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 1.35 2006/11/07 18:50:06 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/cpufunc.h>
70 #include <machine/smp.h>
73 #include <sys/malloc.h>
74 MALLOC_DEFINE(M_LWKTMSG, "lwkt message", "lwkt message");
78 #include <sys/stdint.h>
79 #include <libcaps/thread.h>
80 #include <sys/thread.h>
81 #include <sys/msgport.h>
82 #include <sys/errno.h>
83 #include <libcaps/globaldata.h>
84 #include <machine/cpufunc.h>
85 #include <sys/thread2.h>
86 #include <sys/msgport2.h>
92 /************************************************************************
94 ************************************************************************/
97 static void lwkt_replyport_remote(lwkt_msg_t msg);
98 static void lwkt_putport_remote(lwkt_msg_t msg);
99 static void lwkt_abortmsg_remote(lwkt_msg_t msg);
105 * Send a message asynchronously. This function requests asynchronous
106 * completion and calls lwkt_beginmsg(). If the target port decides to
107 * run the message synchronously this function will automatically queue
108 * the message to the current thread's message queue to present a
109 * consistent interface to the caller.
111 * The message's ms_cmd must be initialized and its ms_flags must
112 * be zero'd out. lwkt_sendmsg() will initialize the ms_abort_port
113 * (abort chasing port). If abort is supported, ms_abort must also be
116 * NOTE: you cannot safely request an abort until lwkt_sendmsg() returns
119 * NOTE: MSGF_DONE is left set. The target port must clear it if the
120 * message is to be handled asynchronously, while the synchronous case
121 * can just ignore it.
124 lwkt_sendmsg(lwkt_port_t port, lwkt_msg_t msg)
128 msg->ms_flags |= MSGF_ASYNC;
129 msg->ms_flags &= ~(MSGF_REPLY1 | MSGF_REPLY2 | MSGF_QUEUED | \
130 MSGF_ABORTED | MSGF_RETRIEVED);
131 KKASSERT(msg->ms_reply_port != NULL);
132 msg->ms_abort_port = msg->ms_reply_port;
133 if ((error = lwkt_beginmsg(port, msg)) != EASYNC) {
134 lwkt_replymsg(msg, error);
141 * Send a message synchronously. This function requests synchronous
142 * completion and calls lwkt_beginmsg(). If the target port decides to
143 * run the message asynchronously this function will block waiting for
144 * the message to complete. Since MSGF_ASYNC is not set the target
145 * will not attempt to queue the reply to a reply port but will simply
146 * wake up anyone waiting on the message.
148 * A synchronous error code is always returned.
150 * The message's ms_cmd must be initialized, and its ms_flags must be
151 * at least zero'd out. lwkt_domsg() will initialize the message's
152 * ms_abort_port (abort chasing port). If abort is supported, ms_abort
153 * must also be initialized.
155 * NOTE: you cannot safely request an abort until lwkt_domsg() blocks.
156 * XXX this probably needs some work.
158 * NOTE: MSGF_DONE is left set. The target port must clear it if the
159 * message is to be handled asynchronously, while the synchronous case
160 * can just ignore it.
163 lwkt_domsg(lwkt_port_t port, lwkt_msg_t msg)
167 msg->ms_flags &= ~(MSGF_ASYNC | MSGF_REPLY1 | MSGF_REPLY2 | \
168 MSGF_QUEUED | MSGF_ABORTED | MSGF_RETRIEVED);
169 KKASSERT(msg->ms_reply_port != NULL);
170 msg->ms_abort_port = msg->ms_reply_port;
171 if ((error = lwkt_beginmsg(port, msg)) == EASYNC) {
172 error = lwkt_waitmsg(msg);
177 /************************************************************************
179 ************************************************************************/
184 * Initialize a port for use and assign it to the specified thread.
185 * The default reply function is to return the message to the originator.
188 lwkt_initport(lwkt_port_t port, thread_t td)
190 bzero(port, sizeof(*port));
191 TAILQ_INIT(&port->mp_msgq);
193 port->mp_putport = lwkt_default_putport;
194 port->mp_waitport = lwkt_default_waitport;
195 port->mp_replyport = lwkt_default_replyport;
196 port->mp_abortport = lwkt_default_abortport;
200 * Similar to the standard initport, this function simply marks the message
201 * as being done and does not attempt to return it to an originating port.
204 lwkt_initport_null_rport(lwkt_port_t port, thread_t td)
206 lwkt_initport(port, td);
207 port->mp_replyport = lwkt_null_replyport;
213 * Retrieve the next message from the port's message queue, return NULL
214 * if no messages are pending. Note that callers CANNOT use the
215 * MSGF_ABORTED flag as a litmus test to determine if a message
216 * was aborted. The flag only indicates that an abort was requested.
217 * The message's error code will indicate whether an abort occured
218 * (typically by returning EINTR).
220 * Note that once a message has been dequeued it is subject to being
221 * requeued via an IPI based abort request if it is not marked MSGF_DONE.
223 * If the message has been aborted we have to guarentee that abort
224 * semantics are properly followed. The target port will always see
225 * the original message at least once, and if it does not reply the
226 * message before looping on its message port again it will then see
227 * the message again with ms_cmd set to ms_abort.
229 * The calling thread MUST own the port.
234 _lwkt_pullmsg(lwkt_port_t port, lwkt_msg_t msg)
236 if ((msg->ms_flags & MSGF_ABORTED) == 0) {
238 * normal case, remove and return the message.
240 TAILQ_REMOVE(&port->mp_msgq, msg, ms_node);
241 msg->ms_flags = (msg->ms_flags & ~MSGF_QUEUED) | MSGF_RETRIEVED;
243 if (msg->ms_flags & MSGF_RETRIEVED) {
245 * abort case, message already returned once, remvoe and
246 * return the aborted message a second time after setting
247 * ms_cmd to ms_abort.
249 TAILQ_REMOVE(&port->mp_msgq, msg, ms_node);
250 msg->ms_flags &= ~MSGF_QUEUED;
251 msg->ms_cmd = msg->ms_abort;
254 * abort case, abort races initial message retrieval. The
255 * message is returned normally but not removed from the
256 * queue. On the next loop the 'aborted' message will be
257 * dequeued and returned. Note that if the caller replies
258 * to the message it will be dequeued (the abort becomes a
261 msg->ms_flags |= MSGF_RETRIEVED;
267 lwkt_getport(lwkt_port_t port)
271 KKASSERT(port->mp_td == curthread);
273 crit_enter_quick(port->mp_td);
274 if ((msg = TAILQ_FIRST(&port->mp_msgq)) != NULL)
275 _lwkt_pullmsg(port, msg);
276 crit_exit_quick(port->mp_td);
281 * This inline helper function completes processing of a reply from an
282 * unknown cpu context.
284 * The message is being returned to the specified port. The port is
285 * owned by the mp_td thread. If we are on the same cpu as the mp_td
286 * thread we can trivially queue the message to the reply port and schedule
287 * the target thread, otherwise we have to send an ipi message to the
290 * This inline must be entered with a critical section already held.
291 * Note that the IPIQ callback function (*_remote) is entered with a
292 * critical section already held, and we obtain one in lwkt_replyport().
296 _lwkt_replyport(lwkt_port_t port, lwkt_msg_t msg, int force)
298 thread_t td = port->mp_td;
300 if (force || td->td_gd == mycpu) {
302 * We can only reply the message if the abort has caught up with us,
303 * or if no abort was issued (same case).
305 if (msg->ms_abort_port == port) {
306 KKASSERT((msg->ms_flags & MSGF_QUEUED) == 0);
307 TAILQ_INSERT_TAIL(&port->mp_msgq, msg, ms_node);
308 msg->ms_flags |= MSGF_DONE | MSGF_QUEUED | MSGF_REPLY2;
309 if (port->mp_flags & MSGPORTF_WAITING)
314 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_replyport_remote, msg);
316 panic("lwkt_replyport: thread %p has bad gd pointer", td);
324 * This function completes reply processing for the default case in the
325 * context of the originating cpu.
329 lwkt_replyport_remote(lwkt_msg_t msg)
331 _lwkt_replyport(msg->ms_reply_port, msg, 1);
337 * This function is called in the context of the target to reply a message.
338 * The critical section protects us from IPIs on the this CPU.
341 lwkt_default_replyport(lwkt_port_t port, lwkt_msg_t msg)
344 msg->ms_flags |= MSGF_REPLY1;
347 * An abort may have caught up to us while we were processing the
348 * message. If this occured we have to dequeue the message from the
349 * target port in the context of our current cpu before we can finish
352 if (msg->ms_flags & MSGF_QUEUED) {
353 KKASSERT(msg->ms_flags & MSGF_ABORTED);
354 TAILQ_REMOVE(&msg->ms_target_port->mp_msgq, msg, ms_node);
355 msg->ms_flags &= ~MSGF_QUEUED;
359 * Do reply port processing for async messages. Just mark the message
360 * done and wakeup the owner of the reply port for synchronous messages.
362 if (msg->ms_flags & MSGF_ASYNC) {
363 _lwkt_replyport(port, msg, 0);
365 msg->ms_flags |= MSGF_DONE;
366 if (port->mp_flags & MSGPORTF_WAITING)
367 lwkt_schedule(port->mp_td);
373 * You can point a port's reply vector at this function if you just want
374 * the message marked done, without any queueing or signaling. This is
375 * often used for structure-embedded messages.
378 lwkt_null_replyport(lwkt_port_t port, lwkt_msg_t msg)
381 msg->ms_flags |= MSGF_DONE|MSGF_REPLY1;
386 * lwkt_default_putport()
388 * This function is typically assigned to the mp_putport port vector.
390 * Queue a message to the target port and wakeup the thread owning it.
391 * This function always returns EASYNC and may be assigned to a
392 * message port's mp_putport function vector. Note that we must set
393 * MSGF_QUEUED prior to sending any IPIs in order to interlock against
394 * ABORT requests and other tests that might be performed.
396 * Note that messages start out as synchronous entities, and as an
397 * optimization MSGF_DONE is usually left set (so in the synchronous path
398 * no modifications to ms_flags are ever required). If a message becomes
399 * async, i.e. you return EASYNC, then MSGF_DONE must be cleared or
400 * lwkt_replymsg() will wind up being a NOP.
402 * The inline must be called from a critical section (the remote function
403 * is called from an IPI and will be in a critical section).
408 _lwkt_putport(lwkt_port_t port, lwkt_msg_t msg, int force)
410 thread_t td = port->mp_td;
412 if (force || td->td_gd == mycpu) {
413 TAILQ_INSERT_TAIL(&port->mp_msgq, msg, ms_node);
414 if (port->mp_flags & MSGPORTF_WAITING)
418 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_putport_remote, msg);
420 panic("lwkt_putport: thread %p has bad gd pointer", td);
429 lwkt_putport_remote(lwkt_msg_t msg)
433 * try to catch a free-after-send issue.
435 if (msg->ms_target_port == (void *)0xdeadc0de) {
437 for (i = 0; i < 1000000; ++i) {
438 if (msg->ms_target_port != (void *)0xdeadc0de)
442 panic("msg %p ms_target_port is bogus: reads %p after %d loops\n", msg, msg->ms_target_port, i);
445 _lwkt_putport(msg->ms_target_port, msg, 1);
451 lwkt_default_putport(lwkt_port_t port, lwkt_msg_t msg)
454 msg->ms_flags |= MSGF_QUEUED; /* abort interlock */
455 msg->ms_flags &= ~MSGF_DONE;
456 msg->ms_target_port = port;
457 _lwkt_putport(port, msg, 0);
465 * Forward a message received on one port to another port. The forwarding
466 * function must deal with a pending abort but othewise essentially just
467 * issues a putport to the target port.
469 * An abort may have two side effects: First, the message may have been
470 * requeued to the current target port. If so, we must dequeue it before
474 lwkt_forwardmsg(lwkt_port_t port, lwkt_msg_t msg)
479 if (msg->ms_flags & MSGF_QUEUED) {
480 KKASSERT(msg->ms_flags & MSGF_ABORTED);
481 TAILQ_REMOVE(&msg->ms_target_port->mp_msgq, msg, ms_node);
482 msg->ms_flags &= ~MSGF_QUEUED;
484 msg->ms_flags &= ~MSGF_RETRIEVED;
485 if ((error = port->mp_putport(port, msg)) != EASYNC)
486 lwkt_replymsg(msg, error);
494 * Aborting a message is a fairly complex task. The first order of
495 * business is to get the message to the cpu that owns the target
496 * port, during which we may have to do some port chasing due to
497 * message forwarding operations.
499 * NOTE! Since an aborted message is requeued all message processing
500 * loops should check the MSGF_ABORTED flag.
504 lwkt_abortmsg(lwkt_msg_t msg)
510 * A critical section protects us from reply IPIs on this cpu. We
511 * can only abort messages that have not yet completed (DONE), are not
512 * in the midst of being replied (REPLY1), and which support the
513 * abort function (ABORTABLE).
516 if ((msg->ms_flags & (MSGF_DONE|MSGF_REPLY1|MSGF_ABORTABLE)) == MSGF_ABORTABLE) {
518 * Chase the message. If REPLY1 is set the message has been replied
519 * all the way back to the originator, otherwise it is sitting on
520 * ms_target_port (but we can only complete processing if we are
521 * on the same cpu as the selected port in order to avoid
522 * SMP cache synchronization issues).
524 * When chasing through multiple ports ms_flags may not be
525 * synchronized to the current cpu, but it WILL be synchronized
526 * with regards to testing the MSGF_REPLY1 bit once we reach the
527 * target port that made the reply and since the cpu owning
528 * some port X stores the new port in ms_target_port if the message
529 * is forwarded, the current port will only ever equal the target
530 * port when we are on the correct cpu.
532 if (msg->ms_flags & MSGF_REPLY1)
533 port = msg->ms_reply_port;
535 port = msg->ms_target_port;
537 cpu_ccfence(); /* don't let the compiler reload ms_*_port */
540 * The chase call must run on the cpu owning the port. Fully
541 * synchronous ports (mp_td == NULL) can run the call on any cpu.
544 if (td && td->td_gd != mycpu) {
546 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_abortmsg_remote, msg);
548 panic("lwkt_abortmsg: thread %p has bad gd pointer", td);
551 port->mp_abortport(port, msg);
561 lwkt_abortmsg_remote(lwkt_msg_t msg)
566 if (msg->ms_flags & MSGF_REPLY1)
567 port = msg->ms_reply_port;
569 port = msg->ms_target_port;
570 cpu_ccfence(); /* don't let the compiler reload ms_*_port */
572 if (td->td_gd != mycpu) {
573 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)lwkt_abortmsg_remote, msg);
575 port->mp_abortport(port, msg);
582 * The mp_abortport function is called when the abort has finally caught up
583 * to the target port or (if the message has been replied) the reply port.
586 lwkt_default_abortport(lwkt_port_t port, lwkt_msg_t msg)
589 * Set ms_abort_port to ms_reply_port to indicate the completion of
590 * the messaging chasing portion of the abort request. Note that
591 * the passed port is the port that we finally caught up to, not
592 * necessarily the reply port.
594 msg->ms_abort_port = msg->ms_reply_port;
596 if (msg->ms_flags & MSGF_REPLY2) {
598 * If REPLY2 is set we must have chased it all the way back to
599 * the reply port, but the replyport code has not queued the message
600 * (because it was waiting for the abort to catch up). We become
601 * responsible for queueing the message to the reply port.
603 KKASSERT((msg->ms_flags & MSGF_QUEUED) == 0);
604 KKASSERT(port == msg->ms_reply_port);
605 TAILQ_INSERT_TAIL(&port->mp_msgq, msg, ms_node);
606 msg->ms_flags |= MSGF_DONE | MSGF_QUEUED;
607 if (port->mp_flags & MSGPORTF_WAITING)
608 lwkt_schedule(port->mp_td);
609 } else if ((msg->ms_flags & (MSGF_QUEUED|MSGF_REPLY1)) == 0) {
611 * Abort on the target port. The message has not yet been replied
612 * and must be requeued to the target port.
614 msg->ms_flags |= MSGF_ABORTED | MSGF_QUEUED;
615 TAILQ_INSERT_TAIL(&port->mp_msgq, msg, ms_node);
616 if (port->mp_flags & MSGPORTF_WAITING)
617 lwkt_schedule(port->mp_td);
618 } else if ((msg->ms_flags & MSGF_REPLY1) == 0) {
620 * The message has not yet been retrieved by the target port, set
621 * MSGF_ABORTED so the target port can requeue the message abort after
624 msg->ms_flags |= MSGF_ABORTED;
629 * lwkt_default_waitport()
631 * If msg is NULL, dequeue the next message from the port's message
632 * queue, block until a message is ready. This function never
635 * If msg is non-NULL, block until the requested message has been returned
636 * to the port then dequeue and return it. DO NOT USE THIS TO WAIT FOR
637 * INCOMING REQUESTS, ONLY USE THIS TO WAIT FOR REPLIES.
639 * Note that the API does not currently support multiple threads waiting
640 * on a port. By virtue of owning the port it is controlled by our
641 * cpu and we can safely manipulate it's contents.
644 lwkt_default_waitport(lwkt_port_t port, lwkt_msg_t msg)
646 thread_t td = curthread;
649 KKASSERT(port->mp_td == td);
650 crit_enter_quick(td);
652 if ((msg = TAILQ_FIRST(&port->mp_msgq)) == NULL) {
653 port->mp_flags |= MSGPORTF_WAITING;
654 td->td_flags |= TDF_BLOCKED;
656 lwkt_deschedule_self(td);
658 } while ((msg = TAILQ_FIRST(&port->mp_msgq)) == NULL);
659 td->td_flags &= ~TDF_BLOCKED;
660 port->mp_flags &= ~MSGPORTF_WAITING;
662 _lwkt_pullmsg(port, msg);
665 * If a message is not marked done, or if it is queued, we have work
666 * to do. Note that MSGF_DONE is always set in the context of the
669 if ((msg->ms_flags & (MSGF_DONE|MSGF_QUEUED)) != MSGF_DONE) {
671 * We must own the reply port to safely mess with it's contents.
673 port = msg->ms_reply_port;
674 KKASSERT(port->mp_td == td);
676 if ((msg->ms_flags & MSGF_DONE) == 0) {
677 port->mp_flags |= MSGPORTF_WAITING; /* saved by the BGL */
682 * MSGF_PCATCH is only set by processes which wish to
683 * abort the message they are blocked on when a signal
684 * occurs. Note that we still must wait for message
685 * completion after sending an abort request.
687 if (msg->ms_flags & MSGF_PCATCH) {
688 if (sentabort == 0 && CURSIG(port->mp_td->td_proc)) {
696 * XXX set TDF_SINTR so 'ps' knows the difference between
697 * an interruptable wait and a disk wait. YYY eventually
698 * move P_SINTR to TDF_SINTR to reduce duplication.
700 td->td_flags |= TDF_SINTR | TDF_BLOCKED;
701 lwkt_deschedule_self(td);
703 td->td_flags &= ~(TDF_SINTR | TDF_BLOCKED);
704 } while ((msg->ms_flags & MSGF_DONE) == 0);
705 port->mp_flags &= ~MSGPORTF_WAITING; /* saved by the BGL */
708 * We own the message now.
710 if (msg->ms_flags & MSGF_QUEUED) {
711 msg->ms_flags &= ~MSGF_QUEUED;
712 TAILQ_REMOVE(&port->mp_msgq, msg, ms_node);