Merge branch 'vendor/OPENSSL'
[dragonfly.git] / sys / kern / kern_dmsg.c
blob017576a8ee67ad6016c5fa84134d1670aa7578c9
1 /*-
2 * Copyright (c) 2012 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
35 * TODO: txcmd CREATE state is deferred by txmsgq, need to calculate
36 * a streaming response. See subr_diskiocom()'s diskiodone().
38 #include <sys/param.h>
39 #include <sys/types.h>
40 #include <sys/kernel.h>
41 #include <sys/conf.h>
42 #include <sys/systm.h>
43 #include <sys/queue.h>
44 #include <sys/tree.h>
45 #include <sys/malloc.h>
46 #include <sys/mount.h>
47 #include <sys/socket.h>
48 #include <sys/vnode.h>
49 #include <sys/file.h>
50 #include <sys/proc.h>
51 #include <sys/priv.h>
52 #include <sys/thread.h>
53 #include <sys/globaldata.h>
54 #include <sys/limits.h>
56 #include <sys/dmsg.h>
58 RB_GENERATE(kdmsg_state_tree, kdmsg_state, rbnode, kdmsg_state_cmp);
59 RB_GENERATE(kdmsg_circuit_tree, kdmsg_circuit, rbnode, kdmsg_circuit_cmp);
61 static int kdmsg_msg_receive_handling(kdmsg_msg_t *msg);
62 static int kdmsg_circ_msgrx(kdmsg_msg_t *msg);
63 static int kdmsg_state_msgrx(kdmsg_msg_t *msg);
64 static int kdmsg_state_msgtx(kdmsg_msg_t *msg);
65 static void kdmsg_state_cleanuprx(kdmsg_msg_t *msg);
66 static void kdmsg_state_cleanuptx(kdmsg_msg_t *msg);
67 static void kdmsg_state_abort(kdmsg_state_t *state);
68 static void kdmsg_state_free(kdmsg_state_t *state);
70 static void kdmsg_iocom_thread_rd(void *arg);
71 static void kdmsg_iocom_thread_wr(void *arg);
72 static int kdmsg_autorxmsg(kdmsg_msg_t *msg);
73 static void kdmsg_autocirc(kdmsg_msg_t *msg);
74 static int kdmsg_autocirc_reply(kdmsg_state_t *state, kdmsg_msg_t *msg);
76 static struct lwkt_token kdmsg_token = LWKT_TOKEN_INITIALIZER(kdmsg_token);
78 void
79 kdmsg_circ_hold(kdmsg_circuit_t *circ)
81 atomic_add_int(&circ->refs, 1);
84 void
85 kdmsg_circ_drop(kdmsg_circuit_t *circ)
87 kdmsg_iocom_t *iocom;
89 if (atomic_fetchadd_int(&circ->refs, -1) == 1) {
90 KKASSERT(circ->span_state == NULL &&
91 circ->circ_state == NULL &&
92 circ->rcirc_state == NULL &&
93 circ->recorded == 0);
94 iocom = circ->iocom;
95 circ->iocom = NULL;
96 kfree(circ, iocom->mmsg);
102 * Initialize the roll-up communications structure for a network
103 * messaging session. This function does not install the socket.
105 void
106 kdmsg_iocom_init(kdmsg_iocom_t *iocom, void *handle, uint32_t flags,
107 struct malloc_type *mmsg,
108 int (*rcvmsg)(kdmsg_msg_t *msg))
110 bzero(iocom, sizeof(*iocom));
111 iocom->handle = handle;
112 iocom->mmsg = mmsg;
113 iocom->rcvmsg = rcvmsg;
114 iocom->flags = flags;
115 lockinit(&iocom->msglk, "h2msg", 0, 0);
116 TAILQ_INIT(&iocom->msgq);
117 RB_INIT(&iocom->circ_tree);
118 RB_INIT(&iocom->staterd_tree);
119 RB_INIT(&iocom->statewr_tree);
123 * [Re]connect using the passed file pointer. The caller must ref the
124 * fp for us. We own that ref now.
126 void
127 kdmsg_iocom_reconnect(kdmsg_iocom_t *iocom, struct file *fp,
128 const char *subsysname)
131 * Destroy the current connection
133 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
134 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILL);
135 while (iocom->msgrd_td || iocom->msgwr_td) {
136 wakeup(&iocom->msg_ctl);
137 lksleep(iocom, &iocom->msglk, 0, "clstrkl", hz);
141 * Drop communications descriptor
143 if (iocom->msg_fp) {
144 fdrop(iocom->msg_fp);
145 iocom->msg_fp = NULL;
149 * Setup new communications descriptor
151 iocom->msg_ctl = 0;
152 iocom->msg_fp = fp;
153 iocom->msg_seq = 0;
154 iocom->flags &= ~KDMSG_IOCOMF_EXITNOACC;
156 lwkt_create(kdmsg_iocom_thread_rd, iocom, &iocom->msgrd_td,
157 NULL, 0, -1, "%s-msgrd", subsysname);
158 lwkt_create(kdmsg_iocom_thread_wr, iocom, &iocom->msgwr_td,
159 NULL, 0, -1, "%s-msgwr", subsysname);
160 lockmgr(&iocom->msglk, LK_RELEASE);
164 * Caller sets up iocom->auto_lnk_conn and iocom->auto_lnk_span, then calls
165 * this function to handle the state machine for LNK_CONN and LNK_SPAN.
167 * NOTE: Caller typically also sets the IOCOMF_AUTOCONN, IOCOMF_AUTOSPAN,
168 * and IOCOMF_AUTOCIRC in the kdmsg_iocom_init() call. Clients
169 * typically set IOCOMF_AUTOFORGE to automatically forged circuits
170 * for received SPANs.
172 static int kdmsg_lnk_conn_reply(kdmsg_state_t *state, kdmsg_msg_t *msg);
173 static int kdmsg_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg);
175 void
176 kdmsg_iocom_autoinitiate(kdmsg_iocom_t *iocom,
177 void (*auto_callback)(kdmsg_msg_t *msg))
179 kdmsg_msg_t *msg;
181 iocom->auto_callback = auto_callback;
183 msg = kdmsg_msg_alloc(iocom, NULL,
184 DMSG_LNK_CONN | DMSGF_CREATE,
185 kdmsg_lnk_conn_reply, NULL);
186 iocom->auto_lnk_conn.head = msg->any.head;
187 msg->any.lnk_conn = iocom->auto_lnk_conn;
188 iocom->conn_state = msg->state;
189 kdmsg_msg_write(msg);
192 static
194 kdmsg_lnk_conn_reply(kdmsg_state_t *state, kdmsg_msg_t *msg)
196 kdmsg_iocom_t *iocom = state->iocom;
197 kdmsg_msg_t *rmsg;
199 if (msg->any.head.cmd & DMSGF_CREATE) {
200 rmsg = kdmsg_msg_alloc(iocom, NULL,
201 DMSG_LNK_SPAN | DMSGF_CREATE,
202 kdmsg_lnk_span_reply, NULL);
203 iocom->auto_lnk_span.head = rmsg->any.head;
204 rmsg->any.lnk_span = iocom->auto_lnk_span;
205 kdmsg_msg_write(rmsg);
209 * Process shim after the CONN is acknowledged and before the CONN
210 * transaction is deleted. For deletions this gives device drivers
211 * the ability to interlock new operations on the circuit before
212 * it becomes illegal and panics.
214 if (iocom->auto_callback)
215 iocom->auto_callback(msg);
217 if ((state->txcmd & DMSGF_DELETE) == 0 &&
218 (msg->any.head.cmd & DMSGF_DELETE)) {
219 iocom->conn_state = NULL;
220 kdmsg_msg_reply(msg, 0);
223 return (0);
226 static
228 kdmsg_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg)
231 * Be sure to process shim before terminating the SPAN
232 * transaction. Gives device drivers the ability to
233 * interlock new operations on the circuit before it
234 * becomes illegal and panics.
236 if (state->iocom->auto_callback)
237 state->iocom->auto_callback(msg);
239 if ((state->txcmd & DMSGF_DELETE) == 0 &&
240 (msg->any.head.cmd & DMSGF_DELETE)) {
241 kdmsg_msg_reply(msg, 0);
243 return (0);
247 * Disconnect and clean up
249 void
250 kdmsg_iocom_uninit(kdmsg_iocom_t *iocom)
252 kdmsg_state_t *state;
255 * Ask the cluster controller to go away
257 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
258 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILL);
260 while (iocom->msgrd_td || iocom->msgwr_td) {
261 wakeup(&iocom->msg_ctl);
262 lksleep(iocom, &iocom->msglk, 0, "clstrkl", hz);
266 * Cleanup caches
268 if ((state = iocom->freerd_state) != NULL) {
269 iocom->freerd_state = NULL;
270 kdmsg_state_free(state);
273 if ((state = iocom->freewr_state) != NULL) {
274 iocom->freewr_state = NULL;
275 kdmsg_state_free(state);
279 * Drop communications descriptor
281 if (iocom->msg_fp) {
282 fdrop(iocom->msg_fp);
283 iocom->msg_fp = NULL;
285 lockmgr(&iocom->msglk, LK_RELEASE);
289 * Cluster controller thread. Perform messaging functions. We have one
290 * thread for the reader and one for the writer. The writer handles
291 * shutdown requests (which should break the reader thread).
293 static
294 void
295 kdmsg_iocom_thread_rd(void *arg)
297 kdmsg_iocom_t *iocom = arg;
298 dmsg_hdr_t hdr;
299 kdmsg_msg_t *msg = NULL;
300 size_t hbytes;
301 size_t abytes;
302 int error = 0;
304 while ((iocom->msg_ctl & KDMSG_CLUSTERCTL_KILL) == 0) {
306 * Retrieve the message from the pipe or socket.
308 error = fp_read(iocom->msg_fp, &hdr, sizeof(hdr),
309 NULL, 1, UIO_SYSSPACE);
310 if (error)
311 break;
312 if (hdr.magic != DMSG_HDR_MAGIC) {
313 kprintf("kdmsg: bad magic: %04x\n", hdr.magic);
314 error = EINVAL;
315 break;
317 hbytes = (hdr.cmd & DMSGF_SIZE) * DMSG_ALIGN;
318 if (hbytes < sizeof(hdr) || hbytes > DMSG_AUX_MAX) {
319 kprintf("kdmsg: bad header size %zd\n", hbytes);
320 error = EINVAL;
321 break;
323 /* XXX messy: mask cmd to avoid allocating state */
324 msg = kdmsg_msg_alloc(iocom, NULL,
325 hdr.cmd & DMSGF_BASECMDMASK,
326 NULL, NULL);
327 msg->any.head = hdr;
328 msg->hdr_size = hbytes;
329 if (hbytes > sizeof(hdr)) {
330 error = fp_read(iocom->msg_fp, &msg->any.head + 1,
331 hbytes - sizeof(hdr),
332 NULL, 1, UIO_SYSSPACE);
333 if (error) {
334 kprintf("kdmsg: short msg received\n");
335 error = EINVAL;
336 break;
339 msg->aux_size = hdr.aux_bytes;
340 if (msg->aux_size > DMSG_AUX_MAX) {
341 kprintf("kdmsg: illegal msg payload size %zd\n",
342 msg->aux_size);
343 error = EINVAL;
344 break;
346 if (msg->aux_size) {
347 abytes = DMSG_DOALIGN(msg->aux_size);
348 msg->aux_data = kmalloc(abytes, iocom->mmsg, M_WAITOK);
349 msg->flags |= KDMSG_FLAG_AUXALLOC;
350 error = fp_read(iocom->msg_fp, msg->aux_data,
351 abytes, NULL, 1, UIO_SYSSPACE);
352 if (error) {
353 kprintf("kdmsg: short msg payload received\n");
354 break;
358 (void)kdmsg_circ_msgrx(msg);
359 error = kdmsg_msg_receive_handling(msg);
360 msg = NULL;
363 if (error)
364 kprintf("kdmsg: read failed error %d\n", error);
366 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
367 if (msg)
368 kdmsg_msg_free(msg);
371 * Shutdown the socket before waiting for the transmit side.
373 * If we are dying due to e.g. a socket disconnect verses being
374 * killed explicity we have to set KILL in order to kick the tx
375 * side when it might not have any other work to do. KILL might
376 * already be set if we are in an unmount or reconnect.
378 fp_shutdown(iocom->msg_fp, SHUT_RDWR);
380 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILL);
381 wakeup(&iocom->msg_ctl);
384 * Wait for the transmit side to drain remaining messages
385 * before cleaning up the rx state. The transmit side will
386 * set KILLTX and wait for the rx side to completely finish
387 * (set msgrd_td to NULL) before cleaning up any remaining
388 * tx states.
390 lockmgr(&iocom->msglk, LK_RELEASE);
391 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILLRX);
392 wakeup(&iocom->msg_ctl);
393 while ((iocom->msg_ctl & KDMSG_CLUSTERCTL_KILLTX) == 0) {
394 wakeup(&iocom->msg_ctl);
395 tsleep(iocom, 0, "clstrkw", hz);
398 iocom->msgrd_td = NULL;
401 * iocom can be ripped out from under us at this point but
402 * wakeup() is safe.
404 wakeup(iocom);
405 lwkt_exit();
408 static
409 void
410 kdmsg_iocom_thread_wr(void *arg)
412 kdmsg_iocom_t *iocom = arg;
413 kdmsg_msg_t *msg;
414 kdmsg_state_t *state;
415 ssize_t res;
416 size_t abytes;
417 int error = 0;
418 int retries = 20;
421 * Transmit loop
423 msg = NULL;
424 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
426 while ((iocom->msg_ctl & KDMSG_CLUSTERCTL_KILL) == 0 && error == 0) {
428 * Sleep if no messages pending. Interlock with flag while
429 * holding msglk.
431 if (TAILQ_EMPTY(&iocom->msgq)) {
432 atomic_set_int(&iocom->msg_ctl,
433 KDMSG_CLUSTERCTL_SLEEPING);
434 lksleep(&iocom->msg_ctl, &iocom->msglk, 0, "msgwr", hz);
435 atomic_clear_int(&iocom->msg_ctl,
436 KDMSG_CLUSTERCTL_SLEEPING);
439 while ((msg = TAILQ_FIRST(&iocom->msgq)) != NULL) {
441 * Remove msg from the transmit queue and do
442 * persist and half-closed state handling.
444 TAILQ_REMOVE(&iocom->msgq, msg, qentry);
445 lockmgr(&iocom->msglk, LK_RELEASE);
447 error = kdmsg_state_msgtx(msg);
448 if (error == EALREADY) {
449 error = 0;
450 kdmsg_msg_free(msg);
451 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
452 continue;
454 if (error) {
455 kdmsg_msg_free(msg);
456 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
457 break;
461 * Dump the message to the pipe or socket.
463 * We have to clean up the message as if the transmit
464 * succeeded even if it failed.
466 error = fp_write(iocom->msg_fp, &msg->any,
467 msg->hdr_size, &res, UIO_SYSSPACE);
468 if (error || res != msg->hdr_size) {
469 if (error == 0)
470 error = EINVAL;
471 kdmsg_state_cleanuptx(msg);
472 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
473 break;
475 if (msg->aux_size) {
476 abytes = DMSG_DOALIGN(msg->aux_size);
477 error = fp_write(iocom->msg_fp,
478 msg->aux_data, abytes,
479 &res, UIO_SYSSPACE);
480 if (error || res != abytes) {
481 if (error == 0)
482 error = EINVAL;
483 kdmsg_state_cleanuptx(msg);
484 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
485 break;
488 kdmsg_state_cleanuptx(msg);
489 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
494 * Cleanup messages pending transmission and release msgq lock.
496 if (error)
497 kprintf("kdmsg: write failed error %d\n", error);
498 kprintf("thread_wr: Terminating iocom\n");
501 * Shutdown the socket. This will cause the rx thread to get an
502 * EOF and ensure that both threads get to a termination state.
504 fp_shutdown(iocom->msg_fp, SHUT_RDWR);
507 * Set KILLTX (which the rx side waits for), then wait for the RX
508 * side to completely finish before we clean out any remaining
509 * command states.
511 lockmgr(&iocom->msglk, LK_RELEASE);
512 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILLTX);
513 wakeup(&iocom->msg_ctl);
514 while (iocom->msgrd_td) {
515 wakeup(&iocom->msg_ctl);
516 tsleep(iocom, 0, "clstrkw", hz);
518 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
521 * Simulate received MSGF_DELETE's for any remaining states.
522 * (For remote masters).
524 * Drain the message queue to handle any device initiated writes
525 * due to state callbacks.
527 cleanuprd:
528 kdmsg_drain_msgq(iocom);
529 RB_FOREACH(state, kdmsg_state_tree, &iocom->staterd_tree) {
530 if ((state->rxcmd & DMSGF_DELETE) == 0) {
531 lockmgr(&iocom->msglk, LK_RELEASE);
532 kdmsg_state_abort(state);
533 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
534 goto cleanuprd;
539 * Simulate received MSGF_DELETE's for any remaining states.
540 * (For local masters).
542 cleanupwr:
543 kdmsg_drain_msgq(iocom);
544 RB_FOREACH(state, kdmsg_state_tree, &iocom->statewr_tree) {
545 if ((state->rxcmd & DMSGF_DELETE) == 0) {
546 lockmgr(&iocom->msglk, LK_RELEASE);
547 kdmsg_state_abort(state);
548 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
549 goto cleanupwr;
554 * Retry until all work is done
556 if (--retries == 0)
557 panic("kdmsg: comm thread shutdown couldn't drain");
558 if (TAILQ_FIRST(&iocom->msgq) ||
559 RB_ROOT(&iocom->staterd_tree) ||
560 RB_ROOT(&iocom->statewr_tree)) {
561 goto cleanuprd;
563 iocom->flags |= KDMSG_IOCOMF_EXITNOACC;
565 lockmgr(&iocom->msglk, LK_RELEASE);
568 * The state trees had better be empty now
570 KKASSERT(RB_EMPTY(&iocom->staterd_tree));
571 KKASSERT(RB_EMPTY(&iocom->statewr_tree));
572 KKASSERT(iocom->conn_state == NULL);
574 if (iocom->exit_func) {
576 * iocom is invalid after we call the exit function.
578 iocom->msgwr_td = NULL;
579 iocom->exit_func(iocom);
580 } else {
582 * iocom can be ripped out from under us once msgwr_td is
583 * set to NULL. The wakeup is safe.
585 iocom->msgwr_td = NULL;
586 wakeup(iocom);
588 lwkt_exit();
592 * This cleans out the pending transmit message queue, adjusting any
593 * persistent states properly in the process.
595 * Caller must hold pmp->iocom.msglk
597 void
598 kdmsg_drain_msgq(kdmsg_iocom_t *iocom)
600 kdmsg_msg_t *msg;
603 * Clean out our pending transmit queue, executing the
604 * appropriate state adjustments. If this tries to open
605 * any new outgoing transactions we have to loop up and
606 * clean them out.
608 while ((msg = TAILQ_FIRST(&iocom->msgq)) != NULL) {
609 TAILQ_REMOVE(&iocom->msgq, msg, qentry);
610 lockmgr(&iocom->msglk, LK_RELEASE);
611 if (kdmsg_state_msgtx(msg))
612 kdmsg_msg_free(msg);
613 else
614 kdmsg_state_cleanuptx(msg);
615 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
620 * Do all processing required to handle a freshly received message
621 * after its low level header has been validated.
623 static
625 kdmsg_msg_receive_handling(kdmsg_msg_t *msg)
627 kdmsg_iocom_t *iocom = msg->iocom;
628 int error;
631 * State machine tracking, state assignment for msg,
632 * returns error and discard status. Errors are fatal
633 * to the connection except for EALREADY which forces
634 * a discard without execution.
636 error = kdmsg_state_msgrx(msg);
637 if (error) {
639 * Raw protocol or connection error
641 kdmsg_msg_free(msg);
642 if (error == EALREADY)
643 error = 0;
644 } else if (msg->state && msg->state->func) {
646 * Message related to state which already has a
647 * handling function installed for it.
649 error = msg->state->func(msg->state, msg);
650 kdmsg_state_cleanuprx(msg);
651 } else if (iocom->flags & KDMSG_IOCOMF_AUTOANY) {
652 error = kdmsg_autorxmsg(msg);
653 kdmsg_state_cleanuprx(msg);
654 } else {
655 error = iocom->rcvmsg(msg);
656 kdmsg_state_cleanuprx(msg);
658 return error;
662 * Process circuit tracking (NEEDS WORK)
664 static
666 kdmsg_circ_msgrx(kdmsg_msg_t *msg)
668 kdmsg_circuit_t dummy;
669 kdmsg_circuit_t *circ;
670 int error = 0;
672 if (msg->any.head.circuit) {
673 dummy.msgid = msg->any.head.circuit;
674 lwkt_gettoken(&kdmsg_token);
675 circ = RB_FIND(kdmsg_circuit_tree, &msg->iocom->circ_tree,
676 &dummy);
677 if (circ) {
678 msg->circ = circ;
679 kdmsg_circ_hold(circ);
681 if (circ == NULL) {
682 kprintf("KDMSG_CIRC_MSGRX CMD %08x: IOCOM %p "
683 "Bad circuit %016jx\n",
684 msg->any.head.cmd,
685 msg->iocom,
686 (intmax_t)msg->any.head.circuit);
687 kprintf("KDMSG_CIRC_MSGRX: Avail circuits: ");
688 RB_FOREACH(circ, kdmsg_circuit_tree,
689 &msg->iocom->circ_tree) {
690 kprintf(" %016jx", (intmax_t)circ->msgid);
692 kprintf("\n");
693 error = EINVAL;
695 lwkt_reltoken(&kdmsg_token);
697 return (error);
701 * Process state tracking for a message after reception, prior to
702 * execution.
704 * Called with msglk held and the msg dequeued.
706 * All messages are called with dummy state and return actual state.
707 * (One-off messages often just return the same dummy state).
709 * May request that caller discard the message by setting *discardp to 1.
710 * The returned state is not used in this case and is allowed to be NULL.
712 * --
714 * These routines handle persistent and command/reply message state via the
715 * CREATE and DELETE flags. The first message in a command or reply sequence
716 * sets CREATE, the last message in a command or reply sequence sets DELETE.
718 * There can be any number of intermediate messages belonging to the same
719 * sequence sent inbetween the CREATE message and the DELETE message,
720 * which set neither flag. This represents a streaming command or reply.
722 * Any command message received with CREATE set expects a reply sequence to
723 * be returned. Reply sequences work the same as command sequences except the
724 * REPLY bit is also sent. Both the command side and reply side can
725 * degenerate into a single message with both CREATE and DELETE set. Note
726 * that one side can be streaming and the other side not, or neither, or both.
728 * The msgid is unique for the initiator. That is, two sides sending a new
729 * message can use the same msgid without colliding.
731 * --
733 * ABORT sequences work by setting the ABORT flag along with normal message
734 * state. However, ABORTs can also be sent on half-closed messages, that is
735 * even if the command or reply side has already sent a DELETE, as long as
736 * the message has not been fully closed it can still send an ABORT+DELETE
737 * to terminate the half-closed message state.
739 * Since ABORT+DELETEs can race we silently discard ABORT's for message
740 * state which has already been fully closed. REPLY+ABORT+DELETEs can
741 * also race, and in this situation the other side might have already
742 * initiated a new unrelated command with the same message id. Since
743 * the abort has not set the CREATE flag the situation can be detected
744 * and the message will also be discarded.
746 * Non-blocking requests can be initiated with ABORT+CREATE[+DELETE].
747 * The ABORT request is essentially integrated into the command instead
748 * of being sent later on. In this situation the command implementation
749 * detects that CREATE and ABORT are both set (vs ABORT alone) and can
750 * special-case non-blocking operation for the command.
752 * NOTE! Messages with ABORT set without CREATE or DELETE are considered
753 * to be mid-stream aborts for command/reply sequences. ABORTs on
754 * one-way messages are not supported.
756 * NOTE! If a command sequence does not support aborts the ABORT flag is
757 * simply ignored.
759 * --
761 * One-off messages (no reply expected) are sent with neither CREATE or DELETE
762 * set. One-off messages cannot be aborted and typically aren't processed
763 * by these routines. The REPLY bit can be used to distinguish whether a
764 * one-off message is a command or reply. For example, one-off replies
765 * will typically just contain status updates.
767 static
769 kdmsg_state_msgrx(kdmsg_msg_t *msg)
771 kdmsg_iocom_t *iocom = msg->iocom;
772 kdmsg_state_t *state;
773 int error;
776 * Make sure a state structure is ready to go in case we need a new
777 * one. This is the only routine which uses freerd_state so no
778 * races are possible.
780 if ((state = iocom->freerd_state) == NULL) {
781 state = kmalloc(sizeof(*state), iocom->mmsg, M_WAITOK | M_ZERO);
782 state->flags = KDMSG_STATE_DYNAMIC;
783 iocom->freerd_state = state;
787 * Lock RB tree and locate existing persistent state, if any.
789 * If received msg is a command state is on staterd_tree.
790 * If received msg is a reply state is on statewr_tree.
792 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
794 state->msgid = msg->any.head.msgid;
795 state->circ = msg->circ;
796 state->iocom = iocom;
797 if (msg->any.head.cmd & DMSGF_REPLY)
798 state = RB_FIND(kdmsg_state_tree, &iocom->statewr_tree, state);
799 else
800 state = RB_FIND(kdmsg_state_tree, &iocom->staterd_tree, state);
801 msg->state = state;
804 * Short-cut one-off or mid-stream messages (state may be NULL).
806 if ((msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE |
807 DMSGF_ABORT)) == 0) {
808 lockmgr(&iocom->msglk, LK_RELEASE);
809 return(0);
813 * Switch on CREATE, DELETE, REPLY, and also handle ABORT from
814 * inside the case statements.
816 switch(msg->any.head.cmd & (DMSGF_CREATE|DMSGF_DELETE|DMSGF_REPLY)) {
817 case DMSGF_CREATE:
818 case DMSGF_CREATE | DMSGF_DELETE:
820 * New persistant command received.
822 if (state) {
823 kprintf("kdmsg_state_msgrx: duplicate transaction\n");
824 error = EINVAL;
825 break;
827 state = iocom->freerd_state;
828 iocom->freerd_state = NULL;
829 msg->state = state;
830 state->msg = msg;
831 state->icmd = msg->any.head.cmd & DMSGF_BASECMDMASK;
832 state->rxcmd = msg->any.head.cmd & ~DMSGF_DELETE;
833 state->txcmd = DMSGF_REPLY;
834 state->msgid = msg->any.head.msgid;
835 if ((state->circ = msg->circ) != NULL)
836 kdmsg_circ_hold(state->circ);
837 RB_INSERT(kdmsg_state_tree, &iocom->staterd_tree, state);
838 state->flags |= KDMSG_STATE_INSERTED;
839 error = 0;
840 break;
841 case DMSGF_DELETE:
843 * Persistent state is expected but might not exist if an
844 * ABORT+DELETE races the close.
846 if (state == NULL) {
847 if (msg->any.head.cmd & DMSGF_ABORT) {
848 error = EALREADY;
849 } else {
850 kprintf("kdmsg_state_msgrx: "
851 "no state for DELETE\n");
852 error = EINVAL;
854 break;
858 * Handle another ABORT+DELETE case if the msgid has already
859 * been reused.
861 if ((state->rxcmd & DMSGF_CREATE) == 0) {
862 if (msg->any.head.cmd & DMSGF_ABORT) {
863 error = EALREADY;
864 } else {
865 kprintf("kdmsg_state_msgrx: "
866 "state reused for DELETE\n");
867 error = EINVAL;
869 break;
871 error = 0;
872 break;
873 default:
875 * Check for mid-stream ABORT command received, otherwise
876 * allow.
878 if (msg->any.head.cmd & DMSGF_ABORT) {
879 if (state == NULL ||
880 (state->rxcmd & DMSGF_CREATE) == 0) {
881 error = EALREADY;
882 break;
885 error = 0;
886 break;
887 case DMSGF_REPLY | DMSGF_CREATE:
888 case DMSGF_REPLY | DMSGF_CREATE | DMSGF_DELETE:
890 * When receiving a reply with CREATE set the original
891 * persistent state message should already exist.
893 if (state == NULL) {
894 kprintf("kdmsg_state_msgrx: no state match for "
895 "REPLY cmd=%08x msgid=%016jx\n",
896 msg->any.head.cmd,
897 (intmax_t)msg->any.head.msgid);
898 error = EINVAL;
899 break;
901 state->rxcmd = msg->any.head.cmd & ~DMSGF_DELETE;
902 error = 0;
903 break;
904 case DMSGF_REPLY | DMSGF_DELETE:
906 * Received REPLY+ABORT+DELETE in case where msgid has
907 * already been fully closed, ignore the message.
909 if (state == NULL) {
910 if (msg->any.head.cmd & DMSGF_ABORT) {
911 error = EALREADY;
912 } else {
913 kprintf("kdmsg_state_msgrx: no state match "
914 "for REPLY|DELETE\n");
915 error = EINVAL;
917 break;
921 * Received REPLY+ABORT+DELETE in case where msgid has
922 * already been reused for an unrelated message,
923 * ignore the message.
925 if ((state->rxcmd & DMSGF_CREATE) == 0) {
926 if (msg->any.head.cmd & DMSGF_ABORT) {
927 error = EALREADY;
928 } else {
929 kprintf("kdmsg_state_msgrx: state reused "
930 "for REPLY|DELETE\n");
931 error = EINVAL;
933 break;
935 error = 0;
936 break;
937 case DMSGF_REPLY:
939 * Check for mid-stream ABORT reply received to sent command.
941 if (msg->any.head.cmd & DMSGF_ABORT) {
942 if (state == NULL ||
943 (state->rxcmd & DMSGF_CREATE) == 0) {
944 error = EALREADY;
945 break;
948 error = 0;
949 break;
951 lockmgr(&iocom->msglk, LK_RELEASE);
952 return (error);
956 * Called instead of iocom->rcvmsg() if any of the AUTO flags are set.
957 * This routine must call iocom->rcvmsg() for anything not automatically
958 * handled.
960 static int
961 kdmsg_autorxmsg(kdmsg_msg_t *msg)
963 kdmsg_iocom_t *iocom = msg->iocom;
964 kdmsg_circuit_t *circ;
965 int error = 0;
966 uint32_t cmd;
969 * Process a combination of the transaction command and the message
970 * flags. For the purposes of this routine, the message command is
971 * only relevant when it initiates a transaction (where it is
972 * recorded in icmd).
974 cmd = (msg->state ? msg->state->icmd : msg->any.head.cmd) &
975 DMSGF_BASECMDMASK;
976 cmd |= msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE | DMSGF_REPLY);
978 switch(cmd) {
979 case DMSG_LNK_CONN | DMSGF_CREATE:
980 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE:
982 * Received LNK_CONN transaction. Transmit response and
983 * leave transaction open, which allows the other end to
984 * start to the SPAN protocol.
986 * Handle shim after acknowledging the CONN.
988 if ((msg->any.head.cmd & DMSGF_DELETE) == 0) {
989 if (iocom->flags & KDMSG_IOCOMF_AUTOCONN) {
990 kdmsg_msg_result(msg, 0);
991 if (iocom->auto_callback)
992 iocom->auto_callback(msg);
993 } else {
994 error = iocom->rcvmsg(msg);
996 break;
998 /* fall through */
999 case DMSG_LNK_CONN | DMSGF_DELETE:
1001 * This message is usually simulated after a link is lost
1002 * to clean up the transaction.
1004 if (iocom->flags & KDMSG_IOCOMF_AUTOCONN) {
1005 if (iocom->auto_callback)
1006 iocom->auto_callback(msg);
1007 kdmsg_msg_reply(msg, 0);
1008 } else {
1009 error = iocom->rcvmsg(msg);
1011 break;
1012 case DMSG_LNK_SPAN | DMSGF_CREATE:
1013 case DMSG_LNK_SPAN | DMSGF_CREATE | DMSGF_DELETE:
1015 * Received LNK_SPAN transaction. We do not have to respond
1016 * but we must leave the transaction open.
1018 * If AUTOCIRC is set automatically initiate a virtual circuit
1019 * to the received span. This will attach a kdmsg_circuit
1020 * to the SPAN state. The circuit is lost when the span is
1021 * lost.
1023 * Handle shim after acknowledging the SPAN.
1025 if (iocom->flags & KDMSG_IOCOMF_AUTOSPAN) {
1026 if ((msg->any.head.cmd & DMSGF_DELETE) == 0) {
1027 if (iocom->flags & KDMSG_IOCOMF_AUTOFORGE)
1028 kdmsg_autocirc(msg);
1029 if (iocom->auto_callback)
1030 iocom->auto_callback(msg);
1031 break;
1033 /* fall through */
1034 } else {
1035 error = iocom->rcvmsg(msg);
1036 break;
1038 /* fall through */
1039 case DMSG_LNK_SPAN | DMSGF_DELETE:
1041 * Process shims (auto_callback) before cleaning up the
1042 * circuit structure and closing the transactions. Device
1043 * driver should ensure that the circuit is not used after
1044 * the auto_callback() returns.
1046 * Handle shim before closing the SPAN transaction.
1048 if (iocom->flags & KDMSG_IOCOMF_AUTOSPAN) {
1049 if (iocom->auto_callback)
1050 iocom->auto_callback(msg);
1051 if (iocom->flags & KDMSG_IOCOMF_AUTOFORGE)
1052 kdmsg_autocirc(msg);
1053 kdmsg_msg_reply(msg, 0);
1054 } else {
1055 error = iocom->rcvmsg(msg);
1057 break;
1058 case DMSG_LNK_CIRC | DMSGF_CREATE:
1059 case DMSG_LNK_CIRC | DMSGF_CREATE | DMSGF_DELETE:
1061 * Received LNK_CIRC transaction. We must respond and should
1062 * leave the transaction open, allowing the circuit. The
1063 * remote can start issuing commands to us over the circuit
1064 * even before we respond.
1066 if (iocom->flags & KDMSG_IOCOMF_AUTOCIRC) {
1067 if ((msg->any.head.cmd & DMSGF_DELETE) == 0) {
1068 circ = kmalloc(sizeof(*circ), iocom->mmsg,
1069 M_WAITOK | M_ZERO);
1070 lwkt_gettoken(&kdmsg_token);
1071 msg->state->any.circ = circ;
1072 circ->iocom = iocom;
1073 circ->rcirc_state = msg->state;
1074 kdmsg_circ_hold(circ); /* for rcirc_state */
1075 circ->weight = 0;
1076 circ->msgid = circ->rcirc_state->msgid;
1077 /* XXX no span link for received circuits */
1078 kdmsg_circ_hold(circ); /* for circ_state */
1080 if (RB_INSERT(kdmsg_circuit_tree,
1081 &iocom->circ_tree, circ)) {
1082 panic("duplicate circuitid allocated");
1084 lwkt_reltoken(&kdmsg_token);
1085 kdmsg_msg_result(msg, 0);
1088 * Handle shim after adding the circuit and
1089 * after acknowledging the CIRC.
1091 if (iocom->auto_callback)
1092 iocom->auto_callback(msg);
1093 break;
1095 /* fall through */
1096 } else {
1097 error = iocom->rcvmsg(msg);
1098 break;
1100 /* fall through */
1101 case DMSG_LNK_CIRC | DMSGF_DELETE:
1102 if (iocom->flags & KDMSG_IOCOMF_AUTOCIRC) {
1103 circ = msg->state->any.circ;
1104 if (circ == NULL)
1105 break;
1108 * Handle shim before terminating the circuit.
1110 #if 0
1111 kprintf("KDMSG VC: RECEIVE CIRC DELETE "
1112 "IOCOM %p MSGID %016jx\n",
1113 msg->iocom, circ->msgid);
1114 #endif
1115 if (iocom->auto_callback)
1116 iocom->auto_callback(msg);
1118 KKASSERT(circ->rcirc_state == msg->state);
1119 lwkt_gettoken(&kdmsg_token);
1120 circ->rcirc_state = NULL;
1121 msg->state->any.circ = NULL;
1122 RB_REMOVE(kdmsg_circuit_tree, &iocom->circ_tree, circ);
1123 lwkt_reltoken(&kdmsg_token);
1124 kdmsg_circ_drop(circ); /* for rcirc_state */
1125 kdmsg_msg_reply(msg, 0);
1126 } else {
1127 error = iocom->rcvmsg(msg);
1129 break;
1130 default:
1132 * Anything unhandled goes into rcvmsg.
1134 * NOTE: Replies to link-level messages initiated by our side
1135 * are handled by the state callback, they are NOT
1136 * handled here.
1138 error = iocom->rcvmsg(msg);
1139 break;
1141 return (error);
1145 * Handle automatic forging of virtual circuits based on received SPANs.
1146 * (AUTOFORGE). Note that other code handles tracking received circuit
1147 * transactions (AUTOCIRC).
1149 * We can ignore non-transactions here. Use trans->icmd to test the
1150 * transactional command (once past the CREATE the individual message
1151 * commands are not usually the icmd).
1153 * XXX locks
1155 static
1156 void
1157 kdmsg_autocirc(kdmsg_msg_t *msg)
1159 kdmsg_iocom_t *iocom = msg->iocom;
1160 kdmsg_circuit_t *circ;
1161 kdmsg_msg_t *xmsg; /* CIRC */
1163 if (msg->state == NULL)
1164 return;
1167 * Gaining the SPAN, automatically forge a circuit to the target.
1169 * NOTE!! The shim is not executed until we receive an acknowlegement
1170 * to our forged LNK_CIRC (see kdmsg_autocirc_reply()).
1172 if (msg->state->icmd == DMSG_LNK_SPAN &&
1173 (msg->any.head.cmd & DMSGF_CREATE)) {
1174 circ = kmalloc(sizeof(*circ), iocom->mmsg, M_WAITOK | M_ZERO);
1175 lwkt_gettoken(&kdmsg_token);
1176 msg->state->any.circ = circ;
1177 circ->iocom = iocom;
1178 circ->span_state = msg->state;
1179 kdmsg_circ_hold(circ); /* for span_state */
1180 xmsg = kdmsg_msg_alloc(iocom, NULL,
1181 DMSG_LNK_CIRC | DMSGF_CREATE,
1182 kdmsg_autocirc_reply, circ);
1183 circ->circ_state = xmsg->state;
1184 circ->weight = msg->any.lnk_span.dist;
1185 circ->msgid = circ->circ_state->msgid;
1186 kdmsg_circ_hold(circ); /* for circ_state */
1187 #if 0
1188 kprintf("KDMSG VC: CREATE SPAN->CIRC IOCOM %p MSGID %016jx\n",
1189 msg->iocom, circ->msgid);
1190 #endif
1192 if (RB_INSERT(kdmsg_circuit_tree, &iocom->circ_tree, circ))
1193 panic("duplicate circuitid allocated");
1194 lwkt_reltoken(&kdmsg_token);
1196 xmsg->any.lnk_circ.target = msg->any.head.msgid;
1197 kdmsg_msg_write(xmsg);
1201 * Losing the SPAN
1203 * NOTE: When losing a SPAN, any circuits using the span should be
1204 * deleted by the remote end first. XXX might not be ordered
1205 * on actual loss of connection.
1207 if (msg->state->icmd == DMSG_LNK_SPAN &&
1208 (msg->any.head.cmd & DMSGF_DELETE) &&
1209 msg->state->any.circ) {
1210 circ = msg->state->any.circ;
1211 lwkt_gettoken(&kdmsg_token);
1212 circ->span_state = NULL;
1213 msg->state->any.circ = NULL;
1214 RB_REMOVE(kdmsg_circuit_tree, &iocom->circ_tree, circ);
1215 #if 0
1216 kprintf("KDMSG VC: DELETE SPAN->CIRC IOCOM %p MSGID %016jx\n",
1217 msg->iocom, (intmax_t)circ->msgid);
1218 #endif
1219 kdmsg_circ_drop(circ); /* for span_state */
1220 lwkt_reltoken(&kdmsg_token);
1224 static
1226 kdmsg_autocirc_reply(kdmsg_state_t *state, kdmsg_msg_t *msg)
1228 kdmsg_iocom_t *iocom = state->iocom;
1229 kdmsg_circuit_t *circ = state->any.circ;
1232 * Call shim after receiving an acknowlegement to our forged
1233 * circuit and before processing a received termination.
1235 if (iocom->auto_callback)
1236 iocom->auto_callback(msg);
1239 * If the remote is terminating the VC we terminate our side
1241 if ((state->txcmd & DMSGF_DELETE) == 0 &&
1242 (msg->any.head.cmd & DMSGF_DELETE)) {
1243 #if 0
1244 kprintf("KDMSG VC: DELETE CIRC FROM REMOTE\n");
1245 #endif
1246 lwkt_gettoken(&kdmsg_token);
1247 circ->circ_state = NULL;
1248 state->any.circ = NULL;
1249 kdmsg_circ_drop(circ); /* for circ_state */
1250 lwkt_reltoken(&kdmsg_token);
1251 kdmsg_msg_reply(msg, 0);
1253 return (0);
1257 * Post-receive-handling message and state cleanup. This routine is called
1258 * after the state function handling/callback to properly dispose of the
1259 * message and update or dispose of the state.
1261 static
1262 void
1263 kdmsg_state_cleanuprx(kdmsg_msg_t *msg)
1265 kdmsg_iocom_t *iocom = msg->iocom;
1266 kdmsg_state_t *state;
1268 if ((state = msg->state) == NULL) {
1269 kdmsg_msg_free(msg);
1270 } else if (msg->any.head.cmd & DMSGF_DELETE) {
1271 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
1272 KKASSERT((state->rxcmd & DMSGF_DELETE) == 0);
1273 state->rxcmd |= DMSGF_DELETE;
1274 if (state->txcmd & DMSGF_DELETE) {
1275 KKASSERT(state->flags & KDMSG_STATE_INSERTED);
1276 if (state->rxcmd & DMSGF_REPLY) {
1277 KKASSERT(msg->any.head.cmd &
1278 DMSGF_REPLY);
1279 RB_REMOVE(kdmsg_state_tree,
1280 &iocom->statewr_tree, state);
1281 } else {
1282 KKASSERT((msg->any.head.cmd &
1283 DMSGF_REPLY) == 0);
1284 RB_REMOVE(kdmsg_state_tree,
1285 &iocom->staterd_tree, state);
1287 state->flags &= ~KDMSG_STATE_INSERTED;
1288 if (msg != state->msg)
1289 kdmsg_msg_free(msg);
1290 lockmgr(&iocom->msglk, LK_RELEASE);
1291 kdmsg_state_free(state);
1292 } else {
1293 if (msg != state->msg)
1294 kdmsg_msg_free(msg);
1295 lockmgr(&iocom->msglk, LK_RELEASE);
1297 } else if (msg != state->msg) {
1298 kdmsg_msg_free(msg);
1303 * Simulate receiving a message which terminates an active transaction
1304 * state. Our simulated received message must set DELETE and may also
1305 * have to set CREATE. It must also ensure that all fields are set such
1306 * that the receive handling code can find the state (kdmsg_state_msgrx())
1307 * or an endless loop will ensue.
1309 * This is used when the other end of the link or virtual circuit is dead
1310 * so the device driver gets a completed transaction for all pending states.
1312 static
1313 void
1314 kdmsg_state_abort(kdmsg_state_t *state)
1316 kdmsg_iocom_t *iocom = state->iocom;
1317 kdmsg_msg_t *msg;
1320 * Prevent recursive aborts which could otherwise occur if the
1321 * simulated message reception runs state->func which then turns
1322 * around and tries to reply to a broken circuit when then calls
1323 * the state abort code again.
1325 if (state->flags & KDMSG_STATE_ABORTING)
1326 return;
1327 state->flags |= KDMSG_STATE_ABORTING;
1330 * Simulatem essage reception
1332 msg = kdmsg_msg_alloc(iocom, state->circ,
1333 DMSG_LNK_ERROR,
1334 NULL, NULL);
1335 if ((state->rxcmd & DMSGF_CREATE) == 0)
1336 msg->any.head.cmd |= DMSGF_CREATE;
1337 msg->any.head.cmd |= DMSGF_DELETE | (state->rxcmd & DMSGF_REPLY);
1338 msg->any.head.error = DMSG_ERR_LOSTLINK;
1339 msg->any.head.msgid = state->msgid;
1340 msg->state = state;
1341 kdmsg_msg_receive_handling(msg);
1345 * Process state tracking for a message prior to transmission.
1347 * Called with msglk held and the msg dequeued. Returns non-zero if
1348 * the message is bad and should be deleted by the caller.
1350 * One-off messages are usually with dummy state and msg->state may be NULL
1351 * in this situation.
1353 * New transactions (when CREATE is set) will insert the state.
1355 * May request that caller discard the message by setting *discardp to 1.
1356 * A NULL state may be returned in this case.
1358 static
1360 kdmsg_state_msgtx(kdmsg_msg_t *msg)
1362 kdmsg_iocom_t *iocom = msg->iocom;
1363 kdmsg_state_t *state;
1364 int error;
1367 * Make sure a state structure is ready to go in case we need a new
1368 * one. This is the only routine which uses freewr_state so no
1369 * races are possible.
1371 if ((state = iocom->freewr_state) == NULL) {
1372 state = kmalloc(sizeof(*state), iocom->mmsg, M_WAITOK | M_ZERO);
1373 state->flags = KDMSG_STATE_DYNAMIC;
1374 state->iocom = iocom;
1375 iocom->freewr_state = state;
1379 * Lock RB tree. If persistent state is present it will have already
1380 * been assigned to msg.
1382 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
1383 state = msg->state;
1386 * Short-cut one-off or mid-stream messages (state may be NULL).
1388 if ((msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE |
1389 DMSGF_ABORT)) == 0) {
1390 lockmgr(&iocom->msglk, LK_RELEASE);
1391 return(0);
1396 * Switch on CREATE, DELETE, REPLY, and also handle ABORT from
1397 * inside the case statements.
1399 switch(msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE |
1400 DMSGF_REPLY)) {
1401 case DMSGF_CREATE:
1402 case DMSGF_CREATE | DMSGF_DELETE:
1404 * Insert the new persistent message state and mark
1405 * half-closed if DELETE is set. Since this is a new
1406 * message it isn't possible to transition into the fully
1407 * closed state here.
1409 * XXX state must be assigned and inserted by
1410 * kdmsg_msg_write(). txcmd is assigned by us
1411 * on-transmit.
1413 KKASSERT(state != NULL);
1414 state->icmd = msg->any.head.cmd & DMSGF_BASECMDMASK;
1415 state->txcmd = msg->any.head.cmd & ~DMSGF_DELETE;
1416 state->rxcmd = DMSGF_REPLY;
1417 error = 0;
1418 break;
1419 case DMSGF_DELETE:
1421 * Sent ABORT+DELETE in case where msgid has already
1422 * been fully closed, ignore the message.
1424 if (state == NULL) {
1425 if (msg->any.head.cmd & DMSGF_ABORT) {
1426 error = EALREADY;
1427 } else {
1428 kprintf("kdmsg_state_msgtx: no state match "
1429 "for DELETE cmd=%08x msgid=%016jx\n",
1430 msg->any.head.cmd,
1431 (intmax_t)msg->any.head.msgid);
1432 error = EINVAL;
1434 break;
1438 * Sent ABORT+DELETE in case where msgid has
1439 * already been reused for an unrelated message,
1440 * ignore the message.
1442 if ((state->txcmd & DMSGF_CREATE) == 0) {
1443 if (msg->any.head.cmd & DMSGF_ABORT) {
1444 error = EALREADY;
1445 } else {
1446 kprintf("kdmsg_state_msgtx: state reused "
1447 "for DELETE\n");
1448 error = EINVAL;
1450 break;
1452 error = 0;
1453 break;
1454 default:
1456 * Check for mid-stream ABORT command sent
1458 if (msg->any.head.cmd & DMSGF_ABORT) {
1459 if (state == NULL ||
1460 (state->txcmd & DMSGF_CREATE) == 0) {
1461 error = EALREADY;
1462 break;
1465 error = 0;
1466 break;
1467 case DMSGF_REPLY | DMSGF_CREATE:
1468 case DMSGF_REPLY | DMSGF_CREATE | DMSGF_DELETE:
1470 * When transmitting a reply with CREATE set the original
1471 * persistent state message should already exist.
1473 if (state == NULL) {
1474 kprintf("kdmsg_state_msgtx: no state match "
1475 "for REPLY | CREATE\n");
1476 error = EINVAL;
1477 break;
1479 state->txcmd = msg->any.head.cmd & ~DMSGF_DELETE;
1480 error = 0;
1481 break;
1482 case DMSGF_REPLY | DMSGF_DELETE:
1484 * When transmitting a reply with DELETE set the original
1485 * persistent state message should already exist.
1487 * This is very similar to the REPLY|CREATE|* case except
1488 * txcmd is already stored, so we just add the DELETE flag.
1490 * Sent REPLY+ABORT+DELETE in case where msgid has
1491 * already been fully closed, ignore the message.
1493 if (state == NULL) {
1494 if (msg->any.head.cmd & DMSGF_ABORT) {
1495 error = EALREADY;
1496 } else {
1497 kprintf("kdmsg_state_msgtx: no state match "
1498 "for REPLY | DELETE\n");
1499 error = EINVAL;
1501 break;
1505 * Sent REPLY+ABORT+DELETE in case where msgid has already
1506 * been reused for an unrelated message, ignore the message.
1508 if ((state->txcmd & DMSGF_CREATE) == 0) {
1509 if (msg->any.head.cmd & DMSGF_ABORT) {
1510 error = EALREADY;
1511 } else {
1512 kprintf("kdmsg_state_msgtx: state reused "
1513 "for REPLY | DELETE\n");
1514 error = EINVAL;
1516 break;
1518 error = 0;
1519 break;
1520 case DMSGF_REPLY:
1522 * Check for mid-stream ABORT reply sent.
1524 * One-off REPLY messages are allowed for e.g. status updates.
1526 if (msg->any.head.cmd & DMSGF_ABORT) {
1527 if (state == NULL ||
1528 (state->txcmd & DMSGF_CREATE) == 0) {
1529 error = EALREADY;
1530 break;
1533 error = 0;
1534 break;
1536 lockmgr(&iocom->msglk, LK_RELEASE);
1537 return (error);
1540 static
1541 void
1542 kdmsg_state_cleanuptx(kdmsg_msg_t *msg)
1544 kdmsg_iocom_t *iocom = msg->iocom;
1545 kdmsg_state_t *state;
1547 if ((state = msg->state) == NULL) {
1548 kdmsg_msg_free(msg);
1549 } else if (msg->any.head.cmd & DMSGF_DELETE) {
1550 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
1551 KKASSERT((state->txcmd & DMSGF_DELETE) == 0);
1552 state->txcmd |= DMSGF_DELETE;
1553 if (state->rxcmd & DMSGF_DELETE) {
1554 KKASSERT(state->flags & KDMSG_STATE_INSERTED);
1555 if (state->txcmd & DMSGF_REPLY) {
1556 KKASSERT(msg->any.head.cmd &
1557 DMSGF_REPLY);
1558 RB_REMOVE(kdmsg_state_tree,
1559 &iocom->staterd_tree, state);
1560 } else {
1561 KKASSERT((msg->any.head.cmd &
1562 DMSGF_REPLY) == 0);
1563 RB_REMOVE(kdmsg_state_tree,
1564 &iocom->statewr_tree, state);
1566 state->flags &= ~KDMSG_STATE_INSERTED;
1567 if (msg != state->msg)
1568 kdmsg_msg_free(msg);
1569 lockmgr(&iocom->msglk, LK_RELEASE);
1570 kdmsg_state_free(state);
1571 } else {
1572 if (msg != state->msg)
1573 kdmsg_msg_free(msg);
1574 lockmgr(&iocom->msglk, LK_RELEASE);
1576 } else if (msg != state->msg) {
1577 kdmsg_msg_free(msg);
1581 static
1582 void
1583 kdmsg_state_free(kdmsg_state_t *state)
1585 kdmsg_iocom_t *iocom = state->iocom;
1586 kdmsg_msg_t *msg;
1588 KKASSERT((state->flags & KDMSG_STATE_INSERTED) == 0);
1589 msg = state->msg;
1590 state->msg = NULL;
1591 kfree(state, iocom->mmsg);
1592 if (msg) {
1593 msg->state = NULL;
1594 kdmsg_msg_free(msg);
1598 kdmsg_msg_t *
1599 kdmsg_msg_alloc(kdmsg_iocom_t *iocom, kdmsg_circuit_t *circ, uint32_t cmd,
1600 int (*func)(kdmsg_state_t *, kdmsg_msg_t *), void *data)
1602 kdmsg_msg_t *msg;
1603 kdmsg_state_t *state;
1604 size_t hbytes;
1606 KKASSERT(iocom != NULL);
1607 hbytes = (cmd & DMSGF_SIZE) * DMSG_ALIGN;
1608 msg = kmalloc(offsetof(struct kdmsg_msg, any) + hbytes,
1609 iocom->mmsg, M_WAITOK | M_ZERO);
1610 msg->hdr_size = hbytes;
1611 msg->iocom = iocom;
1612 msg->any.head.magic = DMSG_HDR_MAGIC;
1613 msg->any.head.cmd = cmd;
1614 if (circ) {
1615 kdmsg_circ_hold(circ);
1616 msg->circ = circ;
1617 msg->any.head.circuit = circ->msgid;
1620 if (cmd & DMSGF_CREATE) {
1622 * New transaction, requires tracking state and a unique
1623 * msgid to be allocated.
1625 KKASSERT(msg->state == NULL);
1626 state = kmalloc(sizeof(*state), iocom->mmsg, M_WAITOK | M_ZERO);
1627 state->flags = KDMSG_STATE_DYNAMIC;
1628 state->func = func;
1629 state->any.any = data;
1630 state->msg = msg;
1631 state->msgid = (uint64_t)(uintptr_t)state;
1632 state->circ = circ;
1633 state->iocom = iocom;
1634 msg->state = state;
1635 if (circ)
1636 kdmsg_circ_hold(circ);
1637 /*msg->any.head.msgid = state->msgid;XXX*/
1639 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
1640 if (RB_INSERT(kdmsg_state_tree, &iocom->statewr_tree, state))
1641 panic("duplicate msgid allocated");
1642 state->flags |= KDMSG_STATE_INSERTED;
1643 msg->any.head.msgid = state->msgid;
1644 lockmgr(&iocom->msglk, LK_RELEASE);
1646 return (msg);
1649 kdmsg_msg_t *
1650 kdmsg_msg_alloc_state(kdmsg_state_t *state, uint32_t cmd,
1651 int (*func)(kdmsg_state_t *, kdmsg_msg_t *), void *data)
1653 kdmsg_iocom_t *iocom = state->iocom;
1654 kdmsg_msg_t *msg;
1655 size_t hbytes;
1657 KKASSERT(iocom != NULL);
1658 hbytes = (cmd & DMSGF_SIZE) * DMSG_ALIGN;
1659 msg = kmalloc(offsetof(struct kdmsg_msg, any) + hbytes,
1660 iocom->mmsg, M_WAITOK | M_ZERO);
1661 msg->hdr_size = hbytes;
1662 msg->iocom = iocom;
1663 msg->any.head.magic = DMSG_HDR_MAGIC;
1664 msg->any.head.cmd = cmd;
1665 msg->state = state;
1666 if (state->circ) {
1667 kdmsg_circ_hold(state->circ);
1668 msg->circ = state->circ;
1669 msg->any.head.circuit = state->circ->msgid;
1671 return(msg);
1674 void
1675 kdmsg_msg_free(kdmsg_msg_t *msg)
1677 kdmsg_iocom_t *iocom = msg->iocom;
1679 if ((msg->flags & KDMSG_FLAG_AUXALLOC) &&
1680 msg->aux_data && msg->aux_size) {
1681 kfree(msg->aux_data, iocom->mmsg);
1682 msg->flags &= ~KDMSG_FLAG_AUXALLOC;
1684 if (msg->circ) {
1685 kdmsg_circ_drop(msg->circ);
1686 msg->circ = NULL;
1688 if (msg->state) {
1689 if (msg->state->msg == msg)
1690 msg->state->msg = NULL;
1691 msg->state = NULL;
1693 msg->aux_data = NULL;
1694 msg->aux_size = 0;
1695 msg->iocom = NULL;
1696 kfree(msg, iocom->mmsg);
1700 * Circuits are tracked in a red-black tree by their circuit id (msgid).
1703 kdmsg_circuit_cmp(kdmsg_circuit_t *circ1, kdmsg_circuit_t *circ2)
1705 if (circ1->msgid < circ2->msgid)
1706 return(-1);
1707 if (circ1->msgid > circ2->msgid)
1708 return(1);
1709 return (0);
1713 * Indexed messages are stored in a red-black tree indexed by their
1714 * msgid. Only persistent messages are indexed.
1717 kdmsg_state_cmp(kdmsg_state_t *state1, kdmsg_state_t *state2)
1719 if (state1->iocom < state2->iocom)
1720 return(-1);
1721 if (state1->iocom > state2->iocom)
1722 return(1);
1723 if (state1->circ < state2->circ)
1724 return(-1);
1725 if (state1->circ > state2->circ)
1726 return(1);
1727 if (state1->msgid < state2->msgid)
1728 return(-1);
1729 if (state1->msgid > state2->msgid)
1730 return(1);
1731 return(0);
1735 * Write a message. All requisit command flags have been set.
1737 * If msg->state is non-NULL the message is written to the existing
1738 * transaction. msgid will be set accordingly.
1740 * If msg->state is NULL and CREATE is set new state is allocated and
1741 * (func, data) is installed. A msgid is assigned.
1743 * If msg->state is NULL and CREATE is not set the message is assumed
1744 * to be a one-way message. The originator must assign the msgid
1745 * (or leave it 0, which is typical.
1747 * This function merely queues the message to the management thread, it
1748 * does not write to the message socket/pipe.
1750 void
1751 kdmsg_msg_write(kdmsg_msg_t *msg)
1753 kdmsg_iocom_t *iocom = msg->iocom;
1754 kdmsg_state_t *state;
1756 if (msg->state) {
1758 * Continuance or termination of existing transaction.
1759 * The transaction could have been initiated by either end.
1761 * (Function callback and aux data for the receive side can
1762 * be replaced or left alone).
1764 state = msg->state;
1765 msg->any.head.msgid = state->msgid;
1766 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
1767 } else {
1769 * One-off message (always uses msgid 0 to distinguish
1770 * between a possibly lost in-transaction message due to
1771 * competing aborts and a real one-off message?)
1773 state = NULL;
1774 msg->any.head.msgid = 0;
1775 lockmgr(&iocom->msglk, LK_EXCLUSIVE);
1779 * With AUTOCIRC and AUTOFORGE it is possible for the circuit to
1780 * get ripped out in the rxthread while some other thread is
1781 * holding a ref on it inbetween allocating and sending a dmsg.
1783 if (msg->circ && msg->circ->rcirc_state == NULL &&
1784 (msg->circ->span_state == NULL || msg->circ->circ_state == NULL)) {
1785 kprintf("kdmsg_msg_write: Attempt to write message to "
1786 "terminated circuit: msg %08x\n", msg->any.head.cmd);
1787 lockmgr(&iocom->msglk, LK_RELEASE);
1788 if (kdmsg_state_msgtx(msg)) {
1789 if (state == NULL || msg != state->msg)
1790 kdmsg_msg_free(msg);
1791 } else if ((msg->state->rxcmd & DMSGF_DELETE) == 0) {
1792 /* XXX SMP races simulating a response here */
1793 kdmsg_state_t *state = msg->state;
1794 kdmsg_state_cleanuptx(msg);
1795 kdmsg_state_abort(state);
1796 } else {
1797 kdmsg_state_cleanuptx(msg);
1799 return;
1803 * This flag is not set until after the tx thread has drained
1804 * the txmsgq and simulated responses. After that point the
1805 * txthread is dead and can no longer simulate responses.
1807 * Device drivers should never try to send a message once this
1808 * flag is set. They should have detected (through the state
1809 * closures) that the link is in trouble.
1811 if (iocom->flags & KDMSG_IOCOMF_EXITNOACC) {
1812 lockmgr(&iocom->msglk, LK_RELEASE);
1813 panic("kdmsg_msg_write: Attempt to write message to "
1814 "terminated iocom\n");
1818 * Finish up the msg fields. Note that msg->aux_size and the
1819 * aux_bytes stored in the message header represent the unaligned
1820 * (actual) bytes of data, but the buffer is sized to an aligned
1821 * size and the CRC is generated over the aligned length.
1823 msg->any.head.salt = /* (random << 8) | */ (iocom->msg_seq & 255);
1824 ++iocom->msg_seq;
1826 if (msg->aux_data && msg->aux_size) {
1827 uint32_t abytes = DMSG_DOALIGN(msg->aux_size);
1829 msg->any.head.aux_bytes = msg->aux_size;
1830 msg->any.head.aux_crc = iscsi_crc32(msg->aux_data, abytes);
1832 msg->any.head.hdr_crc = 0;
1833 msg->any.head.hdr_crc = iscsi_crc32(msg->any.buf, msg->hdr_size);
1835 TAILQ_INSERT_TAIL(&iocom->msgq, msg, qentry);
1837 if (iocom->msg_ctl & KDMSG_CLUSTERCTL_SLEEPING) {
1838 atomic_clear_int(&iocom->msg_ctl,
1839 KDMSG_CLUSTERCTL_SLEEPING);
1840 wakeup(&iocom->msg_ctl);
1843 lockmgr(&iocom->msglk, LK_RELEASE);
1847 * Reply to a message and terminate our side of the transaction.
1849 * If msg->state is non-NULL we are replying to a one-way message.
1851 void
1852 kdmsg_msg_reply(kdmsg_msg_t *msg, uint32_t error)
1854 kdmsg_state_t *state = msg->state;
1855 kdmsg_msg_t *nmsg;
1856 uint32_t cmd;
1859 * Reply with a simple error code and terminate the transaction.
1861 cmd = DMSG_LNK_ERROR;
1864 * Check if our direction has even been initiated yet, set CREATE.
1866 * Check what direction this is (command or reply direction). Note
1867 * that txcmd might not have been initiated yet.
1869 * If our direction has already been closed we just return without
1870 * doing anything.
1872 if (state) {
1873 if (state->txcmd & DMSGF_DELETE)
1874 return;
1875 if ((state->txcmd & DMSGF_CREATE) == 0)
1876 cmd |= DMSGF_CREATE;
1877 if (state->txcmd & DMSGF_REPLY)
1878 cmd |= DMSGF_REPLY;
1879 cmd |= DMSGF_DELETE;
1880 } else {
1881 if ((msg->any.head.cmd & DMSGF_REPLY) == 0)
1882 cmd |= DMSGF_REPLY;
1885 /* XXX messy mask cmd to avoid allocating state */
1886 nmsg = kdmsg_msg_alloc_state(state, cmd, NULL, NULL);
1887 nmsg->any.head.error = error;
1888 kdmsg_msg_write(nmsg);
1892 * Reply to a message and continue our side of the transaction.
1894 * If msg->state is non-NULL we are replying to a one-way message and this
1895 * function degenerates into the same as kdmsg_msg_reply().
1897 void
1898 kdmsg_msg_result(kdmsg_msg_t *msg, uint32_t error)
1900 kdmsg_state_t *state = msg->state;
1901 kdmsg_msg_t *nmsg;
1902 uint32_t cmd;
1905 * Return a simple result code, do NOT terminate the transaction.
1907 cmd = DMSG_LNK_ERROR;
1910 * Check if our direction has even been initiated yet, set CREATE.
1912 * Check what direction this is (command or reply direction). Note
1913 * that txcmd might not have been initiated yet.
1915 * If our direction has already been closed we just return without
1916 * doing anything.
1918 if (state) {
1919 if (state->txcmd & DMSGF_DELETE)
1920 return;
1921 if ((state->txcmd & DMSGF_CREATE) == 0)
1922 cmd |= DMSGF_CREATE;
1923 if (state->txcmd & DMSGF_REPLY)
1924 cmd |= DMSGF_REPLY;
1925 /* continuing transaction, do not set MSGF_DELETE */
1926 } else {
1927 if ((msg->any.head.cmd & DMSGF_REPLY) == 0)
1928 cmd |= DMSGF_REPLY;
1931 /* XXX messy mask cmd to avoid allocating state */
1932 nmsg = kdmsg_msg_alloc_state(state, cmd, NULL, NULL);
1933 nmsg->any.head.error = error;
1934 kdmsg_msg_write(nmsg);
1938 * Reply to a message and terminate our side of the transaction.
1940 * If msg->state is non-NULL we are replying to a one-way message.
1942 void
1943 kdmsg_state_reply(kdmsg_state_t *state, uint32_t error)
1945 kdmsg_msg_t *nmsg;
1946 uint32_t cmd;
1949 * Reply with a simple error code and terminate the transaction.
1951 cmd = DMSG_LNK_ERROR;
1954 * Check if our direction has even been initiated yet, set CREATE.
1956 * Check what direction this is (command or reply direction). Note
1957 * that txcmd might not have been initiated yet.
1959 * If our direction has already been closed we just return without
1960 * doing anything.
1962 if (state) {
1963 if (state->txcmd & DMSGF_DELETE)
1964 return;
1965 if ((state->txcmd & DMSGF_CREATE) == 0)
1966 cmd |= DMSGF_CREATE;
1967 if (state->txcmd & DMSGF_REPLY)
1968 cmd |= DMSGF_REPLY;
1969 cmd |= DMSGF_DELETE;
1970 } else {
1971 if ((state->txcmd & DMSGF_REPLY) == 0)
1972 cmd |= DMSGF_REPLY;
1975 /* XXX messy mask cmd to avoid allocating state */
1976 nmsg = kdmsg_msg_alloc_state(state, cmd, NULL, NULL);
1977 nmsg->any.head.error = error;
1978 kdmsg_msg_write(nmsg);
1982 * Reply to a message and continue our side of the transaction.
1984 * If msg->state is non-NULL we are replying to a one-way message and this
1985 * function degenerates into the same as kdmsg_msg_reply().
1987 void
1988 kdmsg_state_result(kdmsg_state_t *state, uint32_t error)
1990 kdmsg_msg_t *nmsg;
1991 uint32_t cmd;
1994 * Return a simple result code, do NOT terminate the transaction.
1996 cmd = DMSG_LNK_ERROR;
1999 * Check if our direction has even been initiated yet, set CREATE.
2001 * Check what direction this is (command or reply direction). Note
2002 * that txcmd might not have been initiated yet.
2004 * If our direction has already been closed we just return without
2005 * doing anything.
2007 if (state) {
2008 if (state->txcmd & DMSGF_DELETE)
2009 return;
2010 if ((state->txcmd & DMSGF_CREATE) == 0)
2011 cmd |= DMSGF_CREATE;
2012 if (state->txcmd & DMSGF_REPLY)
2013 cmd |= DMSGF_REPLY;
2014 /* continuing transaction, do not set MSGF_DELETE */
2015 } else {
2016 if ((state->txcmd & DMSGF_REPLY) == 0)
2017 cmd |= DMSGF_REPLY;
2020 /* XXX messy mask cmd to avoid allocating state */
2021 nmsg = kdmsg_msg_alloc_state(state, cmd, NULL, NULL);
2022 nmsg->any.head.error = error;
2023 kdmsg_msg_write(nmsg);