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 fprintf(stderr, "receive %s\n",
388 hammer2_msg_str(msg));
390 iocom->rcvmsg_callback(iocom, msg);
391 hammer2_state_cleanuprx(iocom, msg);
395 if (iocom->flags & HAMMER2_IOCOMF_ARWORK) {
396 iocom->flags &= ~HAMMER2_IOCOMF_ARWORK;
397 iocom->altmsg_callback(iocom);
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.
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.
412 * Thread localized, iocom->mtx not held.
415 hammer2_ioq_read(hammer2_iocom_t *iocom)
417 hammer2_ioq_t *ioq = &iocom->ioq_rx;
419 hammer2_msg_hdr_t *head;
420 hammer2_state_t *state;
428 iocom->flags &= ~(HAMMER2_IOCOMF_RREQ | HAMMER2_IOCOMF_RWORK);
431 * If a message is already pending we can just remove and
432 * return it. Message state has already been processed.
434 if ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
435 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
440 * Message read in-progress (msg is NULL at the moment). We don't
441 * allocate a msg until we have its core header.
443 bytes = ioq->fifo_end - ioq->fifo_beg;
444 nmax = sizeof(ioq->buf) - ioq->fifo_end;
448 case HAMMER2_MSGQ_STATE_HEADER1:
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.
455 if (bytes < (int)sizeof(msg->any.head)) {
456 n = read(iocom->sock_fd,
457 ioq->buf + ioq->fifo_end,
461 ioq->error = HAMMER2_IOQ_ERROR_EOF;
464 if (errno != EINTR &&
465 errno != EINPROGRESS &&
467 ioq->error = HAMMER2_IOQ_ERROR_SOCK;
479 * Insufficient data accumulated (msg is NULL, caller will
483 if (bytes < (int)sizeof(msg->any.head))
487 * Calculate the header, decrypt data received so far.
488 * Data will be decrypted in-place. Partial blocks are
489 * not immediately decrypted.
491 * WARNING! The header might be in the wrong endian, we
492 * do not fix it up until we get the entire
495 hammer2_crypto_decrypt(iocom, ioq);
496 head = (void *)(ioq->buf + ioq->fifo_beg);
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
504 if (head->magic != HAMMER2_MSGHDR_MAGIC &&
505 head->magic != HAMMER2_MSGHDR_MAGIC_REV) {
506 ioq->error = HAMMER2_IOQ_ERROR_SYNC;
511 * Calculate the full header size and aux data size
513 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV) {
514 ioq->hbytes = (bswap32(head->cmd) & HAMMER2_MSGF_SIZE) *
516 ioq->abytes = bswap32(head->aux_bytes) *
519 ioq->hbytes = (head->cmd & HAMMER2_MSGF_SIZE) *
521 ioq->abytes = head->aux_bytes * HAMMER2_MSG_ALIGN;
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;
531 * Allocate the message, the next state will fill it in.
533 msg = hammer2_msg_alloc(iocom, ioq->abytes, 0);
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.
542 if (bytes + nmax < ioq->hbytes) {
543 bcopy(ioq->buf + ioq->fifo_beg, ioq->buf, bytes);
544 ioq->fifo_cdx -= ioq->fifo_beg;
546 ioq->fifo_end = bytes;
547 nmax = sizeof(ioq->buf) - ioq->fifo_end;
549 ioq->state = HAMMER2_MSGQ_STATE_HEADER2;
551 case HAMMER2_MSGQ_STATE_HEADER2:
553 * Fill out the extended header.
556 if (bytes < ioq->hbytes) {
557 n = read(iocom->sock_fd,
558 ioq->buf + ioq->fifo_end,
562 ioq->error = HAMMER2_IOQ_ERROR_EOF;
565 if (errno != EINTR &&
566 errno != EINPROGRESS &&
568 ioq->error = HAMMER2_IOQ_ERROR_SOCK;
580 * Insufficient data accumulated (set msg NULL so caller will
583 if (bytes < ioq->hbytes) {
589 * Calculate the extended header, decrypt data received
590 * so far. Handle endian-conversion for the entire extended
593 hammer2_crypto_decrypt(iocom, ioq);
594 head = (void *)(ioq->buf + ioq->fifo_beg);
599 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV)
600 xcrc32 = bswap32(head->hdr_crc);
602 xcrc32 = head->hdr_crc;
604 if (hammer2_icrc32(head, ioq->hbytes) != xcrc32) {
605 ioq->error = HAMMER2_IOQ_ERROR_XCRC;
606 fprintf(stderr, "BAD-XCRC(%08x,%08x) %s\n",
607 xcrc32, hammer2_icrc32(head, ioq->hbytes),
608 hammer2_msg_str(msg));
612 head->hdr_crc = xcrc32;
614 if (head->magic == HAMMER2_MSGHDR_MAGIC_REV) {
615 hammer2_bswap_head(head);
619 * Copy the extended header into the msg and adjust the
622 bcopy(head, &msg->any, ioq->hbytes);
625 * We are either done or we fall-through.
627 if (ioq->abytes == 0) {
628 ioq->fifo_beg += ioq->hbytes;
633 * Must adjust nmax and bytes (and the state) when falling
636 ioq->fifo_beg += ioq->hbytes;
638 bytes -= ioq->hbytes;
639 ioq->state = HAMMER2_MSGQ_STATE_AUXDATA1;
641 case HAMMER2_MSGQ_STATE_AUXDATA1:
643 * Copy the partial or complete payload from remaining
644 * bytes in the FIFO. We have to fall-through either
645 * way so we can check the crc.
647 * Adjust msg->aux_size to the final actual value.
649 ioq->already = ioq->fifo_cdx - ioq->fifo_beg;
650 if (ioq->already > ioq->abytes)
651 ioq->already = ioq->abytes;
652 if (bytes >= ioq->abytes) {
653 bcopy(ioq->buf + ioq->fifo_beg, msg->aux_data,
655 msg->aux_size = ioq->abytes;
656 ioq->fifo_beg += ioq->abytes;
657 if (ioq->fifo_cdx < ioq->fifo_beg)
658 ioq->fifo_cdx = ioq->fifo_beg;
659 bytes -= ioq->abytes;
661 bcopy(ioq->buf + ioq->fifo_beg, msg->aux_data,
663 msg->aux_size = bytes;
664 ioq->fifo_beg += bytes;
665 if (ioq->fifo_cdx < ioq->fifo_beg)
666 ioq->fifo_cdx = ioq->fifo_beg;
671 ioq->state = HAMMER2_MSGQ_STATE_AUXDATA2;
673 case HAMMER2_MSGQ_STATE_AUXDATA2:
675 * Read the remainder of the payload directly into the
676 * msg->aux_data buffer.
679 if (msg->aux_size < ioq->abytes) {
681 n = read(iocom->sock_fd,
682 msg->aux_data + msg->aux_size,
683 ioq->abytes - msg->aux_size);
686 ioq->error = HAMMER2_IOQ_ERROR_EOF;
689 if (errno != EINTR &&
690 errno != EINPROGRESS &&
692 ioq->error = HAMMER2_IOQ_ERROR_SOCK;
702 * Insufficient data accumulated (set msg NULL so caller will
705 if (msg->aux_size < ioq->abytes) {
709 assert(msg->aux_size == ioq->abytes);
710 hammer2_crypto_decrypt_aux(iocom, ioq, msg, ioq->already);
713 * Check aux_crc, then we are done.
715 xcrc32 = hammer2_icrc32(msg->aux_data, msg->aux_size);
716 if (xcrc32 != msg->any.head.aux_crc) {
717 ioq->error = HAMMER2_IOQ_ERROR_ACRC;
721 case HAMMER2_MSGQ_STATE_ERROR:
723 * Continued calls to drain recorded transactions (returning
724 * a LNK_ERROR for each one), before we return the final
731 * We don't double-return errors, the caller should not
732 * have called us again after getting an error msg.
739 * Check the message sequence. The iv[] should prevent any
740 * possibility of a replay but we add this check anyway.
742 if (msg && ioq->error == 0) {
743 if ((msg->any.head.salt & 255) != (ioq->seq & 255)) {
744 ioq->error = HAMMER2_IOQ_ERROR_MSGSEQ;
751 * Process transactional state for the message.
753 if (msg && ioq->error == 0) {
754 error = hammer2_state_msgrx(iocom, msg);
756 if (error == HAMMER2_IOQ_ERROR_EALREADY) {
757 hammer2_msg_free(iocom, msg);
765 * Handle error, RREQ, or completion
767 * NOTE: nmax and bytes are invalid at this point, we don't bother
768 * to update them when breaking out.
772 * An unrecoverable error causes all active receive
773 * transactions to be terminated with a LNK_ERROR message.
775 * Once all active transactions are exhausted we set the
776 * iocom ERROR flag and return a non-transactional LNK_ERROR
777 * message, which should cause master processing loops to
780 assert(ioq->msg == msg);
782 hammer2_msg_free(iocom, msg);
787 * No more I/O read processing
789 ioq->state = HAMMER2_MSGQ_STATE_ERROR;
792 * Simulate a remote LNK_ERROR DELETE msg for any open
793 * transactions, ending with a final non-transactional
794 * LNK_ERROR (that the session can detect) when no
795 * transactions remain.
797 msg = hammer2_msg_alloc(iocom, 0, 0);
798 bzero(&msg->any.head, sizeof(msg->any.head));
799 msg->any.head.magic = HAMMER2_MSGHDR_MAGIC;
800 msg->any.head.cmd = HAMMER2_LNK_ERROR;
801 msg->any.head.error = ioq->error;
803 pthread_mutex_lock(&iocom->mtx);
804 if ((state = RB_ROOT(&iocom->staterd_tree)) != NULL) {
806 * Active remote transactions are still present.
807 * Simulate the other end sending us a DELETE.
809 if (state->rxcmd & HAMMER2_MSGF_DELETE) {
810 hammer2_msg_free(iocom, msg);
813 /*state->txcmd |= HAMMER2_MSGF_DELETE;*/
815 msg->any.head.spanid = state->spanid;
816 msg->any.head.msgid = state->msgid;
817 msg->any.head.cmd |= HAMMER2_MSGF_ABORT |
820 } else if ((state = RB_ROOT(&iocom->statewr_tree)) != NULL) {
822 * Active local transactions are still present.
823 * Simulate the other end sending us a DELETE.
825 if (state->rxcmd & HAMMER2_MSGF_DELETE) {
826 hammer2_msg_free(iocom, msg);
830 msg->any.head.spanid = state->spanid;
831 msg->any.head.msgid = state->msgid;
832 msg->any.head.cmd |= HAMMER2_MSGF_ABORT |
833 HAMMER2_MSGF_DELETE |
835 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
842 * No active local or remote transactions remain.
843 * Generate a final LNK_ERROR and flag EOF.
846 iocom->flags |= HAMMER2_IOCOMF_EOF;
847 fprintf(stderr, "EOF ON SOCKET %d\n", iocom->sock_fd);
849 pthread_mutex_unlock(&iocom->mtx);
852 * For the iocom error case we want to set RWORK to indicate
853 * that more messages might be pending.
855 * It is possible to return NULL when there is more work to
856 * do because each message has to be DELETEd in both
857 * directions before we continue on with the next (though
858 * this could be optimized). The transmit direction will
862 iocom->flags |= HAMMER2_IOCOMF_RWORK;
863 } else if (msg == NULL) {
865 * Insufficient data received to finish building the message,
866 * set RREQ and return NULL.
868 * Leave ioq->msg intact.
869 * Leave the FIFO intact.
871 iocom->flags |= HAMMER2_IOCOMF_RREQ;
876 * The fifo has already been advanced past the message.
877 * Trivially reset the FIFO indices if possible.
879 * clear the FIFO if it is now empty and set RREQ to wait
880 * for more from the socket. If the FIFO is not empty set
881 * TWORK to bypass the poll so we loop immediately.
883 if (ioq->fifo_beg == ioq->fifo_end) {
884 iocom->flags |= HAMMER2_IOCOMF_RREQ;
889 iocom->flags |= HAMMER2_IOCOMF_RWORK;
891 ioq->state = HAMMER2_MSGQ_STATE_HEADER1;
898 * Calculate the header and data crc's and write a low-level message to
899 * the connection. If aux_crc is non-zero the aux_data crc is already
900 * assumed to have been set.
902 * A non-NULL msg is added to the queue but not necessarily flushed.
903 * Calling this function with msg == NULL will get a flush going.
905 * Caller must hold iocom->mtx.
908 hammer2_iocom_flush1(hammer2_iocom_t *iocom)
910 hammer2_ioq_t *ioq = &iocom->ioq_tx;
914 hammer2_msg_queue_t tmpq;
916 iocom->flags &= ~(HAMMER2_IOCOMF_WREQ | HAMMER2_IOCOMF_WWORK);
918 pthread_mutex_lock(&iocom->mtx);
919 while ((msg = TAILQ_FIRST(&iocom->txmsgq)) != NULL) {
920 TAILQ_REMOVE(&iocom->txmsgq, msg, qentry);
921 TAILQ_INSERT_TAIL(&tmpq, msg, qentry);
923 pthread_mutex_unlock(&iocom->mtx);
925 while ((msg = TAILQ_FIRST(&tmpq)) != NULL) {
927 * Process terminal connection errors.
929 TAILQ_REMOVE(&tmpq, msg, qentry);
931 TAILQ_INSERT_TAIL(&ioq->msgq, msg, qentry);
937 * Finish populating the msg fields. The salt ensures that
938 * the iv[] array is ridiculously randomized and we also
939 * re-seed our PRNG every 32768 messages just to be sure.
941 msg->any.head.magic = HAMMER2_MSGHDR_MAGIC;
942 msg->any.head.salt = (random() << 8) | (ioq->seq & 255);
944 if ((ioq->seq & 32767) == 0)
948 * Calculate aux_crc if 0, then calculate hdr_crc.
950 if (msg->aux_size && msg->any.head.aux_crc == 0) {
951 assert((msg->aux_size & HAMMER2_MSG_ALIGNMASK) == 0);
952 xcrc32 = hammer2_icrc32(msg->aux_data, msg->aux_size);
953 msg->any.head.aux_crc = xcrc32;
955 msg->any.head.aux_bytes = msg->aux_size / HAMMER2_MSG_ALIGN;
956 assert((msg->aux_size & HAMMER2_MSG_ALIGNMASK) == 0);
958 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
960 msg->any.head.hdr_crc = 0;
961 msg->any.head.hdr_crc = hammer2_icrc32(&msg->any.head, hbytes);
964 * Enqueue the message (the flush codes handles stream
967 TAILQ_INSERT_TAIL(&ioq->msgq, msg, qentry);
970 hammer2_iocom_flush2(iocom);
974 * Thread localized, iocom->mtx not held by caller.
977 hammer2_iocom_flush2(hammer2_iocom_t *iocom)
979 hammer2_ioq_t *ioq = &iocom->ioq_tx;
984 struct iovec iov[HAMMER2_IOQ_MAXIOVEC];
992 hammer2_iocom_drain(iocom);
997 * Pump messages out the connection by building an iovec.
999 * ioq->hbytes/ioq->abytes tracks how much of the first message
1000 * in the queue has been successfully written out, so we can
1008 TAILQ_FOREACH(msg, &ioq->msgq, qentry) {
1009 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
1011 abytes = msg->aux_size;
1012 assert(hoff <= hbytes && aoff <= abytes);
1014 if (hoff < hbytes) {
1015 iov[n].iov_base = (char *)&msg->any.head + hoff;
1016 iov[n].iov_len = hbytes - hoff;
1017 nmax += hbytes - hoff;
1019 if (n == HAMMER2_IOQ_MAXIOVEC)
1022 if (aoff < abytes) {
1023 assert(msg->aux_data != NULL);
1024 iov[n].iov_base = (char *)msg->aux_data + aoff;
1025 iov[n].iov_len = abytes - aoff;
1026 nmax += abytes - aoff;
1028 if (n == HAMMER2_IOQ_MAXIOVEC)
1038 * Encrypt and write the data. The crypto code will move the
1039 * data into the fifo and adjust the iov as necessary. If
1040 * encryption is disabled the iov is left alone.
1042 * May return a smaller iov (thus a smaller n), with aggregated
1043 * chunks. May reduce nmax to what fits in the FIFO.
1046 n = hammer2_crypto_encrypt(iocom, ioq, iov, n, &nmax);
1049 * Execute the writev() then figure out what happened.
1051 nact = writev(iocom->sock_fd, iov, n);
1053 if (errno != EINTR &&
1054 errno != EINPROGRESS &&
1059 ioq->error = HAMMER2_IOQ_ERROR_SOCK;
1060 hammer2_iocom_drain(iocom);
1063 * Wait for socket buffer space
1065 iocom->flags |= HAMMER2_IOCOMF_WREQ;
1071 * Indicate bytes written successfully.
1073 * If we were unable to write the entire iov array then set WREQ
1074 * to wait for more socket buffer space.
1076 * If the FIFO space was insufficient to fully drain all messages
1077 * set WWORK to cause the core to call us again for the next batch.
1079 hammer2_crypto_encrypt_wrote(iocom, ioq, nact);
1081 iocom->flags |= HAMMER2_IOCOMF_WREQ;
1084 * Clean out the transmit queue based on what we successfully
1085 * sent. ioq->hbytes/abytes represents the portion of the first
1086 * message previously sent.
1088 while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
1089 hbytes = (msg->any.head.cmd & HAMMER2_MSGF_SIZE) *
1091 abytes = msg->aux_size;
1093 if ((size_t)nact < hbytes - ioq->hbytes) {
1094 ioq->hbytes += nact;
1098 nact -= hbytes - ioq->hbytes;
1099 ioq->hbytes = hbytes;
1100 if ((size_t)nact < abytes - ioq->abytes) {
1101 ioq->abytes += nact;
1105 nact -= abytes - ioq->abytes;
1107 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
1112 hammer2_state_cleanuptx(iocom, msg);
1116 * If more messages are pending on WREQ wasn't set we must
1117 * ensure that WWORK gets set.
1119 if (msg && (iocom->flags & HAMMER2_IOCOMF_WREQ) == 0)
1120 iocom->flags |= HAMMER2_IOCOMF_WWORK;
1123 hammer2_iocom_drain(iocom);
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.
1133 * Thread localized, iocom->mtx not held by caller.
1136 hammer2_iocom_drain(hammer2_iocom_t *iocom)
1138 hammer2_ioq_t *ioq = &iocom->ioq_tx;
1141 iocom->flags &= ~(HAMMER2_IOCOMF_WREQ | HAMMER2_IOCOMF_WWORK);
1145 while ((msg = TAILQ_FIRST(&ioq->msgq)) != NULL) {
1146 TAILQ_REMOVE(&ioq->msgq, msg, qentry);
1148 hammer2_state_cleanuptx(iocom, msg);
1153 * Write a message to an iocom, with additional state processing.
1156 hammer2_msg_write(hammer2_iocom_t *iocom, hammer2_msg_t *msg,
1157 void (*func)(hammer2_state_t *, hammer2_msg_t *),
1159 hammer2_state_t **statep)
1161 hammer2_state_t *state;
1165 * Handle state processing, create state if necessary.
1167 pthread_mutex_lock(&iocom->mtx);
1168 if ((state = msg->state) != NULL) {
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.
1174 msg->any.head.msgid = state->msgid;
1175 msg->any.head.spanid = state->spanid;
1178 state->any.any = data;
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) {
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.
1191 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1193 state = malloc(sizeof(*state));
1194 bzero(state, sizeof(*state));
1195 state->iocom = iocom;
1196 state->flags = HAMMER2_STATE_DYNAMIC;
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;
1203 state->any.any = data;
1204 RB_INSERT(hammer2_state_tree, &iocom->statewr_tree, state);
1205 state->flags |= HAMMER2_STATE_INSERTED;
1207 msg->any.head.msgid = state->msgid;
1208 /* spanid set by caller */
1210 msg->any.head.msgid = 0;
1211 /* spanid set by caller */
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.
1221 TAILQ_INSERT_TAIL(&iocom->txmsgq, msg, qentry);
1223 write(iocom->wakeupfds[1], &dummy, 1); /* XXX optimize me */
1224 pthread_mutex_unlock(&iocom->mtx);
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.
1235 * Replies to one-way messages are a bit of an oxymoron but the feature
1236 * is used by the debug (DBG) protocol.
1238 * The reply contains no extended data.
1241 hammer2_msg_reply(hammer2_iocom_t *iocom, hammer2_msg_t *msg, uint32_t error)
1243 hammer2_state_t *state = msg->state;
1244 hammer2_msg_t *nmsg;
1249 * Reply with a simple error code and terminate the transaction.
1251 cmd = HAMMER2_LNK_ERROR;
1254 * Check if our direction has even been initiated yet, set CREATE.
1256 * Check what direction this is (command or reply direction). Note
1257 * that txcmd might not have been initiated yet.
1259 * If our direction has already been closed we just return without
1263 if (state->txcmd & HAMMER2_MSGF_DELETE)
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;
1271 if ((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0)
1272 cmd |= HAMMER2_MSGF_REPLY;
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);
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
1288 hammer2_msg_result(hammer2_iocom_t *iocom, hammer2_msg_t *msg, uint32_t error)
1290 hammer2_state_t *state = msg->state;
1291 hammer2_msg_t *nmsg;
1296 * Reply with a simple error code and terminate the transaction.
1298 cmd = HAMMER2_LNK_ERROR;
1301 * Check if our direction has even been initiated yet, set CREATE.
1303 * Check what direction this is (command or reply direction). Note
1304 * that txcmd might not have been initiated yet.
1306 * If our direction has already been closed we just return without
1310 if (state->txcmd & HAMMER2_MSGF_DELETE)
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 */
1318 if ((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0)
1319 cmd |= HAMMER2_MSGF_REPLY;
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);
1329 * Terminate a transaction given a state structure by issuing a DELETE.
1332 hammer2_state_reply(hammer2_state_t *state, uint32_t error)
1334 hammer2_msg_t *nmsg;
1335 uint32_t cmd = HAMMER2_LNK_ERROR | HAMMER2_MSGF_DELETE;
1338 * Nothing to do if we already transmitted a delete
1340 if (state->txcmd & HAMMER2_MSGF_DELETE)
1344 * We must also set CREATE if this is our first response to a
1347 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
1348 cmd |= HAMMER2_MSGF_CREATE;
1351 * Set REPLY if the other end initiated the command. Otherwise
1352 * we are the command direction.
1354 if (state->txcmd & HAMMER2_MSGF_REPLY)
1355 cmd |= HAMMER2_MSGF_REPLY;
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);
1363 /************************************************************************
1364 * TRANSACTION STATE HANDLING *
1365 ************************************************************************
1369 RB_GENERATE(hammer2_state_tree, hammer2_state, rbnode, hammer2_state_cmp);
1372 * Process state tracking for a message after reception, prior to
1375 * Called with msglk held and the msg dequeued.
1377 * All messages are called with dummy state and return actual state.
1378 * (One-off messages often just return the same dummy state).
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.
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.
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.
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.
1399 * The msgid is unique for the initiator. That is, two sides sending a new
1400 * message can use the same msgid without colliding.
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.
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.
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.
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.
1427 * NOTE! If a command sequence does not support aborts the ABORT flag is
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.
1439 hammer2_state_msgrx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1441 hammer2_state_t *state;
1442 hammer2_state_t dummy;
1446 * Lock RB tree and locate existing persistent state, if any.
1448 * If received msg is a command state is on staterd_tree.
1449 * If received msg is a reply state is on statewr_tree.
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);
1459 state = RB_FIND(hammer2_state_tree,
1460 &iocom->staterd_tree, &dummy);
1463 pthread_mutex_unlock(&iocom->mtx);
1466 * Short-cut one-off or mid-stream messages (state may be NULL).
1468 if ((msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
1469 HAMMER2_MSGF_ABORT)) == 0) {
1474 * Switch on CREATE, DELETE, REPLY, and also handle ABORT from
1475 * inside the case statements.
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:
1482 * New persistant command received.
1485 fprintf(stderr, "duplicate-trans %s\n",
1486 hammer2_msg_str(msg));
1487 error = HAMMER2_IOQ_ERROR_TRANS;
1490 state = malloc(sizeof(*state));
1491 bzero(state, sizeof(*state));
1492 state->iocom = iocom;
1493 state->flags = HAMMER2_STATE_DYNAMIC;
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;
1506 case HAMMER2_MSGF_DELETE:
1508 * Persistent state is expected but might not exist if an
1509 * ABORT+DELETE races the close.
1511 if (state == NULL) {
1512 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1513 error = HAMMER2_IOQ_ERROR_EALREADY;
1515 fprintf(stderr, "missing-state %s\n",
1516 hammer2_msg_str(msg));
1517 error = HAMMER2_IOQ_ERROR_TRANS;
1523 * Handle another ABORT+DELETE case if the msgid has already
1526 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1527 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1528 error = HAMMER2_IOQ_ERROR_EALREADY;
1530 fprintf(stderr, "reused-state %s\n",
1531 hammer2_msg_str(msg));
1532 error = HAMMER2_IOQ_ERROR_TRANS;
1540 * Check for mid-stream ABORT command received, otherwise
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;
1552 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE:
1553 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
1555 * When receiving a reply with CREATE set the original
1556 * persistent state message should already exist.
1558 if (state == NULL) {
1559 fprintf(stderr, "no-state(r) %s\n",
1560 hammer2_msg_str(msg));
1561 error = HAMMER2_IOQ_ERROR_TRANS;
1564 assert(((state->rxcmd ^ msg->any.head.cmd) &
1565 HAMMER2_MSGF_REPLY) == 0);
1566 state->rxcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
1569 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_DELETE:
1571 * Received REPLY+ABORT+DELETE in case where msgid has
1572 * already been fully closed, ignore the message.
1574 if (state == NULL) {
1575 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1576 error = HAMMER2_IOQ_ERROR_EALREADY;
1578 fprintf(stderr, "no-state(r,d) %s\n",
1579 hammer2_msg_str(msg));
1580 error = HAMMER2_IOQ_ERROR_TRANS;
1586 * Received REPLY+ABORT+DELETE in case where msgid has
1587 * already been reused for an unrelated message,
1588 * ignore the message.
1590 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1591 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1592 error = HAMMER2_IOQ_ERROR_EALREADY;
1594 fprintf(stderr, "reused-state(r,d) %s\n",
1595 hammer2_msg_str(msg));
1596 error = HAMMER2_IOQ_ERROR_TRANS;
1602 case HAMMER2_MSGF_REPLY:
1604 * Check for mid-stream ABORT reply received to sent command.
1606 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
1607 if (state == NULL ||
1608 (state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
1609 error = HAMMER2_IOQ_ERROR_EALREADY;
1620 hammer2_state_cleanuprx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1622 hammer2_state_t *state;
1624 if ((state = msg->state) == NULL) {
1626 * Free a non-transactional message, there is no state
1629 hammer2_msg_free(iocom, msg);
1630 } else if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
1632 * Message terminating transaction, destroy the related
1633 * state, the original message, and this message (if it
1634 * isn't the original message due to a CREATE|DELETE).
1636 pthread_mutex_lock(&iocom->mtx);
1637 state->rxcmd |= HAMMER2_MSGF_DELETE;
1638 if (state->txcmd & HAMMER2_MSGF_DELETE) {
1639 if (state->msg == msg)
1641 assert(state->flags & HAMMER2_STATE_INSERTED);
1642 if (state->rxcmd & HAMMER2_MSGF_REPLY) {
1643 assert(msg->any.head.cmd & HAMMER2_MSGF_REPLY);
1644 RB_REMOVE(hammer2_state_tree,
1645 &iocom->statewr_tree, state);
1647 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1648 RB_REMOVE(hammer2_state_tree,
1649 &iocom->staterd_tree, state);
1651 state->flags &= ~HAMMER2_STATE_INSERTED;
1652 hammer2_state_free(state);
1656 pthread_mutex_unlock(&iocom->mtx);
1657 hammer2_msg_free(iocom, msg);
1658 } else if (state->msg != msg) {
1660 * Message not terminating transaction, leave state intact
1661 * and free message if it isn't the CREATE message.
1663 hammer2_msg_free(iocom, msg);
1668 hammer2_state_cleanuptx(hammer2_iocom_t *iocom, hammer2_msg_t *msg)
1670 hammer2_state_t *state;
1672 if ((state = msg->state) == NULL) {
1673 hammer2_msg_free(iocom, msg);
1674 } else if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
1675 pthread_mutex_lock(&iocom->mtx);
1676 state->txcmd |= HAMMER2_MSGF_DELETE;
1677 if (state->rxcmd & HAMMER2_MSGF_DELETE) {
1678 if (state->msg == msg)
1680 assert(state->flags & HAMMER2_STATE_INSERTED);
1681 if (state->txcmd & HAMMER2_MSGF_REPLY) {
1682 assert(msg->any.head.cmd & HAMMER2_MSGF_REPLY);
1683 RB_REMOVE(hammer2_state_tree,
1684 &iocom->staterd_tree, state);
1686 assert((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0);
1687 RB_REMOVE(hammer2_state_tree,
1688 &iocom->statewr_tree, state);
1690 state->flags &= ~HAMMER2_STATE_INSERTED;
1691 hammer2_state_free(state);
1695 pthread_mutex_unlock(&iocom->mtx);
1696 hammer2_msg_free(iocom, msg);
1697 } else if (state->msg != msg) {
1698 hammer2_msg_free(iocom, msg);
1703 * Called with iocom locked
1706 hammer2_state_free(hammer2_state_t *state)
1708 hammer2_iocom_t *iocom = state->iocom;
1713 fprintf(stderr, "terminate state id=%08x\n",
1714 (uint32_t)state->msgid);
1716 assert(state->any.any == NULL);
1720 hammer2_msg_free_locked(iocom, msg);
1724 * When an iocom error is present we are trying to close down the
1725 * iocom, but we have to wait for all states to terminate before
1726 * we can do so. The iocom rx code will terminate the receive side
1727 * for all transactions by simulating incoming DELETE messages,
1728 * but the state doesn't go away until both sides are terminated.
1730 * We may have to wake up the rx code.
1732 if (iocom->ioq_rx.error &&
1733 RB_EMPTY(&iocom->staterd_tree) &&
1734 RB_EMPTY(&iocom->statewr_tree)) {
1736 write(iocom->wakeupfds[1], &dummy, 1);
1741 * Indexed messages are stored in a red-black tree indexed by their
1742 * msgid. Only persistent messages are indexed.
1745 hammer2_state_cmp(hammer2_state_t *state1, hammer2_state_t *state2)
1747 if (state1->spanid < state2->spanid)
1749 if (state1->spanid > state2->spanid)
1751 if (state1->msgid < state2->msgid)
1753 if (state1->msgid > state2->msgid)
1759 hammer2_basecmd_str(uint32_t cmd)
1761 static char buf[64];
1764 const char *protostr;
1767 switch(cmd & HAMMER2_MSGF_PROTOS) {
1768 case HAMMER2_MSG_PROTO_LNK:
1771 case HAMMER2_MSG_PROTO_DBG:
1774 case HAMMER2_MSG_PROTO_DOM:
1777 case HAMMER2_MSG_PROTO_CAC:
1780 case HAMMER2_MSG_PROTO_QRM:
1783 case HAMMER2_MSG_PROTO_BLK:
1786 case HAMMER2_MSG_PROTO_VOP:
1790 snprintf(protobuf, sizeof(protobuf), "%x_",
1791 (cmd & HAMMER2_MSGF_PROTOS) >> 20);
1792 protostr = protobuf;
1796 switch(cmd & (HAMMER2_MSGF_PROTOS |
1798 HAMMER2_MSGF_SIZE)) {
1799 case HAMMER2_LNK_PAD:
1802 case HAMMER2_LNK_PING:
1805 case HAMMER2_LNK_AUTH:
1808 case HAMMER2_LNK_CONN:
1811 case HAMMER2_LNK_SPAN:
1814 case HAMMER2_LNK_ERROR:
1815 if (cmd & HAMMER2_MSGF_DELETE)
1820 case HAMMER2_DBG_SHELL:
1824 snprintf(cmdbuf, sizeof(cmdbuf),
1825 "%06x", (cmd & (HAMMER2_MSGF_PROTOS |
1827 HAMMER2_MSGF_SIZE)));
1831 snprintf(buf, sizeof(buf), "%s%s", protostr, cmdstr);
1836 hammer2_msg_str(hammer2_msg_t *msg)
1838 hammer2_state_t *state;
1839 static char buf[256];
1843 const char *statestr;
1851 if ((state = msg->state) != NULL) {
1852 basecmd = (state->rxcmd & HAMMER2_MSGF_REPLY) ?
1853 state->txcmd : state->rxcmd;
1854 snprintf(statebuf, sizeof(statebuf),
1855 " %s=%s,L=%s%s,R=%s%s",
1856 ((state->txcmd & HAMMER2_MSGF_REPLY) ?
1857 "rcvcmd" : "sndcmd"),
1858 hammer2_basecmd_str(basecmd),
1859 ((state->txcmd & HAMMER2_MSGF_CREATE) ? "C" : ""),
1860 ((state->txcmd & HAMMER2_MSGF_DELETE) ? "D" : ""),
1861 ((state->rxcmd & HAMMER2_MSGF_CREATE) ? "C" : ""),
1862 ((state->rxcmd & HAMMER2_MSGF_DELETE) ? "D" : "")
1864 statestr = statebuf;
1872 switch(msg->any.head.error) {
1876 case HAMMER2_IOQ_ERROR_SYNC:
1877 errstr = "err=IOQ:NOSYNC";
1879 case HAMMER2_IOQ_ERROR_EOF:
1880 errstr = "err=IOQ:STREAMEOF";
1882 case HAMMER2_IOQ_ERROR_SOCK:
1883 errstr = "err=IOQ:SOCKERR";
1885 case HAMMER2_IOQ_ERROR_FIELD:
1886 errstr = "err=IOQ:BADFIELD";
1888 case HAMMER2_IOQ_ERROR_HCRC:
1889 errstr = "err=IOQ:BADHCRC";
1891 case HAMMER2_IOQ_ERROR_XCRC:
1892 errstr = "err=IOQ:BADXCRC";
1894 case HAMMER2_IOQ_ERROR_ACRC:
1895 errstr = "err=IOQ:BADACRC";
1897 case HAMMER2_IOQ_ERROR_STATE:
1898 errstr = "err=IOQ:BADSTATE";
1900 case HAMMER2_IOQ_ERROR_NOPEER:
1901 errstr = "err=IOQ:PEERCONFIG";
1903 case HAMMER2_IOQ_ERROR_NORKEY:
1904 errstr = "err=IOQ:BADRKEY";
1906 case HAMMER2_IOQ_ERROR_NOLKEY:
1907 errstr = "err=IOQ:BADLKEY";
1909 case HAMMER2_IOQ_ERROR_KEYXCHGFAIL:
1910 errstr = "err=IOQ:BADKEYXCHG";
1912 case HAMMER2_IOQ_ERROR_KEYFMT:
1913 errstr = "err=IOQ:BADFMT";
1915 case HAMMER2_IOQ_ERROR_BADURANDOM:
1916 errstr = "err=IOQ:BADRANDOM";
1918 case HAMMER2_IOQ_ERROR_MSGSEQ:
1919 errstr = "err=IOQ:BADSEQ";
1921 case HAMMER2_IOQ_ERROR_EALREADY:
1922 errstr = "err=IOQ:DUPMSG";
1924 case HAMMER2_IOQ_ERROR_TRANS:
1925 errstr = "err=IOQ:BADTRANS";
1927 case HAMMER2_MSG_ERR_NOSUPP:
1928 errstr = "err=NOSUPPORT";
1931 snprintf(errbuf, sizeof(errbuf),
1932 " err=%d", msg->any.head.error);
1941 if (msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
1942 HAMMER2_MSGF_ABORT | HAMMER2_MSGF_REPLY)) {
1944 if (msg->any.head.cmd & HAMMER2_MSGF_CREATE)
1946 if (msg->any.head.cmd & HAMMER2_MSGF_DELETE)
1948 if (msg->any.head.cmd & HAMMER2_MSGF_REPLY)
1950 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT)
1958 snprintf(buf, sizeof(buf),
1959 "msg=%s%s %s id=%08x span=%08x %s",
1960 hammer2_basecmd_str(msg->any.head.cmd),
1963 (uint32_t)(intmax_t)msg->any.head.msgid, /* for brevity */
1964 (uint32_t)(intmax_t)msg->any.head.spanid, /* for brevity */