2 * Copyright (c) 2012 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
35 * LNK_SPAN PROTOCOL SUPPORT FUNCTIONS
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.
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.
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.
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.
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.
73 * The higher level protocols can then issue transactions to the nodes making
74 * up a cluster to perform all actions required.
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.
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.
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.
98 * It is also important to note that several paths to the same PFS can be
99 * propagated along the same link, which allows concurrency and even
100 * redundancy over several network interfaces or via different routes through
101 * the topology. Any given transaction will use only a single route but busy
102 * servers will often have hundreds of transactions active simultaniously,
103 * so having multiple active paths through the network topology for A<->B
104 * will improve performance.
108 * Most protocols consolidate operations rather than simply relaying them.
109 * This is particularly true of LEAF protocols (such as strict HAMMER2
110 * clients), of which there can be millions connecting into the cluster at
111 * various points. The SPAN protocol is not used for these LEAF elements.
113 * Instead the primary service they connect to implements a proxy for the
114 * client protocols so the core topology only has to propagate a couple of
115 * LNK_SPANs and not millions. LNK_SPANs are meant to be used only for
116 * core master nodes and satellite slaves and cache nodes.
122 * RED-BLACK TREE DEFINITIONS
126 * (1) shared fsid's (a cluster).
127 * (2) unique fsid's (a node in a cluster) <--- LNK_SPAN transactions.
129 * We need to aggegate all active LNK_SPANs, aggregate, and create our own
130 * outgoing LNK_SPAN transactions on each of our connections representing
131 * the aggregated state.
133 * h2span_connect - list of iocom connections who wish to receive SPAN
134 * propagation from other connections. Might contain
135 * a filter string. Only iocom's with an open
136 * LNK_CONN transactions are applicable for SPAN
139 * h2span_relay - List of links relayed (via SPAN). Essentially
140 * each relay structure represents a LNK_SPAN
141 * transaction that we initiated, verses h2span_link
142 * which is a LNK_SPAN transaction that we received.
146 * h2span_cluster - Organizes the shared fsid's. One structure for
149 * h2span_node - Organizes the nodes in a cluster. One structure
150 * for each unique {cluster,node}, aka {fsid, pfs_fsid}.
152 * h2span_link - Organizes all incoming and outgoing LNK_SPAN message
153 * transactions related to a node.
155 * One h2span_link structure for each incoming LNK_SPAN
156 * transaction. Links selected for propagation back
157 * out are also where the outgoing LNK_SPAN messages
158 * are indexed into (so we can propagate changes).
160 * The h2span_link's use a red-black tree to sort the
161 * distance hop metric for the incoming LNK_SPAN. We
162 * then select the top N for outgoing. When the
163 * topology changes the top N may also change and cause
164 * new outgoing LNK_SPAN transactions to be opened
165 * and less desireable ones to be closed, causing
166 * transactional aborts within the message flow in
169 * Also note - All outgoing LNK_SPAN message transactions are also
170 * entered into a red-black tree for use by the routing
171 * function. This is handled by msg.c in the state
177 TAILQ_HEAD(h2span_connect_queue, h2span_connect);
178 TAILQ_HEAD(h2span_relay_queue, h2span_relay);
180 RB_HEAD(h2span_cluster_tree, h2span_cluster);
181 RB_HEAD(h2span_node_tree, h2span_node);
182 RB_HEAD(h2span_link_tree, h2span_link);
183 RB_HEAD(h2span_relay_tree, h2span_relay);
186 * Received LNK_CONN transaction enables SPAN protocol over connection.
187 * (may contain filter).
189 struct h2span_connect {
190 TAILQ_ENTRY(h2span_connect) entry;
191 struct h2span_relay_tree tree;
192 hammer2_state_t *state;
196 * All received LNK_SPANs are organized by cluster (pfs_clid),
197 * node (pfs_fsid), and link (received LNK_SPAN transaction).
199 struct h2span_cluster {
200 RB_ENTRY(h2span_cluster) rbnode;
201 struct h2span_node_tree tree;
202 uuid_t pfs_clid; /* shared fsid */
206 RB_ENTRY(h2span_node) rbnode;
207 struct h2span_link_tree tree;
208 struct h2span_cluster *cls;
209 uuid_t pfs_fsid; /* unique fsid */
213 RB_ENTRY(h2span_link) rbnode;
214 hammer2_state_t *state; /* state<->link */
215 struct h2span_node *node; /* related node */
217 struct h2span_relay_queue relayq; /* relay out */
221 * Any LNK_SPAN transactions we receive which are relayed out other
222 * connections utilize this structure to track the LNK_SPAN transaction
223 * we initiate on the other connections, if selected for relay.
225 * In many respects this is the core of the protocol... actually figuring
226 * out what LNK_SPANs to relay. The spanid used for relaying is the
227 * address of the 'state' structure, which is why h2span_relay has to
228 * be entered into a RB-TREE based at h2span_connect (so we can look
229 * up the spanid to validate it).
231 struct h2span_relay {
232 RB_ENTRY(h2span_relay) rbnode; /* from h2span_connect */
233 TAILQ_ENTRY(h2span_relay) entry; /* from link */
234 struct h2span_connect *conn;
235 hammer2_state_t *state; /* transmitted LNK_SPAN */
236 struct h2span_link *link; /* received LNK_SPAN */
240 typedef struct h2span_connect h2span_connect_t;
241 typedef struct h2span_cluster h2span_cluster_t;
242 typedef struct h2span_node h2span_node_t;
243 typedef struct h2span_link h2span_link_t;
244 typedef struct h2span_relay h2span_relay_t;
248 h2span_cluster_cmp(h2span_cluster_t *cls1, h2span_cluster_t *cls2)
250 return(uuid_compare(&cls1->pfs_clid, &cls2->pfs_clid, NULL));
255 h2span_node_cmp(h2span_node_t *node1, h2span_node_t *node2)
257 return(uuid_compare(&node1->pfs_fsid, &node2->pfs_fsid, NULL));
262 h2span_link_cmp(h2span_link_t *link1, h2span_link_t *link2)
264 if (link1->dist < link2->dist)
266 if (link1->dist > link2->dist)
268 if ((intptr_t)link1 < (intptr_t)link2)
270 if ((intptr_t)link1 > (intptr_t)link2)
276 * Relay entries are sorted by node, subsorted by distance and link
277 * address (so we can match up the conn->tree relay topology with
278 * a node's link topology).
282 h2span_relay_cmp(h2span_relay_t *relay1, h2span_relay_t *relay2)
284 if ((intptr_t)relay1->link->node < (intptr_t)relay2->link->node)
286 if ((intptr_t)relay1->link->node > (intptr_t)relay2->link->node)
288 if ((intptr_t)relay1->link->dist < (intptr_t)relay2->link->dist)
290 if ((intptr_t)relay1->link->dist > (intptr_t)relay2->link->dist)
292 if ((intptr_t)relay1->link < (intptr_t)relay2->link)
294 if ((intptr_t)relay1->link > (intptr_t)relay2->link)
299 RB_PROTOTYPE_STATIC(h2span_cluster_tree, h2span_cluster,
300 rbnode, h2span_cluster_cmp);
301 RB_PROTOTYPE_STATIC(h2span_node_tree, h2span_node,
302 rbnode, h2span_node_cmp);
303 RB_PROTOTYPE_STATIC(h2span_link_tree, h2span_link,
304 rbnode, h2span_link_cmp);
305 RB_PROTOTYPE_STATIC(h2span_relay_tree, h2span_relay,
306 rbnode, h2span_relay_cmp);
308 RB_GENERATE_STATIC(h2span_cluster_tree, h2span_cluster,
309 rbnode, h2span_cluster_cmp);
310 RB_GENERATE_STATIC(h2span_node_tree, h2span_node,
311 rbnode, h2span_node_cmp);
312 RB_GENERATE_STATIC(h2span_link_tree, h2span_link,
313 rbnode, h2span_link_cmp);
314 RB_GENERATE_STATIC(h2span_relay_tree, h2span_relay,
315 rbnode, h2span_relay_cmp);
318 * Global mutex protects cluster_tree lookups.
320 static pthread_mutex_t cluster_mtx;
321 static struct h2span_cluster_tree cluster_tree = RB_INITIALIZER(cluster_tree);
322 static struct h2span_connect_queue connq = TAILQ_HEAD_INITIALIZER(connq);
324 static void hammer2_lnk_span(hammer2_state_t *state, hammer2_msg_t *msg);
325 static void hammer2_lnk_conn(hammer2_state_t *state, hammer2_msg_t *msg);
326 static void hammer2_lnk_relay(hammer2_state_t *state, hammer2_msg_t *msg);
327 static void hammer2_relay_scan(h2span_connect_t *conn, h2span_node_t *node);
328 static void hammer2_relay_delete(h2span_relay_t *relay);
331 * Receive a HAMMER2_MSG_PROTO_LNK message. This only called for
332 * one-way and opening-transactions since state->func will be assigned
333 * in all other cases.
336 hammer2_msg_lnk(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
338 switch(msg->any.head.cmd & HAMMER2_MSGF_BASECMDMASK) {
339 case HAMMER2_LNK_CONN:
340 hammer2_lnk_conn(msg->state, msg);
342 case HAMMER2_LNK_SPAN:
343 hammer2_lnk_span(msg->state, msg);
347 "MSG_PROTO_LNK: Unknown msg %08x\n", msg->any.head.cmd);
348 hammer2_msg_reply(iocom, msg, HAMMER2_MSG_ERR_UNKNOWN);
349 /* state invalid after reply */
355 hammer2_lnk_conn(hammer2_state_t *state, hammer2_msg_t *msg)
357 h2span_connect_t *conn;
358 h2span_relay_t *relay;
361 pthread_mutex_lock(&cluster_mtx);
364 * On transaction start we allocate a new h2span_connect and
365 * acknowledge the request, leaving the transaction open.
366 * We then relay priority-selected SPANs.
368 if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
369 state->func = hammer2_lnk_conn;
371 fprintf(stderr, "LNK_CONN(%016jx): %s/%s\n",
372 (intmax_t)msg->any.head.msgid,
373 hammer2_uuid_to_str(&msg->any.lnk_conn.pfs_clid,
375 msg->any.lnk_conn.label);
378 conn = hammer2_alloc(sizeof(*conn));
380 RB_INIT(&conn->tree);
382 state->any.conn = conn;
383 TAILQ_INSERT_TAIL(&connq, conn, entry);
385 hammer2_msg_result(state->iocom, msg, 0);
388 * Span-synchronize all nodes with the new connection
390 hammer2_relay_scan(conn, NULL);
394 * On transaction terminate we clean out our h2span_connect
395 * and acknowledge the request, closing the transaction.
397 if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
398 fprintf(stderr, "LNK_CONN: Terminated\n");
399 conn = state->any.conn;
403 * Clean out all relays. This requires terminating each
406 while ((relay = RB_ROOT(&conn->tree)) != NULL) {
407 hammer2_relay_delete(relay);
414 msg->state->any.conn = NULL;
415 TAILQ_REMOVE(&connq, conn, entry);
418 hammer2_msg_reply(state->iocom, msg, 0);
419 /* state invalid after reply */
421 pthread_mutex_unlock(&cluster_mtx);
425 hammer2_lnk_span(hammer2_state_t *state, hammer2_msg_t *msg)
427 h2span_cluster_t dummy_cls;
428 h2span_node_t dummy_node;
429 h2span_cluster_t *cls;
431 h2span_link_t *slink;
432 h2span_relay_t *relay;
435 pthread_mutex_lock(&cluster_mtx);
438 * On transaction start we initialize the tracking infrastructure
440 if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
441 state->func = hammer2_lnk_span;
443 fprintf(stderr, "LNK_SPAN: %s/%s\n",
444 hammer2_uuid_to_str(&msg->any.lnk_span.pfs_clid,
446 msg->any.lnk_span.label);
452 dummy_cls.pfs_clid = msg->any.lnk_span.pfs_clid;
453 cls = RB_FIND(h2span_cluster_tree, &cluster_tree, &dummy_cls);
455 cls = hammer2_alloc(sizeof(*cls));
456 cls->pfs_clid = msg->any.lnk_span.pfs_clid;
458 RB_INSERT(h2span_cluster_tree, &cluster_tree, cls);
464 dummy_node.pfs_fsid = msg->any.lnk_span.pfs_fsid;
465 node = RB_FIND(h2span_node_tree, &cls->tree, &dummy_node);
467 node = hammer2_alloc(sizeof(*node));
468 node->pfs_fsid = msg->any.lnk_span.pfs_fsid;
470 RB_INIT(&node->tree);
471 RB_INSERT(h2span_node_tree, &cls->tree, node);
477 assert(state->any.link == NULL);
478 slink = hammer2_alloc(sizeof(*slink));
479 TAILQ_INIT(&slink->relayq);
481 slink->dist = msg->any.lnk_span.dist;
482 slink->state = state;
483 state->any.link = slink;
484 RB_INSERT(h2span_link_tree, &node->tree, slink);
486 hammer2_relay_scan(NULL, node);
490 * On transaction terminate we remove the tracking infrastructure.
492 if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
493 slink = state->any.link;
494 assert(slink != NULL);
499 * Clean out all relays. This requires terminating each
502 while ((relay = TAILQ_FIRST(&slink->relayq)) != NULL) {
503 hammer2_relay_delete(relay);
507 * Clean out the topology
509 RB_REMOVE(h2span_link_tree, &node->tree, slink);
510 if (RB_EMPTY(&node->tree)) {
511 RB_REMOVE(h2span_node_tree, &cls->tree, node);
512 if (RB_EMPTY(&cls->tree)) {
513 RB_REMOVE(h2span_cluster_tree,
521 state->any.link = NULL;
527 * We have to terminate the transaction
529 hammer2_state_reply(state, 0);
530 /* state invalid after reply */
533 * If the node still exists issue any required updates. If
534 * it doesn't then all related relays have already been
535 * removed and there's nothing left to do.
538 hammer2_relay_scan(NULL, node);
541 pthread_mutex_unlock(&cluster_mtx);
545 * Messages received on relay SPANs. These are open transactions so it is
546 * in fact possible for the other end to close the transaction.
548 * XXX MPRACE on state structure
551 hammer2_lnk_relay(hammer2_state_t *state, hammer2_msg_t *msg)
553 h2span_relay_t *relay;
555 if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
556 pthread_mutex_lock(&cluster_mtx);
557 if ((relay = state->any.relay) != NULL) {
558 hammer2_relay_delete(relay);
560 hammer2_state_reply(state, 0);
562 pthread_mutex_unlock(&cluster_mtx);
567 * Update relay transactions for SPANs.
569 * Called with cluster_mtx held.
571 static void hammer2_relay_scan_specific(h2span_node_t *node,
572 h2span_connect_t *conn);
575 hammer2_relay_scan(h2span_connect_t *conn, h2span_node_t *node)
577 h2span_cluster_t *cls;
581 * Iterate specific node
583 TAILQ_FOREACH(conn, &connq, entry)
584 hammer2_relay_scan_specific(node, conn);
589 * Iterate cluster ids, nodes, and either a specific connection
590 * or all connections.
592 RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
596 RB_FOREACH(node, h2span_node_tree, &cls->tree) {
598 * Synchronize the node's link (received SPANs)
599 * with each connection's relays.
602 hammer2_relay_scan_specific(node, conn);
604 TAILQ_FOREACH(conn, &connq, entry) {
605 hammer2_relay_scan_specific(node,
608 assert(conn == NULL);
616 * Update the relay'd SPANs for this (node, conn).
618 * Iterate links and adjust relays to match. We only propagate the top link
619 * for now (XXX we want to propagate the top two).
621 * The hammer2_relay_scan_cmp() function locates the first relay element
622 * for any given node. The relay elements will be sub-sorted by dist.
624 struct relay_scan_info {
626 h2span_relay_t *relay;
630 hammer2_relay_scan_cmp(h2span_relay_t *relay, void *arg)
632 struct relay_scan_info *info = arg;
634 if ((intptr_t)relay->link->node < (intptr_t)info->node)
636 if ((intptr_t)relay->link->node > (intptr_t)info->node)
642 hammer2_relay_scan_callback(h2span_relay_t *relay, void *arg)
644 struct relay_scan_info *info = arg;
651 hammer2_relay_scan_specific(h2span_node_t *node, h2span_connect_t *conn)
653 struct relay_scan_info info;
654 h2span_relay_t *relay;
655 h2span_relay_t *next_relay;
656 h2span_link_t *slink;
663 * Locate the first related relay for the connection. relay will
664 * be NULL if there were none.
666 RB_SCAN(h2span_relay_tree, &conn->tree,
667 hammer2_relay_scan_cmp, hammer2_relay_scan_callback, &info);
670 fprintf(stderr, "relay scan for connection %p\n", conn);
673 * Iterate the node's links (received SPANs) in distance order,
674 * lowest (best) dist first.
676 RB_FOREACH(slink, h2span_link_tree, &node->tree) {
678 * PROPAGATE THE BEST RELAYS BY TRANSMITTING SPANs.
680 * Check for match against current best relay.
682 * A match failure means that the current best relay is not
683 * as good as the link, create a new relay for the link.
685 * (If some prior better link was removed it would have also
686 * removed the relay, so the relay can only match exactly or
690 if (relay == NULL || relay->link != slink) {
693 assert(relay == NULL ||
694 relay->link->dist <= slink->dist);
695 relay = hammer2_alloc(sizeof(*relay));
699 msg = hammer2_msg_alloc(conn->state->iocom, 0,
701 HAMMER2_MSGF_CREATE);
702 msg->any.lnk_span = slink->state->msg->any.lnk_span;
703 ++msg->any.lnk_span.dist; /* XXX add weighting */
705 hammer2_msg_write(conn->state->iocom, msg,
706 hammer2_lnk_relay, relay,
708 fprintf(stderr, "RELAY SPAN ON CLS=%p NODE=%p FD %d state %p\n",
710 conn->state->iocom->sock_fd, relay->state);
712 RB_INSERT(h2span_relay_tree, &conn->tree, relay);
713 TAILQ_INSERT_TAIL(&slink->relayq, relay, entry);
717 * Iterate, figure out the next relay.
719 relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
727 * Any remaining relay's belonging to this connection which match
728 * the node are in excess of the current aggregate spanning state
729 * and should be removed.
731 while (relay && relay->link->node == node) {
732 next_relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
733 hammer2_relay_delete(relay);
740 hammer2_relay_delete(h2span_relay_t *relay)
742 fprintf(stderr, "RELAY DELETE ON CLS=%p NODE=%p FD %d STATE %p\n",
743 relay->link->node->cls, relay->link->node,
744 relay->conn->state->iocom->sock_fd, relay->state);
745 fprintf(stderr, "RELAY TX %08x RX %08x\n", relay->state->txcmd, relay->state->rxcmd);
747 RB_REMOVE(h2span_relay_tree, &relay->conn->tree, relay);
748 TAILQ_REMOVE(&relay->link->relayq, relay, entry);
751 relay->state->any.relay = NULL;
752 hammer2_state_reply(relay->state, 0);
753 /* state invalid after reply */