xref: /freebsd/usr.sbin/bhyve/mevent.c (revision c1d255d3) 
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  * Micro event library for FreeBSD, designed for a single i/o thread
33  * using kqueue, and having events be persistent by default.
34  */
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include <assert.h>
40 #ifndef WITHOUT_CAPSICUM
41 #include <capsicum_helpers.h>
42 #endif
43 #include <err.h>
44 #include <errno.h>
45 #include <stdbool.h>
46 #include <stdlib.h>
47 #include <stdio.h>
48 #include <string.h>
49 #include <sysexits.h>
50 #include <unistd.h>
51 
52 #include <sys/types.h>
53 #ifndef WITHOUT_CAPSICUM
54 #include <sys/capsicum.h>
55 #endif
56 #include <sys/event.h>
57 #include <sys/time.h>
58 
59 #include <pthread.h>
60 #include <pthread_np.h>
61 
62 #include "mevent.h"
63 
64 #define	MEVENT_MAX	64
65 
66 static pthread_t mevent_tid;
67 static pthread_once_t mevent_once = PTHREAD_ONCE_INIT;
68 static int mevent_timid = 43;
69 static int mevent_pipefd[2];
70 static int mfd;
71 static pthread_mutex_t mevent_lmutex = PTHREAD_MUTEX_INITIALIZER;
72 
73 struct mevent {
74 	void	(*me_func)(int, enum ev_type, void *);
75 #define me_msecs me_fd
76 	int	me_fd;
77 	int	me_timid;
78 	enum ev_type me_type;
79 	void    *me_param;
80 	int	me_cq;
81 	int	me_state; /* Desired kevent flags. */
82 	int	me_closefd;
83 	int	me_fflags;
84 	LIST_ENTRY(mevent) me_list;
85 };
86 
87 static LIST_HEAD(listhead, mevent) global_head, change_head;
88 
89 static void
90 mevent_qlock(void)
91 {
92 	pthread_mutex_lock(&mevent_lmutex);
93 }
94 
95 static void
96 mevent_qunlock(void)
97 {
98 	pthread_mutex_unlock(&mevent_lmutex);
99 }
100 
101 static void
102 mevent_pipe_read(int fd, enum ev_type type, void *param)
103 {
104 	char buf[MEVENT_MAX];
105 	int status;
106 
107 	/*
108 	 * Drain the pipe read side. The fd is non-blocking so this is
109 	 * safe to do.
110 	 */
111 	do {
112 		status = read(fd, buf, sizeof(buf));
113 	} while (status == MEVENT_MAX);
114 }
115 
116 static void
117 mevent_notify(void)
118 {
119 	char c = '\0';
120 
121 	/*
122 	 * If calling from outside the i/o thread, write a byte on the
123 	 * pipe to force the i/o thread to exit the blocking kevent call.
124 	 */
125 	if (mevent_pipefd[1] != 0 && pthread_self() != mevent_tid) {
126 		write(mevent_pipefd[1], &c, 1);
127 	}
128 }
129 
130 static void
131 mevent_init(void)
132 {
133 #ifndef WITHOUT_CAPSICUM
134 	cap_rights_t rights;
135 #endif
136 
137 	mfd = kqueue();
138 	assert(mfd > 0);
139 
140 #ifndef WITHOUT_CAPSICUM
141 	cap_rights_init(&rights, CAP_KQUEUE);
142 	if (caph_rights_limit(mfd, &rights) == -1)
143 		errx(EX_OSERR, "Unable to apply rights for sandbox");
144 #endif
145 
146 	LIST_INIT(&change_head);
147 	LIST_INIT(&global_head);
148 }
149 
150 static int
151 mevent_kq_filter(struct mevent *mevp)
152 {
153 	int retval;
154 
155 	retval = 0;
156 
157 	if (mevp->me_type == EVF_READ)
158 		retval = EVFILT_READ;
159 
160 	if (mevp->me_type == EVF_WRITE)
161 		retval = EVFILT_WRITE;
162 
163 	if (mevp->me_type == EVF_TIMER)
164 		retval = EVFILT_TIMER;
165 
166 	if (mevp->me_type == EVF_SIGNAL)
167 		retval = EVFILT_SIGNAL;
168 
169 	if (mevp->me_type == EVF_VNODE)
170 		retval = EVFILT_VNODE;
171 
172 	return (retval);
173 }
174 
175 static int
176 mevent_kq_flags(struct mevent *mevp)
177 {
178 	return (mevp->me_state);
179 }
180 
181 static int
182 mevent_kq_fflags(struct mevent *mevp)
183 {
184 	int retval;
185 
186 	retval = 0;
187 
188 	switch (mevp->me_type) {
189 	case EVF_VNODE:
190 		if ((mevp->me_fflags & EVFF_ATTRIB) != 0)
191 			retval |= NOTE_ATTRIB;
192 		break;
193 	}
194 
195 	return (retval);
196 }
197 
198 static void
199 mevent_populate(struct mevent *mevp, struct kevent *kev)
200 {
201 	if (mevp->me_type == EVF_TIMER) {
202 		kev->ident = mevp->me_timid;
203 		kev->data = mevp->me_msecs;
204 	} else {
205 		kev->ident = mevp->me_fd;
206 		kev->data = 0;
207 	}
208 	kev->filter = mevent_kq_filter(mevp);
209 	kev->flags = mevent_kq_flags(mevp);
210 	kev->fflags = mevent_kq_fflags(mevp);
211 	kev->udata = mevp;
212 }
213 
214 static int
215 mevent_build(struct kevent *kev)
216 {
217 	struct mevent *mevp, *tmpp;
218 	int i;
219 
220 	i = 0;
221 
222 	mevent_qlock();
223 
224 	LIST_FOREACH_SAFE(mevp, &change_head, me_list, tmpp) {
225 		if (mevp->me_closefd) {
226 			/*
227 			 * A close of the file descriptor will remove the
228 			 * event
229 			 */
230 			close(mevp->me_fd);
231 		} else {
232 			assert((mevp->me_state & EV_ADD) == 0);
233 			mevent_populate(mevp, &kev[i]);
234 			i++;
235 		}
236 
237 		mevp->me_cq = 0;
238 		LIST_REMOVE(mevp, me_list);
239 
240 		if (mevp->me_state & EV_DELETE) {
241 			free(mevp);
242 		} else {
243 			LIST_INSERT_HEAD(&global_head, mevp, me_list);
244 		}
245 
246 		assert(i < MEVENT_MAX);
247 	}
248 
249 	mevent_qunlock();
250 
251 	return (i);
252 }
253 
254 static void
255 mevent_handle(struct kevent *kev, int numev)
256 {
257 	struct mevent *mevp;
258 	int i;
259 
260 	for (i = 0; i < numev; i++) {
261 		mevp = kev[i].udata;
262 
263 		/* XXX check for EV_ERROR ? */
264 
265 		(*mevp->me_func)(mevp->me_fd, mevp->me_type, mevp->me_param);
266 	}
267 }
268 
269 static struct mevent *
270 mevent_add_state(int tfd, enum ev_type type,
271 	   void (*func)(int, enum ev_type, void *), void *param,
272 	   int state, int fflags)
273 {
274 	struct kevent kev;
275 	struct mevent *lp, *mevp;
276 	int ret;
277 
278 	if (tfd < 0 || func == NULL) {
279 		return (NULL);
280 	}
281 
282 	mevp = NULL;
283 
284 	pthread_once(&mevent_once, mevent_init);
285 
286 	mevent_qlock();
287 
288 	/*
289 	 * Verify that the fd/type tuple is not present in any list
290 	 */
291 	LIST_FOREACH(lp, &global_head, me_list) {
292 		if (type != EVF_TIMER && lp->me_fd == tfd &&
293 		    lp->me_type == type) {
294 			goto exit;
295 		}
296 	}
297 
298 	LIST_FOREACH(lp, &change_head, me_list) {
299 		if (type != EVF_TIMER && lp->me_fd == tfd &&
300 		    lp->me_type == type) {
301 			goto exit;
302 		}
303 	}
304 
305 	/*
306 	 * Allocate an entry and populate it.
307 	 */
308 	mevp = calloc(1, sizeof(struct mevent));
309 	if (mevp == NULL) {
310 		goto exit;
311 	}
312 
313 	if (type == EVF_TIMER) {
314 		mevp->me_msecs = tfd;
315 		mevp->me_timid = mevent_timid++;
316 	} else
317 		mevp->me_fd = tfd;
318 	mevp->me_type = type;
319 	mevp->me_func = func;
320 	mevp->me_param = param;
321 	mevp->me_state = state;
322 	mevp->me_fflags = fflags;
323 
324 	/*
325 	 * Try to add the event.  If this fails, report the failure to
326 	 * the caller.
327 	 */
328 	mevent_populate(mevp, &kev);
329 	ret = kevent(mfd, &kev, 1, NULL, 0, NULL);
330 	if (ret == -1) {
331 		free(mevp);
332 		mevp = NULL;
333 		goto exit;
334 	}
335 
336 	mevp->me_state &= ~EV_ADD;
337 	LIST_INSERT_HEAD(&global_head, mevp, me_list);
338 
339 exit:
340 	mevent_qunlock();
341 
342 	return (mevp);
343 }
344 
345 struct mevent *
346 mevent_add(int tfd, enum ev_type type,
347 	   void (*func)(int, enum ev_type, void *), void *param)
348 {
349 
350 	return (mevent_add_state(tfd, type, func, param, EV_ADD, 0));
351 }
352 
353 struct mevent *
354 mevent_add_flags(int tfd, enum ev_type type, int fflags,
355 		 void (*func)(int, enum ev_type, void *), void *param)
356 {
357 
358 	return (mevent_add_state(tfd, type, func, param, EV_ADD, fflags));
359 }
360 
361 struct mevent *
362 mevent_add_disabled(int tfd, enum ev_type type,
363 		    void (*func)(int, enum ev_type, void *), void *param)
364 {
365 
366 	return (mevent_add_state(tfd, type, func, param, EV_ADD | EV_DISABLE, 0));
367 }
368 
369 static int
370 mevent_update(struct mevent *evp, bool enable)
371 {
372 	int newstate;
373 
374 	mevent_qlock();
375 
376 	/*
377 	 * It's not possible to enable/disable a deleted event
378 	 */
379 	assert((evp->me_state & EV_DELETE) == 0);
380 
381 	newstate = evp->me_state;
382 	if (enable) {
383 		newstate |= EV_ENABLE;
384 		newstate &= ~EV_DISABLE;
385 	} else {
386 		newstate |= EV_DISABLE;
387 		newstate &= ~EV_ENABLE;
388 	}
389 
390 	/*
391 	 * No update needed if state isn't changing
392 	 */
393 	if (evp->me_state != newstate) {
394 		evp->me_state = newstate;
395 
396 		/*
397 		 * Place the entry onto the changed list if not
398 		 * already there.
399 		 */
400 		if (evp->me_cq == 0) {
401 			evp->me_cq = 1;
402 			LIST_REMOVE(evp, me_list);
403 			LIST_INSERT_HEAD(&change_head, evp, me_list);
404 			mevent_notify();
405 		}
406 	}
407 
408 	mevent_qunlock();
409 
410 	return (0);
411 }
412 
413 int
414 mevent_enable(struct mevent *evp)
415 {
416 
417 	return (mevent_update(evp, true));
418 }
419 
420 int
421 mevent_disable(struct mevent *evp)
422 {
423 
424 	return (mevent_update(evp, false));
425 }
426 
427 static int
428 mevent_delete_event(struct mevent *evp, int closefd)
429 {
430 	mevent_qlock();
431 
432 	/*
433          * Place the entry onto the changed list if not already there, and
434 	 * mark as to be deleted.
435          */
436         if (evp->me_cq == 0) {
437 		evp->me_cq = 1;
438 		LIST_REMOVE(evp, me_list);
439 		LIST_INSERT_HEAD(&change_head, evp, me_list);
440 		mevent_notify();
441         }
442 	evp->me_state = EV_DELETE;
443 
444 	if (closefd)
445 		evp->me_closefd = 1;
446 
447 	mevent_qunlock();
448 
449 	return (0);
450 }
451 
452 int
453 mevent_delete(struct mevent *evp)
454 {
455 
456 	return (mevent_delete_event(evp, 0));
457 }
458 
459 int
460 mevent_delete_close(struct mevent *evp)
461 {
462 
463 	return (mevent_delete_event(evp, 1));
464 }
465 
466 static void
467 mevent_set_name(void)
468 {
469 
470 	pthread_set_name_np(mevent_tid, "mevent");
471 }
472 
473 void
474 mevent_dispatch(void)
475 {
476 	struct kevent changelist[MEVENT_MAX];
477 	struct kevent eventlist[MEVENT_MAX];
478 	struct mevent *pipev;
479 	int numev;
480 	int ret;
481 #ifndef WITHOUT_CAPSICUM
482 	cap_rights_t rights;
483 #endif
484 
485 	mevent_tid = pthread_self();
486 	mevent_set_name();
487 
488 	pthread_once(&mevent_once, mevent_init);
489 
490 	/*
491 	 * Open the pipe that will be used for other threads to force
492 	 * the blocking kqueue call to exit by writing to it. Set the
493 	 * descriptor to non-blocking.
494 	 */
495 	ret = pipe(mevent_pipefd);
496 	if (ret < 0) {
497 		perror("pipe");
498 		exit(0);
499 	}
500 
501 #ifndef WITHOUT_CAPSICUM
502 	cap_rights_init(&rights, CAP_EVENT, CAP_READ, CAP_WRITE);
503 	if (caph_rights_limit(mevent_pipefd[0], &rights) == -1)
504 		errx(EX_OSERR, "Unable to apply rights for sandbox");
505 	if (caph_rights_limit(mevent_pipefd[1], &rights) == -1)
506 		errx(EX_OSERR, "Unable to apply rights for sandbox");
507 #endif
508 
509 	/*
510 	 * Add internal event handler for the pipe write fd
511 	 */
512 	pipev = mevent_add(mevent_pipefd[0], EVF_READ, mevent_pipe_read, NULL);
513 	assert(pipev != NULL);
514 
515 	for (;;) {
516 		/*
517 		 * Build changelist if required.
518 		 * XXX the changelist can be put into the blocking call
519 		 * to eliminate the extra syscall. Currently better for
520 		 * debug.
521 		 */
522 		numev = mevent_build(changelist);
523 		if (numev) {
524 			ret = kevent(mfd, changelist, numev, NULL, 0, NULL);
525 			if (ret == -1) {
526 				perror("Error return from kevent change");
527 			}
528 		}
529 
530 		/*
531 		 * Block awaiting events
532 		 */
533 		ret = kevent(mfd, NULL, 0, eventlist, MEVENT_MAX, NULL);
534 		if (ret == -1 && errno != EINTR) {
535 			perror("Error return from kevent monitor");
536 		}
537 
538 		/*
539 		 * Handle reported events
540 		 */
541 		mevent_handle(eventlist, ret);
542 	}
543 }
544