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
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
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
37 RB_GENERATE(hammer2_state_tree, hammer2_state, rbnode, hammer2_state_cmp);
40 * Process state tracking for a message after reception, prior to
43 * Called with msglk held and the msg dequeued.
45 * All messages are called with dummy state and return actual state.
46 * (One-off messages often just return the same dummy state).
48 * May request that caller discard the message by setting *discardp to 1.
49 * The returned state is not used in this case and is allowed to be NULL.
53 * These routines handle persistent and command/reply message state via the
54 * CREATE and DELETE flags. The first message in a command or reply sequence
55 * sets CREATE, the last message in a command or reply sequence sets DELETE.
57 * There can be any number of intermediate messages belonging to the same
58 * sequence sent inbetween the CREATE message and the DELETE message,
59 * which set neither flag. This represents a streaming command or reply.
61 * Any command message received with CREATE set expects a reply sequence to
62 * be returned. Reply sequences work the same as command sequences except the
63 * REPLY bit is also sent. Both the command side and reply side can
64 * degenerate into a single message with both CREATE and DELETE set. Note
65 * that one side can be streaming and the other side not, or neither, or both.
67 * The msgid is unique for the initiator. That is, two sides sending a new
68 * message can use the same msgid without colliding.
72 * ABORT sequences work by setting the ABORT flag along with normal message
73 * state. However, ABORTs can also be sent on half-closed messages, that is
74 * even if the command or reply side has already sent a DELETE, as long as
75 * the message has not been fully closed it can still send an ABORT+DELETE
76 * to terminate the half-closed message state.
78 * Since ABORT+DELETEs can race we silently discard ABORT's for message
79 * state which has already been fully closed. REPLY+ABORT+DELETEs can
80 * also race, and in this situation the other side might have already
81 * initiated a new unrelated command with the same message id. Since
82 * the abort has not set the CREATE flag the situation can be detected
83 * and the message will also be discarded.
85 * Non-blocking requests can be initiated with ABORT+CREATE[+DELETE].
86 * The ABORT request is essentially integrated into the command instead
87 * of being sent later on. In this situation the command implementation
88 * detects that CREATE and ABORT are both set (vs ABORT alone) and can
89 * special-case non-blocking operation for the command.
91 * NOTE! Messages with ABORT set without CREATE or DELETE are considered
92 * to be mid-stream aborts for command/reply sequences. ABORTs on
93 * one-way messages are not supported.
95 * NOTE! If a command sequence does not support aborts the ABORT flag is
100 * One-off messages (no reply expected) are sent with neither CREATE or DELETE
101 * set. One-off messages cannot be aborted and typically aren't processed
102 * by these routines. The REPLY bit can be used to distinguish whether a
103 * one-off message is a command or reply. For example, one-off replies
104 * will typically just contain status updates.
107 hammer2_state_msgrx(hammer2_pfsmount_t *pmp, hammer2_msg_t *msg)
109 hammer2_state_t *state;
113 * Make sure a state structure is ready to go in case we need a new
114 * one. This is the only routine which uses freerd_state so no
115 * races are possible.
117 if ((state = pmp->freerd_state) == NULL) {
118 state = kmalloc(sizeof(*state), pmp->mmsg, M_WAITOK | M_ZERO);
120 state->flags = HAMMER2_STATE_DYNAMIC;
121 pmp->freerd_state = state;
125 * Lock RB tree and locate existing persistent state, if any.
127 * If received msg is a command state is on staterd_tree.
128 * If received msg is a reply state is on statewr_tree.
130 lockmgr(&pmp->msglk, LK_EXCLUSIVE);
132 state->msgid = msg->any.head.msgid;
133 state->spanid = msg->any.head.spanid;
134 kprintf("received msg %08x msgid %jx spanid=%jx\n",
136 (intmax_t)msg->any.head.msgid, (intmax_t)msg->any.head.spanid);
137 if (msg->any.head.cmd & HAMMER2_MSGF_REPLY)
138 state = RB_FIND(hammer2_state_tree, &pmp->statewr_tree, state);
140 state = RB_FIND(hammer2_state_tree, &pmp->staterd_tree, state);
144 * Short-cut one-off or mid-stream messages (state may be NULL).
146 if ((msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
147 HAMMER2_MSGF_ABORT)) == 0) {
148 lockmgr(&pmp->msglk, LK_RELEASE);
153 * Switch on CREATE, DELETE, REPLY, and also handle ABORT from
154 * inside the case statements.
156 switch(msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
157 HAMMER2_MSGF_REPLY)) {
158 case HAMMER2_MSGF_CREATE:
159 case HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
161 * New persistant command received.
164 kprintf("hammer2_state_msgrx: duplicate transaction\n");
168 state = pmp->freerd_state;
169 pmp->freerd_state = NULL;
172 state->rxcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
173 state->txcmd = HAMMER2_MSGF_REPLY;
174 RB_INSERT(hammer2_state_tree, &pmp->staterd_tree, state);
175 state->flags |= HAMMER2_STATE_INSERTED;
178 case HAMMER2_MSGF_DELETE:
180 * Persistent state is expected but might not exist if an
181 * ABORT+DELETE races the close.
184 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
187 kprintf("hammer2_state_msgrx: no state "
195 * Handle another ABORT+DELETE case if the msgid has already
198 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
199 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
202 kprintf("hammer2_state_msgrx: state reused "
212 * Check for mid-stream ABORT command received, otherwise
215 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
217 (state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
224 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE:
225 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
227 * When receiving a reply with CREATE set the original
228 * persistent state message should already exist.
231 kprintf("hammer2_state_msgrx: no state match for "
232 "REPLY cmd=%08x msgid=%016jx\n",
234 (intmax_t)msg->any.head.msgid);
238 state->rxcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
241 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_DELETE:
243 * Received REPLY+ABORT+DELETE in case where msgid has
244 * already been fully closed, ignore the message.
247 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
250 kprintf("hammer2_state_msgrx: no state match "
251 "for REPLY|DELETE\n");
258 * Received REPLY+ABORT+DELETE in case where msgid has
259 * already been reused for an unrelated message,
260 * ignore the message.
262 if ((state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
263 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
266 kprintf("hammer2_state_msgrx: state reused "
267 "for REPLY|DELETE\n");
274 case HAMMER2_MSGF_REPLY:
276 * Check for mid-stream ABORT reply received to sent command.
278 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
280 (state->rxcmd & HAMMER2_MSGF_CREATE) == 0) {
288 lockmgr(&pmp->msglk, LK_RELEASE);
293 hammer2_state_cleanuprx(hammer2_pfsmount_t *pmp, hammer2_msg_t *msg)
295 hammer2_state_t *state;
297 if ((state = msg->state) == NULL) {
298 hammer2_msg_free(pmp, msg);
299 } else if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
300 lockmgr(&pmp->msglk, LK_EXCLUSIVE);
301 state->rxcmd |= HAMMER2_MSGF_DELETE;
302 if (state->txcmd & HAMMER2_MSGF_DELETE) {
303 if (state->msg == msg)
305 KKASSERT(state->flags & HAMMER2_STATE_INSERTED);
306 if (state->rxcmd & HAMMER2_MSGF_REPLY) {
307 KKASSERT(msg->any.head.cmd &
309 RB_REMOVE(hammer2_state_tree,
310 &pmp->statewr_tree, state);
312 KKASSERT((msg->any.head.cmd &
313 HAMMER2_MSGF_REPLY) == 0);
314 RB_REMOVE(hammer2_state_tree,
315 &pmp->staterd_tree, state);
317 state->flags &= ~HAMMER2_STATE_INSERTED;
318 lockmgr(&pmp->msglk, LK_RELEASE);
319 hammer2_state_free(state);
321 lockmgr(&pmp->msglk, LK_RELEASE);
323 hammer2_msg_free(pmp, msg);
324 } else if (state->msg != msg) {
325 hammer2_msg_free(pmp, msg);
330 * Process state tracking for a message prior to transmission.
332 * Called with msglk held and the msg dequeued.
334 * One-off messages are usually with dummy state and msg->state may be NULL
337 * New transactions (when CREATE is set) will insert the state.
339 * May request that caller discard the message by setting *discardp to 1.
340 * A NULL state may be returned in this case.
343 hammer2_state_msgtx(hammer2_pfsmount_t *pmp, hammer2_msg_t *msg)
345 hammer2_state_t *state;
349 * Make sure a state structure is ready to go in case we need a new
350 * one. This is the only routine which uses freewr_state so no
351 * races are possible.
353 if ((state = pmp->freewr_state) == NULL) {
354 state = kmalloc(sizeof(*state), pmp->mmsg, M_WAITOK | M_ZERO);
356 state->flags = HAMMER2_STATE_DYNAMIC;
357 pmp->freewr_state = state;
361 * Lock RB tree. If persistent state is present it will have already
362 * been assigned to msg.
364 lockmgr(&pmp->msglk, LK_EXCLUSIVE);
368 * Short-cut one-off or mid-stream messages (state may be NULL).
370 if ((msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
371 HAMMER2_MSGF_ABORT)) == 0) {
372 lockmgr(&pmp->msglk, LK_RELEASE);
378 * Switch on CREATE, DELETE, REPLY, and also handle ABORT from
379 * inside the case statements.
381 switch(msg->any.head.cmd & (HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE |
382 HAMMER2_MSGF_REPLY)) {
383 case HAMMER2_MSGF_CREATE:
384 case HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
386 * Insert the new persistent message state and mark
387 * half-closed if DELETE is set. Since this is a new
388 * message it isn't possible to transition into the fully
391 * XXX state must be assigned and inserted by
392 * hammer2_msg_write(). txcmd is assigned by us
395 KKASSERT(state != NULL);
396 state->txcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
397 state->rxcmd = HAMMER2_MSGF_REPLY;
400 case HAMMER2_MSGF_DELETE:
402 * Sent ABORT+DELETE in case where msgid has already
403 * been fully closed, ignore the message.
406 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
409 kprintf("hammer2_state_msgtx: no state match "
410 "for DELETE cmd=%08x msgid=%016jx\n",
412 (intmax_t)msg->any.head.msgid);
419 * Sent ABORT+DELETE in case where msgid has
420 * already been reused for an unrelated message,
421 * ignore the message.
423 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0) {
424 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
427 kprintf("hammer2_state_msgtx: state reused "
437 * Check for mid-stream ABORT command sent
439 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
441 (state->txcmd & HAMMER2_MSGF_CREATE) == 0) {
448 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE:
449 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_CREATE | HAMMER2_MSGF_DELETE:
451 * When transmitting a reply with CREATE set the original
452 * persistent state message should already exist.
455 kprintf("hammer2_state_msgtx: no state match "
456 "for REPLY | CREATE\n");
460 state->txcmd = msg->any.head.cmd & ~HAMMER2_MSGF_DELETE;
463 case HAMMER2_MSGF_REPLY | HAMMER2_MSGF_DELETE:
465 * When transmitting a reply with DELETE set the original
466 * persistent state message should already exist.
468 * This is very similar to the REPLY|CREATE|* case except
469 * txcmd is already stored, so we just add the DELETE flag.
471 * Sent REPLY+ABORT+DELETE in case where msgid has
472 * already been fully closed, ignore the message.
475 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
478 kprintf("hammer2_state_msgtx: no state match "
479 "for REPLY | DELETE\n");
486 * Sent REPLY+ABORT+DELETE in case where msgid has already
487 * been reused for an unrelated message, ignore the message.
489 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0) {
490 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
493 kprintf("hammer2_state_msgtx: state reused "
494 "for REPLY | DELETE\n");
501 case HAMMER2_MSGF_REPLY:
503 * Check for mid-stream ABORT reply sent.
505 * One-off REPLY messages are allowed for e.g. status updates.
507 if (msg->any.head.cmd & HAMMER2_MSGF_ABORT) {
509 (state->txcmd & HAMMER2_MSGF_CREATE) == 0) {
517 lockmgr(&pmp->msglk, LK_RELEASE);
522 hammer2_state_cleanuptx(hammer2_pfsmount_t *pmp, hammer2_msg_t *msg)
524 hammer2_state_t *state;
526 if ((state = msg->state) == NULL) {
527 hammer2_msg_free(pmp, msg);
528 } else if (msg->any.head.cmd & HAMMER2_MSGF_DELETE) {
529 lockmgr(&pmp->msglk, LK_EXCLUSIVE);
530 state->txcmd |= HAMMER2_MSGF_DELETE;
531 if (state->rxcmd & HAMMER2_MSGF_DELETE) {
532 if (state->msg == msg)
534 KKASSERT(state->flags & HAMMER2_STATE_INSERTED);
535 if (state->txcmd & HAMMER2_MSGF_REPLY) {
536 KKASSERT(msg->any.head.cmd &
538 RB_REMOVE(hammer2_state_tree,
539 &pmp->staterd_tree, state);
541 KKASSERT((msg->any.head.cmd &
542 HAMMER2_MSGF_REPLY) == 0);
543 RB_REMOVE(hammer2_state_tree,
544 &pmp->statewr_tree, state);
546 state->flags &= ~HAMMER2_STATE_INSERTED;
547 lockmgr(&pmp->msglk, LK_RELEASE);
548 hammer2_state_free(state);
550 lockmgr(&pmp->msglk, LK_RELEASE);
552 hammer2_msg_free(pmp, msg);
553 } else if (state->msg != msg) {
554 hammer2_msg_free(pmp, msg);
559 hammer2_state_free(hammer2_state_t *state)
561 hammer2_pfsmount_t *pmp = state->pmp;
566 kfree(state, pmp->mmsg);
568 hammer2_msg_free(pmp, msg);
572 hammer2_msg_alloc(hammer2_pfsmount_t *pmp, uint64_t spanid, uint32_t cmd)
577 hbytes = (cmd & HAMMER2_MSGF_SIZE) * HAMMER2_MSG_ALIGN;
578 msg = kmalloc(offsetof(struct hammer2_msg, any) + hbytes,
579 pmp->mmsg, M_WAITOK | M_ZERO);
580 msg->hdr_size = hbytes;
581 msg->any.head.magic = HAMMER2_MSGHDR_MAGIC;
582 msg->any.head.spanid = spanid;
583 msg->any.head.cmd = cmd;
589 hammer2_msg_free(hammer2_pfsmount_t *pmp, hammer2_msg_t *msg)
591 if (msg->aux_data && msg->aux_size) {
592 kfree(msg->aux_data, pmp->mmsg);
593 msg->aux_data = NULL;
596 kfree(msg, pmp->mmsg);
600 * Indexed messages are stored in a red-black tree indexed by their
601 * msgid. Only persistent messages are indexed.
604 hammer2_state_cmp(hammer2_state_t *state1, hammer2_state_t *state2)
606 if (state1->spanid < state2->spanid)
608 if (state1->spanid > state2->spanid)
610 if (state1->msgid < state2->msgid)
612 if (state1->msgid > state2->msgid)
618 * Write a message. All requisit command flags have been set.
620 * If msg->state is non-NULL the message is written to the existing
621 * transaction. Both msgid and spanid will be set accordingly.
623 * If msg->state is NULL and CREATE is set new state is allocated and
624 * (func, data) is installed.
626 * If msg->state is NULL and CREATE is not set the message is assumed
627 * to be a one-way message. The msgid and spanid must be set by the
628 * caller (msgid is typically set to 0 for this case).
630 * This function merely queues the message to the management thread, it
631 * does not write to the message socket/pipe.
634 hammer2_msg_write(hammer2_pfsmount_t *pmp, hammer2_msg_t *msg,
635 int (*func)(hammer2_state_t *, hammer2_msg_t *), void *data)
637 hammer2_state_t *state;
641 * Continuance or termination of existing transaction.
642 * The transaction could have been initiated by either end.
644 * (Function callback and aux data for the receive side can
645 * be replaced or left alone).
647 msg->any.head.msgid = msg->state->msgid;
648 msg->any.head.spanid = msg->state->spanid;
649 lockmgr(&pmp->msglk, LK_EXCLUSIVE);
651 msg->state->func = func;
652 msg->state->any.any = data;
654 } else if (msg->any.head.cmd & HAMMER2_MSGF_CREATE) {
656 * New transaction, requires tracking state and a unique
657 * msgid to be allocated.
659 KKASSERT(msg->state == NULL);
660 state = kmalloc(sizeof(*state), pmp->mmsg, M_WAITOK | M_ZERO);
662 state->flags = HAMMER2_STATE_DYNAMIC;
664 state->any.any = data;
666 state->msgid = (uint64_t)(uintptr_t)state;
667 state->spanid = msg->any.head.spanid;
670 lockmgr(&pmp->msglk, LK_EXCLUSIVE);
671 if (RB_INSERT(hammer2_state_tree, &pmp->statewr_tree, state))
672 panic("duplicate msgid allocated");
673 msg->any.head.msgid = state->msgid;
676 * One-off message (always uses msgid 0 to distinguish
677 * between a possibly lost in-transaction message due to
678 * competing aborts and a real one-off message?)
680 msg->any.head.msgid = 0;
681 lockmgr(&pmp->msglk, LK_EXCLUSIVE);
685 * Finish up the msg fields
687 msg->any.head.salt = /* (random << 8) | */ (pmp->msg_seq & 255);
690 msg->any.head.hdr_crc = 0;
691 msg->any.head.hdr_crc = hammer2_icrc32(msg->any.buf, msg->hdr_size);
693 TAILQ_INSERT_TAIL(&pmp->msgq, msg, qentry);
694 lockmgr(&pmp->msglk, LK_RELEASE);
698 * Reply to a message and terminate our side of the transaction.
700 * If msg->state is non-NULL we are replying to a one-way message.
703 hammer2_msg_reply(hammer2_pfsmount_t *pmp, hammer2_msg_t *msg, uint32_t error)
705 hammer2_state_t *state = msg->state;
709 * Reply with a simple error code and terminate the transaction.
711 cmd = HAMMER2_LNK_ERROR;
714 * Check if our direction has even been initiated yet, set CREATE.
716 * Check what direction this is (command or reply direction). Note
717 * that txcmd might not have been initiated yet.
719 * If our direction has already been closed we just return without
723 if (state->txcmd & HAMMER2_MSGF_DELETE) {
724 hammer2_msg_free(pmp, msg);
727 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
728 cmd |= HAMMER2_MSGF_CREATE;
729 if (state->txcmd & HAMMER2_MSGF_REPLY)
730 cmd |= HAMMER2_MSGF_REPLY;
731 cmd |= HAMMER2_MSGF_DELETE;
733 if ((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0)
734 cmd |= HAMMER2_MSGF_REPLY;
737 msg = hammer2_msg_alloc(pmp, 0, cmd);
738 msg->any.head.error = error;
740 hammer2_msg_write(pmp, msg, NULL, NULL);
744 * Reply to a message and continue our side of the transaction.
746 * If msg->state is non-NULL we are replying to a one-way message and this
747 * function degenerates into the same as hammer2_msg_reply().
750 hammer2_msg_result(hammer2_pfsmount_t *pmp, hammer2_msg_t *msg, uint32_t error)
752 hammer2_state_t *state = msg->state;
756 * Return a simple result code, do NOT terminate the transaction.
758 cmd = HAMMER2_LNK_ERROR;
761 * Check if our direction has even been initiated yet, set CREATE.
763 * Check what direction this is (command or reply direction). Note
764 * that txcmd might not have been initiated yet.
766 * If our direction has already been closed we just return without
770 if (state->txcmd & HAMMER2_MSGF_DELETE) {
771 hammer2_msg_free(pmp, msg);
774 if ((state->txcmd & HAMMER2_MSGF_CREATE) == 0)
775 cmd |= HAMMER2_MSGF_CREATE;
776 if (state->txcmd & HAMMER2_MSGF_REPLY)
777 cmd |= HAMMER2_MSGF_REPLY;
778 /* continuing transaction, do not set MSGF_DELETE */
780 if ((msg->any.head.cmd & HAMMER2_MSGF_REPLY) == 0)
781 cmd |= HAMMER2_MSGF_REPLY;
784 msg = hammer2_msg_alloc(pmp, 0, cmd);
785 msg->any.head.error = error;
787 hammer2_msg_write(pmp, msg, NULL, NULL);