xref: /freebsd/usr.sbin/bhyve/mevent.c (revision dc318a4ffabcbfa23bb56a33403aad36e6de30af) 
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 int mevent_timid = 43;
68 static int mevent_pipefd[2];
69 static pthread_mutex_t mevent_lmutex = PTHREAD_MUTEX_INITIALIZER;
70 
71 struct mevent {
72 	void	(*me_func)(int, enum ev_type, void *);
73 #define me_msecs me_fd
74 	int	me_fd;
75 	int	me_timid;
76 	enum ev_type me_type;
77 	void    *me_param;
78 	int	me_cq;
79 	int	me_state; /* Desired kevent flags. */
80 	int	me_closefd;
81 	LIST_ENTRY(mevent) me_list;
82 };
83 
84 static LIST_HEAD(listhead, mevent) global_head, change_head;
85 
86 static void
87 mevent_qlock(void)
88 {
89 	pthread_mutex_lock(&mevent_lmutex);
90 }
91 
92 static void
93 mevent_qunlock(void)
94 {
95 	pthread_mutex_unlock(&mevent_lmutex);
96 }
97 
98 static void
99 mevent_pipe_read(int fd, enum ev_type type, void *param)
100 {
101 	char buf[MEVENT_MAX];
102 	int status;
103 
104 	/*
105 	 * Drain the pipe read side. The fd is non-blocking so this is
106 	 * safe to do.
107 	 */
108 	do {
109 		status = read(fd, buf, sizeof(buf));
110 	} while (status == MEVENT_MAX);
111 }
112 
113 static void
114 mevent_notify(void)
115 {
116 	char c = '\0';
117 
118 	/*
119 	 * If calling from outside the i/o thread, write a byte on the
120 	 * pipe to force the i/o thread to exit the blocking kevent call.
121 	 */
122 	if (mevent_pipefd[1] != 0 && pthread_self() != mevent_tid) {
123 		write(mevent_pipefd[1], &c, 1);
124 	}
125 }
126 
127 static int
128 mevent_kq_filter(struct mevent *mevp)
129 {
130 	int retval;
131 
132 	retval = 0;
133 
134 	if (mevp->me_type == EVF_READ)
135 		retval = EVFILT_READ;
136 
137 	if (mevp->me_type == EVF_WRITE)
138 		retval = EVFILT_WRITE;
139 
140 	if (mevp->me_type == EVF_TIMER)
141 		retval = EVFILT_TIMER;
142 
143 	if (mevp->me_type == EVF_SIGNAL)
144 		retval = EVFILT_SIGNAL;
145 
146 	return (retval);
147 }
148 
149 static int
150 mevent_kq_flags(struct mevent *mevp)
151 {
152 	return (mevp->me_state);
153 }
154 
155 static int
156 mevent_kq_fflags(struct mevent *mevp)
157 {
158 	/* XXX nothing yet, perhaps EV_EOF for reads ? */
159 	return (0);
160 }
161 
162 static int
163 mevent_build(int mfd, struct kevent *kev)
164 {
165 	struct mevent *mevp, *tmpp;
166 	int i;
167 
168 	i = 0;
169 
170 	mevent_qlock();
171 
172 	LIST_FOREACH_SAFE(mevp, &change_head, me_list, tmpp) {
173 		if (mevp->me_closefd) {
174 			/*
175 			 * A close of the file descriptor will remove the
176 			 * event
177 			 */
178 			close(mevp->me_fd);
179 		} else {
180 			if (mevp->me_type == EVF_TIMER) {
181 				kev[i].ident = mevp->me_timid;
182 				kev[i].data = mevp->me_msecs;
183 			} else {
184 				kev[i].ident = mevp->me_fd;
185 				kev[i].data = 0;
186 			}
187 			kev[i].filter = mevent_kq_filter(mevp);
188 			kev[i].flags = mevent_kq_flags(mevp);
189 			kev[i].fflags = mevent_kq_fflags(mevp);
190 			kev[i].udata = mevp;
191 			i++;
192 		}
193 
194 		mevp->me_cq = 0;
195 		LIST_REMOVE(mevp, me_list);
196 
197 		if (mevp->me_state & EV_DELETE) {
198 			free(mevp);
199 		} else {
200 			/*
201 			 * We need to add the event only once, so we can
202 			 * reset the EV_ADD bit after it has been propagated
203 			 * to the kevent() arguments the first time.
204 			 */
205 			mevp->me_state &= ~EV_ADD;
206 			LIST_INSERT_HEAD(&global_head, mevp, me_list);
207 		}
208 
209 		assert(i < MEVENT_MAX);
210 	}
211 
212 	mevent_qunlock();
213 
214 	return (i);
215 }
216 
217 static void
218 mevent_handle(struct kevent *kev, int numev)
219 {
220 	struct mevent *mevp;
221 	int i;
222 
223 	for (i = 0; i < numev; i++) {
224 		mevp = kev[i].udata;
225 
226 		/* XXX check for EV_ERROR ? */
227 
228 		(*mevp->me_func)(mevp->me_fd, mevp->me_type, mevp->me_param);
229 	}
230 }
231 
232 static struct mevent *
233 mevent_add_state(int tfd, enum ev_type type,
234 	   void (*func)(int, enum ev_type, void *), void *param,
235 	   int state)
236 {
237 	struct mevent *lp, *mevp;
238 
239 	if (tfd < 0 || func == NULL) {
240 		return (NULL);
241 	}
242 
243 	mevp = NULL;
244 
245 	mevent_qlock();
246 
247 	/*
248 	 * Verify that the fd/type tuple is not present in any list
249 	 */
250 	LIST_FOREACH(lp, &global_head, me_list) {
251 		if (type != EVF_TIMER && lp->me_fd == tfd &&
252 		    lp->me_type == type) {
253 			goto exit;
254 		}
255 	}
256 
257 	LIST_FOREACH(lp, &change_head, me_list) {
258 		if (type != EVF_TIMER && lp->me_fd == tfd &&
259 		    lp->me_type == type) {
260 			goto exit;
261 		}
262 	}
263 
264 	/*
265 	 * Allocate an entry, populate it, and add it to the change list.
266 	 */
267 	mevp = calloc(1, sizeof(struct mevent));
268 	if (mevp == NULL) {
269 		goto exit;
270 	}
271 
272 	if (type == EVF_TIMER) {
273 		mevp->me_msecs = tfd;
274 		mevp->me_timid = mevent_timid++;
275 	} else
276 		mevp->me_fd = tfd;
277 	mevp->me_type = type;
278 	mevp->me_func = func;
279 	mevp->me_param = param;
280 
281 	LIST_INSERT_HEAD(&change_head, mevp, me_list);
282 	mevp->me_cq = 1;
283 	mevp->me_state = state;
284 	mevent_notify();
285 
286 exit:
287 	mevent_qunlock();
288 
289 	return (mevp);
290 }
291 
292 struct mevent *
293 mevent_add(int tfd, enum ev_type type,
294 	   void (*func)(int, enum ev_type, void *), void *param)
295 {
296 
297 	return (mevent_add_state(tfd, type, func, param, EV_ADD));
298 }
299 
300 struct mevent *
301 mevent_add_disabled(int tfd, enum ev_type type,
302 		    void (*func)(int, enum ev_type, void *), void *param)
303 {
304 
305 	return (mevent_add_state(tfd, type, func, param, EV_ADD | EV_DISABLE));
306 }
307 
308 static int
309 mevent_update(struct mevent *evp, bool enable)
310 {
311 	int newstate;
312 
313 	mevent_qlock();
314 
315 	/*
316 	 * It's not possible to enable/disable a deleted event
317 	 */
318 	assert((evp->me_state & EV_DELETE) == 0);
319 
320 	newstate = evp->me_state;
321 	if (enable) {
322 		newstate |= EV_ENABLE;
323 		newstate &= ~EV_DISABLE;
324 	} else {
325 		newstate |= EV_DISABLE;
326 		newstate &= ~EV_ENABLE;
327 	}
328 
329 	/*
330 	 * No update needed if state isn't changing
331 	 */
332 	if (evp->me_state != newstate) {
333 		evp->me_state = newstate;
334 
335 		/*
336 		 * Place the entry onto the changed list if not
337 		 * already there.
338 		 */
339 		if (evp->me_cq == 0) {
340 			evp->me_cq = 1;
341 			LIST_REMOVE(evp, me_list);
342 			LIST_INSERT_HEAD(&change_head, evp, me_list);
343 			mevent_notify();
344 		}
345 	}
346 
347 	mevent_qunlock();
348 
349 	return (0);
350 }
351 
352 int
353 mevent_enable(struct mevent *evp)
354 {
355 
356 	return (mevent_update(evp, true));
357 }
358 
359 int
360 mevent_disable(struct mevent *evp)
361 {
362 
363 	return (mevent_update(evp, false));
364 }
365 
366 static int
367 mevent_delete_event(struct mevent *evp, int closefd)
368 {
369 	mevent_qlock();
370 
371 	/*
372          * Place the entry onto the changed list if not already there, and
373 	 * mark as to be deleted.
374          */
375         if (evp->me_cq == 0) {
376 		evp->me_cq = 1;
377 		LIST_REMOVE(evp, me_list);
378 		LIST_INSERT_HEAD(&change_head, evp, me_list);
379 		mevent_notify();
380         }
381 	evp->me_state = EV_DELETE;
382 
383 	if (closefd)
384 		evp->me_closefd = 1;
385 
386 	mevent_qunlock();
387 
388 	return (0);
389 }
390 
391 int
392 mevent_delete(struct mevent *evp)
393 {
394 
395 	return (mevent_delete_event(evp, 0));
396 }
397 
398 int
399 mevent_delete_close(struct mevent *evp)
400 {
401 
402 	return (mevent_delete_event(evp, 1));
403 }
404 
405 static void
406 mevent_set_name(void)
407 {
408 
409 	pthread_set_name_np(mevent_tid, "mevent");
410 }
411 
412 void
413 mevent_dispatch(void)
414 {
415 	struct kevent changelist[MEVENT_MAX];
416 	struct kevent eventlist[MEVENT_MAX];
417 	struct mevent *pipev;
418 	int mfd;
419 	int numev;
420 	int ret;
421 #ifndef WITHOUT_CAPSICUM
422 	cap_rights_t rights;
423 #endif
424 
425 	mevent_tid = pthread_self();
426 	mevent_set_name();
427 
428 	mfd = kqueue();
429 	assert(mfd > 0);
430 
431 #ifndef WITHOUT_CAPSICUM
432 	cap_rights_init(&rights, CAP_KQUEUE);
433 	if (caph_rights_limit(mfd, &rights) == -1)
434 		errx(EX_OSERR, "Unable to apply rights for sandbox");
435 #endif
436 
437 	/*
438 	 * Open the pipe that will be used for other threads to force
439 	 * the blocking kqueue call to exit by writing to it. Set the
440 	 * descriptor to non-blocking.
441 	 */
442 	ret = pipe(mevent_pipefd);
443 	if (ret < 0) {
444 		perror("pipe");
445 		exit(0);
446 	}
447 
448 #ifndef WITHOUT_CAPSICUM
449 	cap_rights_init(&rights, CAP_EVENT, CAP_READ, CAP_WRITE);
450 	if (caph_rights_limit(mevent_pipefd[0], &rights) == -1)
451 		errx(EX_OSERR, "Unable to apply rights for sandbox");
452 	if (caph_rights_limit(mevent_pipefd[1], &rights) == -1)
453 		errx(EX_OSERR, "Unable to apply rights for sandbox");
454 #endif
455 
456 	/*
457 	 * Add internal event handler for the pipe write fd
458 	 */
459 	pipev = mevent_add(mevent_pipefd[0], EVF_READ, mevent_pipe_read, NULL);
460 	assert(pipev != NULL);
461 
462 	for (;;) {
463 		/*
464 		 * Build changelist if required.
465 		 * XXX the changelist can be put into the blocking call
466 		 * to eliminate the extra syscall. Currently better for
467 		 * debug.
468 		 */
469 		numev = mevent_build(mfd, changelist);
470 		if (numev) {
471 			ret = kevent(mfd, changelist, numev, NULL, 0, NULL);
472 			if (ret == -1) {
473 				perror("Error return from kevent change");
474 			}
475 		}
476 
477 		/*
478 		 * Block awaiting events
479 		 */
480 		ret = kevent(mfd, NULL, 0, eventlist, MEVENT_MAX, NULL);
481 		if (ret == -1 && errno != EINTR) {
482 			perror("Error return from kevent monitor");
483 		}
484 
485 		/*
486 		 * Handle reported events
487 		 */
488 		mevent_handle(eventlist, ret);
489 	}
490 }
491