2 * Copyright (c) 2012-2014 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 - Please see sys/dmsg.h for an
36 * involved explanation of the protocol.
39 #include "dmsg_local.h"
42 * Maximum spanning tree distance. This has the practical effect of
43 * stopping tail-chasing closed loops when a feeder span is lost.
45 #define DMSG_SPAN_MAXDIST 16
48 * RED-BLACK TREE DEFINITIONS
52 * (1) shared fsid's (a cluster).
53 * (2) unique fsid's (a node in a cluster) <--- LNK_SPAN transactions.
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.
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
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.
72 * h2span_cluster - Organizes the shared fsid's. One structure for
75 * h2span_node - Organizes the nodes in a cluster. One structure
76 * for each unique {cluster,node}, aka {fsid, pfs_fsid}.
78 * h2span_link - Organizes all incoming and outgoing LNK_SPAN message
79 * transactions related to a node.
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).
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
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
103 TAILQ_HEAD(h2span_conn_queue, h2span_conn);
104 TAILQ_HEAD(h2span_relay_queue, h2span_relay);
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);
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.
118 TAILQ_ENTRY(h2span_conn) entry;
119 struct h2span_relay_tree tree;
121 dmsg_lnk_conn_t lnk_conn;
125 * All received LNK_SPANs are organized by cluster (pfs_clid),
126 * node (pfs_fsid), and link (received LNK_SPAN transaction).
128 struct h2span_cluster {
129 RB_ENTRY(h2span_cluster) rbnode;
130 struct h2span_node_tree tree;
131 uuid_t pfs_clid; /* shared fsid */
133 char cl_label[128]; /* cluster label (typ PEER_BLOCK) */
134 int refs; /* prevents destruction */
138 RB_ENTRY(h2span_node) rbnode;
139 struct h2span_link_tree tree;
140 struct h2span_cluster *cls;
142 uuid_t pfs_fsid; /* unique fsid */
143 char fs_label[128]; /* fs label (typ PEER_HAMMER2) */
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;
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.
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.
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).
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 */
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;
185 #define dmsg_termstr(array) _dmsg_termstr((array), sizeof(array))
187 static h2span_relay_t *dmsg_generate_relay(h2span_conn_t *conn,
188 h2span_link_t *slink);
189 static uint32_t dmsg_rnss(void);
193 _dmsg_termstr(char *base, size_t size)
199 * Cluster peer_type, uuid, AND label must match for a match
203 h2span_cluster_cmp(h2span_cluster_t *cls1, h2span_cluster_t *cls2)
207 if (cls1->peer_type < cls2->peer_type)
209 if (cls1->peer_type > cls2->peer_type)
211 r = uuid_compare(&cls1->pfs_clid, &cls2->pfs_clid, NULL);
213 r = strcmp(cls1->cl_label, cls2->cl_label);
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.
225 h2span_node_cmp(h2span_node_t *node1, h2span_node_t *node2)
229 r = strcmp(node1->fs_label, node2->fs_label);
231 r = uuid_compare(&node1->pfs_fsid, &node2->pfs_fsid, NULL);
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.
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.
245 h2span_link_cmp(h2span_link_t *link1, h2span_link_t *link2)
247 if (link1->lnk_span.dist < link2->lnk_span.dist)
249 if (link1->lnk_span.dist > link2->lnk_span.dist)
251 if (link1->lnk_span.rnss < link2->lnk_span.rnss)
253 if (link1->lnk_span.rnss > link2->lnk_span.rnss)
256 if ((uintptr_t)link1->state < (uintptr_t)link2->state)
258 if ((uintptr_t)link1->state > (uintptr_t)link2->state)
261 if (link1->state->msgid < link2->state->msgid)
263 if (link1->state->msgid > link2->state->msgid)
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).
276 h2span_relay_cmp(h2span_relay_t *relay1, h2span_relay_t *relay2)
278 h2span_link_t *link1 = relay1->source_rt->any.link;
279 h2span_link_t *link2 = relay2->source_rt->any.link;
281 if ((intptr_t)link1->node < (intptr_t)link2->node)
283 if ((intptr_t)link1->node > (intptr_t)link2->node)
285 if (link1->lnk_span.dist < link2->lnk_span.dist)
287 if (link1->lnk_span.dist > link2->lnk_span.dist)
289 if (link1->lnk_span.rnss < link2->lnk_span.rnss)
291 if (link1->lnk_span.rnss > link2->lnk_span.rnss)
294 if ((uintptr_t)link1->state < (uintptr_t)link2->state)
296 if ((uintptr_t)link1->state > (uintptr_t)link2->state)
299 if (link1->state->msgid < link2->state->msgid)
301 if (link1->state->msgid > link2->state->msgid)
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);
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);
326 * Global mutex protects cluster_tree lookups, connq, mediaq.
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);
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);
341 dmsg_msg_lnk_signal(dmsg_iocom_t *iocom __unused)
343 pthread_mutex_lock(&cluster_mtx);
344 dmsg_relay_scan(NULL, NULL);
345 pthread_mutex_unlock(&cluster_mtx);
349 * DMSG_PROTO_LNK - Generic DMSG_PROTO_LNK.
350 * (incoming iocom lock not held)
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.
357 dmsg_msg_lnk(dmsg_msg_t *msg)
359 dmsg_iocom_t *iocom = msg->state->iocom;
361 switch(msg->tcmd & DMSGF_BASECMDMASK) {
372 iocom->usrmsg_callback(msg, 1);
373 /* state invalid after reply */
379 * LNK_CONN - iocom identify message reception.
380 * (incoming iocom lock not held)
382 * Remote node identifies itself to us, sets up a SPAN filter, and gives us
383 * the ok to start transmitting SPANs.
386 dmsg_lnk_conn(dmsg_msg_t *msg)
388 dmsg_state_t *state = msg->state;
389 dmsg_iocom_t *iocom = state->iocom;
392 h2span_relay_t *relay;
395 pthread_mutex_lock(&cluster_mtx);
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);
403 switch(msg->any.head.cmd & DMSGF_TRANSMASK) {
404 case DMSG_LNK_CONN | DMSGF_CREATE:
405 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE:
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.
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,
415 msg->any.lnk_conn.cl_label,
416 msg->any.lnk_conn.fs_label);
419 conn = dmsg_alloc(sizeof(*conn));
420 assert(state->iocom->conn == NULL);
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);
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;
435 TAILQ_FOREACH(media, &mediaq, entry) {
436 if (uuid_compare(&msg->any.lnk_conn.mediaid,
437 &media->mediaid, NULL) == 0) {
442 media = dmsg_alloc(sizeof(*media));
443 media->mediaid = msg->any.lnk_conn.mediaid;
444 TAILQ_INSERT_TAIL(&mediaq, media, entry);
446 state->media = media;
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);
456 case DMSG_LNK_CONN | DMSGF_DELETE:
457 case DMSG_LNK_ERROR | DMSGF_DELETE:
459 * On transaction terminate we clean out our h2span_conn
460 * and acknowledge the request, closing the transaction.
462 dmio_printf(iocom, 3, "%s\n", "LNK_CONN: Terminated");
463 conn = state->any.conn;
469 * Callback will clean out media config / user-opaque state
471 media = state->media;
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);
484 * Clean out all relays. This requires terminating each
487 while ((relay = RB_ROOT(&conn->tree)) != NULL) {
488 dmsg_relay_delete(relay);
495 msg->state->any.conn = NULL;
496 msg->state->iocom->conn = NULL;
497 TAILQ_REMOVE(&connq, conn, entry);
500 dmsg_msg_reply(msg, 0);
501 dmsg_state_drop(state);
502 /* state invalid after reply */
505 iocom->usrmsg_callback(msg, 1);
507 if (msg->any.head.cmd & DMSGF_DELETE)
509 dmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
513 pthread_mutex_unlock(&cluster_mtx);
517 * LNK_SPAN - Spanning tree protocol message reception
518 * (incoming iocom lock not held)
520 * Receive a spanning tree transactional message, creating or destroying
521 * a SPAN and propagating it to other iocoms.
524 dmsg_lnk_span(dmsg_msg_t *msg)
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;
532 h2span_link_t *slink;
533 h2span_relay_t *relay;
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.
545 if (msg->any.head.cmd & DMSGF_REPLY) {
546 dmio_printf(iocom, 2, "%s\n",
547 "Ignore reply to LNK_SPAN");
551 pthread_mutex_lock(&cluster_mtx);
554 * On transaction start we initialize the tracking infrastructure
556 if (msg->any.head.cmd & DMSGF_CREATE) {
557 assert(state->func == NULL);
558 state->func = dmsg_lnk_span;
560 dmsg_termstr(msg->any.lnk_span.cl_label);
561 dmsg_termstr(msg->any.lnk_span.fs_label);
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,
570 sizeof(dummy_cls.cl_label));
571 cls = RB_FIND(h2span_cluster_tree, &cluster_tree, &dummy_cls);
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,
578 sizeof(cls->cl_label));
580 RB_INSERT(h2span_cluster_tree, &cluster_tree, cls);
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);
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,
596 sizeof(node->fs_label));
598 RB_INIT(&node->tree);
599 RB_INSERT(h2span_node_tree, &cls->tree, node);
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).
611 * NOTE: Later on when we relay the SPAN out the outgoing
612 * SPAN state will be set up to relay back to this
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.
621 assert(state->any.link == NULL);
622 dmsg_state_hold(state);
623 slink = dmsg_alloc(sizeof(*slink));
624 TAILQ_INIT(&slink->relayq);
626 slink->state = state;
627 state->any.link = slink;
628 slink->lnk_span = msg->any.lnk_span;
630 RB_INSERT(h2span_link_tree, &node->tree, slink);
632 dmio_printf(iocom, 3,
633 "LNK_SPAN(thr %p): %p %s cl=%s fs=%s dist=%d\n",
635 dmsg_uuid_to_str(&msg->any.lnk_span.pfs_clid,
637 msg->any.lnk_span.cl_label,
638 msg->any.lnk_span.fs_label,
639 msg->any.lnk_span.dist);
642 dmsg_relay_scan(NULL, node);
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.
649 dmsg_state_result(state, 0);
650 dmsg_iocom_signal(iocom);
654 * On transaction terminate we remove the tracking infrastructure.
656 if (msg->any.head.cmd & DMSGF_DELETE) {
657 slink = state->any.link;
658 assert(slink->state == state);
659 assert(slink != NULL);
663 dmio_printf(iocom, 3,
664 "LNK_DELE(thr %p): %p %s cl=%s fs=%s\n",
666 dmsg_uuid_to_str(&cls->pfs_clid, &alloc),
672 * Clean out all relays. This requires terminating each
675 while ((relay = TAILQ_FIRST(&slink->relayq)) != NULL) {
676 dmsg_relay_delete(relay);
680 * Clean out the topology
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,
694 state->any.link = NULL;
697 dmsg_state_drop(state);
701 * We have to terminate the transaction
703 dmsg_state_reply(state, 0);
704 /* state invalid after reply */
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.
713 dmsg_relay_scan(NULL, node);
716 dmsg_iocom_signal(iocom);
719 pthread_mutex_unlock(&cluster_mtx);
723 * Respond to a PING with a PING|REPLY, forward replies to the usermsg
728 dmsg_lnk_ping(dmsg_msg_t *msg)
732 if (msg->any.head.cmd & DMSGF_REPLY) {
733 msg->state->iocom->usrmsg_callback(msg, 1);
735 rep = dmsg_msg_alloc(msg->state, 0,
736 DMSG_LNK_PING | DMSGF_REPLY,
743 * Update relay transactions for SPANs.
745 * Called with cluster_mtx held.
747 static void dmsg_relay_scan_specific(h2span_node_t *node,
748 h2span_conn_t *conn);
751 dmsg_relay_scan(h2span_conn_t *conn, h2span_node_t *node)
753 h2span_cluster_t *cls;
757 * Iterate specific node
759 TAILQ_FOREACH(conn, &connq, entry)
760 dmsg_relay_scan_specific(node, conn);
765 * Iterate cluster ids, nodes, and either a specific connection
766 * or all connections.
768 RB_FOREACH(cls, h2span_cluster_tree, &cluster_tree) {
772 RB_FOREACH(node, h2span_node_tree, &cls->tree) {
774 * Synchronize the node's link (received SPANs)
775 * with each connection's relays.
778 dmsg_relay_scan_specific(node, conn);
780 TAILQ_FOREACH(conn, &connq, entry) {
781 dmsg_relay_scan_specific(node,
784 assert(conn == NULL);
792 * Update the relay'd SPANs for this (node, conn).
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).
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.
800 struct relay_scan_info {
802 h2span_relay_t *relay;
806 dmsg_relay_scan_cmp(h2span_relay_t *relay, void *arg)
808 struct relay_scan_info *info = arg;
810 if ((intptr_t)relay->source_rt->any.link->node < (intptr_t)info->node)
812 if ((intptr_t)relay->source_rt->any.link->node > (intptr_t)info->node)
818 dmsg_relay_scan_callback(h2span_relay_t *relay, void *arg)
820 struct relay_scan_info *info = arg;
827 dmsg_relay_scan_specific(h2span_node_t *node, h2span_conn_t *conn)
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;
837 #ifdef REQUIRE_SYMMETRICAL
838 uint32_t lastdist = DMSG_SPAN_MAXDIST;
839 uint32_t lastrnss = 0;
846 * Locate the first related relay for the node on this connection.
847 * relay will be NULL if there were none.
849 RB_SCAN(h2span_relay_tree, &conn->tree,
850 dmsg_relay_scan_cmp, dmsg_relay_scan_callback, &info);
854 assert(relay->source_rt->any.link->node == node);
856 dm_printf(9, "relay scan for connection %p\n", conn);
859 * Iterate the node's links (received SPANs) in distance order,
860 * lowest (best) dist first.
862 * PROPAGATE THE BEST LINKS OVER THE SPECIFIED CONNECTION.
864 * Track relays while iterating the best links and construct
865 * missing relays when necessary.
867 * (If some prior better link was removed it would have also
868 * removed the relay, so the relay can only match exactly or
872 RB_FOREACH(slink, h2span_link_tree, &node->tree) {
874 * Increment count of successful relays. This isn't
875 * quite accurate if we break out but nothing after
876 * the loop uses (count).
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.
885 * NOTE: Spanning tree does not have to be symmetrical so
886 * this code is not currently enabled.
888 if (++count >= maxcount) {
889 #ifdef REQUIRE_SYMMETRICAL
890 if (lastdist != slink->lnk_span.dist ||
891 lastrnss != slink->lnk_span.rnss) {
897 /* go beyond the nominal maximum desired relays */
901 * Match, relay already in-place, get the next
902 * relay to match against the next slink.
904 if (relay && relay->source_rt->any.link == slink) {
905 relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
910 * We might want this SLINK, if it passes our filters.
912 * The spanning tree can cause closed loops so we have
913 * to limit slink->dist.
915 if (slink->lnk_span.dist > DMSG_SPAN_MAXDIST)
919 * Don't bother transmitting a LNK_SPAN out the same
920 * connection it came in on. Trivial optimization.
922 if (slink->state->iocom == conn->state->iocom)
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.
930 * Don't bother transmitting if the remote connection
931 * is not accepting this SPAN's peer_type.
933 * pfs_mask is typically used so pure clients can filter
934 * out receiving SPANs for other pure clients.
936 lspan = &slink->lnk_span;
937 lconn = &conn->lnk_conn;
938 if (((1LLU << lspan->peer_type) & lconn->peer_mask) == 0)
940 if (((1LLU << lspan->pfs_type) & lconn->pfs_mask) == 0)
944 * Do not give pure clients visibility to other pure clients
946 if (lconn->pfs_type == DMSG_PFSTYPE_CLIENT &&
947 lspan->pfs_type == DMSG_PFSTYPE_CLIENT) {
952 * Connection filter, if cluster uuid is not NULL it must
953 * match the span cluster uuid. Only applies when the
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)) {
964 * Connection filter, if cluster label is not empty it must
965 * match the span cluster label. Only applies when the
968 if (lspan->peer_type == lconn->peer_type &&
969 lconn->cl_label[0] &&
970 strcmp(lconn->cl_label, slink->node->cls->cl_label)) {
975 * NOTE! pfs_fsid differentiates nodes within the same cluster
976 * so we obviously don't want to match those. Similarly
981 * Ok, we've accepted this SPAN for relaying.
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;
994 * Match (created new relay), get the next relay to
995 * match against the next slink.
997 relay = RB_NEXT(h2span_relay_tree, &conn->tree, relay);
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.
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);
1014 * Find the slink associated with the msgid and return its state,
1015 * so the caller can issue a transaction.
1018 dmsg_findspan(const char *label)
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);
1026 pthread_mutex_lock(&cluster_mtx);
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;
1040 pthread_mutex_unlock(&cluster_mtx);
1042 dm_printf(8, "findspan: %p\n", state);
1049 * Helper function to generate missing relay on target connection.
1051 * cluster_mtx must be held
1055 dmsg_generate_relay(h2span_conn_t *conn, h2span_link_t *slink)
1057 h2span_relay_t *relay;
1060 dmsg_state_hold(slink->state);
1061 relay = dmsg_alloc(sizeof(*relay));
1063 relay->source_rt = slink->state;
1064 /* relay->source_rt->any.link = slink; */
1067 * NOTE: relay->target_rt->any.relay set to relay by alloc.
1069 * NOTE: LNK_SPAN is transmitted as a top-level transaction.
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;
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();
1081 RB_INSERT(h2span_relay_tree, &conn->tree, relay);
1082 TAILQ_INSERT_TAIL(&slink->relayq, relay, entry);
1085 * Seed the relay so new sub-transactions received on the outgoing
1086 * SPAN circuit are relayed back to the originator.
1088 msg->state->relay = relay->source_rt;
1089 dmsg_state_hold(msg->state->relay);
1091 dmsg_msg_write(msg);
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.
1100 * XXX MPRACE on state structure
1103 dmsg_lnk_relay(dmsg_msg_t *msg)
1105 dmsg_state_t *state = msg->state;
1106 h2span_relay_t *relay;
1108 assert(msg->any.head.cmd & DMSGF_REPLY);
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);
1116 dmsg_state_reply(state, 0);
1118 pthread_mutex_unlock(&cluster_mtx);
1123 * cluster_mtx held by caller
1127 dmsg_relay_delete(h2span_relay_t *relay)
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,
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,
1140 RB_REMOVE(h2span_relay_tree, &relay->conn->tree, relay);
1141 TAILQ_REMOVE(&relay->source_rt->any.link->relayq, relay, entry);
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;
1152 * NOTE: relay->source_rt->refs is held by the relay SPAN
1153 * state, not by this relay structure.
1156 if (relay->source_rt) {
1157 dmsg_state_drop(relay->source_rt);
1158 relay->source_rt = NULL;
1163 /************************************************************************
1164 * ROUTER AND MESSAGING HANDLES *
1165 ************************************************************************
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.
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
1182 dmsg_cluster_get(uuid_t *pfs_clid)
1184 h2span_cluster_t dummy_cls;
1185 h2span_cluster_t *cls;
1187 dummy_cls.pfs_clid = *pfs_clid;
1188 pthread_mutex_lock(&cluster_mtx);
1189 cls = RB_FIND(h2span_cluster_tree, &cluster_tree, &dummy_cls);
1192 pthread_mutex_unlock(&cluster_mtx);
1197 dmsg_cluster_put(h2span_cluster_t *cls)
1199 pthread_mutex_lock(&cluster_mtx);
1200 assert(cls->refs > 0);
1202 if (RB_EMPTY(&cls->tree) && cls->refs == 0) {
1203 RB_REMOVE(h2span_cluster_tree,
1204 &cluster_tree, cls);
1207 pthread_mutex_unlock(&cluster_mtx);
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
1217 dmsg_node_get(h2span_cluster_t *cls, uuid_t *pfs_fsid)
1224 * Dumps the spanning tree
1229 dmsg_shell_tree(dmsg_iocom_t *iocom, char *cmdbuf __unused)
1231 h2span_cluster_t *cls;
1232 h2span_node_t *node;
1233 h2span_link_t *slink;
1234 h2span_relay_t *relay;
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),
1243 RB_FOREACH(node, h2span_node_tree, &cls->tree) {
1244 dmsg_printf(iocom, " Node %02x %s (%s)\n",
1246 dmsg_uuid_to_str(&node->pfs_fsid, &uustr),
1248 RB_FOREACH(slink, h2span_link_tree, &node->tree) {
1250 "\tSLink msgid %016jx "
1252 (intmax_t)slink->state->msgid,
1253 slink->lnk_span.dist,
1254 slink->state->iocom->sock_fd);
1255 TAILQ_FOREACH(relay, &slink->relayq, entry) {
1257 "\t Relay-out msgid %016jx "
1259 (intmax_t)relay->target_rt->msgid,
1260 relay->target_rt->iocom->sock_fd);
1265 pthread_mutex_unlock(&cluster_mtx);
1269 TAILQ_FOREACH(conn, &connq, entry) {
1277 * Locate the state representing an incoming LNK_SPAN given its msgid.
1280 dmsg_debug_findspan(uint64_t msgid, dmsg_state_t **statep)
1282 h2span_cluster_t *cls;
1283 h2span_node_t *node;
1284 h2span_link_t *slink;
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;
1297 pthread_mutex_unlock(&cluster_mtx);
1301 pthread_mutex_unlock(&cluster_mtx);
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.
1318 if (DMsgRNSS == 0) {
1319 pthread_mutex_lock(&cluster_mtx);
1320 while (DMsgRNSS == 0) {
1322 DMsgRNSS = random();
1324 pthread_mutex_unlock(&cluster_mtx);