2 * Copyright (c) 2003,2004 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
34 * NOTE! This file may be compiled for userland libraries as well as for
37 * $DragonFly: src/sys/kern/lwkt_msgport.c,v 1.30 2004/09/10 18:23:55 dillon Exp $
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/rtprio.h>
47 #include <sys/queue.h>
48 #include <sys/sysctl.h>
49 #include <sys/kthread.h>
50 #include <sys/signalvar.h>
51 #include <machine/cpu.h>
55 #include <vm/vm_param.h>
56 #include <vm/vm_kern.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_map.h>
60 #include <vm/vm_pager.h>
61 #include <vm/vm_extern.h>
62 #include <vm/vm_zone.h>
64 #include <sys/thread2.h>
65 #include <sys/msgport2.h>
67 #include <machine/stdarg.h>
68 #include <machine/ipl.h>
69 #include <machine/cpufunc.h>
71 #include <machine/smp.h>
74 #include <sys/malloc.h>
75 MALLOC_DEFINE(M_LWKTMSG, "lwkt message", "lwkt message");
79 #include <sys/stdint.h>
80 #include <libcaps/thread.h>
81 #include <sys/thread.h>
82 #include <sys/msgport.h>
83 #include <sys/errno.h>
84 #include <libcaps/globaldata.h>
85 #include <machine/cpufunc.h>
86 #include <sys/thread2.h>
87 #include <sys/msgport2.h>
93 /************************************************************************
95 ************************************************************************/
97 static void lwkt_replyport_remote(lwkt_msg_t msg);
98 static void lwkt_putport_remote(lwkt_msg_t msg);
103 * Send a message asynchronously. This function requests asynchronous
104 * completion and calls lwkt_beginmsg(). If the target port decides to
105 * run the message synchronously this function will automatically queue
106 * the message to the current thread's message queue to present a
107 * consistent interface to the caller.
109 * The message's ms_cmd must be initialized and its ms_flags must
110 * be zero'd out. lwkt_sendmsg() will initialize the ms_abort_port
111 * (abort chasing port). If abort is supported, ms_abort must also be
114 * NOTE: you cannot safely request an abort until lwkt_sendmsg() returns
117 * NOTE: MSGF_DONE is left set. The target port must clear it if the
118 * message is to be handled asynchronously, while the synchronous case
119 * can just ignore it.
122 lwkt_sendmsg(lwkt_port_t port, lwkt_msg_t msg)
126 msg->ms_flags |= MSGF_ASYNC;
127 msg->ms_flags &= ~(MSGF_REPLY1 | MSGF_REPLY2 | MSGF_QUEUED | \
128 MSGF_ABORTED | MSGF_RETRIEVED);
129 KKASSERT(msg->ms_reply_port != NULL);
130 msg->ms_abort_port = msg->ms_reply_port;
131 if ((error = lwkt_beginmsg(port, msg)) != EASYNC) {
132 lwkt_replymsg(msg, error);
139 * Send a message synchronously. This function requests synchronous
140 * completion and calls lwkt_beginmsg(). If the target port decides to
141 * run the message asynchronously this function will block waiting for
142 * the message to complete. Since MSGF_ASYNC is not set the target
143 * will not attempt to queue the reply to a reply port but will simply
144 * wake up anyone waiting on the message.
146 * A synchronous error code is always returned.
148 * The message's ms_cmd must be initialized, and its ms_flags must be
149 * at least zero'd out. lwkt_domsg() will initialize the message's
150 * ms_abort_port (abort chasing port). If abort is supported, ms_abort
151 * must also be initialized.
153 * NOTE: you cannot safely request an abort until lwkt_domsg() blocks.
154 * XXX this probably needs some work.
156 * NOTE: MSGF_DONE is left set. The target port must clear it if the
157 * message is to be handled asynchronously, while the synchronous case
158 * can just ignore it.
161 lwkt_domsg(lwkt_port_t port, lwkt_msg_t msg)
165 msg->ms_flags &= ~(MSGF_ASYNC | MSGF_REPLY1 | MSGF_REPLY2 | \
166 MSGF_QUEUED | MSGF_ABORTED | MSGF_RETRIEVED);
167 KKASSERT(msg->ms_reply_port != NULL);
168 msg->ms_abort_port = msg->ms_reply_port;
169 if ((error = lwkt_beginmsg(port, msg)) == EASYNC) {
170 error = lwkt_waitmsg(msg);
175 /************************************************************************
177 ************************************************************************/
182 * Initialize a port for use and assign it to the specified thread.
185 lwkt_initport(lwkt_port_t port, thread_t td)
187 bzero(port, sizeof(*port));
188 TAILQ_INIT(&port->mp_msgq);
190 port->mp_putport = lwkt_default_putport;
191 port->mp_waitport = lwkt_default_waitport;
192 port->mp_replyport = lwkt_default_replyport;
193 port->mp_abortport = lwkt_default_abortport;
197 lwkt_initport_null_rport(lwkt_port_t port, thread_t td)
199 lwkt_initport(port, td);
200 port->mp_replyport = lwkt_null_replyport;
206 * Retrieve the next message from the port's message queue, return NULL
207 * if no messages are pending. Note that callers CANNOT use the
208 * MSGF_ABORTED flag as a litmus test to determine if a message
209 * was aborted. The flag only indicates that an abort was requested.
210 * The message's error code will indicate whether an abort occured
211 * (typically by returning EINTR).
213 * Note that once a message has been dequeued it is subject to being
214 * requeued via an IPI based abort request if it is not marked MSGF_DONE.
216 * If the message has been aborted we have to guarentee that abort
217 * semantics are properly followed. The target port will always see
218 * the original message at least once, and if it does not reply the
219 * message before looping on its message port again it will then see
220 * the message again with ms_cmd set to ms_abort.
222 * The calling thread MUST own the port.
227 _lwkt_pullmsg(lwkt_port_t port, lwkt_msg_t msg)
229 if ((msg->ms_flags & MSGF_ABORTED) == 0) {
231 * normal case, remove and return the message.
233 TAILQ_REMOVE(&port->mp_msgq, msg, ms_node);
234 msg->ms_flags = (msg->ms_flags & ~MSGF_QUEUED) | MSGF_RETRIEVED;
236 if (msg->ms_flags & MSGF_RETRIEVED) {
238 * abort case, message already returned once, remvoe and
239 * return the aborted message a second time after setting
240 * ms_cmd to ms_abort.
242 TAILQ_REMOVE(&port->mp_msgq, msg, ms_node);
243 msg->ms_flags &= ~MSGF_QUEUED;
244 msg->ms_cmd = msg->ms_abort;
247 * abort case, abort races initial message retrieval. The
248 * message is returned normally but not removed from the
249 * queue. On the next loop the 'aborted' message will be
250 * dequeued and returned. Note that if the caller replies
251 * to the message it will be dequeued (the abort becomes a
254 msg->ms_flags |= MSGF_RETRIEVED;
260 lwkt_getport(lwkt_port_t port)
264 KKASSERT(port->mp_td == curthread);
266 crit_enter_quick(port->mp_td);
267 if ((msg = TAILQ_FIRST(&port->mp_msgq)) != NULL)
268 _lwkt_pullmsg(port, msg);
269 crit_exit_quick(port->mp_td);
274 * This inline helper function completes processing of a reply from an
275 * unknown cpu context.
277 * The message is being returned to the specified port. The port is
278 * owned by the mp_td thread. If we are on the same cpu as the mp_td
279 * thread we can trivially queue the message to the reply port and schedule
280 * the target thread, otherwise we have to send an ipi message to the
283 * This inline must be entered with a critical section already held.
284 * Note that the IPIQ callback function (*_remote) is entered with a
285 * critical section already held, and we obtain one in lwkt_replyport().
289 _lwkt_replyport(lwkt_port_t port, lwkt_msg_t msg, int force)
291 thread_t td = port->mp_td;
293 if (force || td->td_gd == mycpu) {
295 * We can only reply the message if the abort has caught up with us,
296 * or if no abort was issued (same case).
298 if (msg->ms_abort_port == port) {
299 KKASSERT((msg->ms_flags & MSGF_QUEUED) == 0);
300 TAILQ_INSERT_TAIL(&port->mp_msgq, msg, ms_node);
301 msg->ms_flags |= MSGF_DONE | MSGF_QUEUED | MSGF_REPLY2;
302 if (port->mp_flags & MSGPORTF_WAITING)
306 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_replyport_remote, msg);
311 * This function completes reply processing for the default case in the
312 * context of the originating cpu.
316 lwkt_replyport_remote(lwkt_msg_t msg)
318 _lwkt_replyport(msg->ms_reply_port, msg, 1);
322 * This function is called in the context of the target to reply a message.
323 * The critical section protects us from IPIs on the this CPU.
326 lwkt_default_replyport(lwkt_port_t port, lwkt_msg_t msg)
329 msg->ms_flags |= MSGF_REPLY1;
332 * An abort may have caught up to us while we were processing the
333 * message. If this occured we have to dequeue the message from the
334 * target port in the context of our current cpu before we can finish
337 if (msg->ms_flags & MSGF_QUEUED) {
338 KKASSERT(msg->ms_flags & MSGF_ABORTED);
339 TAILQ_REMOVE(&msg->ms_target_port->mp_msgq, msg, ms_node);
340 msg->ms_flags &= ~MSGF_QUEUED;
344 * Do reply port processing for async messages. Just mark the message
345 * done and wakeup the owner of the reply port for synchronous messages.
347 if (msg->ms_flags & MSGF_ASYNC) {
348 _lwkt_replyport(port, msg, 0);
350 msg->ms_flags |= MSGF_DONE;
351 if (port->mp_flags & MSGPORTF_WAITING)
352 lwkt_schedule(port->mp_td);
358 * You can point a port's reply vector at this function if you just want
359 * the message marked done, without any queueing or signaling. This is
360 * often used for structure-embedded messages.
363 lwkt_null_replyport(lwkt_port_t port, lwkt_msg_t msg)
366 msg->ms_flags |= MSGF_DONE|MSGF_REPLY1;
371 * lwkt_default_putport()
373 * This function is typically assigned to the mp_putport port vector.
375 * Queue a message to the target port and wakeup the thread owning it.
376 * This function always returns EASYNC and may be assigned to a
377 * message port's mp_putport function vector. Note that we must set
378 * MSGF_QUEUED prior to sending any IPIs in order to interlock against
379 * ABORT requests and other tests that might be performed.
381 * Note that messages start out as synchronous entities, and as an
382 * optimization MSGF_DONE is usually left set (so in the synchronous path
383 * no modifications to ms_flags are ever required). If a message becomes
384 * async, i.e. you return EASYNC, then MSGF_DONE must be cleared or
385 * lwkt_replymsg() will wind up being a NOP.
387 * The inline must be called from a critical section (the remote function
388 * is called from an IPI and will be in a critical section).
393 _lwkt_putport(lwkt_port_t port, lwkt_msg_t msg, int force)
395 thread_t td = port->mp_td;
397 if (force || td->td_gd == mycpu) {
398 TAILQ_INSERT_TAIL(&port->mp_msgq, msg, ms_node);
399 if (port->mp_flags & MSGPORTF_WAITING)
402 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_putport_remote, msg);
408 lwkt_putport_remote(lwkt_msg_t msg)
410 _lwkt_putport(msg->ms_target_port, msg, 1);
414 lwkt_default_putport(lwkt_port_t port, lwkt_msg_t msg)
417 msg->ms_flags |= MSGF_QUEUED; /* abort interlock */
418 msg->ms_flags &= ~MSGF_DONE;
419 msg->ms_target_port = port;
420 _lwkt_putport(port, msg, 0);
428 * Forward a message received on one port to another port. The forwarding
429 * function must deal with a pending abort but othewise essentially just
430 * issues a putport to the target port.
432 * An abort may have two side effects: First, the message may have been
433 * requeued to the current target port. If so, we must dequeue it before
437 lwkt_forwardmsg(lwkt_port_t port, lwkt_msg_t msg)
442 if (msg->ms_flags & MSGF_QUEUED) {
443 KKASSERT(msg->ms_flags & MSGF_ABORTED);
444 TAILQ_REMOVE(&msg->ms_target_port->mp_msgq, msg, ms_node);
445 msg->ms_flags &= ~MSGF_QUEUED;
447 msg->ms_flags &= ~MSGF_RETRIEVED;
448 if ((error = port->mp_putport(port, msg)) != EASYNC)
449 lwkt_replymsg(msg, error);
457 * Aborting a message is a fairly complex task. The first order of
458 * business is to get the message to the cpu that owns the target
459 * port, during which we may have to do some port chasing due to
460 * message forwarding operations.
462 * NOTE! Since an aborted message is requeued all message processing
463 * loops should check the MSGF_ABORTED flag.
465 static void lwkt_abortmsg_remote(lwkt_msg_t msg);
468 lwkt_abortmsg(lwkt_msg_t msg)
474 * A critical section protects us from reply IPIs on this cpu. We
475 * can only abort messages that have not yet completed (DONE), are not
476 * in the midst of being replied (REPLY1), and which support the
477 * abort function (ABORTABLE).
480 if ((msg->ms_flags & (MSGF_DONE|MSGF_REPLY1|MSGF_ABORTABLE)) == MSGF_ABORTABLE) {
482 * Chase the message. If REPLY1 is set the message has been replied
483 * all the way back to the originator, otherwise it is sitting on
484 * ms_target_port (but we can only complete processing if we are
485 * on the same cpu as the selected port in order to avoid
486 * SMP cache synchronization issues).
488 * When chasing through multiple ports ms_flags may not be
489 * synchronized to the current cpu, but it WILL be synchronized
490 * with regards to testing the MSGF_REPLY1 bit once we reach the
491 * target port that made the reply and since the cpu owning
492 * some port X stores the new port in ms_target_port if the message
493 * is forwarded, the current port will only ever equal the target
494 * port when we are on the correct cpu.
496 if (msg->ms_flags & MSGF_REPLY1)
497 port = msg->ms_reply_port;
499 port = msg->ms_target_port;
503 * The chase call must run on the cpu owning the port. Fully
504 * synchronous ports (mp_td == NULL) can run the call on any cpu.
507 if (td && td->td_gd != mycpu) {
508 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_abortmsg_remote, msg);
510 port->mp_abortport(port, msg);
518 lwkt_abortmsg_remote(lwkt_msg_t msg)
523 if (msg->ms_flags & MSGF_REPLY1)
524 port = msg->ms_reply_port;
526 port = msg->ms_target_port;
529 if (td->td_gd != mycpu) {
530 lwkt_send_ipiq(td->td_gd, (ipifunc_t)lwkt_abortmsg_remote, msg);
532 port->mp_abortport(port, msg);
537 * The mp_abortport function is called when the abort has finally caught up
538 * to the target port or (if the message has been replied) the reply port.
541 lwkt_default_abortport(lwkt_port_t port, lwkt_msg_t msg)
544 * Set ms_abort_port to ms_reply_port to indicate the completion of
545 * the messaging chasing portion of the abort request. Note that
546 * the passed port is the port that we finally caught up to, not
547 * necessarily the reply port.
549 msg->ms_abort_port = msg->ms_reply_port;
551 if (msg->ms_flags & MSGF_REPLY2) {
553 * If REPLY2 is set we must have chased it all the way back to
554 * the reply port, but the replyport code has not queued the message
555 * (because it was waiting for the abort to catch up). We become
556 * responsible for queueing the message to the reply port.
558 KKASSERT((msg->ms_flags & MSGF_QUEUED) == 0);
559 KKASSERT(port == msg->ms_reply_port);
560 TAILQ_INSERT_TAIL(&port->mp_msgq, msg, ms_node);
561 msg->ms_flags |= MSGF_DONE | MSGF_QUEUED;
562 if (port->mp_flags & MSGPORTF_WAITING)
563 lwkt_schedule(port->mp_td);
564 } else if ((msg->ms_flags & (MSGF_QUEUED|MSGF_REPLY1)) == 0) {
566 * Abort on the target port. The message has not yet been replied
567 * and must be requeued to the target port.
569 msg->ms_flags |= MSGF_ABORTED | MSGF_QUEUED;
570 TAILQ_INSERT_TAIL(&port->mp_msgq, msg, ms_node);
571 if (port->mp_flags & MSGPORTF_WAITING)
572 lwkt_schedule(port->mp_td);
573 } else if ((msg->ms_flags & MSGF_REPLY1) == 0) {
575 * The message has not yet been retrieved by the target port, set
576 * MSGF_ABORTED so the target port can requeue the message abort after
579 msg->ms_flags |= MSGF_ABORTED;
584 * lwkt_default_waitport()
586 * If msg is NULL, dequeue the next message from the port's message
587 * queue, block until a message is ready. This function never
590 * If msg is non-NULL, block until the requested message has been returned
591 * to the port then dequeue and return it. DO NOT USE THIS TO WAIT FOR
592 * INCOMING REQUESTS, ONLY USE THIS TO WAIT FOR REPLIES.
594 * Note that the API does not currently support multiple threads waiting
595 * on a port. By virtue of owning the port it is controlled by our
596 * cpu and we can safely manipulate it's contents.
599 lwkt_default_waitport(lwkt_port_t port, lwkt_msg_t msg)
601 thread_t td = curthread;
604 KKASSERT(port->mp_td == td);
605 crit_enter_quick(td);
607 if ((msg = TAILQ_FIRST(&port->mp_msgq)) == NULL) {
608 port->mp_flags |= MSGPORTF_WAITING;
609 td->td_flags |= TDF_BLOCKED;
611 lwkt_deschedule_self(td);
613 } while ((msg = TAILQ_FIRST(&port->mp_msgq)) == NULL);
614 td->td_flags &= ~TDF_BLOCKED;
615 port->mp_flags &= ~MSGPORTF_WAITING;
617 _lwkt_pullmsg(port, msg);
620 * If a message is not marked done, or if it is queued, we have work
621 * to do. Note that MSGF_DONE is always set in the context of the
624 if ((msg->ms_flags & (MSGF_DONE|MSGF_QUEUED)) != MSGF_DONE) {
626 * We must own the reply port to safely mess with it's contents.
628 port = msg->ms_reply_port;
629 KKASSERT(port->mp_td == td);
631 if ((msg->ms_flags & MSGF_DONE) == 0) {
632 port->mp_flags |= MSGPORTF_WAITING; /* saved by the BGL */
637 * MSGF_PCATCH is only set by processes which wish to
638 * abort the message they are blocked on when a signal
639 * occurs. Note that we still must wait for message
640 * completion after sending an abort request.
642 if (msg->ms_flags & MSGF_PCATCH) {
643 if (sentabort == 0 && CURSIG(port->mp_td->td_proc)) {
651 * XXX set TDF_SINTR so 'ps' knows the difference between
652 * an interruptable wait and a disk wait. YYY eventually
653 * move P_SINTR to TDF_SINTR to reduce duplication.
655 td->td_flags |= TDF_SINTR | TDF_BLOCKED;
656 lwkt_deschedule_self(td);
658 td->td_flags &= ~(TDF_SINTR | TDF_BLOCKED);
659 } while ((msg->ms_flags & MSGF_DONE) == 0);
660 port->mp_flags &= ~MSGPORTF_WAITING; /* saved by the BGL */
663 * We own the message now.
665 if (msg->ms_flags & MSGF_QUEUED) {
666 msg->ms_flags &= ~MSGF_QUEUED;
667 TAILQ_REMOVE(&port->mp_msgq, msg, ms_node);