ee9a8e9b7c7f9c94bb9fc7e3518ebfc127f2703b
[dragonfly.git] / lib / libdmsg / msg.c
1 /*
2  * Copyright (c) 2011-2012 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  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in
16  *    the documentation and/or other materials provided with the
17  *    distribution.
18  * 3. Neither the name of The DragonFly Project nor the names of its
19  *    contributors may be used to endorse or promote products derived
20  *    from this software without specific, prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35
36 #include "dmsg_local.h"
37
38 int DMsgDebugOpt;
39
40 static int dmsg_state_msgrx(dmsg_msg_t *msg);
41 static void dmsg_state_cleanuptx(dmsg_msg_t *msg);
42
43 /*
44  * ROUTER TREE - Represents available routes for message routing, indexed
45  *               by their span transaction id.  The router structure is
46  *               embedded in either an iocom, h2span_link (incoming),
47  *               or h2span_relay (outgoing) (see msg_lnk.c).
48  */
49 int
50 dmsg_router_cmp(dmsg_router_t *router1, dmsg_router_t *router2)
51 {
52         if (router1->target < router2->target)
53                 return(-1);
54         if (router1->target > router2->target)
55                 return(1);
56         return(0);
57 }
58
59 RB_GENERATE(dmsg_router_tree, dmsg_router, rbnode, dmsg_router_cmp);
60
61 static pthread_mutex_t router_mtx;
62 static struct dmsg_router_tree router_ltree = RB_INITIALIZER(router_ltree);
63 static struct dmsg_router_tree router_rtree = RB_INITIALIZER(router_rtree);
64
65 /*
66  * STATE TREE - Represents open transactions which are indexed by their
67  *              {router,msgid} relative to the governing iocom.
68  *
69  *              router is usually iocom->router since state isn't stored
70  *              for relayed messages.
71  */
72 int
73 dmsg_state_cmp(dmsg_state_t *state1, dmsg_state_t *state2)
74 {
75 #if 0
76         if (state1->router < state2->router)
77                 return(-1);
78         if (state1->router > state2->router)
79                 return(1);
80 #endif
81         if (state1->msgid < state2->msgid)
82                 return(-1);
83         if (state1->msgid > state2->msgid)
84                 return(1);
85         return(0);
86 }
87
88 RB_GENERATE(dmsg_state_tree, dmsg_state, rbnode, dmsg_state_cmp);
89
90 /*
91  * Initialize a low-level ioq
92  */
93 void
94 dmsg_ioq_init(dmsg_iocom_t *iocom __unused, dmsg_ioq_t *ioq)
95 {
96         bzero(ioq, sizeof(*ioq));
97         ioq->state = DMSG_MSGQ_STATE_HEADER1;
98         TAILQ_INIT(&ioq->msgq);
99 }
100
101 /*
102  * Cleanup queue.
103  *
104  * caller holds iocom->mtx.
105  */
106 void
107 dmsg_ioq_done(dmsg_iocom_t *iocom __unused, dmsg_ioq_t *ioq)
108 {
109         dmsg_msg_t *msg;
110
111         while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
112                 assert(0);      /* shouldn't happen */
113                 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
114                 dmsg_msg_free(msg);
115         }
116         if ((msg = ioq->msg) != NULL) {
117                 ioq->msg = NULL;
118                 dmsg_msg_free(msg);
119         }
120 }
121
122 /*
123  * Initialize a low-level communications channel.
124  *
125  * NOTE: The signal_func() is called at least once from the loop and can be
126  *       re-armed via dmsg_iocom_restate().
127  */
128 void
129 dmsg_iocom_init(dmsg_iocom_t *iocom, int sock_fd, int alt_fd,
130                    void (*signal_func)(dmsg_router_t *),
131                    void (*rcvmsg_func)(dmsg_msg_t *),
132                    void (*dbgmsg_func)(dmsg_msg_t *),
133                    void (*altmsg_func)(dmsg_iocom_t *))
134 {
135         struct stat st;
136
137         bzero(iocom, sizeof(*iocom));
138
139         iocom->router = dmsg_router_alloc();
140         iocom->router->signal_callback = signal_func;
141         iocom->router->rcvmsg_callback = rcvmsg_func;
142         iocom->router->altmsg_callback = altmsg_func;
143         iocom->router->dbgmsg_callback = dbgmsg_func;
144         /* we do not call dmsg_router_connect() for iocom routers */
145
146         pthread_mutex_init(&iocom->mtx, NULL);
147         RB_INIT(&iocom->router->staterd_tree);
148         RB_INIT(&iocom->router->statewr_tree);
149         TAILQ_INIT(&iocom->freeq);
150         TAILQ_INIT(&iocom->freeq_aux);
151         TAILQ_INIT(&iocom->router->txmsgq);
152         iocom->router->iocom = iocom;
153         iocom->sock_fd = sock_fd;
154         iocom->alt_fd = alt_fd;
155         iocom->flags = DMSG_IOCOMF_RREQ;
156         if (signal_func)
157                 iocom->flags |= DMSG_IOCOMF_SWORK;
158         dmsg_ioq_init(iocom, &iocom->ioq_rx);
159         dmsg_ioq_init(iocom, &iocom->ioq_tx);
160         if (pipe(iocom->wakeupfds) < 0)
161                 assert(0);
162         fcntl(iocom->wakeupfds[0], F_SETFL, O_NONBLOCK);
163         fcntl(iocom->wakeupfds[1], F_SETFL, O_NONBLOCK);
164
165         /*
166          * Negotiate session crypto synchronously.  This will mark the
167          * connection as error'd if it fails.  If this is a pipe it's
168          * a linkage that we set up ourselves to the filesystem and there
169          * is no crypto.
170          */
171         if (fstat(sock_fd, &st) < 0)
172                 assert(0);
173         if (S_ISSOCK(st.st_mode))
174                 dmsg_crypto_negotiate(iocom);
175
176         /*
177          * Make sure our fds are set to non-blocking for the iocom core.
178          */
179         if (sock_fd >= 0)
180                 fcntl(sock_fd, F_SETFL, O_NONBLOCK);
181 #if 0
182         /* if line buffered our single fgets() should be fine */
183         if (alt_fd >= 0)
184                 fcntl(alt_fd, F_SETFL, O_NONBLOCK);
185 #endif
186 }
187
188 /*
189  * May only be called from a callback from iocom_core.
190  *
191  * Adjust state machine functions, set flags to guarantee that both
192  * the recevmsg_func and the sendmsg_func is called at least once.
193  */
194 void
195 dmsg_router_restate(dmsg_router_t *router,
196                    void (*signal_func)(dmsg_router_t *),
197                    void (*rcvmsg_func)(dmsg_msg_t *msg),
198                    void (*altmsg_func)(dmsg_iocom_t *))
199 {
200         router->signal_callback = signal_func;
201         router->rcvmsg_callback = rcvmsg_func;
202         router->altmsg_callback = altmsg_func;
203         if (signal_func)
204                 router->iocom->flags |= DMSG_IOCOMF_SWORK;
205         else
206                 router->iocom->flags &= ~DMSG_IOCOMF_SWORK;
207 }
208
209 void
210 dmsg_router_signal(dmsg_router_t *router)
211 {
212         if (router->signal_callback)
213                 router->iocom->flags |= DMSG_IOCOMF_SWORK;
214 }
215
216 /*
217  * Cleanup a terminating iocom.
218  *
219  * Caller should not hold iocom->mtx.  The iocom has already been disconnected
220  * from all possible references to it.
221  */
222 void
223 dmsg_iocom_done(dmsg_iocom_t *iocom)
224 {
225         dmsg_msg_t *msg;
226
227         if (iocom->sock_fd >= 0) {
228                 close(iocom->sock_fd);
229                 iocom->sock_fd = -1;
230         }
231         if (iocom->alt_fd >= 0) {
232                 close(iocom->alt_fd);
233                 iocom->alt_fd = -1;
234         }
235         dmsg_ioq_done(iocom, &iocom->ioq_rx);
236         dmsg_ioq_done(iocom, &iocom->ioq_tx);
237         if ((msg = TAILQ_FIRST(&iocom->freeq)) != NULL) {
238                 TAILQ_REMOVE(&iocom->freeq, msg, qentry);
239                 free(msg);
240         }
241         if ((msg = TAILQ_FIRST(&iocom->freeq_aux)) != NULL) {
242                 TAILQ_REMOVE(&iocom->freeq_aux, msg, qentry);
243                 free(msg->aux_data);
244                 msg->aux_data = NULL;
245                 free(msg);
246         }
247         if (iocom->wakeupfds[0] >= 0) {
248                 close(iocom->wakeupfds[0]);
249                 iocom->wakeupfds[0] = -1;
250         }
251         if (iocom->wakeupfds[1] >= 0) {
252                 close(iocom->wakeupfds[1]);
253                 iocom->wakeupfds[1] = -1;
254         }
255         pthread_mutex_destroy(&iocom->mtx);
256 }
257
258 /*
259  * Allocate a new one-way message.
260  */
261 dmsg_msg_t *
262 dmsg_msg_alloc(dmsg_router_t *router, size_t aux_size, uint32_t cmd,
263                   void (*func)(dmsg_msg_t *), void *data)
264 {
265         dmsg_state_t *state = NULL;
266         dmsg_iocom_t *iocom = router->iocom;
267         dmsg_msg_t *msg;
268         int hbytes;
269
270         pthread_mutex_lock(&iocom->mtx);
271         if (aux_size) {
272                 aux_size = (aux_size + DMSG_ALIGNMASK) &
273                            ~DMSG_ALIGNMASK;
274                 if ((msg = TAILQ_FIRST(&iocom->freeq_aux)) != NULL)
275                         TAILQ_REMOVE(&iocom->freeq_aux, msg, qentry);
276         } else {
277                 if ((msg = TAILQ_FIRST(&iocom->freeq)) != NULL)
278                         TAILQ_REMOVE(&iocom->freeq, msg, qentry);
279         }
280         if ((cmd & (DMSGF_CREATE | DMSGF_REPLY)) == DMSGF_CREATE) {
281                 /*
282                  * Create state when CREATE is set without REPLY.
283                  *
284                  * NOTE: CREATE in txcmd handled by dmsg_msg_write()
285                  * NOTE: DELETE in txcmd handled by dmsg_state_cleanuptx()
286                  */
287                 state = malloc(sizeof(*state));
288                 bzero(state, sizeof(*state));
289                 state->iocom = iocom;
290                 state->flags = DMSG_STATE_DYNAMIC;
291                 state->msgid = (uint64_t)(uintptr_t)state;
292                 state->router = router;
293                 state->txcmd = cmd & ~(DMSGF_CREATE | DMSGF_DELETE);
294                 state->rxcmd = DMSGF_REPLY;
295                 state->func = func;
296                 state->any.any = data;
297                 pthread_mutex_lock(&iocom->mtx);
298                 RB_INSERT(dmsg_state_tree,
299                           &iocom->router->statewr_tree,
300                           state);
301                 pthread_mutex_unlock(&iocom->mtx);
302                 state->flags |= DMSG_STATE_INSERTED;
303         }
304         pthread_mutex_unlock(&iocom->mtx);
305         if (msg == NULL) {
306                 msg = malloc(sizeof(*msg));
307                 bzero(msg, sizeof(*msg));
308                 msg->aux_data = NULL;
309                 msg->aux_size = 0;
310         }
311         if (msg->aux_size != aux_size) {
312                 if (msg->aux_data) {
313                         free(msg->aux_data);
314                         msg->aux_data = NULL;
315                         msg->aux_size = 0;
316                 }
317                 if (aux_size) {
318                         msg->aux_data = malloc(aux_size);
319                         msg->aux_size = aux_size;
320                 }
321         }
322         hbytes = (cmd & DMSGF_SIZE) * DMSG_ALIGN;
323         if (hbytes)
324                 bzero(&msg->any.head, hbytes);
325         msg->hdr_size = hbytes;
326         msg->any.head.cmd = cmd;
327         msg->any.head.aux_descr = 0;
328         msg->any.head.aux_crc = 0;
329         msg->router = router;
330         if (state) {
331                 msg->state = state;
332                 state->msg = msg;
333                 msg->any.head.msgid = state->msgid;
334         }
335         return (msg);
336 }
337
338 /*
339  * Free a message so it can be reused afresh.
340  *
341  * NOTE: aux_size can be 0 with a non-NULL aux_data.
342  */
343 static
344 void
345 dmsg_msg_free_locked(dmsg_msg_t *msg)
346 {
347         dmsg_iocom_t *iocom = msg->router->iocom;
348
349         msg->state = NULL;
350         if (msg->aux_data)
351                 TAILQ_INSERT_TAIL(&iocom->freeq_aux, msg, qentry);
352         else
353                 TAILQ_INSERT_TAIL(&iocom->freeq, msg, qentry);
354 }
355
356 void
357 dmsg_msg_free(dmsg_msg_t *msg)
358 {
359         dmsg_iocom_t *iocom = msg->router->iocom;
360
361         pthread_mutex_lock(&iocom->mtx);
362         dmsg_msg_free_locked(msg);
363         pthread_mutex_unlock(&iocom->mtx);
364 }
365
366 /*
367  * I/O core loop for an iocom.
368  *
369  * Thread localized, iocom->mtx not held.
370  */
371 void
372 dmsg_iocom_core(dmsg_iocom_t *iocom)
373 {
374         struct pollfd fds[3];
375         char dummybuf[256];
376         dmsg_msg_t *msg;
377         int timeout;
378         int count;
379         int wi; /* wakeup pipe */
380         int si; /* socket */
381         int ai; /* alt bulk path socket */
382
383         while ((iocom->flags & DMSG_IOCOMF_EOF) == 0) {
384                 if ((iocom->flags & (DMSG_IOCOMF_RWORK |
385                                      DMSG_IOCOMF_WWORK |
386                                      DMSG_IOCOMF_PWORK |
387                                      DMSG_IOCOMF_SWORK |
388                                      DMSG_IOCOMF_ARWORK |
389                                      DMSG_IOCOMF_AWWORK)) == 0) {
390                         /*
391                          * Only poll if no immediate work is pending.
392                          * Otherwise we are just wasting our time calling
393                          * poll.
394                          */
395                         timeout = 5000;
396
397                         count = 0;
398                         wi = -1;
399                         si = -1;
400                         ai = -1;
401
402                         /*
403                          * Always check the inter-thread pipe, e.g.
404                          * for iocom->txmsgq work.
405                          */
406                         wi = count++;
407                         fds[wi].fd = iocom->wakeupfds[0];
408                         fds[wi].events = POLLIN;
409                         fds[wi].revents = 0;
410
411                         /*
412                          * Check the socket input/output direction as
413                          * requested
414                          */
415                         if (iocom->flags & (DMSG_IOCOMF_RREQ |
416                                             DMSG_IOCOMF_WREQ)) {
417                                 si = count++;
418                                 fds[si].fd = iocom->sock_fd;
419                                 fds[si].events = 0;
420                                 fds[si].revents = 0;
421
422                                 if (iocom->flags & DMSG_IOCOMF_RREQ)
423                                         fds[si].events |= POLLIN;
424                                 if (iocom->flags & DMSG_IOCOMF_WREQ)
425                                         fds[si].events |= POLLOUT;
426                         }
427
428                         /*
429                          * Check the alternative fd for work.
430                          */
431                         if (iocom->alt_fd >= 0) {
432                                 ai = count++;
433                                 fds[ai].fd = iocom->alt_fd;
434                                 fds[ai].events = POLLIN;
435                                 fds[ai].revents = 0;
436                         }
437                         poll(fds, count, timeout);
438
439                         if (wi >= 0 && (fds[wi].revents & POLLIN))
440                                 iocom->flags |= DMSG_IOCOMF_PWORK;
441                         if (si >= 0 && (fds[si].revents & POLLIN))
442                                 iocom->flags |= DMSG_IOCOMF_RWORK;
443                         if (si >= 0 && (fds[si].revents & POLLOUT))
444                                 iocom->flags |= DMSG_IOCOMF_WWORK;
445                         if (wi >= 0 && (fds[wi].revents & POLLOUT))
446                                 iocom->flags |= DMSG_IOCOMF_WWORK;
447                         if (ai >= 0 && (fds[ai].revents & POLLIN))
448                                 iocom->flags |= DMSG_IOCOMF_ARWORK;
449                 } else {
450                         /*
451                          * Always check the pipe
452                          */
453                         iocom->flags |= DMSG_IOCOMF_PWORK;
454                 }
455
456                 if (iocom->flags & DMSG_IOCOMF_SWORK) {
457                         iocom->flags &= ~DMSG_IOCOMF_SWORK;
458                         iocom->router->signal_callback(iocom->router);
459                 }
460
461                 /*
462                  * Pending message queues from other threads wake us up
463                  * with a write to the wakeupfds[] pipe.  We have to clear
464                  * the pipe with a dummy read.
465                  */
466                 if (iocom->flags & DMSG_IOCOMF_PWORK) {
467                         iocom->flags &= ~DMSG_IOCOMF_PWORK;
468                         read(iocom->wakeupfds[0], dummybuf, sizeof(dummybuf));
469                         iocom->flags |= DMSG_IOCOMF_RWORK;
470                         iocom->flags |= DMSG_IOCOMF_WWORK;
471                         if (TAILQ_FIRST(&iocom->router->txmsgq))
472                                 dmsg_iocom_flush1(iocom);
473                 }
474
475                 /*
476                  * Message write sequencing
477                  */
478                 if (iocom->flags & DMSG_IOCOMF_WWORK)
479                         dmsg_iocom_flush1(iocom);
480
481                 /*
482                  * Message read sequencing.  Run this after the write
483                  * sequencing in case the write sequencing allowed another
484                  * auto-DELETE to occur on the read side.
485                  */
486                 if (iocom->flags & DMSG_IOCOMF_RWORK) {
487                         while ((iocom->flags & DMSG_IOCOMF_EOF) == 0 &&
488                                (msg = dmsg_ioq_read(iocom)) != NULL) {
489                                 if (DMsgDebugOpt) {
490                                         fprintf(stderr, "receive %s\n",
491                                                 dmsg_msg_str(msg));
492                                 }
493                                 iocom->router->rcvmsg_callback(msg);
494                                 dmsg_state_cleanuprx(iocom, msg);
495                         }
496                 }
497
498                 if (iocom->flags & DMSG_IOCOMF_ARWORK) {
499                         iocom->flags &= ~DMSG_IOCOMF_ARWORK;
500                         iocom->router->altmsg_callback(iocom);
501                 }
502         }
503 }
504
505 /*
506  * Make sure there's enough room in the FIFO to hold the
507  * needed data.
508  *
509  * Assume worst case encrypted form is 2x the size of the
510  * plaintext equivalent.
511  */
512 static
513 size_t
514 dmsg_ioq_makeroom(dmsg_ioq_t *ioq, size_t needed)
515 {
516         size_t bytes;
517         size_t nmax;
518
519         bytes = ioq->fifo_cdx - ioq->fifo_beg;
520         nmax = sizeof(ioq->buf) - ioq->fifo_end;
521         if (bytes + nmax / 2 < needed) {
522                 if (bytes) {
523                         bcopy(ioq->buf + ioq->fifo_beg,
524                               ioq->buf,
525                               bytes);
526                 }
527                 ioq->fifo_cdx -= ioq->fifo_beg;
528                 ioq->fifo_beg = 0;
529                 if (ioq->fifo_cdn < ioq->fifo_end) {
530                         bcopy(ioq->buf + ioq->fifo_cdn,
531                               ioq->buf + ioq->fifo_cdx,
532                               ioq->fifo_end - ioq->fifo_cdn);
533                 }
534                 ioq->fifo_end -= ioq->fifo_cdn - ioq->fifo_cdx;
535                 ioq->fifo_cdn = ioq->fifo_cdx;
536                 nmax = sizeof(ioq->buf) - ioq->fifo_end;
537         }
538         return(nmax);
539 }
540
541 /*
542  * Read the next ready message from the ioq, issuing I/O if needed.
543  * Caller should retry on a read-event when NULL is returned.
544  *
545  * If an error occurs during reception a DMSG_LNK_ERROR msg will
546  * be returned for each open transaction, then the ioq and iocom
547  * will be errored out and a non-transactional DMSG_LNK_ERROR
548  * msg will be returned as the final message.  The caller should not call
549  * us again after the final message is returned.
550  *
551  * Thread localized, iocom->mtx not held.
552  */
553 dmsg_msg_t *
554 dmsg_ioq_read(dmsg_iocom_t *iocom)
555 {
556         dmsg_ioq_t *ioq = &iocom->ioq_rx;
557         dmsg_msg_t *msg;
558         dmsg_state_t *state;
559         dmsg_hdr_t *head;
560         ssize_t n;
561         size_t bytes;
562         size_t nmax;
563         uint32_t xcrc32;
564         int error;
565
566 again:
567         iocom->flags &= ~(DMSG_IOCOMF_RREQ | DMSG_IOCOMF_RWORK);
568
569         /*
570          * If a message is already pending we can just remove and
571          * return it.  Message state has already been processed.
572          * (currently not implemented)
573          */
574         if ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
575                 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
576                 return (msg);
577         }
578
579         /*
580          * If the stream is errored out we stop processing it.
581          */
582         if (ioq->error)
583                 goto skip;
584
585         /*
586          * Message read in-progress (msg is NULL at the moment).  We don't
587          * allocate a msg until we have its core header.
588          */
589         nmax = sizeof(ioq->buf) - ioq->fifo_end;
590         bytes = ioq->fifo_cdx - ioq->fifo_beg;          /* already decrypted */
591         msg = ioq->msg;
592
593         switch(ioq->state) {
594         case DMSG_MSGQ_STATE_HEADER1:
595                 /*
596                  * Load the primary header, fail on any non-trivial read
597                  * error or on EOF.  Since the primary header is the same
598                  * size is the message alignment it will never straddle
599                  * the end of the buffer.
600                  */
601                 nmax = dmsg_ioq_makeroom(ioq, sizeof(msg->any.head));
602                 if (bytes < sizeof(msg->any.head)) {
603                         n = read(iocom->sock_fd,
604                                  ioq->buf + ioq->fifo_end,
605                                  nmax);
606                         if (n <= 0) {
607                                 if (n == 0) {
608                                         ioq->error = DMSG_IOQ_ERROR_EOF;
609                                         break;
610                                 }
611                                 if (errno != EINTR &&
612                                     errno != EINPROGRESS &&
613                                     errno != EAGAIN) {
614                                         ioq->error = DMSG_IOQ_ERROR_SOCK;
615                                         break;
616                                 }
617                                 n = 0;
618                                 /* fall through */
619                         }
620                         ioq->fifo_end += (size_t)n;
621                         nmax -= (size_t)n;
622                 }
623
624                 /*
625                  * Decrypt data received so far.  Data will be decrypted
626                  * in-place but might create gaps in the FIFO.  Partial
627                  * blocks are not immediately decrypted.
628                  *
629                  * WARNING!  The header might be in the wrong endian, we
630                  *           do not fix it up until we get the entire
631                  *           extended header.
632                  */
633                 if (iocom->flags & DMSG_IOCOMF_CRYPTED) {
634                         dmsg_crypto_decrypt(iocom, ioq);
635                 } else {
636                         ioq->fifo_cdx = ioq->fifo_end;
637                         ioq->fifo_cdn = ioq->fifo_end;
638                 }
639                 bytes = ioq->fifo_cdx - ioq->fifo_beg;
640
641                 /*
642                  * Insufficient data accumulated (msg is NULL, caller will
643                  * retry on event).
644                  */
645                 assert(msg == NULL);
646                 if (bytes < sizeof(msg->any.head))
647                         break;
648
649                 /*
650                  * Check and fixup the core header.  Note that the icrc
651                  * has to be calculated before any fixups, but the crc
652                  * fields in the msg may have to be swapped like everything
653                  * else.
654                  */
655                 head = (void *)(ioq->buf + ioq->fifo_beg);
656                 if (head->magic != DMSG_HDR_MAGIC &&
657                     head->magic != DMSG_HDR_MAGIC_REV) {
658                         ioq->error = DMSG_IOQ_ERROR_SYNC;
659                         break;
660                 }
661
662                 /*
663                  * Calculate the full header size and aux data size
664                  */
665                 if (head->magic == DMSG_HDR_MAGIC_REV) {
666                         ioq->hbytes = (bswap32(head->cmd) & DMSGF_SIZE) *
667                                       DMSG_ALIGN;
668                         ioq->abytes = bswap32(head->aux_bytes) *
669                                       DMSG_ALIGN;
670                 } else {
671                         ioq->hbytes = (head->cmd & DMSGF_SIZE) *
672                                       DMSG_ALIGN;
673                         ioq->abytes = head->aux_bytes * DMSG_ALIGN;
674                 }
675                 if (ioq->hbytes < sizeof(msg->any.head) ||
676                     ioq->hbytes > sizeof(msg->any) ||
677                     ioq->abytes > DMSG_AUX_MAX) {
678                         ioq->error = DMSG_IOQ_ERROR_FIELD;
679                         break;
680                 }
681
682                 /*
683                  * Allocate the message, the next state will fill it in.
684                  */
685                 msg = dmsg_msg_alloc(iocom->router, ioq->abytes, 0,
686                                         NULL, NULL);
687                 ioq->msg = msg;
688
689                 /*
690                  * Fall through to the next state.  Make sure that the
691                  * extended header does not straddle the end of the buffer.
692                  * We still want to issue larger reads into our buffer,
693                  * book-keeping is easier if we don't bcopy() yet.
694                  *
695                  * Make sure there is enough room for bloated encrypt data.
696                  */
697                 nmax = dmsg_ioq_makeroom(ioq, ioq->hbytes);
698                 ioq->state = DMSG_MSGQ_STATE_HEADER2;
699                 /* fall through */
700         case DMSG_MSGQ_STATE_HEADER2:
701                 /*
702                  * Fill out the extended header.
703                  */
704                 assert(msg != NULL);
705                 if (bytes < ioq->hbytes) {
706                         n = read(iocom->sock_fd,
707                                  ioq->buf + ioq->fifo_end,
708                                  nmax);
709                         if (n <= 0) {
710                                 if (n == 0) {
711                                         ioq->error = DMSG_IOQ_ERROR_EOF;
712                                         break;
713                                 }
714                                 if (errno != EINTR &&
715                                     errno != EINPROGRESS &&
716                                     errno != EAGAIN) {
717                                         ioq->error = DMSG_IOQ_ERROR_SOCK;
718                                         break;
719                                 }
720                                 n = 0;
721                                 /* fall through */
722                         }
723                         ioq->fifo_end += (size_t)n;
724                         nmax -= (size_t)n;
725                 }
726
727                 if (iocom->flags & DMSG_IOCOMF_CRYPTED) {
728                         dmsg_crypto_decrypt(iocom, ioq);
729                 } else {
730                         ioq->fifo_cdx = ioq->fifo_end;
731                         ioq->fifo_cdn = ioq->fifo_end;
732                 }
733                 bytes = ioq->fifo_cdx - ioq->fifo_beg;
734
735                 /*
736                  * Insufficient data accumulated (set msg NULL so caller will
737                  * retry on event).
738                  */
739                 if (bytes < ioq->hbytes) {
740                         msg = NULL;
741                         break;
742                 }
743
744                 /*
745                  * Calculate the extended header, decrypt data received
746                  * so far.  Handle endian-conversion for the entire extended
747                  * header.
748                  */
749                 head = (void *)(ioq->buf + ioq->fifo_beg);
750
751                 /*
752                  * Check the CRC.
753                  */
754                 if (head->magic == DMSG_HDR_MAGIC_REV)
755                         xcrc32 = bswap32(head->hdr_crc);
756                 else
757                         xcrc32 = head->hdr_crc;
758                 head->hdr_crc = 0;
759                 if (dmsg_icrc32(head, ioq->hbytes) != xcrc32) {
760                         ioq->error = DMSG_IOQ_ERROR_XCRC;
761                         fprintf(stderr, "BAD-XCRC(%08x,%08x) %s\n",
762                                 xcrc32, dmsg_icrc32(head, ioq->hbytes),
763                                 dmsg_msg_str(msg));
764                         assert(0);
765                         break;
766                 }
767                 head->hdr_crc = xcrc32;
768
769                 if (head->magic == DMSG_HDR_MAGIC_REV) {
770                         dmsg_bswap_head(head);
771                 }
772
773                 /*
774                  * Copy the extended header into the msg and adjust the
775                  * FIFO.
776                  */
777                 bcopy(head, &msg->any, ioq->hbytes);
778
779                 /*
780                  * We are either done or we fall-through.
781                  */
782                 if (ioq->abytes == 0) {
783                         ioq->fifo_beg += ioq->hbytes;
784                         break;
785                 }
786
787                 /*
788                  * Must adjust bytes (and the state) when falling through.
789                  * nmax doesn't change.
790                  */
791                 ioq->fifo_beg += ioq->hbytes;
792                 bytes -= ioq->hbytes;
793                 ioq->state = DMSG_MSGQ_STATE_AUXDATA1;
794                 /* fall through */
795         case DMSG_MSGQ_STATE_AUXDATA1:
796                 /*
797                  * Copy the partial or complete payload from remaining
798                  * bytes in the FIFO in order to optimize the makeroom call
799                  * in the AUXDATA2 state.  We have to fall-through either
800                  * way so we can check the crc.
801                  *
802                  * msg->aux_size tracks our aux data.
803                  */
804                 if (bytes >= ioq->abytes) {
805                         bcopy(ioq->buf + ioq->fifo_beg, msg->aux_data,
806                               ioq->abytes);
807                         msg->aux_size = ioq->abytes;
808                         ioq->fifo_beg += ioq->abytes;
809                         assert(ioq->fifo_beg <= ioq->fifo_cdx);
810                         assert(ioq->fifo_cdx <= ioq->fifo_cdn);
811                         bytes -= ioq->abytes;
812                 } else if (bytes) {
813                         bcopy(ioq->buf + ioq->fifo_beg, msg->aux_data,
814                               bytes);
815                         msg->aux_size = bytes;
816                         ioq->fifo_beg += bytes;
817                         if (ioq->fifo_cdx < ioq->fifo_beg)
818                                 ioq->fifo_cdx = ioq->fifo_beg;
819                         assert(ioq->fifo_beg <= ioq->fifo_cdx);
820                         assert(ioq->fifo_cdx <= ioq->fifo_cdn);
821                         bytes = 0;
822                 } else {
823                         msg->aux_size = 0;
824                 }
825                 ioq->state = DMSG_MSGQ_STATE_AUXDATA2;
826                 /* fall through */
827         case DMSG_MSGQ_STATE_AUXDATA2:
828                 /*
829                  * Make sure there is enough room for more data.
830                  */
831                 assert(msg);
832                 nmax = dmsg_ioq_makeroom(ioq, ioq->abytes - msg->aux_size);
833
834                 /*
835                  * Read and decrypt more of the payload.
836                  */
837                 if (msg->aux_size < ioq->abytes) {
838                         assert(bytes == 0);
839                         n = read(iocom->sock_fd,
840                                  ioq->buf + ioq->fifo_end,
841                                  nmax);
842                         if (n <= 0) {
843                                 if (n == 0) {
844                                         ioq->error = DMSG_IOQ_ERROR_EOF;
845                                         break;
846                                 }
847                                 if (errno != EINTR &&
848                                     errno != EINPROGRESS &&
849                                     errno != EAGAIN) {
850                                         ioq->error = DMSG_IOQ_ERROR_SOCK;
851                                         break;
852                                 }
853                                 n = 0;
854                                 /* fall through */
855                         }
856                         ioq->fifo_end += (size_t)n;
857                         nmax -= (size_t)n;
858                 }
859
860                 if (iocom->flags & DMSG_IOCOMF_CRYPTED) {
861                         dmsg_crypto_decrypt(iocom, ioq);
862                 } else {
863                         ioq->fifo_cdx = ioq->fifo_end;
864                         ioq->fifo_cdn = ioq->fifo_end;
865                 }
866                 bytes = ioq->fifo_cdx - ioq->fifo_beg;
867
868                 if (bytes > ioq->abytes - msg->aux_size)
869                         bytes = ioq->abytes - msg->aux_size;
870
871                 if (bytes) {
872                         bcopy(ioq->buf + ioq->fifo_beg,
873                               msg->aux_data + msg->aux_size,
874                               bytes);
875                         msg->aux_size += bytes;
876                         ioq->fifo_beg += bytes;
877                 }
878
879                 /*
880                  * Insufficient data accumulated (set msg NULL so caller will
881                  * retry on event).
882                  */
883                 if (msg->aux_size < ioq->abytes) {
884                         msg = NULL;
885                         break;
886                 }
887                 assert(msg->aux_size == ioq->abytes);
888
889                 /*
890                  * Check aux_crc, then we are done.
891                  */
892                 xcrc32 = dmsg_icrc32(msg->aux_data, msg->aux_size);
893                 if (xcrc32 != msg->any.head.aux_crc) {
894                         ioq->error = DMSG_IOQ_ERROR_ACRC;
895                         break;
896                 }
897                 break;
898         case DMSG_MSGQ_STATE_ERROR:
899                 /*
900                  * Continued calls to drain recorded transactions (returning
901                  * a LNK_ERROR for each one), before we return the final
902                  * LNK_ERROR.
903                  */
904                 assert(msg == NULL);
905                 break;
906         default:
907                 /*
908                  * We don't double-return errors, the caller should not
909                  * have called us again after getting an error msg.
910                  */
911                 assert(0);
912                 break;
913         }
914
915         /*
916          * Check the message sequence.  The iv[] should prevent any
917          * possibility of a replay but we add this check anyway.
918          */
919         if (msg && ioq->error == 0) {
920                 if ((msg->any.head.salt & 255) != (ioq->seq & 255)) {
921                         ioq->error = DMSG_IOQ_ERROR_MSGSEQ;
922                 } else {
923                         ++ioq->seq;
924                 }
925         }
926
927         /*
928          * Handle relaying.  Transactional state is not recorded XXX
929          */
930
931         /*
932          * Process transactional state for the message.
933          */
934         if (msg && ioq->error == 0) {
935                 error = dmsg_state_msgrx(msg);
936                 if (error) {
937                         if (error == DMSG_IOQ_ERROR_EALREADY) {
938                                 dmsg_msg_free(msg);
939                                 goto again;
940                         }
941                         ioq->error = error;
942                 }
943         }
944
945         /*
946          * Handle error, RREQ, or completion
947          *
948          * NOTE: nmax and bytes are invalid at this point, we don't bother
949          *       to update them when breaking out.
950          */
951         if (ioq->error) {
952 skip:
953                 /*
954                  * An unrecoverable error causes all active receive
955                  * transactions to be terminated with a LNK_ERROR message.
956                  *
957                  * Once all active transactions are exhausted we set the
958                  * iocom ERROR flag and return a non-transactional LNK_ERROR
959                  * message, which should cause master processing loops to
960                  * terminate.
961                  */
962                 assert(ioq->msg == msg);
963                 if (msg) {
964                         dmsg_msg_free(msg);
965                         ioq->msg = NULL;
966                 }
967
968                 /*
969                  * No more I/O read processing
970                  */
971                 ioq->state = DMSG_MSGQ_STATE_ERROR;
972
973                 /*
974                  * Simulate a remote LNK_ERROR DELETE msg for any open
975                  * transactions, ending with a final non-transactional
976                  * LNK_ERROR (that the session can detect) when no
977                  * transactions remain.
978                  */
979                 msg = dmsg_msg_alloc(iocom->router, 0, 0, NULL, NULL);
980                 bzero(&msg->any.head, sizeof(msg->any.head));
981                 msg->any.head.magic = DMSG_HDR_MAGIC;
982                 msg->any.head.cmd = DMSG_LNK_ERROR;
983                 msg->any.head.error = ioq->error;
984
985                 pthread_mutex_lock(&iocom->mtx);
986                 dmsg_iocom_drain(iocom);
987                 if ((state = RB_ROOT(&iocom->router->staterd_tree)) != NULL) {
988                         /*
989                          * Active remote transactions are still present.
990                          * Simulate the other end sending us a DELETE.
991                          */
992                         if (state->rxcmd & DMSGF_DELETE) {
993                                 dmsg_msg_free(msg);
994                                 msg = NULL;
995                         } else {
996                                 /*state->txcmd |= DMSGF_DELETE;*/
997                                 msg->state = state;
998                                 msg->router = state->router;
999                                 msg->any.head.msgid = state->msgid;
1000                                 msg->any.head.cmd |= DMSGF_ABORT |
1001                                                      DMSGF_DELETE;
1002                         }
1003                 } else if ((state = RB_ROOT(&iocom->router->statewr_tree)) !=
1004                            NULL) {
1005                         /*
1006                          * Active local transactions are still present.
1007                          * Simulate the other end sending us a DELETE.
1008                          */
1009                         if (state->rxcmd & DMSGF_DELETE) {
1010                                 dmsg_msg_free(msg);
1011                                 msg = NULL;
1012                         } else {
1013                                 msg->state = state;
1014                                 msg->router = state->router;
1015                                 msg->any.head.msgid = state->msgid;
1016                                 msg->any.head.cmd |= DMSGF_ABORT |
1017                                                      DMSGF_DELETE |
1018                                                      DMSGF_REPLY;
1019                                 if ((state->rxcmd & DMSGF_CREATE) == 0) {
1020                                         msg->any.head.cmd |=
1021                                                      DMSGF_CREATE;
1022                                 }
1023                         }
1024                 } else {
1025                         /*
1026                          * No active local or remote transactions remain.
1027                          * Generate a final LNK_ERROR and flag EOF.
1028                          */
1029                         msg->state = NULL;
1030                         iocom->flags |= DMSG_IOCOMF_EOF;
1031                         fprintf(stderr, "EOF ON SOCKET %d\n", iocom->sock_fd);
1032                 }
1033                 pthread_mutex_unlock(&iocom->mtx);
1034
1035                 /*
1036                  * For the iocom error case we want to set RWORK to indicate
1037                  * that more messages might be pending.
1038                  *
1039                  * It is possible to return NULL when there is more work to
1040                  * do because each message has to be DELETEd in both
1041                  * directions before we continue on with the next (though
1042                  * this could be optimized).  The transmit direction will
1043                  * re-set RWORK.
1044                  */
1045                 if (msg)
1046                         iocom->flags |= DMSG_IOCOMF_RWORK;
1047         } else if (msg == NULL) {
1048                 /*
1049                  * Insufficient data received to finish building the message,
1050                  * set RREQ and return NULL.
1051                  *
1052                  * Leave ioq->msg intact.
1053                  * Leave the FIFO intact.
1054                  */
1055                 iocom->flags |= DMSG_IOCOMF_RREQ;
1056         } else {
1057                 /*
1058                  * Return msg.
1059                  *
1060                  * The fifo has already been advanced past the message.
1061                  * Trivially reset the FIFO indices if possible.
1062                  *
1063                  * clear the FIFO if it is now empty and set RREQ to wait
1064                  * for more from the socket.  If the FIFO is not empty set
1065                  * TWORK to bypass the poll so we loop immediately.
1066                  */
1067                 if (ioq->fifo_beg == ioq->fifo_cdx &&
1068                     ioq->fifo_cdn == ioq->fifo_end) {
1069                         iocom->flags |= DMSG_IOCOMF_RREQ;
1070                         ioq->fifo_cdx = 0;
1071                         ioq->fifo_cdn = 0;
1072                         ioq->fifo_beg = 0;
1073                         ioq->fifo_end = 0;
1074                 } else {
1075                         iocom->flags |= DMSG_IOCOMF_RWORK;
1076                 }
1077                 ioq->state = DMSG_MSGQ_STATE_HEADER1;
1078                 ioq->msg = NULL;
1079         }
1080         return (msg);
1081 }
1082
1083 /*
1084  * Calculate the header and data crc's and write a low-level message to
1085  * the connection.  If aux_crc is non-zero the aux_data crc is already
1086  * assumed to have been set.
1087  *
1088  * A non-NULL msg is added to the queue but not necessarily flushed.
1089  * Calling this function with msg == NULL will get a flush going.
1090  *
1091  * Caller must hold iocom->mtx.
1092  */
1093 void
1094 dmsg_iocom_flush1(dmsg_iocom_t *iocom)
1095 {
1096         dmsg_ioq_t *ioq = &iocom->ioq_tx;
1097         dmsg_msg_t *msg;
1098         uint32_t xcrc32;
1099         int hbytes;
1100         dmsg_msg_queue_t tmpq;
1101
1102         iocom->flags &= ~(DMSG_IOCOMF_WREQ | DMSG_IOCOMF_WWORK);
1103         TAILQ_INIT(&tmpq);
1104         pthread_mutex_lock(&iocom->mtx);
1105         while ((msg = TAILQ_FIRST(&iocom->router->txmsgq)) != NULL) {
1106                 TAILQ_REMOVE(&iocom->router->txmsgq, msg, qentry);
1107                 TAILQ_INSERT_TAIL(&tmpq, msg, qentry);
1108         }
1109         pthread_mutex_unlock(&iocom->mtx);
1110
1111         while ((msg = TAILQ_FIRST(&tmpq)) != NULL) {
1112                 /*
1113                  * Process terminal connection errors.
1114                  */
1115                 TAILQ_REMOVE(&tmpq, msg, qentry);
1116                 if (ioq->error) {
1117                         TAILQ_INSERT_TAIL(&ioq->msgq, msg, qentry);
1118                         ++ioq->msgcount;
1119                         continue;
1120                 }
1121
1122                 /*
1123                  * Finish populating the msg fields.  The salt ensures that
1124                  * the iv[] array is ridiculously randomized and we also
1125                  * re-seed our PRNG every 32768 messages just to be sure.
1126                  */
1127                 msg->any.head.magic = DMSG_HDR_MAGIC;
1128                 msg->any.head.salt = (random() << 8) | (ioq->seq & 255);
1129                 ++ioq->seq;
1130                 if ((ioq->seq & 32767) == 0)
1131                         srandomdev();
1132
1133                 /*
1134                  * Calculate aux_crc if 0, then calculate hdr_crc.
1135                  */
1136                 if (msg->aux_size && msg->any.head.aux_crc == 0) {
1137                         assert((msg->aux_size & DMSG_ALIGNMASK) == 0);
1138                         xcrc32 = dmsg_icrc32(msg->aux_data, msg->aux_size);
1139                         msg->any.head.aux_crc = xcrc32;
1140                 }
1141                 msg->any.head.aux_bytes = msg->aux_size / DMSG_ALIGN;
1142                 assert((msg->aux_size & DMSG_ALIGNMASK) == 0);
1143
1144                 hbytes = (msg->any.head.cmd & DMSGF_SIZE) *
1145                          DMSG_ALIGN;
1146                 msg->any.head.hdr_crc = 0;
1147                 msg->any.head.hdr_crc = dmsg_icrc32(&msg->any.head, hbytes);
1148
1149                 /*
1150                  * Enqueue the message (the flush codes handles stream
1151                  * encryption).
1152                  */
1153                 TAILQ_INSERT_TAIL(&ioq->msgq, msg, qentry);
1154                 ++ioq->msgcount;
1155         }
1156         dmsg_iocom_flush2(iocom);
1157 }
1158
1159 /*
1160  * Thread localized, iocom->mtx not held by caller.
1161  */
1162 void
1163 dmsg_iocom_flush2(dmsg_iocom_t *iocom)
1164 {
1165         dmsg_ioq_t *ioq = &iocom->ioq_tx;
1166         dmsg_msg_t *msg;
1167         ssize_t n;
1168         struct iovec iov[DMSG_IOQ_MAXIOVEC];
1169         size_t nact;
1170         size_t hbytes;
1171         size_t abytes;
1172         size_t hoff;
1173         size_t aoff;
1174         int iovcnt;
1175
1176         if (ioq->error) {
1177                 dmsg_iocom_drain(iocom);
1178                 return;
1179         }
1180
1181         /*
1182          * Pump messages out the connection by building an iovec.
1183          *
1184          * ioq->hbytes/ioq->abytes tracks how much of the first message
1185          * in the queue has been successfully written out, so we can
1186          * resume writing.
1187          */
1188         iovcnt = 0;
1189         nact = 0;
1190         hoff = ioq->hbytes;
1191         aoff = ioq->abytes;
1192
1193         TAILQ_FOREACH(msg, &ioq->msgq, qentry) {
1194                 hbytes = (msg->any.head.cmd & DMSGF_SIZE) *
1195                          DMSG_ALIGN;
1196                 abytes = msg->aux_size;
1197                 assert(hoff <= hbytes && aoff <= abytes);
1198
1199                 if (hoff < hbytes) {
1200                         iov[iovcnt].iov_base = (char *)&msg->any.head + hoff;
1201                         iov[iovcnt].iov_len = hbytes - hoff;
1202                         nact += hbytes - hoff;
1203                         ++iovcnt;
1204                         if (iovcnt == DMSG_IOQ_MAXIOVEC)
1205                                 break;
1206                 }
1207                 if (aoff < abytes) {
1208                         assert(msg->aux_data != NULL);
1209                         iov[iovcnt].iov_base = (char *)msg->aux_data + aoff;
1210                         iov[iovcnt].iov_len = abytes - aoff;
1211                         nact += abytes - aoff;
1212                         ++iovcnt;
1213                         if (iovcnt == DMSG_IOQ_MAXIOVEC)
1214                                 break;
1215                 }
1216                 hoff = 0;
1217                 aoff = 0;
1218         }
1219         if (iovcnt == 0)
1220                 return;
1221
1222         /*
1223          * Encrypt and write the data.  The crypto code will move the
1224          * data into the fifo and adjust the iov as necessary.  If
1225          * encryption is disabled the iov is left alone.
1226          *
1227          * May return a smaller iov (thus a smaller n), with aggregated
1228          * chunks.  May reduce nmax to what fits in the FIFO.
1229          *
1230          * This function sets nact to the number of original bytes now
1231          * encrypted, adding to the FIFO some number of bytes that might
1232          * be greater depending on the crypto mechanic.  iov[] is adjusted
1233          * to point at the FIFO if necessary.
1234          *
1235          * NOTE: The return value from the writev() is the post-encrypted
1236          *       byte count, not the plaintext count.
1237          */
1238         if (iocom->flags & DMSG_IOCOMF_CRYPTED) {
1239                 /*
1240                  * Make sure the FIFO has a reasonable amount of space
1241                  * left (if not completely full).
1242                  */
1243                 if (ioq->fifo_beg > sizeof(ioq->buf) / 2 &&
1244                     sizeof(ioq->buf) - ioq->fifo_end >= DMSG_ALIGN * 2) {
1245                         bcopy(ioq->buf + ioq->fifo_beg, ioq->buf,
1246                               ioq->fifo_end - ioq->fifo_beg);
1247                         ioq->fifo_cdx -= ioq->fifo_beg;
1248                         ioq->fifo_cdn -= ioq->fifo_beg;
1249                         ioq->fifo_end -= ioq->fifo_beg;
1250                         ioq->fifo_beg = 0;
1251                 }
1252
1253                 iovcnt = dmsg_crypto_encrypt(iocom, ioq, iov, iovcnt, &nact);
1254                 n = writev(iocom->sock_fd, iov, iovcnt);
1255                 if (n > 0) {
1256                         ioq->fifo_beg += n;
1257                         ioq->fifo_cdn += n;
1258                         ioq->fifo_cdx += n;
1259                         if (ioq->fifo_beg == ioq->fifo_end) {
1260                                 ioq->fifo_beg = 0;
1261                                 ioq->fifo_cdn = 0;
1262                                 ioq->fifo_cdx = 0;
1263                                 ioq->fifo_end = 0;
1264                         }
1265                 }
1266         } else {
1267                 n = writev(iocom->sock_fd, iov, iovcnt);
1268                 if (n > 0)
1269                         nact = n;
1270                 else
1271                         nact = 0;
1272         }
1273
1274         /*
1275          * Clean out the transmit queue based on what we successfully
1276          * sent (nact is the plaintext count).  ioq->hbytes/abytes
1277          * represents the portion of the first message previously sent.
1278          */
1279         while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
1280                 hbytes = (msg->any.head.cmd & DMSGF_SIZE) *
1281                          DMSG_ALIGN;
1282                 abytes = msg->aux_size;
1283
1284                 if ((size_t)nact < hbytes - ioq->hbytes) {
1285                         ioq->hbytes += nact;
1286                         nact = 0;
1287                         break;
1288                 }
1289                 nact -= hbytes - ioq->hbytes;
1290                 ioq->hbytes = hbytes;
1291                 if ((size_t)nact < abytes - ioq->abytes) {
1292                         ioq->abytes += nact;
1293                         nact = 0;
1294                         break;
1295                 }
1296                 nact -= abytes - ioq->abytes;
1297
1298                 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
1299                 --ioq->msgcount;
1300                 ioq->hbytes = 0;
1301                 ioq->abytes = 0;
1302
1303                 dmsg_state_cleanuptx(msg);
1304         }
1305         assert(nact == 0);
1306
1307         /*
1308          * Process the return value from the write w/regards to blocking.
1309          */
1310         if (n < 0) {
1311                 if (errno != EINTR &&
1312                     errno != EINPROGRESS &&
1313                     errno != EAGAIN) {
1314                         /*
1315                          * Fatal write error
1316                          */
1317                         ioq->error = DMSG_IOQ_ERROR_SOCK;
1318                         dmsg_iocom_drain(iocom);
1319                 } else {
1320                         /*
1321                          * Wait for socket buffer space
1322                          */
1323                         iocom->flags |= DMSG_IOCOMF_WREQ;
1324                 }
1325         } else {
1326                 iocom->flags |= DMSG_IOCOMF_WREQ;
1327         }
1328         if (ioq->error) {
1329                 dmsg_iocom_drain(iocom);
1330         }
1331 }
1332
1333 /*
1334  * Kill pending msgs on ioq_tx and adjust the flags such that no more
1335  * write events will occur.  We don't kill read msgs because we want
1336  * the caller to pull off our contrived terminal error msg to detect
1337  * the connection failure.
1338  *
1339  * Thread localized, iocom->mtx not held by caller.
1340  */
1341 void
1342 dmsg_iocom_drain(dmsg_iocom_t *iocom)
1343 {
1344         dmsg_ioq_t *ioq = &iocom->ioq_tx;
1345         dmsg_msg_t *msg;
1346
1347         iocom->flags &= ~(DMSG_IOCOMF_WREQ | DMSG_IOCOMF_WWORK);
1348         ioq->hbytes = 0;
1349         ioq->abytes = 0;
1350
1351         while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
1352                 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
1353                 --ioq->msgcount;
1354                 dmsg_state_cleanuptx(msg);
1355         }
1356 }
1357
1358 /*
1359  * Write a message to an iocom, with additional state processing.
1360  */
1361 void
1362 dmsg_msg_write(dmsg_msg_t *msg)
1363 {
1364         dmsg_iocom_t *iocom = msg->router->iocom;
1365         dmsg_state_t *state;
1366         char dummy;
1367
1368         /*
1369          * Handle state processing, create state if necessary.
1370          */
1371         pthread_mutex_lock(&iocom->mtx);
1372         if ((state = msg->state) != NULL) {
1373                 /*
1374                  * Existing transaction (could be reply).  It is also
1375                  * possible for this to be the first reply (CREATE is set),
1376                  * in which case we populate state->txcmd.
1377                  *
1378                  * state->txcmd is adjusted to hold the final message cmd,
1379                  * and we also be sure to set the CREATE bit here.  We did
1380                  * not set it in dmsg_msg_alloc() because that would have
1381                  * not been serialized (state could have gotten ripped out
1382                  * from under the message prior to it being transmitted).
1383                  */
1384                 if ((msg->any.head.cmd & (DMSGF_CREATE | DMSGF_REPLY)) ==
1385                     DMSGF_CREATE) {
1386                         state->txcmd = msg->any.head.cmd & ~DMSGF_DELETE;
1387                 }
1388                 msg->any.head.msgid = state->msgid;
1389                 assert(((state->txcmd ^ msg->any.head.cmd) & DMSGF_REPLY) == 0);
1390                 if (msg->any.head.cmd & DMSGF_CREATE)
1391                         state->txcmd = msg->any.head.cmd & ~DMSGF_DELETE;
1392         } else {
1393                 msg->any.head.msgid = 0;
1394                 /* XXX set spanid by router */
1395         }
1396         msg->any.head.source = 0;
1397         msg->any.head.target = msg->router->target;
1398
1399         /*
1400          * Queue it for output, wake up the I/O pthread.  Note that the
1401          * I/O thread is responsible for generating the CRCs and encryption.
1402          */
1403         TAILQ_INSERT_TAIL(&iocom->router->txmsgq, msg, qentry);
1404         dummy = 0;
1405         write(iocom->wakeupfds[1], &dummy, 1);  /* XXX optimize me */
1406         pthread_mutex_unlock(&iocom->mtx);
1407 }
1408
1409 /*
1410  * This is a shortcut to formulate a reply to msg with a simple error code,
1411  * It can reply to and terminate a transaction, or it can reply to a one-way
1412  * messages.  A DMSG_LNK_ERROR command code is utilized to encode
1413  * the error code (which can be 0).  Not all transactions are terminated
1414  * with DMSG_LNK_ERROR status (the low level only cares about the
1415  * MSGF_DELETE flag), but most are.
1416  *
1417  * Replies to one-way messages are a bit of an oxymoron but the feature
1418  * is used by the debug (DBG) protocol.
1419  *
1420  * The reply contains no extended data.
1421  */
1422 void
1423 dmsg_msg_reply(dmsg_msg_t *msg, uint32_t error)
1424 {
1425         dmsg_iocom_t *iocom = msg->router->iocom;
1426         dmsg_state_t *state = msg->state;
1427         dmsg_msg_t *nmsg;
1428         uint32_t cmd;
1429
1430
1431         /*
1432          * Reply with a simple error code and terminate the transaction.
1433          */
1434         cmd = DMSG_LNK_ERROR;
1435
1436         /*
1437          * Check if our direction has even been initiated yet, set CREATE.
1438          *
1439          * Check what direction this is (command or reply direction).  Note
1440          * that txcmd might not have been initiated yet.
1441          *
1442          * If our direction has already been closed we just return without
1443          * doing anything.
1444          */
1445         if (state) {
1446                 if (state->txcmd & DMSGF_DELETE)
1447                         return;
1448                 if (state->txcmd & DMSGF_REPLY)
1449                         cmd |= DMSGF_REPLY;
1450                 cmd |= DMSGF_DELETE;
1451         } else {
1452                 if ((msg->any.head.cmd & DMSGF_REPLY) == 0)
1453                         cmd |= DMSGF_REPLY;
1454         }
1455
1456         /*
1457          * Allocate the message and associate it with the existing state.
1458          * We cannot pass MSGF_CREATE to msg_alloc() because that may
1459          * allocate new state.  We have our state already.
1460          */
1461         nmsg = dmsg_msg_alloc(iocom->router, 0, cmd, NULL, NULL);
1462         if (state) {
1463                 if ((state->txcmd & DMSGF_CREATE) == 0)
1464                         nmsg->any.head.cmd |= DMSGF_CREATE;
1465         }
1466         nmsg->any.head.error = error;
1467         nmsg->state = state;
1468         dmsg_msg_write(nmsg);
1469 }
1470
1471 /*
1472  * Similar to dmsg_msg_reply() but leave the transaction open.  That is,
1473  * we are generating a streaming reply or an intermediate acknowledgement
1474  * of some sort as part of the higher level protocol, with more to come
1475  * later.
1476  */
1477 void
1478 dmsg_msg_result(dmsg_msg_t *msg, uint32_t error)
1479 {
1480         dmsg_iocom_t *iocom = msg->router->iocom;
1481         dmsg_state_t *state = msg->state;
1482         dmsg_msg_t *nmsg;
1483         uint32_t cmd;
1484
1485
1486         /*
1487          * Reply with a simple error code and terminate the transaction.
1488          */
1489         cmd = DMSG_LNK_ERROR;
1490
1491         /*
1492          * Check if our direction has even been initiated yet, set CREATE.
1493          *
1494          * Check what direction this is (command or reply direction).  Note
1495          * that txcmd might not have been initiated yet.
1496          *
1497          * If our direction has already been closed we just return without
1498          * doing anything.
1499          */
1500         if (state) {
1501                 if (state->txcmd & DMSGF_DELETE)
1502                         return;
1503                 if (state->txcmd & DMSGF_REPLY)
1504                         cmd |= DMSGF_REPLY;
1505                 /* continuing transaction, do not set MSGF_DELETE */
1506         } else {
1507                 if ((msg->any.head.cmd & DMSGF_REPLY) == 0)
1508                         cmd |= DMSGF_REPLY;
1509         }
1510
1511         nmsg = dmsg_msg_alloc(iocom->router, 0, cmd, NULL, NULL);
1512         if (state) {
1513                 if ((state->txcmd & DMSGF_CREATE) == 0)
1514                         nmsg->any.head.cmd |= DMSGF_CREATE;
1515         }
1516         nmsg->any.head.error = error;
1517         nmsg->state = state;
1518         dmsg_msg_write(nmsg);
1519 }
1520
1521 /*
1522  * Terminate a transaction given a state structure by issuing a DELETE.
1523  */
1524 void
1525 dmsg_state_reply(dmsg_state_t *state, uint32_t error)
1526 {
1527         dmsg_msg_t *nmsg;
1528         uint32_t cmd = DMSG_LNK_ERROR | DMSGF_DELETE;
1529
1530         /*
1531          * Nothing to do if we already transmitted a delete
1532          */
1533         if (state->txcmd & DMSGF_DELETE)
1534                 return;
1535
1536         /*
1537          * Set REPLY if the other end initiated the command.  Otherwise
1538          * we are the command direction.
1539          */
1540         if (state->txcmd & DMSGF_REPLY)
1541                 cmd |= DMSGF_REPLY;
1542
1543         nmsg = dmsg_msg_alloc(state->iocom->router, 0, cmd, NULL, NULL);
1544         if (state) {
1545                 if ((state->txcmd & DMSGF_CREATE) == 0)
1546                         nmsg->any.head.cmd |= DMSGF_CREATE;
1547         }
1548         nmsg->any.head.error = error;
1549         nmsg->state = state;
1550         dmsg_msg_write(nmsg);
1551 }
1552
1553 /************************************************************************
1554  *                      TRANSACTION STATE HANDLING                      *
1555  ************************************************************************
1556  *
1557  */
1558
1559 /*
1560  * Process state tracking for a message after reception, prior to
1561  * execution.
1562  *
1563  * Called with msglk held and the msg dequeued.
1564  *
1565  * All messages are called with dummy state and return actual state.
1566  * (One-off messages often just return the same dummy state).
1567  *
1568  * May request that caller discard the message by setting *discardp to 1.
1569  * The returned state is not used in this case and is allowed to be NULL.
1570  *
1571  * --
1572  *
1573  * These routines handle persistent and command/reply message state via the
1574  * CREATE and DELETE flags.  The first message in a command or reply sequence
1575  * sets CREATE, the last message in a command or reply sequence sets DELETE.
1576  *
1577  * There can be any number of intermediate messages belonging to the same
1578  * sequence sent inbetween the CREATE message and the DELETE message,
1579  * which set neither flag.  This represents a streaming command or reply.
1580  *
1581  * Any command message received with CREATE set expects a reply sequence to
1582  * be returned.  Reply sequences work the same as command sequences except the
1583  * REPLY bit is also sent.  Both the command side and reply side can
1584  * degenerate into a single message with both CREATE and DELETE set.  Note
1585  * that one side can be streaming and the other side not, or neither, or both.
1586  *
1587  * The msgid is unique for the initiator.  That is, two sides sending a new
1588  * message can use the same msgid without colliding.
1589  *
1590  * --
1591  *
1592  * ABORT sequences work by setting the ABORT flag along with normal message
1593  * state.  However, ABORTs can also be sent on half-closed messages, that is
1594  * even if the command or reply side has already sent a DELETE, as long as
1595  * the message has not been fully closed it can still send an ABORT+DELETE
1596  * to terminate the half-closed message state.
1597  *
1598  * Since ABORT+DELETEs can race we silently discard ABORT's for message
1599  * state which has already been fully closed.  REPLY+ABORT+DELETEs can
1600  * also race, and in this situation the other side might have already
1601  * initiated a new unrelated command with the same message id.  Since
1602  * the abort has not set the CREATE flag the situation can be detected
1603  * and the message will also be discarded.
1604  *
1605  * Non-blocking requests can be initiated with ABORT+CREATE[+DELETE].
1606  * The ABORT request is essentially integrated into the command instead
1607  * of being sent later on.  In this situation the command implementation
1608  * detects that CREATE and ABORT are both set (vs ABORT alone) and can
1609  * special-case non-blocking operation for the command.
1610  *
1611  * NOTE!  Messages with ABORT set without CREATE or DELETE are considered
1612  *        to be mid-stream aborts for command/reply sequences.  ABORTs on
1613  *        one-way messages are not supported.
1614  *
1615  * NOTE!  If a command sequence does not support aborts the ABORT flag is
1616  *        simply ignored.
1617  *
1618  * --
1619  *
1620  * One-off messages (no reply expected) are sent with neither CREATE or DELETE
1621  * set.  One-off messages cannot be aborted and typically aren't processed
1622  * by these routines.  The REPLY bit can be used to distinguish whether a
1623  * one-off message is a command or reply.  For example, one-off replies
1624  * will typically just contain status updates.
1625  */
1626 static int
1627 dmsg_state_msgrx(dmsg_msg_t *msg)
1628 {
1629         dmsg_iocom_t *iocom = msg->router->iocom;
1630         dmsg_state_t *state;
1631         dmsg_state_t dummy;
1632         int error;
1633
1634         /*
1635          * Lock RB tree and locate existing persistent state, if any.
1636          *
1637          * If received msg is a command state is on staterd_tree.
1638          * If received msg is a reply state is on statewr_tree.
1639          */
1640         dummy.msgid = msg->any.head.msgid;
1641         pthread_mutex_lock(&iocom->mtx);
1642         if (msg->any.head.cmd & DMSGF_REPLY) {
1643                 state = RB_FIND(dmsg_state_tree,
1644                                 &iocom->router->statewr_tree, &dummy);
1645         } else {
1646                 state = RB_FIND(dmsg_state_tree,
1647                                 &iocom->router->staterd_tree, &dummy);
1648         }
1649         msg->state = state;
1650         pthread_mutex_unlock(&iocom->mtx);
1651
1652         /*
1653          * Short-cut one-off or mid-stream messages (state may be NULL).
1654          */
1655         if ((msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE |
1656                                   DMSGF_ABORT)) == 0) {
1657                 return(0);
1658         }
1659
1660         /*
1661          * Switch on CREATE, DELETE, REPLY, and also handle ABORT from
1662          * inside the case statements.
1663          */
1664         switch(msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE |
1665                                     DMSGF_REPLY)) {
1666         case DMSGF_CREATE:
1667         case DMSGF_CREATE | DMSGF_DELETE:
1668                 /*
1669                  * New persistant command received.
1670                  */
1671                 if (state) {
1672                         fprintf(stderr, "duplicate-trans %s\n",
1673                                 dmsg_msg_str(msg));
1674                         error = DMSG_IOQ_ERROR_TRANS;
1675                         assert(0);
1676                         break;
1677                 }
1678                 state = malloc(sizeof(*state));
1679                 bzero(state, sizeof(*state));
1680                 state->iocom = iocom;
1681                 state->flags = DMSG_STATE_DYNAMIC;
1682                 state->msg = msg;
1683                 state->txcmd = DMSGF_REPLY;
1684                 state->rxcmd = msg->any.head.cmd & ~DMSGF_DELETE;
1685                 state->flags |= DMSG_STATE_INSERTED;
1686                 state->msgid = msg->any.head.msgid;
1687                 state->router = msg->router;
1688                 msg->state = state;
1689                 pthread_mutex_lock(&iocom->mtx);
1690                 RB_INSERT(dmsg_state_tree,
1691                           &iocom->router->staterd_tree, state);
1692                 pthread_mutex_unlock(&iocom->mtx);
1693                 error = 0;
1694                 if (DMsgDebugOpt) {
1695                         fprintf(stderr, "create state %p id=%08x on iocom staterd %p\n",
1696                                 state, (uint32_t)state->msgid, iocom);
1697                 }
1698                 break;
1699         case DMSGF_DELETE:
1700                 /*
1701                  * Persistent state is expected but might not exist if an
1702                  * ABORT+DELETE races the close.
1703                  */
1704                 if (state == NULL) {
1705                         if (msg->any.head.cmd & DMSGF_ABORT) {
1706                                 error = DMSG_IOQ_ERROR_EALREADY;
1707                         } else {
1708                                 fprintf(stderr, "missing-state %s\n",
1709                                         dmsg_msg_str(msg));
1710                                 error = DMSG_IOQ_ERROR_TRANS;
1711                         assert(0);
1712                         }
1713                         break;
1714                 }
1715
1716                 /*
1717                  * Handle another ABORT+DELETE case if the msgid has already
1718                  * been reused.
1719                  */
1720                 if ((state->rxcmd & DMSGF_CREATE) == 0) {
1721                         if (msg->any.head.cmd & DMSGF_ABORT) {
1722                                 error = DMSG_IOQ_ERROR_EALREADY;
1723                         } else {
1724                                 fprintf(stderr, "reused-state %s\n",
1725                                         dmsg_msg_str(msg));
1726                                 error = DMSG_IOQ_ERROR_TRANS;
1727                         assert(0);
1728                         }
1729                         break;
1730                 }
1731                 error = 0;
1732                 break;
1733         default:
1734                 /*
1735                  * Check for mid-stream ABORT command received, otherwise
1736                  * allow.
1737                  */
1738                 if (msg->any.head.cmd & DMSGF_ABORT) {
1739                         if (state == NULL ||
1740                             (state->rxcmd & DMSGF_CREATE) == 0) {
1741                                 error = DMSG_IOQ_ERROR_EALREADY;
1742                                 break;
1743                         }
1744                 }
1745                 error = 0;
1746                 break;
1747         case DMSGF_REPLY | DMSGF_CREATE:
1748         case DMSGF_REPLY | DMSGF_CREATE | DMSGF_DELETE:
1749                 /*
1750                  * When receiving a reply with CREATE set the original
1751                  * persistent state message should already exist.
1752                  */
1753                 if (state == NULL) {
1754                         fprintf(stderr, "no-state(r) %s\n",
1755                                 dmsg_msg_str(msg));
1756                         error = DMSG_IOQ_ERROR_TRANS;
1757                         assert(0);
1758                         break;
1759                 }
1760                 assert(((state->rxcmd ^ msg->any.head.cmd) &
1761                         DMSGF_REPLY) == 0);
1762                 state->rxcmd = msg->any.head.cmd & ~DMSGF_DELETE;
1763                 error = 0;
1764                 break;
1765         case DMSGF_REPLY | DMSGF_DELETE:
1766                 /*
1767                  * Received REPLY+ABORT+DELETE in case where msgid has
1768                  * already been fully closed, ignore the message.
1769                  */
1770                 if (state == NULL) {
1771                         if (msg->any.head.cmd & DMSGF_ABORT) {
1772                                 error = DMSG_IOQ_ERROR_EALREADY;
1773                         } else {
1774                                 fprintf(stderr, "no-state(r,d) %s\n",
1775                                         dmsg_msg_str(msg));
1776                                 error = DMSG_IOQ_ERROR_TRANS;
1777                         assert(0);
1778                         }
1779                         break;
1780                 }
1781
1782                 /*
1783                  * Received REPLY+ABORT+DELETE in case where msgid has
1784                  * already been reused for an unrelated message,
1785                  * ignore the message.
1786                  */
1787                 if ((state->rxcmd & DMSGF_CREATE) == 0) {
1788                         if (msg->any.head.cmd & DMSGF_ABORT) {
1789                                 error = DMSG_IOQ_ERROR_EALREADY;
1790                         } else {
1791                                 fprintf(stderr, "reused-state(r,d) %s\n",
1792                                         dmsg_msg_str(msg));
1793                                 error = DMSG_IOQ_ERROR_TRANS;
1794                         assert(0);
1795                         }
1796                         break;
1797                 }
1798                 error = 0;
1799                 break;
1800         case DMSGF_REPLY:
1801                 /*
1802                  * Check for mid-stream ABORT reply received to sent command.
1803                  */
1804                 if (msg->any.head.cmd & DMSGF_ABORT) {
1805                         if (state == NULL ||
1806                             (state->rxcmd & DMSGF_CREATE) == 0) {
1807                                 error = DMSG_IOQ_ERROR_EALREADY;
1808                                 break;
1809                         }
1810                 }
1811                 error = 0;
1812                 break;
1813         }
1814         return (error);
1815 }
1816
1817 void
1818 dmsg_state_cleanuprx(dmsg_iocom_t *iocom, dmsg_msg_t *msg)
1819 {
1820         dmsg_state_t *state;
1821
1822         if ((state = msg->state) == NULL) {
1823                 /*
1824                  * Free a non-transactional message, there is no state
1825                  * to worry about.
1826                  */
1827                 dmsg_msg_free(msg);
1828         } else if (msg->any.head.cmd & DMSGF_DELETE) {
1829                 /*
1830                  * Message terminating transaction, destroy the related
1831                  * state, the original message, and this message (if it
1832                  * isn't the original message due to a CREATE|DELETE).
1833                  */
1834                 pthread_mutex_lock(&iocom->mtx);
1835                 state->rxcmd |= DMSGF_DELETE;
1836                 if (state->txcmd & DMSGF_DELETE) {
1837                         if (state->msg == msg)
1838                                 state->msg = NULL;
1839                         assert(state->flags & DMSG_STATE_INSERTED);
1840                         if (state->rxcmd & DMSGF_REPLY) {
1841                                 assert(msg->any.head.cmd & DMSGF_REPLY);
1842                                 RB_REMOVE(dmsg_state_tree,
1843                                           &iocom->router->statewr_tree, state);
1844                         } else {
1845                                 assert((msg->any.head.cmd & DMSGF_REPLY) == 0);
1846                                 RB_REMOVE(dmsg_state_tree,
1847                                           &iocom->router->staterd_tree, state);
1848                         }
1849                         state->flags &= ~DMSG_STATE_INSERTED;
1850                         dmsg_state_free(state);
1851                 } else {
1852                         ;
1853                 }
1854                 pthread_mutex_unlock(&iocom->mtx);
1855                 dmsg_msg_free(msg);
1856         } else if (state->msg != msg) {
1857                 /*
1858                  * Message not terminating transaction, leave state intact
1859                  * and free message if it isn't the CREATE message.
1860                  */
1861                 dmsg_msg_free(msg);
1862         }
1863 }
1864
1865 static void
1866 dmsg_state_cleanuptx(dmsg_msg_t *msg)
1867 {
1868         dmsg_iocom_t *iocom = msg->router->iocom;
1869         dmsg_state_t *state;
1870
1871         if ((state = msg->state) == NULL) {
1872                 dmsg_msg_free(msg);
1873         } else if (msg->any.head.cmd & DMSGF_DELETE) {
1874                 pthread_mutex_lock(&iocom->mtx);
1875                 state->txcmd |= DMSGF_DELETE;
1876                 if (state->rxcmd & DMSGF_DELETE) {
1877                         if (state->msg == msg)
1878                                 state->msg = NULL;
1879                         assert(state->flags & DMSG_STATE_INSERTED);
1880                         if (state->txcmd & DMSGF_REPLY) {
1881                                 assert(msg->any.head.cmd & DMSGF_REPLY);
1882                                 RB_REMOVE(dmsg_state_tree,
1883                                           &iocom->router->staterd_tree, state);
1884                         } else {
1885                                 assert((msg->any.head.cmd & DMSGF_REPLY) == 0);
1886                                 RB_REMOVE(dmsg_state_tree,
1887                                           &iocom->router->statewr_tree, state);
1888                         }
1889                         state->flags &= ~DMSG_STATE_INSERTED;
1890                         dmsg_state_free(state);
1891                 } else {
1892                         ;
1893                 }
1894                 pthread_mutex_unlock(&iocom->mtx);
1895                 dmsg_msg_free(msg);
1896         } else if (state->msg != msg) {
1897                 dmsg_msg_free(msg);
1898         }
1899 }
1900
1901 /*
1902  * Called with iocom locked
1903  */
1904 void
1905 dmsg_state_free(dmsg_state_t *state)
1906 {
1907         dmsg_iocom_t *iocom = state->iocom;
1908         dmsg_msg_t *msg;
1909         char dummy;
1910
1911         if (DMsgDebugOpt) {
1912                 fprintf(stderr, "terminate state %p id=%08x\n",
1913                         state, (uint32_t)state->msgid);
1914         }
1915         assert(state->any.any == NULL);
1916         msg = state->msg;
1917         state->msg = NULL;
1918         if (msg)
1919                 dmsg_msg_free_locked(msg);
1920         free(state);
1921
1922         /*
1923          * When an iocom error is present we are trying to close down the
1924          * iocom, but we have to wait for all states to terminate before
1925          * we can do so.  The iocom rx code will terminate the receive side
1926          * for all transactions by simulating incoming DELETE messages,
1927          * but the state doesn't go away until both sides are terminated.
1928          *
1929          * We may have to wake up the rx code.
1930          */
1931         if (iocom->ioq_rx.error &&
1932             RB_EMPTY(&iocom->router->staterd_tree) &&
1933             RB_EMPTY(&iocom->router->statewr_tree)) {
1934                 dummy = 0;
1935                 write(iocom->wakeupfds[1], &dummy, 1);
1936         }
1937 }
1938
1939 /************************************************************************
1940  *                              ROUTING                                 *
1941  ************************************************************************
1942  *
1943  * Incoming messages are routed by their spanid, matched up against
1944  * outgoing LNK_SPANs managed by h2span_relay structures (see msg_lnk.c).
1945  * Any replies run through the same router.
1946  *
1947  * Originated messages are routed by their spanid, matched up against
1948  * incoming LNK_SPANs managed by h2span_link structures (see msg_lnk.c).
1949  * Replies come back through the same route.
1950  *
1951  * Keep in mind that ALL MESSAGE TRAFFIC pertaining to a particular
1952  * transaction runs through the same route.  Commands and replies both.
1953  *
1954  * An originated message will use a different routing spanid to
1955  * reach a target node than a message which originates from that node.
1956  * They might use the same physical pipes (each pipe can have multiple
1957  * SPANs and RELAYs), but the routes are distinct from the perspective
1958  * of the router.
1959  */
1960 dmsg_router_t *
1961 dmsg_router_alloc(void)
1962 {
1963         dmsg_router_t *router;
1964
1965         router = dmsg_alloc(sizeof(*router));
1966         TAILQ_INIT(&router->txmsgq);
1967         return (router);
1968 }
1969
1970 void
1971 dmsg_router_connect(dmsg_router_t *router)
1972 {
1973         dmsg_router_t *tmp;
1974
1975         assert(router->link || router->relay);
1976         assert((router->flags & DMSG_ROUTER_CONNECTED) == 0);
1977
1978         pthread_mutex_lock(&router_mtx);
1979         if (router->link)
1980                 tmp = RB_INSERT(dmsg_router_tree, &router_ltree, router);
1981         else
1982                 tmp = RB_INSERT(dmsg_router_tree, &router_rtree, router);
1983         assert(tmp == NULL);
1984         router->flags |= DMSG_ROUTER_CONNECTED;
1985         pthread_mutex_unlock(&router_mtx);
1986 }
1987
1988 void
1989 dmsg_router_disconnect(dmsg_router_t **routerp)
1990 {
1991         dmsg_router_t *router;
1992
1993         router = *routerp;
1994         assert(router->link || router->relay);
1995         assert(router->flags & DMSG_ROUTER_CONNECTED);
1996
1997         pthread_mutex_lock(&router_mtx);
1998         if (router->link)
1999                 RB_REMOVE(dmsg_router_tree, &router_ltree, router);
2000         else
2001                 RB_REMOVE(dmsg_router_tree, &router_rtree, router);
2002         router->flags &= ~DMSG_ROUTER_CONNECTED;
2003         *routerp = NULL;
2004         pthread_mutex_unlock(&router_mtx);
2005 }
2006
2007 #if 0
2008 /*
2009  * XXX
2010  */
2011 dmsg_router_t *
2012 dmsg_route_msg(dmsg_msg_t *msg)
2013 {
2014 }
2015 #endif
2016
2017 /*
2018  * This swaps endian for a hammer2_msg_hdr.  Note that the extended
2019  * header is not adjusted, just the core header.
2020  */
2021 void
2022 dmsg_bswap_head(dmsg_hdr_t *head)
2023 {
2024         head->magic     = bswap16(head->magic);
2025         head->reserved02 = bswap16(head->reserved02);
2026         head->salt      = bswap32(head->salt);
2027
2028         head->msgid     = bswap64(head->msgid);
2029         head->source    = bswap64(head->source);
2030         head->target    = bswap64(head->target);
2031
2032         head->cmd       = bswap32(head->cmd);
2033         head->aux_crc   = bswap32(head->aux_crc);
2034         head->aux_bytes = bswap32(head->aux_bytes);
2035         head->error     = bswap32(head->error);
2036         head->aux_descr = bswap64(head->aux_descr);
2037         head->reserved38= bswap32(head->reserved38);
2038         head->hdr_crc   = bswap32(head->hdr_crc);
2039 }