xref: /illumos-gate/usr/src/cmd/bhyve/mevent.c (revision cf7690eb)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 NetApp, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * $FreeBSD$
29  */
30 
31 /*
32  * Copyright 2018 Joyent, Inc.
33  */
34 
35 /*
36  * Micro event library for FreeBSD, designed for a single i/o thread
37  * using kqueue, and having events be persistent by default.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include <assert.h>
44 #ifndef WITHOUT_CAPSICUM
45 #include <capsicum_helpers.h>
46 #endif
47 #include <err.h>
48 #include <errno.h>
49 #include <stdbool.h>
50 #include <stdlib.h>
51 #include <stdio.h>
52 #include <string.h>
53 #include <sysexits.h>
54 #include <unistd.h>
55 
56 #include <sys/types.h>
57 #ifndef WITHOUT_CAPSICUM
58 #include <sys/capsicum.h>
59 #endif
60 #ifdef __FreeBSD__
61 #include <sys/event.h>
62 #else
63 #include <port.h>
64 #include <sys/poll.h>
65 #include <sys/siginfo.h>
66 #include <sys/queue.h>
67 #include <sys/debug.h>
68 #endif
69 #include <sys/time.h>
70 
71 #include <pthread.h>
72 #include <pthread_np.h>
73 
74 #include "mevent.h"
75 
76 #define	MEVENT_MAX	64
77 
78 #ifndef __FreeBSD__
79 #define	EV_ENABLE	0x01
80 #define	EV_ADD		EV_ENABLE
81 #define	EV_DISABLE	0x02
82 #define	EV_DELETE	0x04
83 #endif
84 
85 extern char *vmname;
86 
87 static pthread_t mevent_tid;
88 static int mevent_timid = 43;
89 static int mevent_pipefd[2];
90 static pthread_mutex_t mevent_lmutex = PTHREAD_MUTEX_INITIALIZER;
91 
92 struct mevent {
93 	void	(*me_func)(int, enum ev_type, void *);
94 #define me_msecs me_fd
95 	int	me_fd;
96 #ifdef __FreeBSD__
97 	int	me_timid;
98 #else
99 	timer_t me_timid;
100 #endif
101 	enum ev_type me_type;
102 	void    *me_param;
103 	int	me_cq;
104 	int	me_state; /* Desired kevent flags. */
105 	int	me_closefd;
106 #ifndef __FreeBSD__
107 	port_notify_t	me_notify;
108 	struct sigevent	me_sigev;
109 	boolean_t	me_auto_requeue;
110 #endif
111 	LIST_ENTRY(mevent) me_list;
112 };
113 
114 static LIST_HEAD(listhead, mevent) global_head, change_head;
115 
116 static void
117 mevent_qlock(void)
118 {
119 	pthread_mutex_lock(&mevent_lmutex);
120 }
121 
122 static void
123 mevent_qunlock(void)
124 {
125 	pthread_mutex_unlock(&mevent_lmutex);
126 }
127 
128 static void
129 mevent_pipe_read(int fd, enum ev_type type, void *param)
130 {
131 	char buf[MEVENT_MAX];
132 	int status;
133 
134 	/*
135 	 * Drain the pipe read side. The fd is non-blocking so this is
136 	 * safe to do.
137 	 */
138 	do {
139 		status = read(fd, buf, sizeof(buf));
140 	} while (status == MEVENT_MAX);
141 }
142 
143 static void
144 mevent_notify(void)
145 {
146 	char c = '\0';
147 
148 	/*
149 	 * If calling from outside the i/o thread, write a byte on the
150 	 * pipe to force the i/o thread to exit the blocking kevent call.
151 	 */
152 	if (mevent_pipefd[1] != 0 && pthread_self() != mevent_tid) {
153 		write(mevent_pipefd[1], &c, 1);
154 	}
155 }
156 #ifdef __FreeBSD__
157 static int
158 mevent_kq_filter(struct mevent *mevp)
159 {
160 	int retval;
161 
162 	retval = 0;
163 
164 	if (mevp->me_type == EVF_READ)
165 		retval = EVFILT_READ;
166 
167 	if (mevp->me_type == EVF_WRITE)
168 		retval = EVFILT_WRITE;
169 
170 	if (mevp->me_type == EVF_TIMER)
171 		retval = EVFILT_TIMER;
172 
173 	if (mevp->me_type == EVF_SIGNAL)
174 		retval = EVFILT_SIGNAL;
175 
176 	return (retval);
177 }
178 
179 static int
180 mevent_kq_flags(struct mevent *mevp)
181 {
182 	return (mevp->me_state);
183 }
184 
185 static int
186 mevent_kq_fflags(struct mevent *mevp)
187 {
188 	/* XXX nothing yet, perhaps EV_EOF for reads ? */
189 	return (0);
190 }
191 
192 static int
193 mevent_build(int mfd, struct kevent *kev)
194 {
195 	struct mevent *mevp, *tmpp;
196 	int i;
197 
198 	i = 0;
199 
200 	mevent_qlock();
201 
202 	LIST_FOREACH_SAFE(mevp, &change_head, me_list, tmpp) {
203 		if (mevp->me_closefd) {
204 			/*
205 			 * A close of the file descriptor will remove the
206 			 * event
207 			 */
208 			close(mevp->me_fd);
209 		} else {
210 			if (mevp->me_type == EVF_TIMER) {
211 				kev[i].ident = mevp->me_timid;
212 				kev[i].data = mevp->me_msecs;
213 			} else {
214 				kev[i].ident = mevp->me_fd;
215 				kev[i].data = 0;
216 			}
217 			kev[i].filter = mevent_kq_filter(mevp);
218 			kev[i].flags = mevent_kq_flags(mevp);
219 			kev[i].fflags = mevent_kq_fflags(mevp);
220 			kev[i].udata = mevp;
221 			i++;
222 		}
223 
224 		mevp->me_cq = 0;
225 		LIST_REMOVE(mevp, me_list);
226 
227 		if (mevp->me_state & EV_DELETE) {
228 			free(mevp);
229 		} else {
230 			/*
231 			 * We need to add the event only once, so we can
232 			 * reset the EV_ADD bit after it has been propagated
233 			 * to the kevent() arguments the first time.
234 			 */
235 			mevp->me_state &= ~EV_ADD;
236 			LIST_INSERT_HEAD(&global_head, mevp, me_list);
237 		}
238 
239 		assert(i < MEVENT_MAX);
240 	}
241 
242 	mevent_qunlock();
243 
244 	return (i);
245 }
246 
247 static void
248 mevent_handle(struct kevent *kev, int numev)
249 {
250 	struct mevent *mevp;
251 	int i;
252 
253 	for (i = 0; i < numev; i++) {
254 		mevp = kev[i].udata;
255 
256 		/* XXX check for EV_ERROR ? */
257 
258 		(*mevp->me_func)(mevp->me_fd, mevp->me_type, mevp->me_param);
259 	}
260 }
261 
262 #else /* __FreeBSD__ */
263 
264 static boolean_t
265 mevent_clarify_state(struct mevent *mevp)
266 {
267 	const int state = mevp->me_state;
268 
269 	if ((state & EV_DELETE) != 0) {
270 		/* All other intents are overriden by delete. */
271 		mevp->me_state = EV_DELETE;
272 		return (B_TRUE);
273 	}
274 
275 	/*
276 	 * Without a distinction between EV_ADD and EV_ENABLE in our emulation,
277 	 * handling the add-disabled case means eliding the portfs operation
278 	 * when both flags are present.
279 	 *
280 	 * This is not a concern for subsequent enable/disable operations, as
281 	 * mevent_update() toggles the flags properly so they are not left in
282 	 * conflict.
283 	 */
284 	if (state == (EV_ENABLE|EV_DISABLE)) {
285 		mevp->me_state = EV_DISABLE;
286 		return (B_FALSE);
287 	}
288 
289 	return (B_TRUE);
290 }
291 
292 static void
293 mevent_update_one(struct mevent *mevp)
294 {
295 	int portfd = mevp->me_notify.portnfy_port;
296 
297 	switch (mevp->me_type) {
298 	case EVF_READ:
299 	case EVF_WRITE:
300 		mevp->me_auto_requeue = B_FALSE;
301 
302 		switch (mevp->me_state) {
303 		case EV_ENABLE:
304 		{
305 			int events;
306 
307 			events = (mevp->me_type == EVF_READ) ? POLLIN : POLLOUT;
308 
309 			if (port_associate(portfd, PORT_SOURCE_FD, mevp->me_fd,
310 			    events, mevp) != 0) {
311 				(void) fprintf(stderr,
312 				    "port_associate fd %d %p failed: %s\n",
313 				    mevp->me_fd, mevp, strerror(errno));
314 			}
315 			return;
316 		}
317 		case EV_DISABLE:
318 		case EV_DELETE:
319 			/*
320 			 * A disable that comes in while an event is being
321 			 * handled will result in an ENOENT.
322 			 */
323 			if (port_dissociate(portfd, PORT_SOURCE_FD,
324 			    mevp->me_fd) != 0 && errno != ENOENT) {
325 				(void) fprintf(stderr, "port_dissociate "
326 				    "portfd %d fd %d mevp %p failed: %s\n",
327 				    portfd, mevp->me_fd, mevp, strerror(errno));
328 			}
329 			return;
330 		default:
331 			goto abort;
332 		}
333 
334 	case EVF_TIMER:
335 		mevp->me_auto_requeue = B_TRUE;
336 
337 		switch (mevp->me_state) {
338 		case EV_ENABLE:
339 		{
340 			struct itimerspec it = { 0 };
341 
342 			mevp->me_sigev.sigev_notify = SIGEV_PORT;
343 			mevp->me_sigev.sigev_value.sival_ptr = &mevp->me_notify;
344 
345 			if (timer_create(CLOCK_REALTIME, &mevp->me_sigev,
346 			    &mevp->me_timid) != 0) {
347 				(void) fprintf(stderr,
348 				    "timer_create failed: %s", strerror(errno));
349 				return;
350 			}
351 
352 			/* The first timeout */
353 			it.it_value.tv_sec = mevp->me_msecs / MILLISEC;
354 			it.it_value.tv_nsec =
355 				MSEC2NSEC(mevp->me_msecs % MILLISEC);
356 			/* Repeat at the same interval */
357 			it.it_interval = it.it_value;
358 
359 			if (timer_settime(mevp->me_timid, 0, &it, NULL) != 0) {
360 				(void) fprintf(stderr, "timer_settime failed: "
361 				    "%s", strerror(errno));
362 			}
363 			return;
364 		}
365 		case EV_DISABLE:
366 		case EV_DELETE:
367 			if (timer_delete(mevp->me_timid) != 0) {
368 				(void) fprintf(stderr, "timer_delete failed: "
369 				    "%s", strerror(errno));
370 			}
371 			return;
372 		default:
373 			goto abort;
374 		}
375 	default:
376 		/* EVF_SIGNAL not yet implemented. */
377 		goto abort;
378 	}
379 
380 abort:
381 	(void) fprintf(stderr, "%s: unhandled type %d state %d\n", __func__,
382 	    mevp->me_type, mevp->me_state);
383 	abort();
384 }
385 
386 static void
387 mevent_update_pending(int portfd)
388 {
389 	struct mevent *mevp, *tmpp;
390 
391 	mevent_qlock();
392 
393 	LIST_FOREACH_SAFE(mevp, &change_head, me_list, tmpp) {
394 		mevp->me_notify.portnfy_port = portfd;
395 		mevp->me_notify.portnfy_user = mevp;
396 		if (mevp->me_closefd) {
397 			/*
398 			 * A close of the file descriptor will remove the
399 			 * event
400 			 */
401 			(void) close(mevp->me_fd);
402 			mevp->me_fd = -1;
403 		} else {
404 			if (mevent_clarify_state(mevp)) {
405 				mevent_update_one(mevp);
406 			}
407 		}
408 
409 		mevp->me_cq = 0;
410 		LIST_REMOVE(mevp, me_list);
411 
412 		if (mevp->me_state & EV_DELETE) {
413 			free(mevp);
414 		} else {
415 			LIST_INSERT_HEAD(&global_head, mevp, me_list);
416 		}
417 	}
418 
419 	mevent_qunlock();
420 }
421 
422 static void
423 mevent_handle_pe(port_event_t *pe)
424 {
425 	struct mevent *mevp = pe->portev_user;
426 
427 	mevent_qunlock();
428 
429 	(*mevp->me_func)(mevp->me_fd, mevp->me_type, mevp->me_param);
430 
431 	mevent_qlock();
432 	if (!mevp->me_cq && !mevp->me_auto_requeue) {
433 		mevent_update_one(mevp);
434 	}
435 	mevent_qunlock();
436 }
437 #endif
438 
439 static struct mevent *
440 mevent_add_state(int tfd, enum ev_type type,
441 	   void (*func)(int, enum ev_type, void *), void *param,
442 	   int state)
443 {
444 	struct mevent *lp, *mevp;
445 
446 	if (tfd < 0 || func == NULL) {
447 		return (NULL);
448 	}
449 
450 	mevp = NULL;
451 
452 	mevent_qlock();
453 
454 	/*
455 	 * Verify that the fd/type tuple is not present in any list
456 	 */
457 	LIST_FOREACH(lp, &global_head, me_list) {
458 		if (type != EVF_TIMER && lp->me_fd == tfd &&
459 		    lp->me_type == type) {
460 			goto exit;
461 		}
462 	}
463 
464 	LIST_FOREACH(lp, &change_head, me_list) {
465 		if (type != EVF_TIMER && lp->me_fd == tfd &&
466 		    lp->me_type == type) {
467 			goto exit;
468 		}
469 	}
470 
471 	/*
472 	 * Allocate an entry, populate it, and add it to the change list.
473 	 */
474 	mevp = calloc(1, sizeof(struct mevent));
475 	if (mevp == NULL) {
476 		goto exit;
477 	}
478 
479 	if (type == EVF_TIMER) {
480 		mevp->me_msecs = tfd;
481 		mevp->me_timid = mevent_timid++;
482 	} else
483 		mevp->me_fd = tfd;
484 	mevp->me_type = type;
485 	mevp->me_func = func;
486 	mevp->me_param = param;
487 
488 	LIST_INSERT_HEAD(&change_head, mevp, me_list);
489 	mevp->me_cq = 1;
490 	mevp->me_state = state;
491 	mevent_notify();
492 
493 exit:
494 	mevent_qunlock();
495 
496 	return (mevp);
497 }
498 
499 struct mevent *
500 mevent_add(int tfd, enum ev_type type,
501 	   void (*func)(int, enum ev_type, void *), void *param)
502 {
503 
504 	return (mevent_add_state(tfd, type, func, param, EV_ADD));
505 }
506 
507 struct mevent *
508 mevent_add_disabled(int tfd, enum ev_type type,
509 		    void (*func)(int, enum ev_type, void *), void *param)
510 {
511 
512 	return (mevent_add_state(tfd, type, func, param, EV_ADD | EV_DISABLE));
513 }
514 
515 static int
516 mevent_update(struct mevent *evp, bool enable)
517 {
518 	int newstate;
519 
520 	mevent_qlock();
521 
522 	/*
523 	 * It's not possible to enable/disable a deleted event
524 	 */
525 	assert((evp->me_state & EV_DELETE) == 0);
526 
527 	newstate = evp->me_state;
528 	if (enable) {
529 		newstate |= EV_ENABLE;
530 		newstate &= ~EV_DISABLE;
531 	} else {
532 		newstate |= EV_DISABLE;
533 		newstate &= ~EV_ENABLE;
534 	}
535 
536 	/*
537 	 * No update needed if state isn't changing
538 	 */
539 	if (evp->me_state != newstate) {
540 		evp->me_state = newstate;
541 
542 		/*
543 		 * Place the entry onto the changed list if not
544 		 * already there.
545 		 */
546 		if (evp->me_cq == 0) {
547 			evp->me_cq = 1;
548 			LIST_REMOVE(evp, me_list);
549 			LIST_INSERT_HEAD(&change_head, evp, me_list);
550 			mevent_notify();
551 		}
552 	}
553 
554 	mevent_qunlock();
555 
556 	return (0);
557 }
558 
559 int
560 mevent_enable(struct mevent *evp)
561 {
562 
563 	return (mevent_update(evp, true));
564 }
565 
566 int
567 mevent_disable(struct mevent *evp)
568 {
569 
570 	return (mevent_update(evp, false));
571 }
572 
573 static int
574 mevent_delete_event(struct mevent *evp, int closefd)
575 {
576 	mevent_qlock();
577 
578 	/*
579          * Place the entry onto the changed list if not already there, and
580 	 * mark as to be deleted.
581          */
582         if (evp->me_cq == 0) {
583 		evp->me_cq = 1;
584 		LIST_REMOVE(evp, me_list);
585 		LIST_INSERT_HEAD(&change_head, evp, me_list);
586 		mevent_notify();
587         }
588 	evp->me_state = EV_DELETE;
589 
590 	if (closefd)
591 		evp->me_closefd = 1;
592 
593 	mevent_qunlock();
594 
595 	return (0);
596 }
597 
598 int
599 mevent_delete(struct mevent *evp)
600 {
601 
602 	return (mevent_delete_event(evp, 0));
603 }
604 
605 int
606 mevent_delete_close(struct mevent *evp)
607 {
608 
609 	return (mevent_delete_event(evp, 1));
610 }
611 
612 static void
613 mevent_set_name(void)
614 {
615 
616 	pthread_set_name_np(mevent_tid, "mevent");
617 }
618 
619 void
620 mevent_dispatch(void)
621 {
622 #ifdef __FreeBSD__
623 	struct kevent changelist[MEVENT_MAX];
624 	struct kevent eventlist[MEVENT_MAX];
625 	struct mevent *pipev;
626 	int mfd;
627 	int numev;
628 #else
629 	struct mevent *pipev;
630 	int portfd;
631 #endif
632 	int ret;
633 #ifndef WITHOUT_CAPSICUM
634 	cap_rights_t rights;
635 #endif
636 
637 	mevent_tid = pthread_self();
638 	mevent_set_name();
639 
640 #ifdef __FreeBSD__
641 	mfd = kqueue();
642 	assert(mfd > 0);
643 #else
644 	portfd = port_create();
645 	assert(portfd >= 0);
646 #endif
647 
648 #ifndef WITHOUT_CAPSICUM
649 	cap_rights_init(&rights, CAP_KQUEUE);
650 	if (caph_rights_limit(mfd, &rights) == -1)
651 		errx(EX_OSERR, "Unable to apply rights for sandbox");
652 #endif
653 
654 	/*
655 	 * Open the pipe that will be used for other threads to force
656 	 * the blocking kqueue call to exit by writing to it. Set the
657 	 * descriptor to non-blocking.
658 	 */
659 	ret = pipe(mevent_pipefd);
660 	if (ret < 0) {
661 		perror("pipe");
662 		exit(0);
663 	}
664 
665 #ifndef WITHOUT_CAPSICUM
666 	cap_rights_init(&rights, CAP_EVENT, CAP_READ, CAP_WRITE);
667 	if (caph_rights_limit(mevent_pipefd[0], &rights) == -1)
668 		errx(EX_OSERR, "Unable to apply rights for sandbox");
669 	if (caph_rights_limit(mevent_pipefd[1], &rights) == -1)
670 		errx(EX_OSERR, "Unable to apply rights for sandbox");
671 #endif
672 
673 	/*
674 	 * Add internal event handler for the pipe write fd
675 	 */
676 	pipev = mevent_add(mevent_pipefd[0], EVF_READ, mevent_pipe_read, NULL);
677 	assert(pipev != NULL);
678 
679 	for (;;) {
680 #ifdef __FreeBSD__
681 		/*
682 		 * Build changelist if required.
683 		 * XXX the changelist can be put into the blocking call
684 		 * to eliminate the extra syscall. Currently better for
685 		 * debug.
686 		 */
687 		numev = mevent_build(mfd, changelist);
688 		if (numev) {
689 			ret = kevent(mfd, changelist, numev, NULL, 0, NULL);
690 			if (ret == -1) {
691 				perror("Error return from kevent change");
692 			}
693 		}
694 
695 		/*
696 		 * Block awaiting events
697 		 */
698 		ret = kevent(mfd, NULL, 0, eventlist, MEVENT_MAX, NULL);
699 		if (ret == -1 && errno != EINTR) {
700 			perror("Error return from kevent monitor");
701 		}
702 
703 		/*
704 		 * Handle reported events
705 		 */
706 		mevent_handle(eventlist, ret);
707 
708 #else /* __FreeBSD__ */
709 		port_event_t pev;
710 
711 		/* Handle any pending updates */
712 		mevent_update_pending(portfd);
713 
714 		/* Block awaiting events */
715 		ret = port_get(portfd, &pev, NULL);
716 		if (ret != 0) {
717 			if (errno != EINTR)
718 				perror("Error return from port_get");
719 			continue;
720 		}
721 
722 		/* Handle reported event */
723 		mevent_handle_pe(&pev);
724 #endif /* __FreeBSD__ */
725 	}
726 }
727