the relay so new sub-transactions received on the outgoing
1086          * SPAN circuit are relayed back to the originator.
1087          */
1088         msg->state->relay = relay->source_rt;
1089         dmsg_state_hold(msg->state->relay);
1090
1091         dmsg_msg_write(msg);
1092
1093         return (relay);
1094 }
1095
1096 /*
1097  * Messages received on relay SPANs.  These are open transactions so it is
1098  * in fact possible for the other end to close the transaction.
1099  *
1100  * XXX MPRACE on state structure
1101  */
1102 static void
1103 dmsg_lnk_relay(dmsg_msg_t *msg)
1104 {
1105         dmsg_state_t *state = msg->state;
1106         h2span_relay_t *relay;
1107
1108         assert(msg->any.head.cmd & DMSGF_REPLY);
1109
1110         if (msg->any.head.cmd & DMSGF_DELETE) {
1111                 pthread_mutex_lock(&cluster_mtx);
1112                 dm_printf(8, "%s\n", "RELAY DELETE FROM LNK_RELAY MSG");
1113                 if ((relay = state->any.relay) != NULL) {
1114                         dmsg_relay_delete(relay);
1115                 } else {
1116                         dmsg_state_reply(state, 0);
1117                 }
1118                 pthread_mutex_unlock(&cluster_mtx);
1119         }
1120 }
1121
1122 /*
1123  * cluster_mtx held by caller
1124  */
1125 static
1126 void
1127 dmsg_relay_delete(h2span_relay_t *relay)
1128 {
1129         dm_printf(8,
1130                   "RELAY DELETE %p RELAY %p ON CLS=%p NODE=%p "
1131                   "DIST=%d FD %d STATE %p\n",
1132                   relay->source_rt->any.link,
1133                   relay,
1134                   relay->source_rt->any.link->node->cls,
1135                   relay->source_rt->any.link->node,
1136                   relay->source_rt->any.link->lnk_span.dist,
1137                   relay->conn->state->iocom->sock_fd,
1138                   relay->target_rt);
1139
1140         RB_REMOVE(h2span_relay_tree, &relay->conn->tree, relay);
1141         TAILQ_REMOVE(&relay->source_rt->any.link->relayq, relay, entry);
1142
1143         if (relay->target_rt) {
1144                 relay->target_rt->any.relay = NULL;
1145                 dmsg_state_reply(relay->target_rt, 0);
1146                 dmsg_state_drop(relay->target_rt);
1147                 /* state invalid after reply */
1148                 relay->target_rt = NULL;
1149         }
1150
1151         /*
1152          * NOTE: relay->source_rt->refs is held by the relay SPAN
1153          *       state, not by this relay structure.
1154          */
1155         relay->conn = NULL;
1156         if (relay->source_rt) {
1157                 dmsg_state_drop(relay->source_rt);
1158                 relay->source_rt = NULL;
1159         }
1160         dmsg_free(relay);
1161 }
1162
1163 /************************************************************************
1164  *                      ROUTER AND MESSAGING HANDLES                    *
1165  ************************************************************************
1166  *
1167  * Basically the idea here is to provide a stable data structure which
1168  * can be localized to the caller for higher level protocols to work with.
1169  * Depends on the context, these dmsg_handle's can be pooled by use-case
1170  * and remain persistent through a client (or mount point's) life.
1171  */
1172
1173 #if 0
1174 /*
1175  * Obtain a stable handle on a cluster given its uuid.  This ties directly
1176  * into the global cluster topology, creating the structure if necessary
1177  * (even if the uuid does not exist or does not exist yet), and preventing
1178  * the structure from getting ripped out from under us while we hold a
1179  * pointer to it.
1180  */
1181 h2span_cluster_t *
1182 dmsg_cluster_get(uuid_t *pfs_clid)
1183 {
1184         h2span_cluster_t dummy_cls;
1185         h2span_cluster_t *cls;
1186
1187         dummy_cls.pfs_clid = *pfs_clid;
1188         pthread_mutex_lock(&cluster_mtx);
1189         cls = RB_FIND(h2span_cluster_tree, &cluster_tree, &dummy_cls);
1190         if (cls)
1191                 ++cls->refs;
1192         pthread_mutex_unlock(&cluster_mtx);
1193         return (cls);
1194 }
1195
1196 void
1197 dmsg_cluster_put(h2span_cluster_t *cls)
1198 {
1199         pthread_mutex_lock(&cluster_mtx);
1200         assert(cls->refs > 0);
1201         --cls->refs;
1202         if (RB_EMPTY(&cls->tree) && cls->refs == 0) {
1203                 RB_REMOVE(h2span_cluster_tree,
1204                           &cluster_tree, cls);
1205                 dmsg_free(cls);
1206         }
1207         pthread_mutex_unlock(&cluster_mtx);
1208 }
1209
1210 /*
1211  * Obtain a stable handle to a specific cluster node given its uuid.
1212  * This handle does NOT lock in the route to the node and is typically
1213  * used as part of the dmsg_handle_*() API to obtain a set of
1214  * stable nodes.
1215  */
1216 h2span_node_t *
1217 dmsg_node_get(h2span_cluster_t *cls, uuid_t *pfs_fsid)
1218 {
1219 }
1220
1221 #endif
1222
1223 /*
1224  * Dumps the spanning tree
1225  *
1226  * DEBUG ONLY
1227  */
1228 void
1229 dmsg_shell_tree(dmsg_iocom_t *iocom, char *cmdbuf __unused)
1230 {
1231         h2span_cluster_t *cls;
1232         h2span_node_t *node;
1233         h2span_link_t *slink;
1234         h2span_relay_t *relay;
1235         char *uustr = NULL;
1236
1237         pthread_mutex_lock(&cluster_mtx);
1238         RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
1239                 dmsg_printf(iocom, "Cluster %s %s (%s)\n",
1240                                   dmsg_peer_type_to_str(cls->peer_type),
1241                                   dmsg_uuid_to_str(&cls->pfs_clid, &uustr),
1242                                   cls->cl_label);
1243                 RB_FOREACH(node, h2span_node_tree, &cls->tree) {
1244                         dmsg_printf(iocom, "    Node %02x %s (%s)\n",
1245                                 node->pfs_type,
1246                                 dmsg_uuid_to_str(&node->pfs_fsid, &uustr),
1247                                 node->fs_label);
1248                         RB_FOREACH(slink, h2span_link_tree, &node->tree) {
1249                                 dmsg_printf(iocom,
1250                                             "\tSLink msgid %016jx "
1251                                             "dist=%d via %d\n",
1252                                             (intmax_t)slink->state->msgid,
1253                                             slink->lnk_span.dist,
1254                                             slink->state->iocom->sock_fd);
1255                                 TAILQ_FOREACH(relay, &slink->relayq, entry) {
1256                                         dmsg_printf(iocom,
1257                                             "\t    Relay-out msgid %016jx "
1258                                             "via %d\n",
1259                                             (intmax_t)relay->target_rt->msgid,
1260                                             relay->target_rt->iocom->sock_fd);
1261                                 }
1262                         }
1263                 }
1264         }
1265         pthread_mutex_unlock(&cluster_mtx);
1266         if (uustr)
1267                 free(uustr);
1268 #if 0
1269         TAILQ_FOREACH(conn, &connq, entry) {
1270         }
1271 #endif
1272 }
1273
1274 /*
1275  * DEBUG ONLY
1276  *
1277  * Locate the state representing an incoming LNK_SPAN given its msgid.
1278  */
1279 int
1280 dmsg_debug_findspan(uint64_t msgid, dmsg_state_t **statep)
1281 {
1282         h2span_cluster_t *cls;
1283         h2span_node_t *node;
1284         h2span_link_t *slink;
1285
1286         pthread_mutex_lock(&cluster_mtx);
1287         RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
1288                 RB_FOREACH(node, h2span_node_tree, &cls->tree) {
1289                         RB_FOREACH(slink, h2span_link_tree, &node->tree) {
1290                                 if (slink->state->msgid == msgid) {
1291                                         *statep = slink->state;
1292                                         goto found;
1293                                 }
1294                         }
1295                 }
1296         }
1297         pthread_mutex_unlock(&cluster_mtx);
1298         *statep = NULL;
1299         return(ENOENT);
1300 found:
1301         pthread_mutex_unlock(&cluster_mtx);
1302         return(0);
1303 }
1304
1305 /*
1306  * Random number sub-sort value to add to SPAN rnss fields on relay.
1307  * This allows us to differentiate spans with the same <dist> field
1308  * for relaying purposes.  We must normally limit the number of relays
1309  * for any given SPAN origination but we must also guarantee that a
1310  * symmetric reverse path exists, so we use the rnss field as a sub-sort
1311  * (since there can be thousands or millions if we only match on <dist>),
1312  * and if there STILL too many spans we go past the limit.
1313  */
1314 static
1315 uint32_t
1316 dmsg_rnss(void)
1317 {
1318         if (DMsgRNSS == 0) {
1319                 pthread_mutex_lock(&cluster_mtx);
1320                 while (DMsgRNSS == 0) {
1321                         srandomdev();
1322                         DMsgRNSS = random();
1323                 }
1324                 pthread_mutex_unlock(&cluster_mtx);
1325         }
1326         return(DMsgRNSS);
1327 }