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  * FULLY ABORTING A ROUTED MESSAGE is handled via link-failure propagation
  99  * back to the originator.  Only the originator keeps tracks of a message.
 100  * Routers just pass it through.  If a route is lost during transit the
 101  * message is simply thrown away.
 102  *
 103  * It is also important to note that several paths to the same PFS can be
 104  * propagated along the same link, which allows concurrency and even
 105  * redundancy over several network interfaces or via different routes through
 106  * the topology.  Any given transaction will use only a single route but busy
 107  * servers will often have hundreds of transactions active simultaniously,
 108  * so having multiple active paths through the network topology for A<->B
 109  * will improve performance.
 110  *
 111  * --
 112  *
 113  * Most protocols consolidate operations rather than simply relaying them.
 114  * This is particularly true of LEAF protocols (such as strict HAMMER2
 115  * clients), of which there can be millions connecting into the cluster at
 116  * various points.  The SPAN protocol is not used for these LEAF elements.
 117  *
 118  * Instead the primary service they connect to implements a proxy for the
 119  * client protocols so the core topology only has to propagate a couple of
 120  * LNK_SPANs and not millions.  LNK_SPANs are meant to be used only for
 121  * core master nodes and satellite slaves and cache nodes.
 122  */
 123
 124 #include "hammer2.h"
 125
 126 /*
 127  * Maximum spanning tree distance.  This has the practical effect of
 128  * stopping tail-chasing closed loops when a feeder span is lost.
 129  */
 130 #define HAMMER2_SPAN_MAXDIST    16
 131
 132 /*
 133  * RED-BLACK TREE DEFINITIONS
 134  *
 135  * We need to track:
 136  *
 137  * (1) shared fsid's (a cluster).
 138  * (2) unique fsid's (a node in a cluster) <--- LNK_SPAN transactions.
 139  *
 140  * We need to aggegate all active LNK_SPANs, aggregate, and create our own
 141  * outgoing LNK_SPAN transactions on each of our connections representing
 142  * the aggregated state.
 143  *
 144  * h2span_connect       - list of iocom connections who wish to receive SPAN
 145  *                        propagation from other connections.  Might contain
 146  *                        a filter string.  Only iocom's with an open
 147  *                        LNK_CONN transactions are applicable for SPAN
 148  *                        propagation.
 149  *
 150  * h2span_relay         - List of links relayed (via SPAN).  Essentially
 151  *                        each relay structure represents a LNK_SPAN
 152  *                        transaction that we initiated, verses h2span_link
 153  *                        which is a LNK_SPAN transaction that we received.
 154  *
 155  * --
 156  *
 157  * h2span_cluster       - Organizes the shared fsid's.  One structure for
 158  *                        each cluster.
 159  *
 160  * h2span_node          - Organizes the nodes in a cluster.  One structure
 161  *                        for each unique {cluster,node}, aka {fsid, pfs_fsid}.
 162  *
 163  * h2span_link          - Organizes all incoming and outgoing LNK_SPAN message
 164  *                        transactions related to a node.
 165  *
 166  *                        One h2span_link structure for each incoming LNK_SPAN
 167  *                        transaction.  Links selected for propagation back
 168  *                        out are also where the outgoing LNK_SPAN messages
 169  *                        are indexed into (so we can propagate changes).
 170  *
 171  *                        The h2span_link's use a red-black tree to sort the
 172  *                        distance hop metric for the incoming LNK_SPAN.  We
 173  *                        then select the top N for outgoing.  When the
 174  *                        topology changes the top N may also change and cause
 175  *                        new outgoing LNK_SPAN transactions to be opened
 176  *                        and less desireable ones to be closed, causing
 177  *                        transactional aborts within the message flow in
 178  *                        the process.
 179  *
 180  * Also note            - All outgoing LNK_SPAN message transactions are also
 181  *                        entered into a red-black tree for use by the routing
 182  *                        function.  This is handled by msg.c in the state
 183  *                        code, not here.
 184  */
 185
 186 struct h2span_link;
 187 struct h2span_relay;
 188 TAILQ_HEAD(h2span_connect_queue, h2span_connect);
 189 TAILQ_HEAD(h2span_relay_queue, h2span_relay);
 190
 191 RB_HEAD(h2span_cluster_tree, h2span_cluster);
 192 RB_HEAD(h2span_node_tree, h2span_node);
 193 RB_HEAD(h2span_link_tree, h2span_link);
 194 RB_HEAD(h2span_relay_tree, h2span_relay);
 195
 196 /*
 197  * Received LNK_CONN transaction enables SPAN protocol over connection.
 198  * (may contain filter).
 199  */
 200 struct h2span_connect {
 201         TAILQ_ENTRY(h2span_connect) entry;
 202         struct h2span_relay_tree tree;
 203         hammer2_state_t *state;
 204 };
 205
 206 /*
 207  * All received LNK_SPANs are organized by cluster (pfs_clid),
 208  * node (pfs_fsid), and link (received LNK_SPAN transaction).
 209  */
 210 struct h2span_cluster {
 211         RB_ENTRY(h2span_cluster) rbnode;
 212         struct h2span_node_tree tree;
 213         uuid_t  pfs_clid;               /* shared fsid */
 214         int     refs;                   /* prevents destruction */
 215 };
 216
 217 struct h2span_node {
 218         RB_ENTRY(h2span_node) rbnode;
 219         struct h2span_link_tree tree;
 220         struct h2span_cluster *cls;
 221         uuid_t  pfs_fsid;               /* unique fsid */
 222         char label[64];
 223 };
 224
 225 struct h2span_link {
 226         RB_ENTRY(h2span_link) rbnode;
 227         hammer2_state_t *state;         /* state<->link */
 228         struct h2span_node *node;       /* related node */
 229         int32_t dist;
 230         struct h2span_relay_queue relayq; /* relay out */
 231         struct hammer2_router *router;  /* route out this link */
 232 };
 233
 234 /*
 235  * Any LNK_SPAN transactions we receive which are relayed out other
 236  * connections utilize this structure to track the LNK_SPAN transaction
 237  * we initiate on the other connections, if selected for relay.
 238  *
 239  * In many respects this is the core of the protocol... actually figuring
 240  * out what LNK_SPANs to relay.  The spanid used for relaying is the
 241  * address of the 'state' structure, which is why h2span_relay has to
 242  * be entered into a RB-TREE based at h2span_connect (so we can look
 243  * up the spanid to validate it).
 244  *
 245  * NOTE: Messages can be received via the LNK_SPAN transaction the
 246  *       relay maintains, and can be replied via relay->router, but
 247  *       messages are NOT initiated via a relay.  Messages are initiated
 248  *       via incoming links (h2span_link's).
 249  *
 250  *       relay->link represents the link being relayed, NOT the LNK_SPAN
 251  *       transaction the relay is holding open.
 252  */
 253 struct h2span_relay {
 254         RB_ENTRY(h2span_relay) rbnode;  /* from h2span_connect */
 255         TAILQ_ENTRY(h2span_relay) entry; /* from link */
 256         struct h2span_connect *conn;
 257         hammer2_state_t *state;         /* transmitted LNK_SPAN */
 258         struct h2span_link *link;       /* LNK_SPAN being relayed */
 259         struct hammer2_router   *router;/* route out this relay */
 260 };
 261
 262
 263 typedef struct h2span_connect h2span_connect_t;
 264 typedef struct h2span_cluster h2span_cluster_t;
 265 typedef struct h2span_node h2span_node_t;
 266 typedef struct h2span_link h2span_link_t;
 267 typedef struct h2span_relay h2span_relay_t;
 268
 269 static
 270 int
 271 h2span_cluster_cmp(h2span_cluster_t *cls1, h2span_cluster_t *cls2)
 272 {
 273         return(uuid_compare(&cls1->pfs_clid, &cls2->pfs_clid, NULL));
 274 }
 275
 276 static
 277 int
 278 h2span_node_cmp(h2span_node_t *node1, h2span_node_t *node2)
 279 {
 280         return(uuid_compare(&node1->pfs_fsid, &node2->pfs_fsid, NULL));
 281 }
 282
 283 /*
 284  * NOTE: Sort/subsort must match h2span_relay_cmp() under any given
 285  *       node.
 286  */
 287 static
 288 int
 289 h2span_link_cmp(h2span_link_t *link1, h2span_link_t *link2)
 290 {
 291         if (link1->dist < link2->dist)
 292                 return(-1);
 293         if (link1->dist > link2->dist)
 294                 return(1);
 295         if (link1->state->msgid < link2->state->msgid)
 296                 return(-1);
 297         if (link1->state->msgid > link2->state->msgid)
 298                 return(1);
 299         return(0);
 300 }
 301
 302 /*
 303  * Relay entries are sorted by node, subsorted by distance and link
 304  * address (so we can match up the conn->tree relay topology with
 305  * a node's link topology).
 306  */
 307 static
 308 int
 309 h2span_relay_cmp(h2span_relay_t *relay1, h2span_relay_t *relay2)
 310 {
 311         h2span_link_t *link1 = relay1->link;
 312         h2span_link_t *link2 = relay2->link;
 313
 314         if ((intptr_t)link1->node < (intptr_t)link2->node)
 315                 return(-1);
 316         if ((intptr_t)link1->node > (intptr_t)link2->node)
 317                 return(1);
 318         if (link1->dist < link2->dist)
 319                 return(-1);
 320         if (link1->dist > link2->dist)
 321                 return(1);
 322         if (link1->state->msgid < link2->state->msgid)
 323                 return(-1);
 324         if (link1->state->msgid > link2->state->msgid)
 325                 return(1);
 326         return(0);
 327 }
 328
 329 RB_PROTOTYPE_STATIC(h2span_cluster_tree, h2span_cluster,
 330              rbnode, h2span_cluster_cmp);
 331 RB_PROTOTYPE_STATIC(h2span_node_tree, h2span_node,
 332              rbnode, h2span_node_cmp);
 333 RB_PROTOTYPE_STATIC(h2span_link_tree, h2span_link,
 334              rbnode, h2span_link_cmp);
 335 RB_PROTOTYPE_STATIC(h2span_relay_tree, h2span_relay,
 336              rbnode, h2span_relay_cmp);
 337
 338 RB_GENERATE_STATIC(h2span_cluster_tree, h2span_cluster,
 339              rbnode, h2span_cluster_cmp);
 340 RB_GENERATE_STATIC(h2span_node_tree, h2span_node,
 341              rbnode, h2span_node_cmp);
 342 RB_GENERATE_STATIC(h2span_link_tree, h2span_link,
 343              rbnode, h2span_link_cmp);
 344 RB_GENERATE_STATIC(h2span_relay_tree, h2span_relay,
 345              rbnode, h2span_relay_cmp);
 346
 347 /*
 348  * Global mutex protects cluster_tree lookups.
 349  */
 350 static pthread_mutex_t cluster_mtx;
 351 static struct h2span_cluster_tree cluster_tree = RB_INITIALIZER(cluster_tree);
 352 static struct h2span_connect_queue connq = TAILQ_HEAD_INITIALIZER(connq);
 353
 354 static void hammer2_lnk_span(hammer2_msg_t *msg);
 355 static void hammer2_lnk_conn(hammer2_msg_t *msg);
 356 static void hammer2_lnk_relay(hammer2_msg_t *msg);
 357 static void hammer2_relay_scan(h2span_connect_t *conn, h2span_node_t *node);
 358 static void hammer2_relay_delete(h2span_relay_t *relay);
 359
 360 void
 361 hammer2_msg_lnk_signal(hammer2_router_t *router __unused)
 362 {
 363         pthread_mutex_lock(&cluster_mtx);
 364         hammer2_relay_scan(NULL, NULL);
 365         pthread_mutex_unlock(&cluster_mtx);
 366 }
 367
 368 /*
 369  * Receive a HAMMER2_MSG_PROTO_LNK message.  This only called for
 370  * one-way and opening-transactions since state->func will be assigned
 371  * in all other cases.
 372  */
 373 void
 374 hammer2_msg_lnk(hammer2_msg_t *msg)
 375 {
 376         switch(msg->any.head.cmd & HAMMER2_MSGF_BASECMDMASK) {
 377         case HAMMER2_LNK_CONN:
 378                 hammer2_lnk_conn(msg);
 379                 break;
 380         case HAMMER2_LNK_SPAN:
 381                 hammer2_lnk_span(msg);
 382                 break;
 383         default:
 384                 fprintf(stderr,
 385                         "MSG_PROTO_LNK: Unknown msg %08x\n", msg->any.head.cmd);
 386                 hammer2_msg_reply(msg, HAMMER2_MSG_ERR_NOSUPP);
 387                 /* state invalid after reply */
 388                 break;
 389         }
 390 }
 391
 392 void
 393 hammer2_lnk_conn(hammer2_msg_t *msg)
 394 {
 395         hammer2_state_t *state = msg->state;
 396         h2span_connect_t *conn;
 397         h2span_relay_t *relay;
 398         char *alloc = NULL;
 399
 400         pthread_mutex_lock(&cluster_mtx);
 401
 402         /*
 403          * On transaction start we allocate a new h2span_connect and
 404          * acknowledge the request, leaving the transaction open.
 405          * We then relay priority-selected SPANs.
 406          */
 407         if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
 408                 state->func = hammer2_lnk_conn;
 409
 410                 fprintf(stderr, "LNK_CONN(%08x): %s/%s\n",
 411                         (uint32_t)msg->any.head.msgid,
 412                         hammer2_uuid_to_str(&msg->any.lnk_conn.pfs_clid,
 413                                             &alloc),
 414                         msg->any.lnk_conn.label);
 415                 free(alloc);
 416
 417                 conn = hammer2_alloc(sizeof(*conn));
 418
 419                 RB_INIT(&conn->tree);
 420                 conn->state = state;
 421                 state->any.conn = conn;
 422                 TAILQ_INSERT_TAIL(&connq, conn, entry);
 423
 424                 hammer2_msg_result(msg, 0);
 425
 426 #if 0
 427                 /*
 428                  * Span-synchronize all nodes with the new connection
 429                  */
 430                 hammer2_relay_scan(conn, NULL);
 431 #endif
 432                 hammer2_router_signal(msg->router);
 433         }
 434
 435         /*
 436          * On transaction terminate we clean out our h2span_connect
 437          * and acknowledge the request, closing the transaction.
 438          */
 439         if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
 440                 fprintf(stderr, "LNK_CONN: Terminated\n");
 441                 conn = state->any.conn;
 442                 assert(conn);
 443
 444                 /*
 445                  * Clean out all relays.  This requires terminating each
 446                  * relay transaction.
 447                  */
 448                 while ((relay = RB_ROOT(&conn->tree)) != NULL) {
 449                         hammer2_relay_delete(relay);
 450                 }
 451
 452                 /*
 453                  * Clean out conn
 454                  */
 455                 conn->state = NULL;
 456                 msg->state->any.conn = NULL;
 457                 TAILQ_REMOVE(&connq, conn, entry);
 458                 hammer2_free(conn);
 459
 460                 hammer2_msg_reply(msg, 0);
 461                 /* state invalid after reply */
 462         }
 463         pthread_mutex_unlock(&cluster_mtx);
 464 }
 465
 466 void
 467 hammer2_lnk_span(hammer2_msg_t *msg)
 468 {
 469         hammer2_state_t *state = msg->state;
 470         h2span_cluster_t dummy_cls;
 471         h2span_node_t dummy_node;
 472         h2span_cluster_t *cls;
 473         h2span_node_t *node;
 474         h2span_link_t *slink;
 475         h2span_relay_t *relay;
 476         char *alloc = NULL;
 477
 478         assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
 479
 480         pthread_mutex_lock(&cluster_mtx);
 481
 482         /*
 483          * On transaction start we initialize the tracking infrastructure
 484          */
 485         if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
 486                 assert(state->func == NULL);
 487                 state->func = hammer2_lnk_span;
 488
 489                 msg->any.lnk_span.label[sizeof(msg->any.lnk_span.label)-1] = 0;
 490
 491                 /*
 492                  * Find the cluster
 493                  */
 494                 dummy_cls.pfs_clid = msg->any.lnk_span.pfs_clid;
 495                 cls = RB_FIND(h2span_cluster_tree, &cluster_tree, &dummy_cls);
 496                 if (cls == NULL) {
 497                         cls = hammer2_alloc(sizeof(*cls));
 498                         cls->pfs_clid = msg->any.lnk_span.pfs_clid;
 499                         RB_INIT(&cls->tree);
 500                         RB_INSERT(h2span_cluster_tree, &cluster_tree, cls);
 501                 }
 502
 503                 /*
 504                  * Find the node
 505                  */
 506                 dummy_node.pfs_fsid = msg->any.lnk_span.pfs_fsid;
 507                 node = RB_FIND(h2span_node_tree, &cls->tree, &dummy_node);
 508                 if (node == NULL) {
 509                         node = hammer2_alloc(sizeof(*node));
 510                         node->pfs_fsid = msg->any.lnk_span.pfs_fsid;
 511                         node->cls = cls;
 512                         RB_INIT(&node->tree);
 513                         RB_INSERT(h2span_node_tree, &cls->tree, node);
 514                         snprintf(node->label, sizeof(node->label),
 515                                  "%s", msg->any.lnk_span.label);
 516                 }
 517
 518                 /*
 519                  * Create the link
 520                  */
 521                 assert(state->any.link == NULL);
 522                 slink = hammer2_alloc(sizeof(*slink));
 523                 TAILQ_INIT(&slink->relayq);
 524                 slink->node = node;
 525                 slink->dist = msg->any.lnk_span.dist;
 526                 slink->state = state;
 527                 state->any.link = slink;
 528
 529                 /*
 530                  * Embedded router structure in link for message forwarding.
 531                  *
 532                  * The spanning id for the router is the message id of
 533                  * the SPAN link it is embedded in, allowing messages to
 534                  * be routed via &slink->router.
 535                  */
 536                 slink->router = hammer2_router_alloc();
 537                 slink->router->iocom = state->iocom;
 538                 slink->router->link = slink;
 539                 slink->router->spanid = state->msgid;
 540                 hammer2_router_connect(slink->router);
 541
 542                 RB_INSERT(h2span_link_tree, &node->tree, slink);
 543
 544                 fprintf(stderr, "LNK_SPAN(thr %p): %p %s/%s dist=%d\n",
 545                         msg->router->iocom,
 546                         slink,
 547                         hammer2_uuid_to_str(&msg->any.lnk_span.pfs_clid,
 548                                             &alloc),
 549                         msg->any.lnk_span.label,
 550                         msg->any.lnk_span.dist);
 551                 free(alloc);
 552 #if 0
 553                 hammer2_relay_scan(NULL, node);
 554 #endif
 555                 hammer2_router_signal(msg->router);
 556         }
 557
 558         /*
 559          * On transaction terminate we remove the tracking infrastructure.
 560          */
 561         if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
 562                 slink = state->any.link;
 563                 assert(slink != NULL);
 564                 node = slink->node;
 565                 cls = node->cls;
 566
 567                 fprintf(stderr, "LNK_DELE(thr %p): %p %s/%s dist=%d\n",
 568                         msg->router->iocom,
 569                         slink,
 570                         hammer2_uuid_to_str(&cls->pfs_clid, &alloc),
 571                         state->msg->any.lnk_span.label,
 572                         state->msg->any.lnk_span.dist);
 573                 free(alloc);
 574
 575                 /*
 576                  * Remove the router from consideration
 577                  */
 578                 hammer2_router_disconnect(&slink->router);
 579
 580                 /*
 581                  * Clean out all relays.  This requires terminating each
 582                  * relay transaction.
 583                  */
 584                 while ((relay = TAILQ_FIRST(&slink->relayq)) != NULL) {
 585                         hammer2_relay_delete(relay);
 586                 }
 587
 588                 /*
 589                  * Clean out the topology
 590                  */
 591                 RB_REMOVE(h2span_link_tree, &node->tree, slink);
 592                 if (RB_EMPTY(&node->tree)) {
 593                         RB_REMOVE(h2span_node_tree, &cls->tree, node);
 594                         if (RB_EMPTY(&cls->tree) && cls->refs == 0) {
 595                                 RB_REMOVE(h2span_cluster_tree,
 596                                           &cluster_tree, cls);
 597                                 hammer2_free(cls);
 598                         }
 599                         node->cls = NULL;
 600                         hammer2_free(node);
 601                         node = NULL;
 602                 }
 603                 state->any.link = NULL;
 604                 slink->state = NULL;
 605                 slink->node = NULL;
 606                 hammer2_free(slink);
 607
 608                 /*
 609                  * We have to terminate the transaction
 610                  */
 611                 hammer2_state_reply(state, 0);
 612                 /* state invalid after reply */
 613
 614                 /*
 615                  * If the node still exists issue any required updates.  If
 616                  * it doesn't then all related relays have already been
 617                  * removed and there's nothing left to do.
 618                  */
 619 #if 0
 620                 if (node)
 621                         hammer2_relay_scan(NULL, node);
 622 #endif
 623                 if (node)
 624                         hammer2_router_signal(msg->router);
 625         }
 626
 627         pthread_mutex_unlock(&cluster_mtx);
 628 }
 629
 630 /*
 631  * Messages received on relay SPANs.  These are open transactions so it is
 632  * in fact possible for the other end to close the transaction.
 633  *
 634  * XXX MPRACE on state structure
 635  */
 636 static void
 637 hammer2_lnk_relay(hammer2_msg_t *msg)
 638 {
 639         hammer2_state_t *state = msg->state;
 640         h2span_relay_t *relay;
 641
 642         assert(msg->any.head.cmd & HAMMER2_MSGF_REPLY);
 643
 644         if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
 645                 pthread_mutex_lock(&cluster_mtx);
 646                 if ((relay = state->any.relay) != NULL) {
 647                         hammer2_relay_delete(relay);
 648                 } else {
 649                         hammer2_state_reply(state, 0);
 650                 }
 651                 pthread_mutex_unlock(&cluster_mtx);
 652         }
 653 }
 654
 655 /*
 656  * Update relay transactions for SPANs.
 657  *
 658  * Called with cluster_mtx held.
 659  */
 660 static void hammer2_relay_scan_specific(h2span_node_t *node,
 661                                         h2span_connect_t *conn);
 662
 663 static void
 664 hammer2_relay_scan(h2span_connect_t *conn, h2span_node_t *node)
 665 {
 666         h2span_cluster_t *cls;
 667
 668         if (node) {
 669                 /*
 670                  * Iterate specific node
 671                  */
 672                 TAILQ_FOREACH(conn, &connq, entry)
 673                         hammer2_relay_scan_specific(node, conn);
 674         } else {
 675                 /*
 676                  * Full iteration.
 677                  *
 678                  * Iterate cluster ids, nodes, and either a specific connection
 679                  * or all connections.
 680                  */
 681                 RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
 682                         /*
 683                          * Iterate node ids
 684                          */
 685                         RB_FOREACH(node, h2span_node_tree, &cls->tree) {
 686                                 /*
 687                                  * Synchronize the node's link (received SPANs)
 688                                  * with each connection's relays.
 689                                  */
 690                                 if (conn) {
 691                                         hammer2_relay_scan_specific(node, conn);
 692                                 } else {
 693                                         TAILQ_FOREACH(conn, &connq, entry) {
 694                                             hammer2_relay_scan_specific(node,
 695                                                                         conn);
 696                                         }
 697                                         assert(conn == NULL);
 698                                 }
 699                         }
 700                 }
 701         }
 702 }
 703
 704 /*
 705  * Update the relay'd SPANs for this (node, conn).
 706  *
 707  * Iterate links and adjust relays to match.  We only propagate the top link
 708  * for now (XXX we want to propagate the top two).
 709  *
 710  * The hammer2_relay_scan_cmp() function locates the first relay element
 711  * for any given node.  The relay elements will be sub-sorted by dist.
 712  */
 713 struct relay_scan_info {
 714         h2span_node_t *node;
 715         h2span_relay_t *relay;
 716 };
 717
 718 static int
 719 hammer2_relay_scan_cmp(h2span_relay_t *relay, void *arg)
 720 {
 721         struct relay_scan_info *info = arg;
 722
 723         if ((intptr_t)relay->link->node < (intptr_t)info->node)
 724                 return(-1);
 725         if ((intptr_t)relay->link->node > (intptr_t)info->node)
 726                 return(1);
 727         return(0);
 728 }
 729
 730 static int
 731 hammer2_relay_scan_callback(h2span_relay_t *relay, void *arg)
 732 {
 733         struct relay_scan_info *info = arg;
 734
 735         info->relay = relay;
 736         return(-1);
 737 }
 738
 739 static void
 740 hammer2_relay_scan_specific(h2span_node_t *node, h2span_connect_t *conn)
 741 {
 742         struct relay_scan_info info;
 743         h2span_relay_t *relay;
 744         h2span_relay_t *next_relay;
 745         h2span_link_t *slink;
 746         int count = 2;
 747
 748         info.node = node;
 749         info.relay = NULL;
 750
 751         /*
 752          * Locate the first related relay for the node on this connection.
 753          * relay will be NULL if there were none.
 754          */
 755         RB_SCAN(h2span_relay_tree, &conn->tree,
 756                 hammer2_relay_scan_cmp, hammer2_relay_scan_callback, &info);
 757         relay = info.relay;
 758         info.relay = NULL;
 759         if (relay)
 760                 assert(relay->link->node == node);
 761
 762         if (DebugOpt > 8)
 763                 fprintf(stderr, "relay scan for connection %p\n", conn);
 764
 765         /*
 766          * Iterate the node's links (received SPANs) in distance order,
 767          * lowest (best) dist first.
 768          */
 769         /* fprintf(stderr, "LOOP\n"); */
 770         RB_FOREACH(slink, h2span_link_tree, &node->tree) {
 771                 /*
 772                 fprintf(stderr, "SLINK %p RELAY %p(%p)\n",
 773                         slink, relay, relay ? relay->link : NULL);
 774                 */
 775                 /*
 776                  * PROPAGATE THE BEST LINKS OVER THE SPECIFIED CONNECTION.
 777                  *
 778                  * Track relays while iterating the best links and construct
 779                  * missing relays when necessary.
 780                  *
 781                  * (If some prior better link was removed it would have also
 782                  *  removed the relay, so the relay can only match exactly or
 783                  *  be worse).
 784                  */
 785                 if (relay && relay->link == slink) {
 786                         /*
 787                          * Match, relay already in-place, get the next
 788                          * relay to match against the next slink.
 789                          */
 790                         relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
 791                         if (--count == 0)
 792                                 break;
 793                 } else if (slink->dist > HAMMER2_SPAN_MAXDIST) {
 794                         /*
 795                          * No match but span distance is too great,
 796                          * do not relay.  This prevents endless closed
 797                          * loops with ever-incrementing distances when
 798                          * the seed span is lost in the graph.
 799                          *
 800                          * All later spans will also be too far away so
 801                          * we can break out of the loop.
 802                          */
 803                         break;
 804                 } else {
 805                         /*
 806                          * No match, distance is ok, construct a new relay.
 807                          * (slink is better than relay).
 808                          */
 809                         hammer2_msg_t *msg;
 810
 811                         assert(relay == NULL ||
 812                                relay->link->node != slink->node ||
 813                                relay->link->dist >= slink->dist);
 814                         relay = hammer2_alloc(sizeof(*relay));
 815                         relay->conn = conn;
 816                         relay->link = slink;
 817
 818                         msg = hammer2_msg_alloc(conn->state->iocom->router, 0,
 819                                                 HAMMER2_LNK_SPAN |
 820                                                 HAMMER2_MSGF_CREATE,
 821                                                 hammer2_lnk_relay, relay);
 822                         relay->state = msg->state;
 823                         relay->router = hammer2_router_alloc();
 824                         relay->router->iocom = relay->state->iocom;
 825                         relay->router->relay = relay;
 826                         relay->router->spanid = relay->state->msgid;
 827
 828                         msg->any.lnk_span = slink->state->msg->any.lnk_span;
 829                         msg->any.lnk_span.dist = slink->dist + 1;
 830
 831                         hammer2_router_connect(relay->router);
 832
 833                         RB_INSERT(h2span_relay_tree, &conn->tree, relay);
 834                         TAILQ_INSERT_TAIL(&slink->relayq, relay, entry);
 835
 836                         hammer2_msg_write(msg);
 837
 838                         fprintf(stderr,
 839                                 "RELAY SPAN %p RELAY %p ON CLS=%p NODE=%p DIST=%d "
 840                                 "FD %d state %p\n",
 841                                 slink,
 842                                 relay,
 843                                 node->cls, node, slink->dist,
 844                                 conn->state->iocom->sock_fd, relay->state);
 845
 846                         /*
 847                          * Match (created new relay), get the next relay to
 848                          * match against the next slink.
 849                          */
 850                         relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
 851                         if (--count == 0)
 852                                 break;
 853                 }
 854         }
 855
 856         /*
 857          * Any remaining relay's belonging to this connection which match
 858          * the node are in excess of the current aggregate spanning state
 859          * and should be removed.
 860          */
 861         while (relay && relay->link->node == node) {
 862                 next_relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
 863                 hammer2_relay_delete(relay);
 864                 relay = next_relay;
 865         }
 866 }
 867
 868 static
 869 void
 870 hammer2_relay_delete(h2span_relay_t *relay)
 871 {
 872         fprintf(stderr,
 873                 "RELAY DELETE %p RELAY %p ON CLS=%p NODE=%p DIST=%d FD %d STATE %p\n",
 874                 relay->link,
 875                 relay,
 876                 relay->link->node->cls, relay->link->node,
 877                 relay->link->dist,
 878                 relay->conn->state->iocom->sock_fd, relay->state);
 879
 880         hammer2_router_disconnect(&relay->router);
 881
 882         RB_REMOVE(h2span_relay_tree, &relay->conn->tree, relay);
 883         TAILQ_REMOVE(&relay->link->relayq, relay, entry);
 884
 885         if (relay->state) {
 886                 relay->state->any.relay = NULL;
 887                 hammer2_state_reply(relay->state, 0);
 888                 /* state invalid after reply */
 889                 relay->state = NULL;
 890         }
 891         relay->conn = NULL;
 892         relay->link = NULL;
 893         hammer2_free(relay);
 894 }
 895
 896 /************************************************************************
 897  *                      ROUTER AND MESSAGING HANDLES                    *
 898  ************************************************************************
 899  *
 900  * Basically the idea here is to provide a stable data structure which
 901  * can be localized to the caller for higher level protocols to work with.
 902  * Depends on the context, these hammer2_handle's can be pooled by use-case
 903  * and remain persistent through a client (or mount point's) life.
 904  */
 905
 906 #if 0
 907 /*
 908  * Obtain a stable handle on a cluster given its uuid.  This ties directly
 909  * into the global cluster topology, creating the structure if necessary
 910  * (even if the uuid does not exist or does not exist yet), and preventing
 911  * the structure from getting ripped out from under us while we hold a
 912  * pointer to it.
 913  */
 914 h2span_cluster_t *
 915 hammer2_cluster_get(uuid_t *pfs_clid)
 916 {
 917         h2span_cluster_t dummy_cls;
 918         h2span_cluster_t *cls;
 919
 920         dummy_cls.pfs_clid = *pfs_clid;
 921         pthread_mutex_lock(&cluster_mtx);
 922         cls = RB_FIND(h2span_cluster_tree, &cluster_tree, &dummy_cls);
 923         if (cls)
 924                 ++cls->refs;
 925         pthread_mutex_unlock(&cluster_mtx);
 926         return (cls);
 927 }
 928
 929 void
 930 hammer2_cluster_put(h2span_cluster_t *cls)
 931 {
 932         pthread_mutex_lock(&cluster_mtx);
 933         assert(cls->refs > 0);
 934         --cls->refs;
 935         if (RB_EMPTY(&cls->tree) && cls->refs == 0) {
 936                 RB_REMOVE(h2span_cluster_tree,
 937                           &cluster_tree, cls);
 938                 hammer2_free(cls);
 939         }
 940         pthread_mutex_unlock(&cluster_mtx);
 941 }
 942
 943 /*
 944  * Obtain a stable handle to a specific cluster node given its uuid.
 945  * This handle does NOT lock in the route to the node and is typically
 946  * used as part of the hammer2_handle_*() API to obtain a set of
 947  * stable nodes.
 948  */
 949 h2span_node_t *
 950 hammer2_node_get(h2span_cluster_t *cls, uuid_t *pfs_fsid)
 951 {
 952 }
 953
 954 #endif
 955
 956 #if 0
 957 /*
 958  * Acquire a persistent router structure given the cluster and node ids.
 959  * Messages can be transacted via this structure while held.  If the route
 960  * is lost messages will return failure.
 961  */
 962 hammer2_router_t *
 963 hammer2_router_get(uuid_t *pfs_clid, uuid_t *pfs_fsid)
 964 {
 965 }
 966
 967 /*
 968  * Release previously acquired router.
 969  */
 970 void
 971 hammer2_router_put(hammer2_router_t *router)
 972 {
 973 }
 974 #endif
 975
 976 /************************************************************************
 977  *                              DEBUGGER                                *
 978  ************************************************************************/
 979 /*
 980  * Dumps the spanning tree
 981  */
 982 void
 983 shell_tree(hammer2_router_t *router, char *cmdbuf __unused)
 984 {
 985         h2span_cluster_t *cls;
 986         h2span_node_t *node;
 987         h2span_link_t *slink;
 988         char *uustr = NULL;
 989
 990         pthread_mutex_lock(&cluster_mtx);
 991         RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
 992                 router_printf(router, "Cluster %s\n",
 993                              hammer2_uuid_to_str(&cls->pfs_clid, &uustr));
 994                 RB_FOREACH(node, h2span_node_tree, &cls->tree) {
 995                         router_printf(router, "    Node %s (%s)\n",
 996                                  hammer2_uuid_to_str(&node->pfs_fsid, &uustr),
 997                                  node->label);
 998                         RB_FOREACH(slink, h2span_link_tree, &node->tree) {
 999                                 router_printf(router, "\tLink dist=%d via %d\n",
1000                                              slink->dist,
1001                                              slink->state->iocom->sock_fd);
1002                         }
1003                 }
1004         }
1005         pthread_mutex_unlock(&cluster_mtx);
1006         if (uustr)
1007                 free(uustr);
1008 #if 0
1009         TAILQ_FOREACH(conn, &connq, entry) {
1010         }
1011 #endif
1012 }