2 * Copyright (c) 2011-2012 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@dragonflybsd.org>
6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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
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.
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
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);
42 * Initialize a low-level ioq
45 hammer2_ioq_init(hammer2_iocom_t *iocom __unused, hammer2_ioq_t *ioq)
47 bzero(ioq, sizeof(*ioq));
48 ioq->state = HAMMER2_MSGQ_STATE_HEADER1;
49 TAILQ_INIT(&ioq->msgq);
55 * caller holds iocom->mtx.
58 hammer2_ioq_done(hammer2_iocom_t *iocom __unused, hammer2_ioq_t *ioq)
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);
67 if ((msg = ioq->msg) != NULL) {
69 hammer2_msg_free(iocom, msg);
74 * Initialize a low-level communications channel.
76 * NOTE: The state_func() is called at least once from the loop and can be
77 * re-armed via hammer2_iocom_restate().
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 *))
85 bzero(iocom, sizeof(*iocom));
87 iocom->state_callback = state_func;
88 iocom->rcvmsg_callback = rcvmsg_func;
89 iocom->altmsg_callback = altmsg_func;
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;
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)
107 fcntl(iocom->wakeupfds[0], F_SETFL, O_NONBLOCK);
108 fcntl(iocom->wakeupfds[1], F_SETFL, O_NONBLOCK);
111 * Negotiate session crypto synchronously. This will mark the
112 * connection as error'd if it fails.
114 hammer2_crypto_negotiate(iocom);
117 * Make sure our fds are set to non-blocking for the iocom core.
120 fcntl(sock_fd, F_SETFL, O_NONBLOCK);
122 /* if line buffered our single fgets() should be fine */
124 fcntl(alt_fd, F_SETFL, O_NONBLOCK);
129 * May only be called from a callback from iocom_core.
131 * Adjust state machine functions, set flags to guarantee that both
132 * the recevmsg_func and the sendmsg_func is called at least once.
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 *))
140 iocom->state_callback = state_func;
141 iocom->rcvmsg_callback = rcvmsg_func;
142 iocom->altmsg_callback = altmsg_func;
144 iocom->flags |= HAMMER2_IOCOMF_SWORK;
146 iocom->flags &= ~HAMMER2_IOCOMF_SWORK;
150 * Cleanup a terminating iocom.
152 * Caller should not hold iocom->mtx. The iocom has already been disconnected
153 * from all possible references to it.
156 hammer2_iocom_done(hammer2_iocom_t *iocom)
160 if (iocom->sock_fd >= 0) {
161 close(iocom->sock_fd);
164 if (iocom->alt_fd >= 0) {
165 close(iocom->alt_fd);
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);
174 if ((msg = TAILQ_FIRST(&iocom->freeq_aux)) != NULL) {
175 TAILQ_REMOVE(&iocom->freeq_aux, msg, qentry);
177 msg->aux_data = NULL;
180 if (iocom->wakeupfds[0] >= 0) {
181 close(iocom->wakeupfds[0]);
182 iocom->wakeupfds[0] = -1;
184 if (iocom->wakeupfds[1] >= 0) {
185 close(iocom->wakeupfds[1]);
186 iocom->wakeupfds[1] = -1;
188 pthread_mutex_destroy(&iocom->mtx);
192 * Allocate a new one-way message.
195 hammer2_msg_alloc(hammer2_iocom_t *iocom, size_t aux_size, uint32_t cmd)
200 pthread_mutex_lock(&iocom->mtx);
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);
207 if ((msg = TAILQ_FIRST(&iocom->freeq)) != NULL)
208 TAILQ_REMOVE(&iocom->freeq, msg, qentry);
210 pthread_mutex_unlock(&iocom->mtx);
212 msg = malloc(sizeof(*msg));
213 bzero(msg, sizeof(*msg));
214 msg->aux_data = NULL;
217 if (msg->aux_size != aux_size) {
220 msg->aux_data = NULL;
224 msg->aux_data = malloc(aux_size);
225 msg->aux_size = aux_size;
228 hbytes = (cmd & HAMMER2_MSGF_SIZE) * HAMMER2_MSG_ALIGN;
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;
240 * Free a message so it can be reused afresh.
242 * NOTE: aux_size can be 0 with a non-NULL aux_data.
246 hammer2_msg_free_locked(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
250 TAILQ_INSERT_TAIL(&iocom->freeq_aux, msg, qentry);
252 TAILQ_INSERT_TAIL(&iocom->freeq, msg, qentry);
256 hammer2_msg_free(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
258 pthread_mutex_lock(&iocom->mtx);
259 hammer2_msg_free_locked(iocom, msg);
260 pthread_mutex_unlock(&iocom->mtx);
264 * I/O core loop for an iocom.
266 * Thread localized, iocom->mtx not held.
269 hammer2_iocom_core(hammer2_iocom_t *iocom)
271 struct pollfd fds[3];
276 int wi; /* wakeup pipe */
278 int ai; /* alt bulk path socket */
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) {
288 * Only poll if no immediate work is pending.
289 * Otherwise we are just wasting our time calling
300 * Always check the inter-thread pipe, e.g.
301 * for iocom->txmsgq work.
304 fds[wi].fd = iocom->wakeupfds[0];
305 fds[wi].events = POLLIN;
309 * Check the socket input/output direction as
312 if (iocom->flags & (HAMMER2_IOCOMF_RREQ |
313 HAMMER2_IOCOMF_WREQ)) {
315 fds[si].fd = iocom->sock_fd;
319 if (iocom->flags & HAMMER2_IOCOMF_RREQ)
320 fds[si].events |= POLLIN;
321 if (iocom->flags & HAMMER2_IOCOMF_WREQ)
322 fds[si].events |= POLLOUT;
326 * Check the alternative fd for work.
328 if (iocom->alt_fd >= 0) {
330 fds[ai].fd = iocom->alt_fd;
331 fds[ai].events = POLLIN;
334 poll(fds, count, timeout);
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;
348 * Always check the pipe
350 iocom->flags |= HAMMER2_IOCOMF_PWORK;
353 if (iocom->flags & HAMMER2_IOCOMF_SWORK) {
354 iocom->flags &= ~HAMMER2_IOCOMF_SWORK;
355 iocom->state_callback(iocom);
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.
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);
373 * Message write sequencing
375 if (iocom->flags & HAMMER2_IOCOMF_WWORK)
376 hammer2_iocom_flush1(iocom);
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.
383 if (iocom->flags & HAMMER2_IOCOMF_RWORK) {
384 while ((iocom->flags & HAMMER2_IOCOMF_EOF) == 0 &&
385 (msg = hammer2_ioq_read(iocom)) != NULL) {
387 "receive msg cmd=%08x msgid=%016jx\n",
389 (intmax_t)msg->any.head.msgid);
390 iocom->rcvmsg_callback(iocom, msg);
391 hammer2_state_cleanuprx(iocom, msg);
395 if (iocom->flags & HAMMER2_IOCOMF_ARWORK)
396 iocom->altmsg_callback(iocom);
401 * Read the next ready message from the ioq, issuing I/O if needed.
402 * Caller should retry on a read-event when NULL is returned.
404 * If an error occurs during reception a HAMMER2_LNK_ERROR msg will
405 * be returned for each open transaction, then the ioq and iocom
406 * will be errored out and a non-transactional HAMMER2_LNK_ERROR
407 * msg will be returned as the final message. The caller should not call
408 * us again after the final message is returned.
410 * Thread localized, iocom->mtx not held.
413 hammer2_ioq_read(hammer2_iocom_t *iocom)
415 hammer2_ioq_t *ioq = &iocom->ioq_rx;
417 hammer2_msg_hdr_t *head;
418 hammer2_state_t *state;
426 iocom->flags &= ~(HAMMER2_IOCOMF_RREQ | HAMMER2_IOCOMF_RWORK);
429 * If a message is already pending we can just remove and
430 * return it. Message state has already been processed.
432 if ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
433 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
438 * Message read in-progress (msg is NULL at the moment). We don't
439 * allocate a msg until we have its core header.
441 bytes = ioq->fifo_end - ioq->fifo_beg;
442 nmax = sizeof(ioq->buf) - ioq->fifo_end;
446 case HAMMER2_MSGQ_STATE_HEADER1:
448 * Load the primary header, fail on any non-trivial read
449 * error or on EOF. Since the primary header is the same
450 * size is the message alignment it will never straddle
451 * the end of the buffer.
453 if (bytes < (int)sizeof(msg->any.head)) {
454 n = read(iocom->sock_fd,
455 ioq->buf + ioq->fifo_end,
459 ioq->error = HAMMER2_IOQ_ERROR_EOF;
462 if (errno != EINTR &&
463 errno != EINPROGRESS &&
465 ioq->error = HAMMER2_IOQ_ERROR_SOCK;
477 * Insufficient data accumulated (msg is NULL, caller will
481 if (bytes < (int)sizeof(msg->any.head))
485 * Calculate the header, decrypt data received so far.
486 * Data will be decrypted in-place. Partial blocks are
487 * not immediately decrypted.
489 * WARNING! The header might be in the wrong endian, we
490 * do not fix it up until we get the entire
493 hammer2_crypto_decrypt(iocom, ioq);
494 head = (void *)(ioq->buf + ioq->fifo_beg);
497 * Check and fixup the core header. Note that the icrc
498 * has to be calculated before any fixups, but the crc
499 * fields in the msg may have to be swapped like everything
502 if (head->magic != HAMMER2_MSGHDR_MAGIC &&
503 head->magic != HAMMER2_MSGHDR_MAGIC_REV) {
504 ioq->error = HAMMER2_IOQ_ERROR_SYNC;
509 * Calculate the full header size and aux data size
511 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV) {
512 ioq->hbytes = (bswap32(head->cmd) & HAMMER2_MSGF_SIZE) *
514 ioq->abytes = bswap32(head->aux_bytes) *
517 ioq->hbytes = (head->cmd & HAMMER2_MSGF_SIZE) *
519 ioq->abytes = head->aux_bytes * HAMMER2_MSG_ALIGN;
521 if (ioq->hbytes < sizeof(msg->any.head) ||
522 ioq->hbytes > sizeof(msg->any) ||
523 ioq->abytes > HAMMER2_MSGAUX_MAX) {
524 ioq->error = HAMMER2_IOQ_ERROR_FIELD;
529 * Finally allocate the message and copy the core header
530 * to the embedded extended header.
532 * Initialize msg->aux_size to 0 and use it to track
533 * the amount of data copied from the stream.
535 msg = hammer2_msg_alloc(iocom, ioq->abytes, 0);
539 * We are either done or we fall-through
541 if (ioq->hbytes == sizeof(msg->any.head) && ioq->abytes == 0) {
542 bcopy(head, &msg->any.head, sizeof(msg->any.head));
543 ioq->fifo_beg += ioq->hbytes;
548 * Fall through to the next state. Make sure that the
549 * extended header does not straddle the end of the buffer.
550 * We still want to issue larger reads into our buffer,
551 * book-keeping is easier if we don't bcopy() yet.
553 if (bytes + nmax < ioq->hbytes) {
554 bcopy(ioq->buf + ioq->fifo_beg, ioq->buf, bytes);
555 ioq->fifo_cdx -= ioq->fifo_beg;
557 ioq->fifo_end = bytes;
558 nmax = sizeof(ioq->buf) - ioq->fifo_end;
560 ioq->state = HAMMER2_MSGQ_STATE_HEADER2;
562 case HAMMER2_MSGQ_STATE_HEADER2:
564 * Fill out the extended header.
567 if (bytes < ioq->hbytes) {
568 n = read(iocom->sock_fd,
569 msg->any.buf + ioq->fifo_end,
573 ioq->error = HAMMER2_IOQ_ERROR_EOF;
576 if (errno != EINTR &&
577 errno != EINPROGRESS &&
579 ioq->error = HAMMER2_IOQ_ERROR_SOCK;
591 * Insufficient data accumulated (set msg NULL so caller will
594 if (bytes < ioq->hbytes) {
600 * Calculate the extended header, decrypt data received
601 * so far. Handle endian-conversion for the entire extended
604 hammer2_crypto_decrypt(iocom, ioq);
605 head = (void *)(ioq->buf + ioq->fifo_beg);
610 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV)
611 xcrc32 = bswap32(head->hdr_crc);
613 xcrc32 = head->hdr_crc;
615 if (hammer2_icrc32(head, ioq->hbytes) != xcrc32) {
616 ioq->error = HAMMER2_IOQ_ERROR_XCRC;
619 head->hdr_crc = xcrc32;
621 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV) {
622 hammer2_bswap_head(head);
626 * Copy the extended header into the msg and adjust the
629 bcopy(head, &msg->any, ioq->hbytes);
632 * We are either done or we fall-through.
634 if (ioq->abytes == 0) {
635 ioq->fifo_beg += ioq->hbytes;
640 * Must adjust nmax and bytes (and the state) when falling
643 ioq->fifo_beg += ioq->hbytes;
645 bytes -= ioq->hbytes;
646 ioq->state = HAMMER2_MSGQ_STATE_AUXDATA1;
648 case HAMMER2_MSGQ_STATE_AUXDATA1:
650 * Copy the partial or complete payload from remaining
651 * bytes in the FIFO. We have to fall-through either
652 * way so we can check the crc.
654 * Adjust msg->aux_size to the final actual value.
656 ioq->already = ioq->fifo_cdx - ioq->fifo_beg;
657 if (ioq->already > ioq->abytes)
658 ioq->already = ioq->abytes;
659 if (bytes >= ioq->abytes) {
660 bcopy(ioq->buf + ioq->fifo_beg, msg->aux_data,
662 msg->aux_size = ioq->abytes;
663 ioq->fifo_beg += ioq->abytes;
664 if (ioq->fifo_cdx < ioq->fifo_beg)
665 ioq->fifo_cdx = ioq->fifo_beg;
666 bytes -= ioq->abytes;
668 bcopy(ioq->buf + ioq->fifo_beg, msg->aux_data,
670 msg->aux_size = bytes;
671 ioq->fifo_beg += bytes;
672 if (ioq->fifo_cdx < ioq->fifo_beg)
673 ioq->fifo_cdx = ioq->fifo_beg;
678 ioq->state = HAMMER2_MSGQ_STATE_AUXDATA2;
680 case HAMMER2_MSGQ_STATE_AUXDATA2:
682 * Read the remainder of the payload directly into the
683 * msg->aux_data buffer.
686 if (msg->aux_size < ioq->abytes) {
688 n = read(iocom->sock_fd,
689 msg->aux_data + msg->aux_size,
690 ioq->abytes - msg->aux_size);
693 ioq->error = HAMMER2_IOQ_ERROR_EOF;
696 if (errno != EINTR &&
697 errno != EINPROGRESS &&
699 ioq->error = HAMMER2_IOQ_ERROR_SOCK;
709 * Insufficient data accumulated (set msg NULL so caller will
712 if (msg->aux_size < ioq->abytes) {
716 assert(msg->aux_size == ioq->abytes);
717 hammer2_crypto_decrypt_aux(iocom, ioq, msg, ioq->already);
720 * Check aux_crc, then we are done.
722 xcrc32 = hammer2_icrc32(msg->aux_data, msg->aux_size);
723 if (xcrc32 != msg->any.head.aux_crc) {
724 ioq->error = HAMMER2_IOQ_ERROR_ACRC;
728 case HAMMER2_MSGQ_STATE_ERROR:
730 * Continued calls to drain recorded transactions (returning
731 * a LNK_ERROR for each one), before we return the final
738 * We don't double-return errors, the caller should not
739 * have called us again after getting an error msg.
746 * Check the message sequence. The iv[] should prevent any
747 * possibility of a replay but we add this check anyway.
749 if (msg && ioq->error == 0) {
750 if ((msg->any.head.salt & 255) != (ioq->seq & 255)) {
751 ioq->error = HAMMER2_IOQ_ERROR_MSGSEQ;
758 * Process transactional state for the message.
760 if (msg && ioq->error == 0) {
761 error = hammer2_state_msgrx(iocom, msg);
763 if (error == HAMMER2_IOQ_ERROR_EALREADY) {
764 hammer2_msg_free(iocom, msg);
772 * Handle error, RREQ, or completion
774 * NOTE: nmax and bytes are invalid at this point, we don't bother
775 * to update them when breaking out.
779 * An unrecoverable error causes all active receive
780 * transactions to be terminated with a LNK_ERROR message.
782 * Once all active transactions are exhausted we set the
783 * iocom ERROR flag and return a non-transactional LNK_ERROR
784 * message, which should cause master processing loops to
787 assert(ioq->msg == msg);
789 hammer2_msg_free(iocom, msg);
794 * No more I/O read processing
796 ioq->state = HAMMER2_MSGQ_STATE_ERROR;
799 * Simulate a remote LNK_ERROR DELETE msg for any open
800 * transactions, ending with a final non-transactional
801 * LNK_ERROR (that the session can detect) when no
802 * transactions remain.
804 msg = hammer2_msg_alloc(iocom, 0, 0);
805 bzero(&msg->any.head, sizeof(msg->any.head));
806 msg->any.head.magic = HAMMER2_MSGHDR_MAGIC;
807 msg->any.head.cmd = HAMMER2_LNK_ERROR;
808 msg->any.head.error = ioq->error;
810 pthread_mutex_lock(&iocom->mtx);
811 fprintf(stderr, "CHECK REMAINING RXMSGS\n");
812 if ((state = RB_ROOT(&iocom->staterd_tree)) != NULL) {
814 * Active remote transactions are still present.
815 * Simulate the other end sending us a DELETE.
817 if (state->rxcmd & HAMMER2_MSGF_DELETE) {
818 fprintf(stderr, "SIMULATE DELETION RCONT %p\n", state);
819 hammer2_msg_free(iocom, msg);
822 fprintf(stderr, "SIMULATE DELETION %p RD RXCMD %08x\n", state, state->rxcmd);
823 /*state->txcmd |= HAMMER2_MSGF_DELETE;*/
825 msg->any.head.spanid = state->spanid;
826 msg->any.head.msgid = state->msgid;
827 msg->any.head.cmd |= HAMMER2_MSGF_ABORT |
830 } else if ((state = RB_ROOT(&iocom->statewr_tree)) != NULL) {
832 * Active local transactions are still present.
833 * Simulate the other end sending us a DELETE.
835 if (state->rxcmd & HAMMER2_MSGF_DELETE) {
836 fprintf(stderr, "SIMULATE DELETION WCONT STATE->txcmd = %08x rxcmd = %08x msgid=%016jx\n", state->txcmd, state->rxcmd, state->msgid );
837 hammer2_msg_free(iocom, msg);
840 fprintf(stderr, "SIMULATE DELETION WD RXCMD %08x\n", state->txcmd);
841 /*state->txcmd |= HAMMER2_MSGF_DELETE;*/
843 msg->any.head.spanid = state->spanid;
844 msg->any.head.msgid = state->msgid;
845 msg->any.head.cmd |= HAMMER2_MSGF_ABORT |
846 HAMMER2_MSGF_DELETE |
848 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
855 * No active local or remote transactions remain.
856 * Generate a final LNK_ERROR and flag EOF.
859 iocom->flags |= HAMMER2_IOCOMF_EOF;
860 fprintf(stderr, "EOF ON SOCKET\n");
862 pthread_mutex_unlock(&iocom->mtx);
865 * For the iocom error case we want to set RWORK to indicate
866 * that more messages might be pending.
868 * It is possible to return NULL when there is more work to
869 * do because each message has to be DELETEd in both
870 * directions before we continue on with the next (though
871 * this could be optimized). The transmit direction will
875 iocom->flags |= HAMMER2_IOCOMF_RWORK;
876 } else if (msg == NULL) {
878 * Insufficient data received to finish building the message,
879 * set RREQ and return NULL.
881 * Leave ioq->msg intact.
882 * Leave the FIFO intact.
884 iocom->flags |= HAMMER2_IOCOMF_RREQ;
889 * The fifo has already been advanced past the message.
890 * Trivially reset the FIFO indices if possible.
892 * clear the FIFO if it is now empty and set RREQ to wait
893 * for more from the socket. If the FIFO is not empty set
894 * TWORK to bypass the poll so we loop immediately.
896 if (ioq->fifo_beg == ioq->fifo_end) {
897 iocom->flags |= HAMMER2_IOCOMF_RREQ;
902 iocom->flags |= HAMMER2_IOCOMF_RWORK;
904 ioq->state = HAMMER2_MSGQ_STATE_HEADER1;
911 * Calculate the header and data crc's and write a low-level message to
912 * the connection. If aux_crc is non-zero the aux_data crc is already
913 * assumed to have been set.
915 * A non-NULL msg is added to the queue but not necessarily flushed.
916 * Calling this function with msg == NULL will get a flush going.
918 * Caller must hold iocom->mtx.
921 hammer2_iocom_flush1(hammer2_iocom_t *iocom)
923 hammer2_ioq_t *ioq = &iocom->ioq_tx;
927 hammer2_msg_queue_t tmpq;
929 iocom->flags &= ~(HAMMER2_IOCOMF_WREQ | HAMMER2_IOCOMF_WWORK);
931 pthread_mutex_lock(&iocom->mtx);
932 while ((msg = TAILQ_FIRST(&iocom->txmsgq)) != NULL) {
933 TAILQ_REMOVE(&iocom->txmsgq, msg, qentry);
934 TAILQ_INSERT_TAIL(&tmpq, msg, qentry);
936 pthread_mutex_unlock(&iocom->mtx);
938 while ((msg = TAILQ_FIRST(&tmpq)) != NULL) {
940 * Process terminal connection errors.
942 TAILQ_REMOVE(&tmpq, msg, qentry);
944 TAILQ_INSERT_TAIL(&ioq->msgq, msg, qentry);
950 * Finish populating the msg fields. The salt ensures that
951 * the iv[] array is ridiculously randomized and we also
952 * re-seed our PRNG every 32768 messages just to be sure.
954 msg->any.head.magic = HAMMER2_MSGHDR_MAGIC;
955 msg->any.head.salt = (random() << 8) | (ioq->seq & 255);
957 if ((ioq->seq & 32767) == 0)
961 * Calculate aux_crc if 0, then calculate hdr_crc.
963 if (msg->aux_size && msg->any.head.aux_crc == 0) {
964 assert((msg->aux_size & HAMMER2_MSG_ALIGNMASK) == 0);
965 xcrc32 = hammer2_icrc32(msg->aux_data, msg->aux_size);
966 msg->any.head.aux_crc = xcrc32;
968 msg->any.head.aux_bytes = msg->aux_size / HAMMER2_MSG_ALIGN;
969 assert((msg->aux_size & HAMMER2_MSG_ALIGNMASK) == 0);
971 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
973 msg->any.head.hdr_crc = 0;
974 msg->any.head.hdr_crc = hammer2_icrc32(&msg->any.head, hbytes);
977 * Enqueue the message (the flush codes handles stream
980 TAILQ_INSERT_TAIL(&ioq->msgq, msg, qentry);
983 hammer2_iocom_flush2(iocom);
987 * Thread localized, iocom->mtx not held by caller.
990 hammer2_iocom_flush2(hammer2_iocom_t *iocom)
992 hammer2_ioq_t *ioq = &iocom->ioq_tx;
996 struct iovec iov[HAMMER2_IOQ_MAXIOVEC];
1004 hammer2_iocom_drain(iocom);
1009 * Pump messages out the connection by building an iovec.
1014 TAILQ_FOREACH(msg, &ioq->msgq, qentry) {
1016 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
1019 abytes = msg->aux_size;
1021 hoff += ioq->hbytes;
1022 aoff += ioq->abytes;
1024 if (hbytes - hoff > 0) {
1025 iov[n].iov_base = (char *)&msg->any.head + hoff;
1026 iov[n].iov_len = hbytes - hoff;
1027 nmax += hbytes - hoff;
1029 if (n == HAMMER2_IOQ_MAXIOVEC)
1032 if (abytes - aoff > 0) {
1033 assert(msg->aux_data != NULL);
1034 iov[n].iov_base = msg->aux_data + aoff;
1035 iov[n].iov_len = abytes - aoff;
1036 nmax += abytes - aoff;
1038 if (n == HAMMER2_IOQ_MAXIOVEC)
1046 * Encrypt and write the data. The crypto code will move the
1047 * data into the fifo and adjust the iov as necessary. If
1048 * encryption is disabled the iov is left alone.
1050 * hammer2_crypto_encrypt_wrote()
1052 n = hammer2_crypto_encrypt(iocom, ioq, iov, n);
1055 * Execute the writev() then figure out what happened.
1057 nact = writev(iocom->sock_fd, iov, n);
1059 if (errno != EINTR &&
1060 errno != EINPROGRESS &&
1065 ioq->error = HAMMER2_IOQ_ERROR_SOCK;
1066 hammer2_iocom_drain(iocom);
1069 * Wait for socket buffer space
1071 iocom->flags |= HAMMER2_IOCOMF_WREQ;
1077 * Indicate bytes written successfully. If we were unable to
1078 * write the entire iov array then set WREQ to wait for more
1079 * socket buffer space.
1081 hammer2_crypto_encrypt_wrote(iocom, ioq, nact);
1083 iocom->flags |= HAMMER2_IOCOMF_WREQ;
1086 * Clean out the transmit queue based on what we successfully
1089 while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
1090 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
1092 abytes = msg->aux_size;
1094 if ((size_t)nact < hbytes - ioq->hbytes) {
1095 ioq->hbytes += nact;
1098 nact -= hbytes - ioq->hbytes;
1099 ioq->hbytes = hbytes;
1100 if ((size_t)nact < abytes - ioq->abytes) {
1101 ioq->abytes += nact;
1104 nact -= abytes - ioq->abytes;
1106 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
1111 hammer2_state_cleanuptx(iocom, msg);
1114 hammer2_iocom_drain(iocom);
1119 * Kill pending msgs on ioq_tx and adjust the flags such that no more
1120 * write events will occur. We don't kill read msgs because we want
1121 * the caller to pull off our contrived terminal error msg to detect
1122 * the connection failure.
1124 * Thread localized, iocom->mtx not held by caller.
1127 hammer2_iocom_drain(hammer2_iocom_t *iocom)
1129 hammer2_ioq_t *ioq = &iocom->ioq_tx;
1132 iocom->flags &= ~(HAMMER2_IOCOMF_WREQ | HAMMER2_IOCOMF_WWORK);
1134 while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
1135 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
1137 hammer2_state_cleanuptx(iocom, msg);
1142 * Write a message to an iocom, with additional state processing.
1145 hammer2_msg_write(hammer2_iocom_t *iocom, hammer2_msg_t *msg,
1146 void (*func)(hammer2_state_t *, hammer2_msg_t *),
1148 hammer2_state_t **statep)
1150 hammer2_state_t *state;
1154 * Handle state processing, create state if necessary.
1156 pthread_mutex_lock(&iocom->mtx);
1157 if ((state = msg->state) != NULL) {
1159 * Existing transaction (could be reply). It is also
1160 * possible for this to be the first reply (CREATE is set),
1161 * in which case we populate state->txcmd.
1163 msg->any.head.msgid = state->msgid;
1164 msg->any.head.spanid = state->spanid;
1167 state->any.any = data;
1169 assert(((state->txcmd ^ msg->any.head.cmd) &
1170 HAMMER2_MSGF_REPLY) == 0);
1171 if (msg->any.head.cmd & HAMMER2_MSGF_CREATE)
1172 state->txcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1173 } else if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
1175 * No existing state and CREATE is set, create new
1176 * state for outgoing command. This can't happen if
1177 * REPLY is set as the state would already exist for
1178 * a transaction reply.
1180 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1182 state = malloc(sizeof(*state));
1183 bzero(state, sizeof(*state));
1184 state->iocom = iocom;
1185 state->flags = HAMMER2_STATE_DYNAMIC;
1187 state->msgid = (uint64_t)(uintptr_t)state;
1188 state->spanid = msg->any.head.spanid;
1189 state->txcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1190 state->rxcmd = HAMMER2_MSGF_REPLY;
1192 state->any.any = data;
1193 RB_INSERT(hammer2_state_tree, &iocom->statewr_tree, state);
1194 state->flags |= HAMMER2_STATE_INSERTED;
1196 msg->any.head.msgid = state->msgid;
1197 /* spanid set by caller */
1199 msg->any.head.msgid = 0;
1200 /* spanid set by caller */
1207 * Queue it for output, wake up the I/O pthread. Note that the
1208 * I/O thread is responsible for generating the CRCs and encryption.
1210 TAILQ_INSERT_TAIL(&iocom->txmsgq, msg, qentry);
1212 write(iocom->wakeupfds[1], &dummy, 1); /* XXX optimize me */
1213 pthread_mutex_unlock(&iocom->mtx);
1217 * This is a shortcut to formulate a reply to msg with a simple error code,
1218 * It can reply to and terminate a transaction, or it can reply to a one-way
1219 * messages. A HAMMER2_LNK_ERROR command code is utilized to encode
1220 * the error code (which can be 0). Not all transactions are terminated
1221 * with HAMMER2_LNK_ERROR status (the low level only cares about the
1222 * MSGF_DELETE flag), but most are.
1224 * Replies to one-way messages are a bit of an oxymoron but the feature
1225 * is used by the debug (DBG) protocol.
1227 * The reply contains no extended data.
1230 hammer2_msg_reply(hammer2_iocom_t *iocom, hammer2_msg_t *msg, uint32_t error)
1232 hammer2_state_t *state = msg->state;
1233 hammer2_msg_t *nmsg;
1238 * Reply with a simple error code and terminate the transaction.
1240 cmd = HAMMER2_LNK_ERROR;
1243 * Check if our direction has even been initiated yet, set CREATE.
1245 * Check what direction this is (command or reply direction). Note
1246 * that txcmd might not have been initiated yet.
1248 * If our direction has already been closed we just return without
1252 if (state->txcmd & HAMMER2_MSGF_DELETE)
1254 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
1255 cmd |= HAMMER2_MSGF_CREATE;
1256 if (state->txcmd & HAMMER2_MSGF_REPLY)
1257 cmd |= HAMMER2_MSGF_REPLY;
1258 cmd |= HAMMER2_MSGF_DELETE;
1260 if ((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0)
1261 cmd |= HAMMER2_MSGF_REPLY;
1264 nmsg = hammer2_msg_alloc(iocom, 0, cmd);
1265 nmsg->any.head.error = error;
1266 nmsg->state = msg->state;
1267 hammer2_msg_write(iocom, nmsg, NULL, NULL, NULL);
1271 * Similar to hammer2_msg_reply() but leave the transaction open. That is,
1272 * we are generating a streaming reply or an intermediate acknowledgement
1273 * of some sort as part of the higher level protocol, with more to come
1277 hammer2_msg_result(hammer2_iocom_t *iocom, hammer2_msg_t *msg, uint32_t error)
1279 hammer2_state_t *state = msg->state;
1280 hammer2_msg_t *nmsg;
1285 * Reply with a simple error code and terminate the transaction.
1287 cmd = HAMMER2_LNK_ERROR;
1290 * Check if our direction has even been initiated yet, set CREATE.
1292 * Check what direction this is (command or reply direction). Note
1293 * that txcmd might not have been initiated yet.
1295 * If our direction has already been closed we just return without
1299 if (state->txcmd & HAMMER2_MSGF_DELETE)
1301 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
1302 cmd |= HAMMER2_MSGF_CREATE;
1303 if (state->txcmd & HAMMER2_MSGF_REPLY)
1304 cmd |= HAMMER2_MSGF_REPLY;
1305 /* continuing transaction, do not set MSGF_DELETE */
1307 if ((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0)
1308 cmd |= HAMMER2_MSGF_REPLY;
1311 nmsg = hammer2_msg_alloc(iocom, 0, cmd);
1312 nmsg->any.head.error = error;
1313 nmsg->state = state;
1314 hammer2_msg_write(iocom, nmsg, NULL, NULL, NULL);
1318 * Terminate a transaction given a state structure by issuing a DELETE.
1321 hammer2_state_reply(hammer2_state_t *state, uint32_t error)
1323 hammer2_msg_t *nmsg;
1324 uint32_t cmd = HAMMER2_LNK_ERROR | HAMMER2_MSGF_DELETE;
1327 * Nothing to do if we already transmitted a delete
1329 if (state->txcmd & HAMMER2_MSGF_DELETE)
1333 * We must also set CREATE if this is our first response to a
1336 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
1337 cmd |= HAMMER2_MSGF_CREATE;
1340 * Set REPLY if the other end initiated the command. Otherwise
1341 * we are the command direction.
1343 if (state->txcmd & HAMMER2_MSGF_REPLY)
1344 cmd |= HAMMER2_MSGF_REPLY;
1346 nmsg = hammer2_msg_alloc(state->iocom, 0, cmd);
1347 nmsg->any.head.error = error;
1348 nmsg->state = state;
1349 hammer2_msg_write(state->iocom, nmsg, NULL, NULL, NULL);
1352 /************************************************************************
1353 * TRANSACTION STATE HANDLING *
1354 ************************************************************************
1358 RB_GENERATE(hammer2_state_tree, hammer2_state, rbnode, hammer2_state_cmp);
1361 * Process state tracking for a message after reception, prior to
1364 * Called with msglk held and the msg dequeued.
1366 * All messages are called with dummy state and return actual state.
1367 * (One-off messages often just return the same dummy state).
1369 * May request that caller discard the message by setting *discardp to 1.
1370 * The returned state is not used in this case and is allowed to be NULL.
1374 * These routines handle persistent and command/reply message state via the
1375 * CREATE and DELETE flags. The first message in a command or reply sequence
1376 * sets CREATE, the last message in a command or reply sequence sets DELETE.
1378 * There can be any number of intermediate messages belonging to the same
1379 * sequence sent inbetween the CREATE message and the DELETE message,
1380 * which set neither flag. This represents a streaming command or reply.
1382 * Any command message received with CREATE set expects a reply sequence to
1383 * be returned. Reply sequences work the same as command sequences except the
1384 * REPLY bit is also sent. Both the command side and reply side can
1385 * degenerate into a single message with both CREATE and DELETE set. Note
1386 * that one side can be streaming and the other side not, or neither, or both.
1388 * The msgid is unique for the initiator. That is, two sides sending a new
1389 * message can use the same msgid without colliding.
1393 * ABORT sequences work by setting the ABORT flag along with normal message
1394 * state. However, ABORTs can also be sent on half-closed messages, that is
1395 * even if the command or reply side has already sent a DELETE, as long as
1396 * the message has not been fully closed it can still send an ABORT+DELETE
1397 * to terminate the half-closed message state.
1399 * Since ABORT+DELETEs can race we silently discard ABORT's for message
1400 * state which has already been fully closed. REPLY+ABORT+DELETEs can
1401 * also race, and in this situation the other side might have already
1402 * initiated a new unrelated command with the same message id. Since
1403 * the abort has not set the CREATE flag the situation can be detected
1404 * and the message will also be discarded.
1406 * Non-blocking requests can be initiated with ABORT+CREATE[+DELETE].
1407 * The ABORT request is essentially integrated into the command instead
1408 * of being sent later on. In this situation the command implementation
1409 * detects that CREATE and ABORT are both set (vs ABORT alone) and can
1410 * special-case non-blocking operation for the command.
1412 * NOTE! Messages with ABORT set without CREATE or DELETE are considered
1413 * to be mid-stream aborts for command/reply sequences. ABORTs on
1414 * one-way messages are not supported.
1416 * NOTE! If a command sequence does not support aborts the ABORT flag is
1421 * One-off messages (no reply expected) are sent with neither CREATE or DELETE
1422 * set. One-off messages cannot be aborted and typically aren't processed
1423 * by these routines. The REPLY bit can be used to distinguish whether a
1424 * one-off message is a command or reply. For example, one-off replies
1425 * will typically just contain status updates.
1428 hammer2_state_msgrx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1430 hammer2_state_t *state;
1431 hammer2_state_t dummy;
1435 * Lock RB tree and locate existing persistent state, if any.
1437 * If received msg is a command state is on staterd_tree.
1438 * If received msg is a reply state is on statewr_tree.
1441 dummy.msgid = msg->any.head.msgid;
1442 dummy.spanid = msg->any.head.spanid;
1443 pthread_mutex_lock(&iocom->mtx);
1444 if (msg->any.head.cmd & HAMMER2_MSGF_REPLY) {
1445 state = RB_FIND(hammer2_state_tree,
1446 &iocom->statewr_tree, &dummy);
1448 state = RB_FIND(hammer2_state_tree,
1449 &iocom->staterd_tree, &dummy);
1452 pthread_mutex_unlock(&iocom->mtx);
1455 * Short-cut one-off or mid-stream messages (state may be NULL).
1457 if ((msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
1458 HAMMER2_MSGF_ABORT)) == 0) {
1463 * Switch on CREATE, DELETE, REPLY, and also handle ABORT from
1464 * inside the case statements.
1466 switch(msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
1467 HAMMER2_MSGF_REPLY)) {
1468 case HAMMER2_MSGF_CREATE:
1469 case HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
1471 * New persistant command received.
1474 fprintf(stderr, "hammer2_state_msgrx: "
1475 "duplicate transaction\n");
1476 error = HAMMER2_IOQ_ERROR_TRANS;
1479 state = malloc(sizeof(*state));
1480 bzero(state, sizeof(*state));
1481 state->iocom = iocom;
1482 state->flags = HAMMER2_STATE_DYNAMIC;
1484 state->txcmd = HAMMER2_MSGF_REPLY;
1485 state->rxcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1486 pthread_mutex_lock(&iocom->mtx);
1487 RB_INSERT(hammer2_state_tree, &iocom->staterd_tree, state);
1488 pthread_mutex_unlock(&iocom->mtx);
1489 state->flags |= HAMMER2_STATE_INSERTED;
1490 state->msgid = msg->any.head.msgid;
1491 state->spanid = msg->any.head.spanid;
1495 case HAMMER2_MSGF_DELETE:
1497 * Persistent state is expected but might not exist if an
1498 * ABORT+DELETE races the close.
1500 if (state == NULL) {
1501 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1502 error = HAMMER2_IOQ_ERROR_EALREADY;
1504 fprintf(stderr, "hammer2_state_msgrx: "
1505 "no state for DELETE\n");
1506 error = HAMMER2_IOQ_ERROR_TRANS;
1512 * Handle another ABORT+DELETE case if the msgid has already
1515 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1516 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1517 error = HAMMER2_IOQ_ERROR_EALREADY;
1519 fprintf(stderr, "hammer2_state_msgrx: "
1520 "state reused for DELETE\n");
1521 error = HAMMER2_IOQ_ERROR_TRANS;
1529 * Check for mid-stream ABORT command received, otherwise
1532 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1533 if (state == NULL ||
1534 (state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1535 error = HAMMER2_IOQ_ERROR_EALREADY;
1541 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE:
1542 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
1544 * When receiving a reply with CREATE set the original
1545 * persistent state message should already exist.
1547 if (state == NULL) {
1549 "hammer2_state_msgrx: no state match for REPLY"
1550 " cmd=%08x msgid=%016jx\n",
1552 (intmax_t)msg->any.head.msgid);
1553 error = HAMMER2_IOQ_ERROR_TRANS;
1556 assert(((state->rxcmd ^ msg->any.head.cmd) &
1557 HAMMER2_MSGF_REPLY) == 0);
1558 state->rxcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1561 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_DELETE:
1563 * Received REPLY+ABORT+DELETE in case where msgid has
1564 * already been fully closed, ignore the message.
1566 if (state == NULL) {
1567 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1568 error = HAMMER2_IOQ_ERROR_EALREADY;
1570 fprintf(stderr, "hammer2_state_msgrx: "
1571 "no state match for "
1573 error = HAMMER2_IOQ_ERROR_TRANS;
1579 * Received REPLY+ABORT+DELETE in case where msgid has
1580 * already been reused for an unrelated message,
1581 * ignore the message.
1583 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1584 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1585 error = HAMMER2_IOQ_ERROR_EALREADY;
1587 fprintf(stderr, "hammer2_state_msgrx: "
1590 error = HAMMER2_IOQ_ERROR_TRANS;
1596 case HAMMER2_MSGF_REPLY:
1598 * Check for mid-stream ABORT reply received to sent command.
1600 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1601 if (state == NULL ||
1602 (state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1603 error = HAMMER2_IOQ_ERROR_EALREADY;
1614 hammer2_state_cleanuprx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1616 hammer2_state_t *state;
1618 if ((state = msg->state) == NULL) {
1620 * Free a non-transactional message, there is no state
1623 hammer2_msg_free(iocom, msg);
1624 } else if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
1626 * Message terminating transaction, destroy the related
1627 * state, the original message, and this message (if it
1628 * isn't the original message due to a CREATE|DELETE).
1630 pthread_mutex_lock(&iocom->mtx);
1631 state->rxcmd |= HAMMER2_MSGF_DELETE;
1632 if (state->txcmd & HAMMER2_MSGF_DELETE) {
1633 if (state->msg == msg)
1635 assert(state->flags & HAMMER2_STATE_INSERTED);
1636 if (state->rxcmd & HAMMER2_MSGF_REPLY) {
1637 assert(msg->any.head.cmd & HAMMER2_MSGF_REPLY);
1638 RB_REMOVE(hammer2_state_tree,
1639 &iocom->statewr_tree, state);
1641 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1642 RB_REMOVE(hammer2_state_tree,
1643 &iocom->staterd_tree, state);
1645 state->flags &= ~HAMMER2_STATE_INSERTED;
1646 hammer2_state_free(state);
1650 pthread_mutex_unlock(&iocom->mtx);
1651 hammer2_msg_free(iocom, msg);
1652 } else if (state->msg != msg) {
1654 * Message not terminating transaction, leave state intact
1655 * and free message if it isn't the CREATE message.
1657 hammer2_msg_free(iocom, msg);
1662 hammer2_state_cleanuptx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1664 hammer2_state_t *state;
1666 if ((state = msg->state) == NULL) {
1667 hammer2_msg_free(iocom, msg);
1668 } else if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
1669 pthread_mutex_lock(&iocom->mtx);
1670 state->txcmd |= HAMMER2_MSGF_DELETE;
1671 if (state->rxcmd & HAMMER2_MSGF_DELETE) {
1672 if (state->msg == msg)
1674 assert(state->flags & HAMMER2_STATE_INSERTED);
1675 if (state->txcmd & HAMMER2_MSGF_REPLY) {
1676 assert(msg->any.head.cmd & HAMMER2_MSGF_REPLY);
1677 RB_REMOVE(hammer2_state_tree,
1678 &iocom->staterd_tree, state);
1680 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1681 RB_REMOVE(hammer2_state_tree,
1682 &iocom->statewr_tree, state);
1684 state->flags &= ~HAMMER2_STATE_INSERTED;
1685 hammer2_state_free(state);
1689 pthread_mutex_unlock(&iocom->mtx);
1690 hammer2_msg_free(iocom, msg);
1691 } else if (state->msg != msg) {
1692 hammer2_msg_free(iocom, msg);
1697 * Called with iocom locked
1700 hammer2_state_free(hammer2_state_t *state)
1702 hammer2_iocom_t *iocom = state->iocom;
1706 fprintf(stderr, "STATE FREE %p\n", state);
1708 assert(state->any.any == NULL);
1712 hammer2_msg_free_locked(iocom, msg);
1716 * When an iocom error is present we are trying to close down the
1717 * iocom, but we have to wait for all states to terminate before
1718 * we can do so. The iocom rx code will terminate the receive side
1719 * for all transactions by simulating incoming DELETE messages,
1720 * but the state doesn't go away until both sides are terminated.
1722 * We may have to wake up the rx code.
1724 if (iocom->ioq_rx.error &&
1725 RB_EMPTY(&iocom->staterd_tree) &&
1726 RB_EMPTY(&iocom->statewr_tree)) {
1728 write(iocom->wakeupfds[1], &dummy, 1);
1733 * Indexed messages are stored in a red-black tree indexed by their
1734 * msgid. Only persistent messages are indexed.
1737 hammer2_state_cmp(hammer2_state_t *state1, hammer2_state_t *state2)
1739 if (state1->spanid < state2->spanid)
1741 if (state1->spanid > state2->spanid)
1743 if (state1->msgid < state2->msgid)
1745 if (state1->msgid > state2->msgid)