29a13e050cec4056642568fc8cd8b45c58016dae
[dragonfly.git] / sbin / hammer2 / 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 "hammer2.h"
37
38 static int hammer2_state_msgrx(hammer2_iocom_t *iocom, hammer2_msg_t *msg);
39 static void hammer2_state_cleanuptx(hammer2_iocom_t *iocom, hammer2_msg_t *msg);
40
41 /*
42  * Initialize a low-level ioq
43  */
44 void
45 hammer2_ioq_init(hammer2_iocom_t *iocom __unused, hammer2_ioq_t *ioq)
46 {
47         bzero(ioq, sizeof(*ioq));
48         ioq->state = HAMMER2_MSGQ_STATE_HEADER1;
49         TAILQ_INIT(&ioq->msgq);
50 }
51
52 /*
53  * Cleanup queue.
54  *
55  * caller holds iocom->mtx.
56  */
57 void
58 hammer2_ioq_done(hammer2_iocom_t *iocom __unused, hammer2_ioq_t *ioq)
59 {
60         hammer2_msg_t *msg;
61
62         while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
63                 assert(0);      /* shouldn't happen */
64                 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
65                 hammer2_msg_free(iocom, msg);
66         }
67         if ((msg = ioq->msg) != NULL) {
68                 ioq->msg = NULL;
69                 hammer2_msg_free(iocom, msg);
70         }
71 }
72
73 /*
74  * Initialize a low-level communications channel.
75  *
76  * NOTE: The state_func() is called at least once from the loop and can be
77  *       re-armed via hammer2_iocom_restate().
78  */
79 void
80 hammer2_iocom_init(hammer2_iocom_t *iocom, int sock_fd, int alt_fd,
81                    void (*state_func)(hammer2_iocom_t *),
82                    void (*rcvmsg_func)(hammer2_iocom_t *, hammer2_msg_t *msg),
83                    void (*altmsg_func)(hammer2_iocom_t *))
84 {
85         bzero(iocom, sizeof(*iocom));
86
87         iocom->state_callback = state_func;
88         iocom->rcvmsg_callback = rcvmsg_func;
89         iocom->altmsg_callback = altmsg_func;
90
91         pthread_mutex_init(&iocom->mtx, NULL);
92         RB_INIT(&iocom->staterd_tree);
93         RB_INIT(&iocom->statewr_tree);
94         TAILQ_INIT(&iocom->freeq);
95         TAILQ_INIT(&iocom->freeq_aux);
96         TAILQ_INIT(&iocom->addrq);
97         TAILQ_INIT(&iocom->txmsgq);
98         iocom->sock_fd = sock_fd;
99         iocom->alt_fd = alt_fd;
100         iocom->flags = HAMMER2_IOCOMF_RREQ;
101         if (state_func)
102                 iocom->flags |= HAMMER2_IOCOMF_SWORK;
103         hammer2_ioq_init(iocom, &iocom->ioq_rx);
104         hammer2_ioq_init(iocom, &iocom->ioq_tx);
105         if (pipe(iocom->wakeupfds) < 0)
106                 assert(0);
107         fcntl(iocom->wakeupfds[0], F_SETFL, O_NONBLOCK);
108         fcntl(iocom->wakeupfds[1], F_SETFL, O_NONBLOCK);
109
110         /*
111          * Negotiate session crypto synchronously.  This will mark the
112          * connection as error'd if it fails.
113          */
114         hammer2_crypto_negotiate(iocom);
115
116         /*
117          * Make sure our fds are set to non-blocking for the iocom core.
118          */
119         if (sock_fd >= 0)
120                 fcntl(sock_fd, F_SETFL, O_NONBLOCK);
121 #if 0
122         /* if line buffered our single fgets() should be fine */
123         if (alt_fd >= 0)
124                 fcntl(alt_fd, F_SETFL, O_NONBLOCK);
125 #endif
126 }
127
128 /*
129  * May only be called from a callback from iocom_core.
130  *
131  * Adjust state machine functions, set flags to guarantee that both
132  * the recevmsg_func and the sendmsg_func is called at least once.
133  */
134 void
135 hammer2_iocom_restate(hammer2_iocom_t *iocom,
136                    void (*state_func)(hammer2_iocom_t *),
137                    void (*rcvmsg_func)(hammer2_iocom_t *, hammer2_msg_t *msg),
138                    void (*altmsg_func)(hammer2_iocom_t *))
139 {
140         iocom->state_callback = state_func;
141         iocom->rcvmsg_callback = rcvmsg_func;
142         iocom->altmsg_callback = altmsg_func;
143         if (state_func)
144                 iocom->flags |= HAMMER2_IOCOMF_SWORK;
145         else
146                 iocom->flags &= ~HAMMER2_IOCOMF_SWORK;
147 }
148
149 /*
150  * Cleanup a terminating iocom.
151  *
152  * Caller should not hold iocom->mtx.  The iocom has already been disconnected
153  * from all possible references to it.
154  */
155 void
156 hammer2_iocom_done(hammer2_iocom_t *iocom)
157 {
158         hammer2_msg_t *msg;
159
160         if (iocom->sock_fd >= 0) {
161                 close(iocom->sock_fd);
162                 iocom->sock_fd = -1;
163         }
164         if (iocom->alt_fd >= 0) {
165                 close(iocom->alt_fd);
166                 iocom->alt_fd = -1;
167         }
168         hammer2_ioq_done(iocom, &iocom->ioq_rx);
169         hammer2_ioq_done(iocom, &iocom->ioq_tx);
170         if ((msg = TAILQ_FIRST(&iocom->freeq)) != NULL) {
171                 TAILQ_REMOVE(&iocom->freeq, msg, qentry);
172                 free(msg);
173         }
174         if ((msg = TAILQ_FIRST(&iocom->freeq_aux)) != NULL) {
175                 TAILQ_REMOVE(&iocom->freeq_aux, msg, qentry);
176                 free(msg->aux_data);
177                 msg->aux_data = NULL;
178                 free(msg);
179         }
180         if (iocom->wakeupfds[0] >= 0) {
181                 close(iocom->wakeupfds[0]);
182                 iocom->wakeupfds[0] = -1;
183         }
184         if (iocom->wakeupfds[1] >= 0) {
185                 close(iocom->wakeupfds[1]);
186                 iocom->wakeupfds[1] = -1;
187         }
188         pthread_mutex_destroy(&iocom->mtx);
189 }
190
191 /*
192  * Allocate a new one-way message.
193  */
194 hammer2_msg_t *
195 hammer2_msg_alloc(hammer2_iocom_t *iocom, size_t aux_size, uint32_t cmd)
196 {
197         hammer2_msg_t *msg;
198         int hbytes;
199
200         pthread_mutex_lock(&iocom->mtx);
201         if (aux_size) {
202                 aux_size = (aux_size + HAMMER2_MSG_ALIGNMASK) &
203                            ~HAMMER2_MSG_ALIGNMASK;
204                 if ((msg = TAILQ_FIRST(&iocom->freeq_aux)) != NULL)
205                         TAILQ_REMOVE(&iocom->freeq_aux, msg, qentry);
206         } else {
207                 if ((msg = TAILQ_FIRST(&iocom->freeq)) != NULL)
208                         TAILQ_REMOVE(&iocom->freeq, msg, qentry);
209         }
210         pthread_mutex_unlock(&iocom->mtx);
211         if (msg == NULL) {
212                 msg = malloc(sizeof(*msg));
213                 bzero(msg, sizeof(*msg));
214                 msg->aux_data = NULL;
215                 msg->aux_size = 0;
216         }
217         if (msg->aux_size != aux_size) {
218                 if (msg->aux_data) {
219                         free(msg->aux_data);
220                         msg->aux_data = NULL;
221                         msg->aux_size = 0;
222                 }
223                 if (aux_size) {
224                         msg->aux_data = malloc(aux_size);
225                         msg->aux_size = aux_size;
226                 }
227         }
228         hbytes = (cmd & HAMMER2_MSGF_SIZE) * HAMMER2_MSG_ALIGN;
229         if (hbytes)
230                 bzero(&msg->any.head, hbytes);
231         msg->hdr_size = hbytes;
232         msg->any.head.cmd = cmd;
233         msg->any.head.aux_descr = 0;
234         msg->any.head.aux_crc = 0;
235
236         return (msg);
237 }
238
239 /*
240  * Free a message so it can be reused afresh.
241  *
242  * NOTE: aux_size can be 0 with a non-NULL aux_data.
243  */
244 static
245 void
246 hammer2_msg_free_locked(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
247 {
248         msg->state = NULL;
249         if (msg->aux_data)
250                 TAILQ_INSERT_TAIL(&iocom->freeq_aux, msg, qentry);
251         else
252                 TAILQ_INSERT_TAIL(&iocom->freeq, msg, qentry);
253 }
254
255 void
256 hammer2_msg_free(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
257 {
258         pthread_mutex_lock(&iocom->mtx);
259         hammer2_msg_free_locked(iocom, msg);
260         pthread_mutex_unlock(&iocom->mtx);
261 }
262
263 /*
264  * I/O core loop for an iocom.
265  *
266  * Thread localized, iocom->mtx not held.
267  */
268 void
269 hammer2_iocom_core(hammer2_iocom_t *iocom)
270 {
271         struct pollfd fds[3];
272         char dummybuf[256];
273         hammer2_msg_t *msg;
274         int timeout;
275         int count;
276         int wi; /* wakeup pipe */
277         int si; /* socket */
278         int ai; /* alt bulk path socket */
279
280         while ((iocom->flags & HAMMER2_IOCOMF_EOF) == 0) {
281                 if ((iocom->flags & (HAMMER2_IOCOMF_RWORK |
282                                      HAMMER2_IOCOMF_WWORK |
283                                      HAMMER2_IOCOMF_PWORK |
284                                      HAMMER2_IOCOMF_SWORK |
285                                      HAMMER2_IOCOMF_ARWORK |
286                                      HAMMER2_IOCOMF_AWWORK)) == 0) {
287                         /*
288                          * Only poll if no immediate work is pending.
289                          * Otherwise we are just wasting our time calling
290                          * poll.
291                          */
292                         timeout = 5000;
293
294                         count = 0;
295                         wi = -1;
296                         si = -1;
297                         ai = -1;
298
299                         /*
300                          * Always check the inter-thread pipe, e.g.
301                          * for iocom->txmsgq work.
302                          */
303                         wi = count++;
304                         fds[wi].fd = iocom->wakeupfds[0];
305                         fds[wi].events = POLLIN;
306                         fds[wi].revents = 0;
307
308                         /*
309                          * Check the socket input/output direction as
310                          * requested
311                          */
312                         if (iocom->flags & (HAMMER2_IOCOMF_RREQ |
313                                             HAMMER2_IOCOMF_WREQ)) {
314                                 si = count++;
315                                 fds[si].fd = iocom->sock_fd;
316                                 fds[si].events = 0;
317                                 fds[si].revents = 0;
318
319                                 if (iocom->flags & HAMMER2_IOCOMF_RREQ)
320                                         fds[si].events |= POLLIN;
321                                 if (iocom->flags & HAMMER2_IOCOMF_WREQ)
322                                         fds[si].events |= POLLOUT;
323                         }
324
325                         /*
326                          * Check the alternative fd for work.
327                          */
328                         if (iocom->alt_fd >= 0) {
329                                 ai = count++;
330                                 fds[ai].fd = iocom->alt_fd;
331                                 fds[ai].events = POLLIN;
332                                 fds[ai].revents = 0;
333                         }
334                         poll(fds, count, timeout);
335
336                         if (wi >= 0 && (fds[wi].revents & POLLIN))
337                                 iocom->flags |= HAMMER2_IOCOMF_PWORK;
338                         if (si >= 0 && (fds[si].revents & POLLIN))
339                                 iocom->flags |= HAMMER2_IOCOMF_RWORK;
340                         if (si >= 0 && (fds[si].revents & POLLOUT))
341                                 iocom->flags |= HAMMER2_IOCOMF_WWORK;
342                         if (wi >= 0 && (fds[wi].revents & POLLOUT))
343                                 iocom->flags |= HAMMER2_IOCOMF_WWORK;
344                         if (ai >= 0 && (fds[ai].revents & POLLIN))
345                                 iocom->flags |= HAMMER2_IOCOMF_ARWORK;
346                 } else {
347                         /*
348                          * Always check the pipe
349                          */
350                         iocom->flags |= HAMMER2_IOCOMF_PWORK;
351                 }
352
353                 if (iocom->flags & HAMMER2_IOCOMF_SWORK) {
354                         iocom->flags &= ~HAMMER2_IOCOMF_SWORK;
355                         iocom->state_callback(iocom);
356                 }
357
358                 /*
359                  * Pending message queues from other threads wake us up
360                  * with a write to the wakeupfds[] pipe.  We have to clear
361                  * the pipe with a dummy read.
362                  */
363                 if (iocom->flags & HAMMER2_IOCOMF_PWORK) {
364                         iocom->flags &= ~HAMMER2_IOCOMF_PWORK;
365                         read(iocom->wakeupfds[0], dummybuf, sizeof(dummybuf));
366                         iocom->flags |= HAMMER2_IOCOMF_RWORK;
367                         iocom->flags |= HAMMER2_IOCOMF_WWORK;
368                         if (TAILQ_FIRST(&iocom->txmsgq))
369                                 hammer2_iocom_flush1(iocom);
370                 }
371
372                 /*
373                  * Message write sequencing
374                  */
375                 if (iocom->flags & HAMMER2_IOCOMF_WWORK)
376                         hammer2_iocom_flush1(iocom);
377
378                 /*
379                  * Message read sequencing.  Run this after the write
380                  * sequencing in case the write sequencing allowed another
381                  * auto-DELETE to occur on the read side.
382                  */
383                 if (iocom->flags & HAMMER2_IOCOMF_RWORK) {
384                         while ((iocom->flags & HAMMER2_IOCOMF_EOF) == 0 &&
385                                (msg = hammer2_ioq_read(iocom)) != NULL) {
386                                 fprintf(stderr,
387                                         "receive msg cmd=%08x msgid=%016jx\n",
388                                         msg->any.head.cmd,
389                                         (intmax_t)msg->any.head.msgid);
390                                 iocom->rcvmsg_callback(iocom, msg);
391                                 hammer2_state_cleanuprx(iocom, msg);
392                         }
393                 }
394
395                 if (iocom->flags & HAMMER2_IOCOMF_ARWORK) {
396                         iocom->flags &= ~HAMMER2_IOCOMF_ARWORK;
397                         iocom->altmsg_callback(iocom);
398                 }
399         }
400 }
401
402 /*
403  * Read the next ready message from the ioq, issuing I/O if needed.
404  * Caller should retry on a read-event when NULL is returned.
405  *
406  * If an error occurs during reception a HAMMER2_LNK_ERROR msg will
407  * be returned for each open transaction, then the ioq and iocom
408  * will be errored out and a non-transactional HAMMER2_LNK_ERROR
409  * msg will be returned as the final message.  The caller should not call
410  * us again after the final message is returned.
411  *
412  * Thread localized, iocom->mtx not held.
413  */
414 hammer2_msg_t *
415 hammer2_ioq_read(hammer2_iocom_t *iocom)
416 {
417         hammer2_ioq_t *ioq = &iocom->ioq_rx;
418         hammer2_msg_t *msg;
419         hammer2_msg_hdr_t *head;
420         hammer2_state_t *state;
421         ssize_t n;
422         size_t bytes;
423         size_t nmax;
424         uint32_t xcrc32;
425         int error;
426
427 again:
428         iocom->flags &= ~(HAMMER2_IOCOMF_RREQ | HAMMER2_IOCOMF_RWORK);
429
430         /*
431          * If a message is already pending we can just remove and
432          * return it.  Message state has already been processed.
433          */
434         if ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
435                 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
436                 return (msg);
437         }
438
439         /*
440          * Message read in-progress (msg is NULL at the moment).  We don't
441          * allocate a msg until we have its core header.
442          */
443         bytes = ioq->fifo_end - ioq->fifo_beg;
444         nmax = sizeof(ioq->buf) - ioq->fifo_end;
445         msg = ioq->msg;
446
447         switch(ioq->state) {
448         case HAMMER2_MSGQ_STATE_HEADER1:
449                 /*
450                  * Load the primary header, fail on any non-trivial read
451                  * error or on EOF.  Since the primary header is the same
452                  * size is the message alignment it will never straddle
453                  * the end of the buffer.
454                  */
455                 if (bytes < (int)sizeof(msg->any.head)) {
456                         n = read(iocom->sock_fd,
457                                  ioq->buf + ioq->fifo_end,
458                                  nmax);
459                         if (n <= 0) {
460                                 if (n == 0) {
461                                         ioq->error = HAMMER2_IOQ_ERROR_EOF;
462                                         break;
463                                 }
464                                 if (errno != EINTR &&
465                                     errno != EINPROGRESS &&
466                                     errno != EAGAIN) {
467                                         ioq->error = HAMMER2_IOQ_ERROR_SOCK;
468                                         break;
469                                 }
470                                 n = 0;
471                                 /* fall through */
472                         }
473                         ioq->fifo_end += n;
474                         bytes += n;
475                         nmax -= n;
476                 }
477
478                 /*
479                  * Insufficient data accumulated (msg is NULL, caller will
480                  * retry on event).
481                  */
482                 assert(msg == NULL);
483                 if (bytes < (int)sizeof(msg->any.head))
484                         break;
485
486                 /*
487                  * Calculate the header, decrypt data received so far.
488                  * Data will be decrypted in-place.  Partial blocks are
489                  * not immediately decrypted.
490                  *
491                  * WARNING!  The header might be in the wrong endian, we
492                  *           do not fix it up until we get the entire
493                  *           extended header.
494                  */
495                 hammer2_crypto_decrypt(iocom, ioq);
496                 head = (void *)(ioq->buf + ioq->fifo_beg);
497
498                 /*
499                  * Check and fixup the core header.  Note that the icrc
500                  * has to be calculated before any fixups, but the crc
501                  * fields in the msg may have to be swapped like everything
502                  * else.
503                  */
504                 if (head->magic != HAMMER2_MSGHDR_MAGIC &&
505                     head->magic != HAMMER2_MSGHDR_MAGIC_REV) {
506                         ioq->error = HAMMER2_IOQ_ERROR_SYNC;
507                         break;
508                 }
509
510                 /*
511                  * Calculate the full header size and aux data size
512                  */
513                 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV) {
514                         ioq->hbytes = (bswap32(head->cmd) & HAMMER2_MSGF_SIZE) *
515                                       HAMMER2_MSG_ALIGN;
516                         ioq->abytes = bswap32(head->aux_bytes) *
517                                       HAMMER2_MSG_ALIGN;
518                 } else {
519                         ioq->hbytes = (head->cmd & HAMMER2_MSGF_SIZE) *
520                                       HAMMER2_MSG_ALIGN;
521                         ioq->abytes = head->aux_bytes * HAMMER2_MSG_ALIGN;
522                 }
523                 if (ioq->hbytes < sizeof(msg->any.head) ||
524                     ioq->hbytes > sizeof(msg->any) ||
525                     ioq->abytes > HAMMER2_MSGAUX_MAX) {
526                         ioq->error = HAMMER2_IOQ_ERROR_FIELD;
527                         break;
528                 }
529
530                 /*
531                  * Allocate the message, the next state will fill it in.
532                  */
533                 msg = hammer2_msg_alloc(iocom, ioq->abytes, 0);
534                 ioq->msg = msg;
535
536                 /*
537                  * Fall through to the next state.  Make sure that the
538                  * extended header does not straddle the end of the buffer.
539                  * We still want to issue larger reads into our buffer,
540                  * book-keeping is easier if we don't bcopy() yet.
541                  */
542                 if (bytes + nmax < ioq->hbytes) {
543                         bcopy(ioq->buf + ioq->fifo_beg, ioq->buf, bytes);
544                         ioq->fifo_cdx -= ioq->fifo_beg;
545                         ioq->fifo_beg = 0;
546                         ioq->fifo_end = bytes;
547                         nmax = sizeof(ioq->buf) - ioq->fifo_end;
548                 }
549                 ioq->state = HAMMER2_MSGQ_STATE_HEADER2;
550                 /* fall through */
551         case HAMMER2_MSGQ_STATE_HEADER2:
552                 /*
553                  * Fill out the extended header.
554                  */
555                 assert(msg != NULL);
556                 if (bytes < ioq->hbytes) {
557                         n = read(iocom->sock_fd,
558                                  ioq->buf + ioq->fifo_end,
559                                  nmax);
560                         if (n <= 0) {
561                                 if (n == 0) {
562                                         ioq->error = HAMMER2_IOQ_ERROR_EOF;
563                                         break;
564                                 }
565                                 if (errno != EINTR &&
566                                     errno != EINPROGRESS &&
567                                     errno != EAGAIN) {
568                                         ioq->error = HAMMER2_IOQ_ERROR_SOCK;
569                                         break;
570                                 }
571                                 n = 0;
572                                 /* fall through */
573                         }
574                         ioq->fifo_end += n;
575                         bytes += n;
576                         nmax -= n;
577                 }
578
579                 /*
580                  * Insufficient data accumulated (set msg NULL so caller will
581                  * retry on event).
582                  */
583                 if (bytes < ioq->hbytes) {
584                         msg = NULL;
585                         break;
586                 }
587
588                 /*
589                  * Calculate the extended header, decrypt data received
590                  * so far.  Handle endian-conversion for the entire extended
591                  * header.
592                  */
593                 hammer2_crypto_decrypt(iocom, ioq);
594                 head = (void *)(ioq->buf + ioq->fifo_beg);
595
596                 /*
597                  * Check the CRC.
598                  */
599                 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV)
600                         xcrc32 = bswap32(head->hdr_crc);
601                 else
602                         xcrc32 = head->hdr_crc;
603                 head->hdr_crc = 0;
604                 if (hammer2_icrc32(head, ioq->hbytes) != xcrc32) {
605                         ioq->error = HAMMER2_IOQ_ERROR_XCRC;
606                         fprintf(stderr, "XCRC FAILED %08x %08x\n",
607                                 xcrc32, hammer2_icrc32(head, ioq->hbytes));
608                         assert(0);
609                         break;
610                 }
611                 head->hdr_crc = xcrc32;
612
613                 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV) {
614                         hammer2_bswap_head(head);
615                 }
616
617                 /*
618                  * Copy the extended header into the msg and adjust the
619                  * FIFO.
620                  */
621                 bcopy(head, &msg->any, ioq->hbytes);
622
623                 /*
624                  * We are either done or we fall-through.
625                  */
626                 if (ioq->abytes == 0) {
627                         ioq->fifo_beg += ioq->hbytes;
628                         break;
629                 }
630
631                 /*
632                  * Must adjust nmax and bytes (and the state) when falling
633                  * through.
634                  */
635                 ioq->fifo_beg += ioq->hbytes;
636                 nmax -= ioq->hbytes;
637                 bytes -= ioq->hbytes;
638                 ioq->state = HAMMER2_MSGQ_STATE_AUXDATA1;
639                 /* fall through */
640         case HAMMER2_MSGQ_STATE_AUXDATA1:
641                 /*
642                  * Copy the partial or complete payload from remaining
643                  * bytes in the FIFO.  We have to fall-through either
644                  * way so we can check the crc.
645                  *
646                  * Adjust msg->aux_size to the final actual value.
647                  */
648                 ioq->already = ioq->fifo_cdx - ioq->fifo_beg;
649                 if (ioq->already > ioq->abytes)
650                         ioq->already = ioq->abytes;
651                 if (bytes >= ioq->abytes) {
652                         bcopy(ioq->buf + ioq->fifo_beg, msg->aux_data,
653                               ioq->abytes);
654                         msg->aux_size = ioq->abytes;
655                         ioq->fifo_beg += ioq->abytes;
656                         if (ioq->fifo_cdx < ioq->fifo_beg)
657                                 ioq->fifo_cdx = ioq->fifo_beg;
658                         bytes -= ioq->abytes;
659                 } else if (bytes) {
660                         bcopy(ioq->buf + ioq->fifo_beg, msg->aux_data,
661                               bytes);
662                         msg->aux_size = bytes;
663                         ioq->fifo_beg += bytes;
664                         if (ioq->fifo_cdx < ioq->fifo_beg)
665                                 ioq->fifo_cdx = ioq->fifo_beg;
666                         bytes = 0;
667                 } else {
668                         msg->aux_size = 0;
669                 }
670                 ioq->state = HAMMER2_MSGQ_STATE_AUXDATA2;
671                 /* fall through */
672         case HAMMER2_MSGQ_STATE_AUXDATA2:
673                 /*
674                  * Read the remainder of the payload directly into the
675                  * msg->aux_data buffer.
676                  */
677                 assert(msg);
678                 if (msg->aux_size < ioq->abytes) {
679                         assert(bytes == 0);
680                         n = read(iocom->sock_fd,
681                                  msg->aux_data + msg->aux_size,
682                                  ioq->abytes - msg->aux_size);
683                         if (n <= 0) {
684                                 if (n == 0) {
685                                         ioq->error = HAMMER2_IOQ_ERROR_EOF;
686                                         break;
687                                 }
688                                 if (errno != EINTR &&
689                                     errno != EINPROGRESS &&
690                                     errno != EAGAIN) {
691                                         ioq->error = HAMMER2_IOQ_ERROR_SOCK;
692                                         break;
693                                 }
694                                 n = 0;
695                                 /* fall through */
696                         }
697                         msg->aux_size += n;
698                 }
699
700                 /*
701                  * Insufficient data accumulated (set msg NULL so caller will
702                  * retry on event).
703                  */
704                 if (msg->aux_size < ioq->abytes) {
705                         msg = NULL;
706                         break;
707                 }
708                 assert(msg->aux_size == ioq->abytes);
709                 hammer2_crypto_decrypt_aux(iocom, ioq, msg, ioq->already);
710
711                 /*
712                  * Check aux_crc, then we are done.
713                  */
714                 xcrc32 = hammer2_icrc32(msg->aux_data, msg->aux_size);
715                 if (xcrc32 != msg->any.head.aux_crc) {
716                         ioq->error = HAMMER2_IOQ_ERROR_ACRC;
717                         break;
718                 }
719                 break;
720         case HAMMER2_MSGQ_STATE_ERROR:
721                 /*
722                  * Continued calls to drain recorded transactions (returning
723                  * a LNK_ERROR for each one), before we return the final
724                  * LNK_ERROR.
725                  */
726                 assert(msg == NULL);
727                 break;
728         default:
729                 /*
730                  * We don't double-return errors, the caller should not
731                  * have called us again after getting an error msg.
732                  */
733                 assert(0);
734                 break;
735         }
736
737         /*
738          * Check the message sequence.  The iv[] should prevent any
739          * possibility of a replay but we add this check anyway.
740          */
741         if (msg && ioq->error == 0) {
742                 if ((msg->any.head.salt & 255) != (ioq->seq & 255)) {
743                         ioq->error = HAMMER2_IOQ_ERROR_MSGSEQ;
744                 } else {
745                         ++ioq->seq;
746                 }
747         }
748
749         /*
750          * Process transactional state for the message.
751          */
752         if (msg && ioq->error == 0) {
753                 error = hammer2_state_msgrx(iocom, msg);
754                 if (error) {
755                         if (error == HAMMER2_IOQ_ERROR_EALREADY) {
756                                 hammer2_msg_free(iocom, msg);
757                                 goto again;
758                         }
759                         ioq->error = error;
760                 }
761         }
762
763         /*
764          * Handle error, RREQ, or completion
765          *
766          * NOTE: nmax and bytes are invalid at this point, we don't bother
767          *       to update them when breaking out.
768          */
769         if (ioq->error) {
770                 /*
771                  * An unrecoverable error causes all active receive
772                  * transactions to be terminated with a LNK_ERROR message.
773                  *
774                  * Once all active transactions are exhausted we set the
775                  * iocom ERROR flag and return a non-transactional LNK_ERROR
776                  * message, which should cause master processing loops to
777                  * terminate.
778                  */
779                 assert(ioq->msg == msg);
780                 if (msg) {
781                         hammer2_msg_free(iocom, msg);
782                         ioq->msg = NULL;
783                 }
784
785                 /*
786                  * No more I/O read processing
787                  */
788                 ioq->state = HAMMER2_MSGQ_STATE_ERROR;
789
790                 /*
791                  * Simulate a remote LNK_ERROR DELETE msg for any open
792                  * transactions, ending with a final non-transactional
793                  * LNK_ERROR (that the session can detect) when no
794                  * transactions remain.
795                  */
796                 msg = hammer2_msg_alloc(iocom, 0, 0);
797                 bzero(&msg->any.head, sizeof(msg->any.head));
798                 msg->any.head.magic = HAMMER2_MSGHDR_MAGIC;
799                 msg->any.head.cmd = HAMMER2_LNK_ERROR;
800                 msg->any.head.error = ioq->error;
801
802                 pthread_mutex_lock(&iocom->mtx);
803                 fprintf(stderr, "CHECK REMAINING RXMSGS\n");
804                 if ((state = RB_ROOT(&iocom->staterd_tree)) != NULL) {
805                         /*
806                          * Active remote transactions are still present.
807                          * Simulate the other end sending us a DELETE.
808                          */
809                         if (state->rxcmd & HAMMER2_MSGF_DELETE) {
810                                 fprintf(stderr, "SIMULATE DELETION RCONT %p\n", state);
811                                 hammer2_msg_free(iocom, msg);
812                                 msg = NULL;
813                         } else {
814                                 fprintf(stderr, "SIMULATE DELETION %p RD RXCMD %08x\n", state, state->rxcmd);
815                                 /*state->txcmd |= HAMMER2_MSGF_DELETE;*/
816                                 msg->state = state;
817                                 msg->any.head.spanid = state->spanid;
818                                 msg->any.head.msgid = state->msgid;
819                                 msg->any.head.cmd |= HAMMER2_MSGF_ABORT |
820                                                      HAMMER2_MSGF_DELETE;
821                         }
822                 } else if ((state = RB_ROOT(&iocom->statewr_tree)) != NULL) {
823                         /*
824                          * Active local transactions are still present.
825                          * Simulate the other end sending us a DELETE.
826                          */
827                         if (state->rxcmd & HAMMER2_MSGF_DELETE) {
828                                 fprintf(stderr, "SIMULATE DELETION WCONT STATE->txcmd = %08x rxcmd = %08x msgid=%016jx\n", state->txcmd, state->rxcmd, state->msgid );
829                                 hammer2_msg_free(iocom, msg);
830                                 msg = NULL;
831                         } else {
832                                 fprintf(stderr, "SIMULATE DELETION WD RXCMD %08x\n", state->txcmd);
833                                 /*state->txcmd |= HAMMER2_MSGF_DELETE;*/
834                                 msg->state = state;
835                                 msg->any.head.spanid = state->spanid;
836                                 msg->any.head.msgid = state->msgid;
837                                 msg->any.head.cmd |= HAMMER2_MSGF_ABORT |
838                                                      HAMMER2_MSGF_DELETE |
839                                                      HAMMER2_MSGF_REPLY;
840                                 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
841                                         msg->any.head.cmd |=
842                                                      HAMMER2_MSGF_CREATE;
843                                 }
844                         }
845                 } else {
846                         /*
847                          * No active local or remote transactions remain.
848                          * Generate a final LNK_ERROR and flag EOF.
849                          */
850                         msg->state = NULL;
851                         iocom->flags |= HAMMER2_IOCOMF_EOF;
852                         fprintf(stderr, "EOF ON SOCKET\n");
853                 }
854                 pthread_mutex_unlock(&iocom->mtx);
855
856                 /*
857                  * For the iocom error case we want to set RWORK to indicate
858                  * that more messages might be pending.
859                  *
860                  * It is possible to return NULL when there is more work to
861                  * do because each message has to be DELETEd in both
862                  * directions before we continue on with the next (though
863                  * this could be optimized).  The transmit direction will
864                  * re-set RWORK.
865                  */
866                 if (msg)
867                         iocom->flags |= HAMMER2_IOCOMF_RWORK;
868         } else if (msg == NULL) {
869                 /*
870                  * Insufficient data received to finish building the message,
871                  * set RREQ and return NULL.
872                  *
873                  * Leave ioq->msg intact.
874                  * Leave the FIFO intact.
875                  */
876                 iocom->flags |= HAMMER2_IOCOMF_RREQ;
877         } else {
878                 /*
879                  * Return msg.
880                  *
881                  * The fifo has already been advanced past the message.
882                  * Trivially reset the FIFO indices if possible.
883                  *
884                  * clear the FIFO if it is now empty and set RREQ to wait
885                  * for more from the socket.  If the FIFO is not empty set
886                  * TWORK to bypass the poll so we loop immediately.
887                  */
888                 if (ioq->fifo_beg == ioq->fifo_end) {
889                         iocom->flags |= HAMMER2_IOCOMF_RREQ;
890                         ioq->fifo_cdx = 0;
891                         ioq->fifo_beg = 0;
892                         ioq->fifo_end = 0;
893                 } else {
894                         iocom->flags |= HAMMER2_IOCOMF_RWORK;
895                 }
896                 ioq->state = HAMMER2_MSGQ_STATE_HEADER1;
897                 ioq->msg = NULL;
898         }
899         return (msg);
900 }
901
902 /*
903  * Calculate the header and data crc's and write a low-level message to
904  * the connection.  If aux_crc is non-zero the aux_data crc is already
905  * assumed to have been set.
906  *
907  * A non-NULL msg is added to the queue but not necessarily flushed.
908  * Calling this function with msg == NULL will get a flush going.
909  *
910  * Caller must hold iocom->mtx.
911  */
912 void
913 hammer2_iocom_flush1(hammer2_iocom_t *iocom)
914 {
915         hammer2_ioq_t *ioq = &iocom->ioq_tx;
916         hammer2_msg_t *msg;
917         uint32_t xcrc32;
918         int hbytes;
919         hammer2_msg_queue_t tmpq;
920
921         iocom->flags &= ~(HAMMER2_IOCOMF_WREQ | HAMMER2_IOCOMF_WWORK);
922         TAILQ_INIT(&tmpq);
923         pthread_mutex_lock(&iocom->mtx);
924         while ((msg = TAILQ_FIRST(&iocom->txmsgq)) != NULL) {
925                 TAILQ_REMOVE(&iocom->txmsgq, msg, qentry);
926                 TAILQ_INSERT_TAIL(&tmpq, msg, qentry);
927         }
928         pthread_mutex_unlock(&iocom->mtx);
929
930         while ((msg = TAILQ_FIRST(&tmpq)) != NULL) {
931                 /*
932                  * Process terminal connection errors.
933                  */
934                 TAILQ_REMOVE(&tmpq, msg, qentry);
935                 if (ioq->error) {
936                         TAILQ_INSERT_TAIL(&ioq->msgq, msg, qentry);
937                         ++ioq->msgcount;
938                         continue;
939                 }
940
941                 /*
942                  * Finish populating the msg fields.  The salt ensures that
943                  * the iv[] array is ridiculously randomized and we also
944                  * re-seed our PRNG every 32768 messages just to be sure.
945                  */
946                 msg->any.head.magic = HAMMER2_MSGHDR_MAGIC;
947                 msg->any.head.salt = (random() << 8) | (ioq->seq & 255);
948                 ++ioq->seq;
949                 if ((ioq->seq & 32767) == 0)
950                         srandomdev();
951
952                 /*
953                  * Calculate aux_crc if 0, then calculate hdr_crc.
954                  */
955                 if (msg->aux_size && msg->any.head.aux_crc == 0) {
956                         assert((msg->aux_size & HAMMER2_MSG_ALIGNMASK) == 0);
957                         xcrc32 = hammer2_icrc32(msg->aux_data, msg->aux_size);
958                         msg->any.head.aux_crc = xcrc32;
959                 }
960                 msg->any.head.aux_bytes = msg->aux_size / HAMMER2_MSG_ALIGN;
961                 assert((msg->aux_size & HAMMER2_MSG_ALIGNMASK) == 0);
962
963                 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
964                          HAMMER2_MSG_ALIGN;
965                 msg->any.head.hdr_crc = 0;
966                 msg->any.head.hdr_crc = hammer2_icrc32(&msg->any.head, hbytes);
967
968                 /*
969                  * Enqueue the message (the flush codes handles stream
970                  * encryption).
971                  */
972                 TAILQ_INSERT_TAIL(&ioq->msgq, msg, qentry);
973                 ++ioq->msgcount;
974         }
975         hammer2_iocom_flush2(iocom);
976 }
977
978 /*
979  * Thread localized, iocom->mtx not held by caller.
980  */
981 void
982 hammer2_iocom_flush2(hammer2_iocom_t *iocom)
983 {
984         hammer2_ioq_t *ioq = &iocom->ioq_tx;
985         hammer2_msg_t *msg;
986         ssize_t nmax;
987         ssize_t omax;
988         ssize_t nact;
989         struct iovec iov[HAMMER2_IOQ_MAXIOVEC];
990         size_t hbytes;
991         size_t abytes;
992         size_t hoff;
993         size_t aoff;
994         int n;
995
996         if (ioq->error) {
997                 hammer2_iocom_drain(iocom);
998                 return;
999         }
1000
1001         /*
1002          * Pump messages out the connection by building an iovec.
1003          *
1004          * ioq->hbytes/ioq->abytes tracks how much of the first message
1005          * in the queue has been successfully written out, so we can
1006          * resume writing.
1007          */
1008         n = 0;
1009         nmax = 0;
1010         hoff = ioq->hbytes;
1011         aoff = ioq->abytes;
1012
1013         TAILQ_FOREACH(msg, &ioq->msgq, qentry) {
1014                 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
1015                          HAMMER2_MSG_ALIGN;
1016                 abytes = msg->aux_size;
1017                 assert(hoff <= hbytes && aoff <= abytes);
1018
1019                 if (hoff < hbytes) {
1020                         iov[n].iov_base = (char *)&msg->any.head + hoff;
1021                         iov[n].iov_len = hbytes - hoff;
1022                         nmax += hbytes - hoff;
1023                         ++n;
1024                         if (n == HAMMER2_IOQ_MAXIOVEC)
1025                                 break;
1026                 }
1027                 if (aoff < abytes) {
1028                         assert(msg->aux_data != NULL);
1029                         iov[n].iov_base = (char *)msg->aux_data + aoff;
1030                         iov[n].iov_len = abytes - aoff;
1031                         nmax += abytes - aoff;
1032                         ++n;
1033                         if (n == HAMMER2_IOQ_MAXIOVEC)
1034                                 break;
1035                 }
1036                 hoff = 0;
1037                 aoff = 0;
1038         }
1039         if (n == 0)
1040                 return;
1041
1042         /*
1043          * Encrypt and write the data.  The crypto code will move the
1044          * data into the fifo and adjust the iov as necessary.  If
1045          * encryption is disabled the iov is left alone.
1046          *
1047          * May return a smaller iov (thus a smaller n), with aggregated
1048          * chunks.  May reduce nmax to what fits in the FIFO.
1049          */
1050         omax = nmax;
1051         n = hammer2_crypto_encrypt(iocom, ioq, iov, n, &nmax);
1052
1053         /*
1054          * Execute the writev() then figure out what happened.
1055          */
1056         nact = writev(iocom->sock_fd, iov, n);
1057         if (nact < 0) {
1058                 if (errno != EINTR &&
1059                     errno != EINPROGRESS &&
1060                     errno != EAGAIN) {
1061                         /*
1062                          * Fatal write error
1063                          */
1064                         ioq->error = HAMMER2_IOQ_ERROR_SOCK;
1065                         hammer2_iocom_drain(iocom);
1066                 } else {
1067                         /*
1068                          * Wait for socket buffer space
1069                          */
1070                         iocom->flags |= HAMMER2_IOCOMF_WREQ;
1071                 }
1072                 return;
1073         }
1074
1075         /*
1076          * Indicate bytes written successfully.
1077          *
1078          * If we were unable to write the entire iov array then set WREQ
1079          * to wait for more socket buffer space.
1080          *
1081          * If the FIFO space was insufficient to fully drain all messages
1082          * set WWORK to cause the core to call us again for the next batch.
1083          */
1084         hammer2_crypto_encrypt_wrote(iocom, ioq, nact);
1085         if (nact != nmax)
1086                 iocom->flags |= HAMMER2_IOCOMF_WREQ;
1087         else if (nmax != omax)
1088                 iocom->flags |= HAMMER2_IOCOMF_WWORK;
1089
1090         /*
1091          * Clean out the transmit queue based on what we successfully
1092          * sent.  ioq->hbytes/abytes represents the portion of the first
1093          * message previously sent.
1094          */
1095         while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
1096                 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
1097                          HAMMER2_MSG_ALIGN;
1098                 abytes = msg->aux_size;
1099
1100                 if ((size_t)nact < hbytes - ioq->hbytes) {
1101                         ioq->hbytes += nact;
1102                         /* nact = 0; */
1103                         break;
1104                 }
1105                 nact -= hbytes - ioq->hbytes;
1106                 ioq->hbytes = hbytes;
1107                 if ((size_t)nact < abytes - ioq->abytes) {
1108                         ioq->abytes += nact;
1109                         /* nact = 0; */
1110                         break;
1111                 }
1112                 nact -= abytes - ioq->abytes;
1113
1114                 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
1115                 --ioq->msgcount;
1116                 ioq->hbytes = 0;
1117                 ioq->abytes = 0;
1118
1119                 hammer2_state_cleanuptx(iocom, msg);
1120         }
1121         assert(nact == 0);
1122         if (ioq->error) {
1123                 hammer2_iocom_drain(iocom);
1124         }
1125 }
1126
1127 /*
1128  * Kill pending msgs on ioq_tx and adjust the flags such that no more
1129  * write events will occur.  We don't kill read msgs because we want
1130  * the caller to pull off our contrived terminal error msg to detect
1131  * the connection failure.
1132  *
1133  * Thread localized, iocom->mtx not held by caller.
1134  */
1135 void
1136 hammer2_iocom_drain(hammer2_iocom_t *iocom)
1137 {
1138         hammer2_ioq_t *ioq = &iocom->ioq_tx;
1139         hammer2_msg_t *msg;
1140
1141         iocom->flags &= ~(HAMMER2_IOCOMF_WREQ | HAMMER2_IOCOMF_WWORK);
1142         ioq->hbytes = 0;
1143         ioq->abytes = 0;
1144
1145         while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
1146                 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
1147                 --ioq->msgcount;
1148                 hammer2_state_cleanuptx(iocom, msg);
1149         }
1150 }
1151
1152 /*
1153  * Write a message to an iocom, with additional state processing.
1154  */
1155 void
1156 hammer2_msg_write(hammer2_iocom_t *iocom, hammer2_msg_t *msg,
1157                   void (*func)(hammer2_state_t *, hammer2_msg_t *),
1158                   void *data,
1159                   hammer2_state_t **statep)
1160 {
1161         hammer2_state_t *state;
1162         char dummy;
1163
1164         /*
1165          * Handle state processing, create state if necessary.
1166          */
1167         pthread_mutex_lock(&iocom->mtx);
1168         if ((state = msg->state) != NULL) {
1169                 /*
1170                  * Existing transaction (could be reply).  It is also
1171                  * possible for this to be the first reply (CREATE is set),
1172                  * in which case we populate state->txcmd.
1173                  */
1174                 msg->any.head.msgid = state->msgid;
1175                 msg->any.head.spanid = state->spanid;
1176                 if (func) {
1177                         state->func = func;
1178                         state->any.any = data;
1179                 }
1180                 assert(((state->txcmd ^ msg->any.head.cmd) &
1181                         HAMMER2_MSGF_REPLY) == 0);
1182                 if (msg->any.head.cmd & HAMMER2_MSGF_CREATE)
1183                         state->txcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1184         } else if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
1185                 /*
1186                  * No existing state and CREATE is set, create new
1187                  * state for outgoing command.  This can't happen if
1188                  * REPLY is set as the state would already exist for
1189                  * a transaction reply.
1190                  */
1191                 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1192
1193                 state = malloc(sizeof(*state));
1194                 bzero(state, sizeof(*state));
1195                 state->iocom = iocom;
1196                 state->flags = HAMMER2_STATE_DYNAMIC;
1197                 state->msg = msg;
1198                 state->msgid = (uint64_t)(uintptr_t)state;
1199                 state->spanid = msg->any.head.spanid;
1200                 state->txcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1201                 state->rxcmd = HAMMER2_MSGF_REPLY;
1202                 state->func = func;
1203                 state->any.any = data;
1204                 RB_INSERT(hammer2_state_tree, &iocom->statewr_tree, state);
1205                 state->flags |= HAMMER2_STATE_INSERTED;
1206                 msg->state = state;
1207                 msg->any.head.msgid = state->msgid;
1208                 /* spanid set by caller */
1209         } else {
1210                 msg->any.head.msgid = 0;
1211                 /* spanid set by caller */
1212         }
1213
1214         if (statep)
1215                 *statep = state;
1216
1217         /*
1218          * Queue it for output, wake up the I/O pthread.  Note that the
1219          * I/O thread is responsible for generating the CRCs and encryption.
1220          */
1221         TAILQ_INSERT_TAIL(&iocom->txmsgq, msg, qentry);
1222         dummy = 0;
1223         write(iocom->wakeupfds[1], &dummy, 1);  /* XXX optimize me */
1224         pthread_mutex_unlock(&iocom->mtx);
1225 }
1226
1227 /*
1228  * This is a shortcut to formulate a reply to msg with a simple error code,
1229  * It can reply to and terminate a transaction, or it can reply to a one-way
1230  * messages.  A HAMMER2_LNK_ERROR command code is utilized to encode
1231  * the error code (which can be 0).  Not all transactions are terminated
1232  * with HAMMER2_LNK_ERROR status (the low level only cares about the
1233  * MSGF_DELETE flag), but most are.
1234  *
1235  * Replies to one-way messages are a bit of an oxymoron but the feature
1236  * is used by the debug (DBG) protocol.
1237  *
1238  * The reply contains no extended data.
1239  */
1240 void
1241 hammer2_msg_reply(hammer2_iocom_t *iocom, hammer2_msg_t *msg, uint32_t error)
1242 {
1243         hammer2_state_t *state = msg->state;
1244         hammer2_msg_t *nmsg;
1245         uint32_t cmd;
1246
1247
1248         /*
1249          * Reply with a simple error code and terminate the transaction.
1250          */
1251         cmd = HAMMER2_LNK_ERROR;
1252
1253         /*
1254          * Check if our direction has even been initiated yet, set CREATE.
1255          *
1256          * Check what direction this is (command or reply direction).  Note
1257          * that txcmd might not have been initiated yet.
1258          *
1259          * If our direction has already been closed we just return without
1260          * doing anything.
1261          */
1262         if (state) {
1263                 if (state->txcmd & HAMMER2_MSGF_DELETE)
1264                         return;
1265                 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
1266                         cmd |= HAMMER2_MSGF_CREATE;
1267                 if (state->txcmd & HAMMER2_MSGF_REPLY)
1268                         cmd |= HAMMER2_MSGF_REPLY;
1269                 cmd |= HAMMER2_MSGF_DELETE;
1270         } else {
1271                 if ((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0)
1272                         cmd |= HAMMER2_MSGF_REPLY;
1273         }
1274
1275         nmsg = hammer2_msg_alloc(iocom, 0, cmd);
1276         nmsg->any.head.error = error;
1277         nmsg->state = msg->state;
1278         hammer2_msg_write(iocom, nmsg, NULL, NULL, NULL);
1279 }
1280
1281 /*
1282  * Similar to hammer2_msg_reply() but leave the transaction open.  That is,
1283  * we are generating a streaming reply or an intermediate acknowledgement
1284  * of some sort as part of the higher level protocol, with more to come
1285  * later.
1286  */
1287 void
1288 hammer2_msg_result(hammer2_iocom_t *iocom, hammer2_msg_t *msg, uint32_t error)
1289 {
1290         hammer2_state_t *state = msg->state;
1291         hammer2_msg_t *nmsg;
1292         uint32_t cmd;
1293
1294
1295         /*
1296          * Reply with a simple error code and terminate the transaction.
1297          */
1298         cmd = HAMMER2_LNK_ERROR;
1299
1300         /*
1301          * Check if our direction has even been initiated yet, set CREATE.
1302          *
1303          * Check what direction this is (command or reply direction).  Note
1304          * that txcmd might not have been initiated yet.
1305          *
1306          * If our direction has already been closed we just return without
1307          * doing anything.
1308          */
1309         if (state) {
1310                 if (state->txcmd & HAMMER2_MSGF_DELETE)
1311                         return;
1312                 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
1313                         cmd |= HAMMER2_MSGF_CREATE;
1314                 if (state->txcmd & HAMMER2_MSGF_REPLY)
1315                         cmd |= HAMMER2_MSGF_REPLY;
1316                 /* continuing transaction, do not set MSGF_DELETE */
1317         } else {
1318                 if ((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0)
1319                         cmd |= HAMMER2_MSGF_REPLY;
1320         }
1321
1322         nmsg = hammer2_msg_alloc(iocom, 0, cmd);
1323         nmsg->any.head.error = error;
1324         nmsg->state = state;
1325         hammer2_msg_write(iocom, nmsg, NULL, NULL, NULL);
1326 }
1327
1328 /*
1329  * Terminate a transaction given a state structure by issuing a DELETE.
1330  */
1331 void
1332 hammer2_state_reply(hammer2_state_t *state, uint32_t error)
1333 {
1334         hammer2_msg_t *nmsg;
1335         uint32_t cmd = HAMMER2_LNK_ERROR | HAMMER2_MSGF_DELETE;
1336
1337         /*
1338          * Nothing to do if we already transmitted a delete
1339          */
1340         if (state->txcmd & HAMMER2_MSGF_DELETE)
1341                 return;
1342
1343         /*
1344          * We must also set CREATE if this is our first response to a
1345          * remote command.
1346          */
1347         if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
1348                 cmd |= HAMMER2_MSGF_CREATE;
1349
1350         /*
1351          * Set REPLY if the other end initiated the command.  Otherwise
1352          * we are the command direction.
1353          */
1354         if (state->txcmd & HAMMER2_MSGF_REPLY)
1355                 cmd |= HAMMER2_MSGF_REPLY;
1356
1357         nmsg = hammer2_msg_alloc(state->iocom, 0, cmd);
1358         nmsg->any.head.error = error;
1359         nmsg->state = state;
1360         hammer2_msg_write(state->iocom, nmsg, NULL, NULL, NULL);
1361 }
1362
1363 /************************************************************************
1364  *                      TRANSACTION STATE HANDLING                      *
1365  ************************************************************************
1366  *
1367  */
1368
1369 RB_GENERATE(hammer2_state_tree, hammer2_state, rbnode, hammer2_state_cmp);
1370
1371 /*
1372  * Process state tracking for a message after reception, prior to
1373  * execution.
1374  *
1375  * Called with msglk held and the msg dequeued.
1376  *
1377  * All messages are called with dummy state and return actual state.
1378  * (One-off messages often just return the same dummy state).
1379  *
1380  * May request that caller discard the message by setting *discardp to 1.
1381  * The returned state is not used in this case and is allowed to be NULL.
1382  *
1383  * --
1384  *
1385  * These routines handle persistent and command/reply message state via the
1386  * CREATE and DELETE flags.  The first message in a command or reply sequence
1387  * sets CREATE, the last message in a command or reply sequence sets DELETE.
1388  *
1389  * There can be any number of intermediate messages belonging to the same
1390  * sequence sent inbetween the CREATE message and the DELETE message,
1391  * which set neither flag.  This represents a streaming command or reply.
1392  *
1393  * Any command message received with CREATE set expects a reply sequence to
1394  * be returned.  Reply sequences work the same as command sequences except the
1395  * REPLY bit is also sent.  Both the command side and reply side can
1396  * degenerate into a single message with both CREATE and DELETE set.  Note
1397  * that one side can be streaming and the other side not, or neither, or both.
1398  *
1399  * The msgid is unique for the initiator.  That is, two sides sending a new
1400  * message can use the same msgid without colliding.
1401  *
1402  * --
1403  *
1404  * ABORT sequences work by setting the ABORT flag along with normal message
1405  * state.  However, ABORTs can also be sent on half-closed messages, that is
1406  * even if the command or reply side has already sent a DELETE, as long as
1407  * the message has not been fully closed it can still send an ABORT+DELETE
1408  * to terminate the half-closed message state.
1409  *
1410  * Since ABORT+DELETEs can race we silently discard ABORT's for message
1411  * state which has already been fully closed.  REPLY+ABORT+DELETEs can
1412  * also race, and in this situation the other side might have already
1413  * initiated a new unrelated command with the same message id.  Since
1414  * the abort has not set the CREATE flag the situation can be detected
1415  * and the message will also be discarded.
1416  *
1417  * Non-blocking requests can be initiated with ABORT+CREATE[+DELETE].
1418  * The ABORT request is essentially integrated into the command instead
1419  * of being sent later on.  In this situation the command implementation
1420  * detects that CREATE and ABORT are both set (vs ABORT alone) and can
1421  * special-case non-blocking operation for the command.
1422  *
1423  * NOTE!  Messages with ABORT set without CREATE or DELETE are considered
1424  *        to be mid-stream aborts for command/reply sequences.  ABORTs on
1425  *        one-way messages are not supported.
1426  *
1427  * NOTE!  If a command sequence does not support aborts the ABORT flag is
1428  *        simply ignored.
1429  *
1430  * --
1431  *
1432  * One-off messages (no reply expected) are sent with neither CREATE or DELETE
1433  * set.  One-off messages cannot be aborted and typically aren't processed
1434  * by these routines.  The REPLY bit can be used to distinguish whether a
1435  * one-off message is a command or reply.  For example, one-off replies
1436  * will typically just contain status updates.
1437  */
1438 static int
1439 hammer2_state_msgrx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1440 {
1441         hammer2_state_t *state;
1442         hammer2_state_t dummy;
1443         int error;
1444
1445         /*
1446          * Lock RB tree and locate existing persistent state, if any.
1447          *
1448          * If received msg is a command state is on staterd_tree.
1449          * If received msg is a reply state is on statewr_tree.
1450          */
1451
1452         dummy.msgid = msg->any.head.msgid;
1453         dummy.spanid = msg->any.head.spanid;
1454         pthread_mutex_lock(&iocom->mtx);
1455         if (msg->any.head.cmd & HAMMER2_MSGF_REPLY) {
1456                 state = RB_FIND(hammer2_state_tree,
1457                                 &iocom->statewr_tree, &dummy);
1458         } else {
1459                 state = RB_FIND(hammer2_state_tree,
1460                                 &iocom->staterd_tree, &dummy);
1461         }
1462         msg->state = state;
1463         pthread_mutex_unlock(&iocom->mtx);
1464
1465         /*
1466          * Short-cut one-off or mid-stream messages (state may be NULL).
1467          */
1468         if ((msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
1469                                   HAMMER2_MSGF_ABORT)) == 0) {
1470                 return(0);
1471         }
1472
1473         /*
1474          * Switch on CREATE, DELETE, REPLY, and also handle ABORT from
1475          * inside the case statements.
1476          */
1477         switch(msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
1478                                     HAMMER2_MSGF_REPLY)) {
1479         case HAMMER2_MSGF_CREATE:
1480         case HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
1481                 /*
1482                  * New persistant command received.
1483                  */
1484                 if (state) {
1485                         fprintf(stderr, "hammer2_state_msgrx: "
1486                                         "duplicate transaction\n");
1487                         error = HAMMER2_IOQ_ERROR_TRANS;
1488                         break;
1489                 }
1490                 state = malloc(sizeof(*state));
1491                 bzero(state, sizeof(*state));
1492                 state->iocom = iocom;
1493                 state->flags = HAMMER2_STATE_DYNAMIC;
1494                 state->msg = msg;
1495                 state->txcmd = HAMMER2_MSGF_REPLY;
1496                 state->rxcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1497                 pthread_mutex_lock(&iocom->mtx);
1498                 RB_INSERT(hammer2_state_tree, &iocom->staterd_tree, state);
1499                 pthread_mutex_unlock(&iocom->mtx);
1500                 state->flags |= HAMMER2_STATE_INSERTED;
1501                 state->msgid = msg->any.head.msgid;
1502                 state->spanid = msg->any.head.spanid;
1503                 msg->state = state;
1504                 error = 0;
1505                 break;
1506         case HAMMER2_MSGF_DELETE:
1507                 /*
1508                  * Persistent state is expected but might not exist if an
1509                  * ABORT+DELETE races the close.
1510                  */
1511                 if (state == NULL) {
1512                         if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1513                                 error = HAMMER2_IOQ_ERROR_EALREADY;
1514                         } else {
1515                                 fprintf(stderr, "hammer2_state_msgrx: "
1516                                                 "no state for DELETE\n");
1517                                 error = HAMMER2_IOQ_ERROR_TRANS;
1518                         }
1519                         break;
1520                 }
1521
1522                 /*
1523                  * Handle another ABORT+DELETE case if the msgid has already
1524                  * been reused.
1525                  */
1526                 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1527                         if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1528                                 error = HAMMER2_IOQ_ERROR_EALREADY;
1529                         } else {
1530                                 fprintf(stderr, "hammer2_state_msgrx: "
1531                                                 "state reused for DELETE\n");
1532                                 error = HAMMER2_IOQ_ERROR_TRANS;
1533                         }
1534                         break;
1535                 }
1536                 error = 0;
1537                 break;
1538         default:
1539                 /*
1540                  * Check for mid-stream ABORT command received, otherwise
1541                  * allow.
1542                  */
1543                 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1544                         if (state == NULL ||
1545                             (state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1546                                 error = HAMMER2_IOQ_ERROR_EALREADY;
1547                                 break;
1548                         }
1549                 }
1550                 error = 0;
1551                 break;
1552         case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE:
1553         case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
1554                 /*
1555                  * When receiving a reply with CREATE set the original
1556                  * persistent state message should already exist.
1557                  */
1558                 if (state == NULL) {
1559                         fprintf(stderr,
1560                                 "hammer2_state_msgrx: no state match for REPLY"
1561                                 " cmd=%08x msgid=%016jx\n",
1562                                 msg->any.head.cmd,
1563                                 (intmax_t)msg->any.head.msgid);
1564                         error = HAMMER2_IOQ_ERROR_TRANS;
1565                         break;
1566                 }
1567                 assert(((state->rxcmd ^ msg->any.head.cmd) &
1568                         HAMMER2_MSGF_REPLY) == 0);
1569                 state->rxcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1570                 error = 0;
1571                 break;
1572         case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_DELETE:
1573                 /*
1574                  * Received REPLY+ABORT+DELETE in case where msgid has
1575                  * already been fully closed, ignore the message.
1576                  */
1577                 if (state == NULL) {
1578                         if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1579                                 error = HAMMER2_IOQ_ERROR_EALREADY;
1580                         } else {
1581                                 fprintf(stderr, "hammer2_state_msgrx: "
1582                                                 "no state match for "
1583                                                 "REPLY|DELETE\n");
1584                                 error = HAMMER2_IOQ_ERROR_TRANS;
1585                         }
1586                         break;
1587                 }
1588
1589                 /*
1590                  * Received REPLY+ABORT+DELETE in case where msgid has
1591                  * already been reused for an unrelated message,
1592                  * ignore the message.
1593                  */
1594                 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1595                         if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1596                                 error = HAMMER2_IOQ_ERROR_EALREADY;
1597                         } else {
1598                                 fprintf(stderr, "hammer2_state_msgrx: "
1599                                                 "state reused for "
1600                                                 "REPLY|DELETE\n");
1601                                 error = HAMMER2_IOQ_ERROR_TRANS;
1602                         }
1603                         break;
1604                 }
1605                 error = 0;
1606                 break;
1607         case HAMMER2_MSGF_REPLY:
1608                 /*
1609                  * Check for mid-stream ABORT reply received to sent command.
1610                  */
1611                 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1612                         if (state == NULL ||
1613                             (state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1614                                 error = HAMMER2_IOQ_ERROR_EALREADY;
1615                                 break;
1616                         }
1617                 }
1618                 error = 0;
1619                 break;
1620         }
1621         return (error);
1622 }
1623
1624 void
1625 hammer2_state_cleanuprx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1626 {
1627         hammer2_state_t *state;
1628
1629         if ((state = msg->state) == NULL) {
1630                 /*
1631                  * Free a non-transactional message, there is no state
1632                  * to worry about.
1633                  */
1634                 hammer2_msg_free(iocom, msg);
1635         } else if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
1636                 /*
1637                  * Message terminating transaction, destroy the related
1638                  * state, the original message, and this message (if it
1639                  * isn't the original message due to a CREATE|DELETE).
1640                  */
1641                 pthread_mutex_lock(&iocom->mtx);
1642                 state->rxcmd |= HAMMER2_MSGF_DELETE;
1643                 if (state->txcmd & HAMMER2_MSGF_DELETE) {
1644                         if (state->msg == msg)
1645                                 state->msg = NULL;
1646                         assert(state->flags & HAMMER2_STATE_INSERTED);
1647                         if (state->rxcmd & HAMMER2_MSGF_REPLY) {
1648                                 assert(msg->any.head.cmd & HAMMER2_MSGF_REPLY);
1649                                 RB_REMOVE(hammer2_state_tree,
1650                                           &iocom->statewr_tree, state);
1651                         } else {
1652                                 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1653                                 RB_REMOVE(hammer2_state_tree,
1654                                           &iocom->staterd_tree, state);
1655                         }
1656                         state->flags &= ~HAMMER2_STATE_INSERTED;
1657                         hammer2_state_free(state);
1658                 } else {
1659                         ;
1660                 }
1661                 pthread_mutex_unlock(&iocom->mtx);
1662                 hammer2_msg_free(iocom, msg);
1663         } else if (state->msg != msg) {
1664                 /*
1665                  * Message not terminating transaction, leave state intact
1666                  * and free message if it isn't the CREATE message.
1667                  */
1668                 hammer2_msg_free(iocom, msg);
1669         }
1670 }
1671
1672 static void
1673 hammer2_state_cleanuptx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1674 {
1675         hammer2_state_t *state;
1676
1677         if ((state = msg->state) == NULL) {
1678                 hammer2_msg_free(iocom, msg);
1679         } else if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
1680                 pthread_mutex_lock(&iocom->mtx);
1681                 state->txcmd |= HAMMER2_MSGF_DELETE;
1682                 if (state->rxcmd & HAMMER2_MSGF_DELETE) {
1683                         if (state->msg == msg)
1684                                 state->msg = NULL;
1685                         assert(state->flags & HAMMER2_STATE_INSERTED);
1686                         if (state->txcmd & HAMMER2_MSGF_REPLY) {
1687                                 assert(msg->any.head.cmd & HAMMER2_MSGF_REPLY);
1688                                 RB_REMOVE(hammer2_state_tree,
1689                                           &iocom->staterd_tree, state);
1690                         } else {
1691                                 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1692                                 RB_REMOVE(hammer2_state_tree,
1693                                           &iocom->statewr_tree, state);
1694                         }
1695                         state->flags &= ~HAMMER2_STATE_INSERTED;
1696                         hammer2_state_free(state);
1697                 } else {
1698                         ;
1699                 }
1700                 pthread_mutex_unlock(&iocom->mtx);
1701                 hammer2_msg_free(iocom, msg);
1702         } else if (state->msg != msg) {
1703                 hammer2_msg_free(iocom, msg);
1704         }
1705 }
1706
1707 /*
1708  * Called with iocom locked
1709  */
1710 void
1711 hammer2_state_free(hammer2_state_t *state)
1712 {
1713         hammer2_iocom_t *iocom = state->iocom;
1714         hammer2_msg_t *msg;
1715         char dummy;
1716
1717         fprintf(stderr, "STATE FREE %p\n", state);
1718
1719         assert(state->any.any == NULL);
1720         msg = state->msg;
1721         state->msg = NULL;
1722         if (msg)
1723                 hammer2_msg_free_locked(iocom, msg);
1724         free(state);
1725
1726         /*
1727          * When an iocom error is present we are trying to close down the
1728          * iocom, but we have to wait for all states to terminate before
1729          * we can do so.  The iocom rx code will terminate the receive side
1730          * for all transactions by simulating incoming DELETE messages,
1731          * but the state doesn't go away until both sides are terminated.
1732          *
1733          * We may have to wake up the rx code.
1734          */
1735         if (iocom->ioq_rx.error &&
1736             RB_EMPTY(&iocom->staterd_tree) &&
1737             RB_EMPTY(&iocom->statewr_tree)) {
1738                 dummy = 0;
1739                 write(iocom->wakeupfds[1], &dummy, 1);
1740         }
1741 }
1742
1743 /*
1744  * Indexed messages are stored in a red-black tree indexed by their
1745  * msgid.  Only persistent messages are indexed.
1746  */
1747 int
1748 hammer2_state_cmp(hammer2_state_t *state1, hammer2_state_t *state2)
1749 {
1750         if (state1->spanid < state2->spanid)
1751                 return(-1);
1752         if (state1->spanid > state2->spanid)
1753                 return(1);
1754         if (state1->msgid < state2->msgid)
1755                 return(-1);
1756         if (state1->msgid > state2->msgid)
1757                 return(1);
1758         return(0);
1759 }