xref: /freebsd/sys/kern/kern_fail.c (revision fd5e3f3ec6c6248e892c9e7b2f17da3bfe7b6837)
1 /*-
2  * Copyright (c) 2009 Isilon Inc http://www.isilon.com/
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  */
25 /**
26  * @file
27  *
28  * fail(9) Facility.
29  *
30  * @ingroup failpoint_private
31  */
32 /**
33  * @defgroup failpoint fail(9) Facility
34  *
35  * Failpoints allow for injecting fake errors into running code on the fly,
36  * without modifying code or recompiling with flags.  Failpoints are always
37  * present, and are very efficient when disabled.  Failpoints are described
38  * in man fail(9).
39  */
40 /**
41  * @defgroup failpoint_private Private fail(9) Implementation functions
42  *
43  * Private implementations for the actual failpoint code.
44  *
45  * @ingroup failpoint
46  */
47 /**
48  * @addtogroup failpoint_private
49  * @{
50  */
51 
52 #include <sys/cdefs.h>
53 __FBSDID("$FreeBSD$");
54 
55 #include "opt_stack.h"
56 
57 #include <sys/ctype.h>
58 #include <sys/errno.h>
59 #include <sys/fail.h>
60 #include <sys/kernel.h>
61 #include <sys/libkern.h>
62 #include <sys/limits.h>
63 #include <sys/lock.h>
64 #include <sys/malloc.h>
65 #include <sys/mutex.h>
66 #include <sys/proc.h>
67 #include <sys/sbuf.h>
68 #include <sys/sleepqueue.h>
69 #include <sys/sx.h>
70 #include <sys/sysctl.h>
71 #include <sys/types.h>
72 
73 #include <machine/atomic.h>
74 #include <machine/stdarg.h>
75 
76 #ifdef ILOG_DEFINE_FOR_FILE
77 ILOG_DEFINE_FOR_FILE(L_ISI_FAIL_POINT, L_ILOG, fail_point);
78 #endif
79 
80 static MALLOC_DEFINE(M_FAIL_POINT, "Fail Points", "fail points system");
81 #define fp_free(ptr) free(ptr, M_FAIL_POINT)
82 #define fp_malloc(size, flags) malloc((size), M_FAIL_POINT, (flags))
83 #define fs_free(ptr) fp_free(ptr)
84 #define fs_malloc() fp_malloc(sizeof(struct fail_point_setting), \
85     M_WAITOK | M_ZERO)
86 
87 /**
88  * These define the wchans that are used for sleeping, pausing respectively.
89  * They are chosen arbitrarily but need to be distinct to the failpoint and
90  * the sleep/pause distinction.
91  */
92 #define FP_SLEEP_CHANNEL(fp) (void*)(fp)
93 #define FP_PAUSE_CHANNEL(fp) __DEVOLATILE(void*, &fp->fp_setting)
94 
95 /**
96  * Don't allow more than this many entries in a fail point set by sysctl.
97  * The 99.99...% case is to have 1 entry.  I can't imagine having this many
98  * entries, so it should not limit us.  Saves on re-mallocs while holding
99  * a non-sleepable lock.
100  */
101 #define FP_MAX_ENTRY_COUNT 20
102 
103 /* Used to drain sbufs to the sysctl output */
104 int fail_sysctl_drain_func(void *, const char *, int);
105 
106 /* Head of tailq of struct fail_point_entry */
107 TAILQ_HEAD(fail_point_entry_queue, fail_point_entry);
108 
109 /**
110  * fp entries garbage list; outstanding entries are cleaned up in the
111  * garbage collector
112  */
113 STAILQ_HEAD(fail_point_setting_garbage, fail_point_setting);
114 static struct fail_point_setting_garbage fp_setting_garbage =
115         STAILQ_HEAD_INITIALIZER(fp_setting_garbage);
116 static struct mtx mtx_garbage_list;
117 MTX_SYSINIT(mtx_garbage_list, &mtx_garbage_list, "fail point garbage mtx",
118         MTX_SPIN);
119 
120 static struct sx sx_fp_set;
121 SX_SYSINIT(sx_fp_set, &sx_fp_set, "fail point set sx");
122 
123 /**
124  * Failpoint types.
125  * Don't change these without changing fail_type_strings in fail.c.
126  * @ingroup failpoint_private
127  */
128 enum fail_point_t {
129 	FAIL_POINT_OFF,		/**< don't fail */
130 	FAIL_POINT_PANIC,	/**< panic */
131 	FAIL_POINT_RETURN,	/**< return an errorcode */
132 	FAIL_POINT_BREAK,	/**< break into the debugger */
133 	FAIL_POINT_PRINT,	/**< print a message */
134 	FAIL_POINT_SLEEP,	/**< sleep for some msecs */
135 	FAIL_POINT_PAUSE,	/**< sleep until failpoint is set to off */
136 	FAIL_POINT_YIELD,	/**< yield the cpu */
137 	FAIL_POINT_DELAY,	/**< busy wait the cpu */
138 	FAIL_POINT_NUMTYPES,
139 	FAIL_POINT_INVALID = -1
140 };
141 
142 static struct {
143 	const char *name;
144 	int	nmlen;
145 } fail_type_strings[] = {
146 #define	FP_TYPE_NM_LEN(s)	{ s, sizeof(s) - 1 }
147 	[FAIL_POINT_OFF] =	FP_TYPE_NM_LEN("off"),
148 	[FAIL_POINT_PANIC] =	FP_TYPE_NM_LEN("panic"),
149 	[FAIL_POINT_RETURN] =	FP_TYPE_NM_LEN("return"),
150 	[FAIL_POINT_BREAK] =	FP_TYPE_NM_LEN("break"),
151 	[FAIL_POINT_PRINT] =	FP_TYPE_NM_LEN("print"),
152 	[FAIL_POINT_SLEEP] =	FP_TYPE_NM_LEN("sleep"),
153 	[FAIL_POINT_PAUSE] =	FP_TYPE_NM_LEN("pause"),
154 	[FAIL_POINT_YIELD] =	FP_TYPE_NM_LEN("yield"),
155 	[FAIL_POINT_DELAY] =	FP_TYPE_NM_LEN("delay"),
156 };
157 
158 #define FE_COUNT_UNTRACKED (INT_MIN)
159 
160 /**
161  * Internal structure tracking a single term of a complete failpoint.
162  * @ingroup failpoint_private
163  */
164 struct fail_point_entry {
165 	volatile bool	fe_stale;
166 	enum fail_point_t	fe_type;	/**< type of entry */
167 	int		fe_arg;		/**< argument to type (e.g. return value) */
168 	int		fe_prob;	/**< likelihood of firing in millionths */
169 	int32_t		fe_count;	/**< number of times to fire, -1 means infinite */
170 	pid_t		fe_pid;		/**< only fail for this process */
171 	struct fail_point	*fe_parent;	/**< backpointer to fp */
172 	TAILQ_ENTRY(fail_point_entry)	fe_entries; /**< next entry ptr */
173 };
174 
175 struct fail_point_setting {
176 	STAILQ_ENTRY(fail_point_setting) fs_garbage_link;
177 	struct fail_point_entry_queue fp_entry_queue;
178 	struct fail_point * fs_parent;
179 	struct mtx feq_mtx; /* Gives fail_point_pause something to do.  */
180 };
181 
182 /**
183  * Defines stating the equivalent of probablilty one (100%)
184  */
185 enum {
186 	PROB_MAX = 1000000,	/* probability between zero and this number */
187 	PROB_DIGITS = 6		/* number of zero's in above number */
188 };
189 
190 /* Get a ref on an fp's fp_setting */
191 static inline struct fail_point_setting *fail_point_setting_get_ref(
192         struct fail_point *fp);
193 /* Release a ref on an fp_setting */
194 static inline void fail_point_setting_release_ref(struct fail_point *fp);
195 /* Allocate and initialize a struct fail_point_setting */
196 static struct fail_point_setting *fail_point_setting_new(struct
197         fail_point *);
198 /* Free a struct fail_point_setting */
199 static void fail_point_setting_destroy(struct fail_point_setting *fp_setting);
200 /* Allocate and initialize a struct fail_point_entry */
201 static struct fail_point_entry *fail_point_entry_new(struct
202         fail_point_setting *);
203 /* Free a struct fail_point_entry */
204 static void fail_point_entry_destroy(struct fail_point_entry *fp_entry);
205 /* Append fp setting to garbage list */
206 static inline void fail_point_setting_garbage_append(
207         struct fail_point_setting *fp_setting);
208 /* Swap fp's setting with fp_setting_new */
209 static inline struct fail_point_setting *
210         fail_point_swap_settings(struct fail_point *fp,
211         struct fail_point_setting *fp_setting_new);
212 /* Free up any zero-ref setting in the garbage queue */
213 static void fail_point_garbage_collect(void);
214 /* If this fail point's setting are empty, then swap it out to NULL. */
215 static inline void fail_point_eval_swap_out(struct fail_point *fp,
216         struct fail_point_setting *fp_setting);
217 
218 bool
219 fail_point_is_off(struct fail_point *fp)
220 {
221 	bool return_val;
222 	struct fail_point_setting *fp_setting;
223 	struct fail_point_entry *ent;
224 
225 	return_val = true;
226 
227 	fp_setting = fail_point_setting_get_ref(fp);
228 	if (fp_setting != NULL) {
229 		TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue,
230 		    fe_entries) {
231 			if (!ent->fe_stale) {
232 				return_val = false;
233 				break;
234 			}
235 		}
236 	}
237 	fail_point_setting_release_ref(fp);
238 
239 	return (return_val);
240 }
241 
242 /* Allocate and initialize a struct fail_point_setting */
243 static struct fail_point_setting *
244 fail_point_setting_new(struct fail_point *fp)
245 {
246 	struct fail_point_setting *fs_new;
247 
248 	fs_new = fs_malloc();
249 	fs_new->fs_parent = fp;
250 	TAILQ_INIT(&fs_new->fp_entry_queue);
251 	mtx_init(&fs_new->feq_mtx, "fail point entries", NULL, MTX_SPIN);
252 
253 	fail_point_setting_garbage_append(fs_new);
254 
255 	return (fs_new);
256 }
257 
258 /* Free a struct fail_point_setting */
259 static void
260 fail_point_setting_destroy(struct fail_point_setting *fp_setting)
261 {
262 	struct fail_point_entry *ent;
263 
264 	while (!TAILQ_EMPTY(&fp_setting->fp_entry_queue)) {
265 		ent = TAILQ_FIRST(&fp_setting->fp_entry_queue);
266 		TAILQ_REMOVE(&fp_setting->fp_entry_queue, ent, fe_entries);
267 		fail_point_entry_destroy(ent);
268 	}
269 
270 	fs_free(fp_setting);
271 }
272 
273 /* Allocate and initialize a struct fail_point_entry */
274 static struct fail_point_entry *
275 fail_point_entry_new(struct fail_point_setting *fp_setting)
276 {
277 	struct fail_point_entry *fp_entry;
278 
279 	fp_entry = fp_malloc(sizeof(struct fail_point_entry),
280 	        M_WAITOK | M_ZERO);
281 	fp_entry->fe_parent = fp_setting->fs_parent;
282 	fp_entry->fe_prob = PROB_MAX;
283 	fp_entry->fe_pid = NO_PID;
284 	fp_entry->fe_count = FE_COUNT_UNTRACKED;
285 	TAILQ_INSERT_TAIL(&fp_setting->fp_entry_queue, fp_entry,
286 	        fe_entries);
287 
288 	return (fp_entry);
289 }
290 
291 /* Free a struct fail_point_entry */
292 static void
293 fail_point_entry_destroy(struct fail_point_entry *fp_entry)
294 {
295 
296 	fp_free(fp_entry);
297 }
298 
299 /* Get a ref on an fp's fp_setting */
300 static inline struct fail_point_setting *
301 fail_point_setting_get_ref(struct fail_point *fp)
302 {
303 	struct fail_point_setting *fp_setting;
304 
305 	/* Invariant: if we have a ref, our pointer to fp_setting is safe */
306 	atomic_add_acq_32(&fp->fp_ref_cnt, 1);
307 	fp_setting = fp->fp_setting;
308 
309 	return (fp_setting);
310 }
311 
312 /* Release a ref on an fp_setting */
313 static inline void
314 fail_point_setting_release_ref(struct fail_point *fp)
315 {
316 
317 	KASSERT(&fp->fp_ref_cnt > 0, ("Attempting to deref w/no refs"));
318 	atomic_subtract_rel_32(&fp->fp_ref_cnt, 1);
319 }
320 
321 /* Append fp entries to fp garbage list */
322 static inline void
323 fail_point_setting_garbage_append(struct fail_point_setting *fp_setting)
324 {
325 
326 	mtx_lock_spin(&mtx_garbage_list);
327 	STAILQ_INSERT_TAIL(&fp_setting_garbage, fp_setting,
328 	        fs_garbage_link);
329 	mtx_unlock_spin(&mtx_garbage_list);
330 }
331 
332 /* Swap fp's entries with fp_setting_new */
333 static struct fail_point_setting *
334 fail_point_swap_settings(struct fail_point *fp,
335         struct fail_point_setting *fp_setting_new)
336 {
337 	struct fail_point_setting *fp_setting_old;
338 
339 	fp_setting_old = fp->fp_setting;
340 	fp->fp_setting = fp_setting_new;
341 
342 	return (fp_setting_old);
343 }
344 
345 static inline void
346 fail_point_eval_swap_out(struct fail_point *fp,
347         struct fail_point_setting *fp_setting)
348 {
349 
350 	/* We may have already been swapped out and replaced; ignore. */
351 	if (fp->fp_setting == fp_setting)
352 		fail_point_swap_settings(fp, NULL);
353 }
354 
355 /* Free up any zero-ref entries in the garbage queue */
356 static void
357 fail_point_garbage_collect(void)
358 {
359 	struct fail_point_setting *fs_current, *fs_next;
360 	struct fail_point_setting_garbage fp_ents_free_list;
361 
362 	/**
363 	  * We will transfer the entries to free to fp_ents_free_list while holding
364 	  * the spin mutex, then free it after we drop the lock. This avoids
365 	  * triggering witness due to sleepable mutexes in the memory
366 	  * allocator.
367 	  */
368 	STAILQ_INIT(&fp_ents_free_list);
369 
370 	mtx_lock_spin(&mtx_garbage_list);
371 	STAILQ_FOREACH_SAFE(fs_current, &fp_setting_garbage, fs_garbage_link,
372 	    fs_next) {
373 		if (fs_current->fs_parent->fp_setting != fs_current &&
374 		        fs_current->fs_parent->fp_ref_cnt == 0) {
375 			STAILQ_REMOVE(&fp_setting_garbage, fs_current,
376 			        fail_point_setting, fs_garbage_link);
377 			STAILQ_INSERT_HEAD(&fp_ents_free_list, fs_current,
378 			        fs_garbage_link);
379 		}
380 	}
381 	mtx_unlock_spin(&mtx_garbage_list);
382 
383 	STAILQ_FOREACH_SAFE(fs_current, &fp_ents_free_list, fs_garbage_link,
384 	        fs_next)
385 		fail_point_setting_destroy(fs_current);
386 }
387 
388 /* Drain out all refs from this fail point */
389 static inline void
390 fail_point_drain(struct fail_point *fp, int expected_ref)
391 {
392 	struct fail_point_setting *entries;
393 
394 	entries = fail_point_swap_settings(fp, NULL);
395 	/**
396 	 * We have unpaused all threads; so we will wait no longer
397 	 * than the time taken for the longest remaining sleep, or
398 	 * the length of time of a long-running code block.
399 	 */
400 	while (fp->fp_ref_cnt > expected_ref) {
401 		wakeup(FP_PAUSE_CHANNEL(fp));
402 		tsleep(&fp, PWAIT, "fail_point_drain", hz / 100);
403 	}
404 	fail_point_swap_settings(fp, entries);
405 }
406 
407 static inline void
408 fail_point_pause(struct fail_point *fp, enum fail_point_return_code *pret,
409         struct mtx *mtx_sleep)
410 {
411 
412 	if (fp->fp_pre_sleep_fn)
413 		fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
414 
415 	msleep_spin(FP_PAUSE_CHANNEL(fp), mtx_sleep, "failpt", 0);
416 
417 	if (fp->fp_post_sleep_fn)
418 		fp->fp_post_sleep_fn(fp->fp_post_sleep_arg);
419 }
420 
421 static inline void
422 fail_point_sleep(struct fail_point *fp, int msecs,
423         enum fail_point_return_code *pret)
424 {
425 	int timo;
426 
427 	/* Convert from millisecs to ticks, rounding up */
428 	timo = howmany((int64_t)msecs * hz, 1000L);
429 
430 	if (timo > 0) {
431 		if (!(fp->fp_flags & FAIL_POINT_USE_TIMEOUT_PATH)) {
432 			if (fp->fp_pre_sleep_fn)
433 				fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
434 
435 			tsleep(FP_SLEEP_CHANNEL(fp), PWAIT, "failpt", timo);
436 
437 			if (fp->fp_post_sleep_fn)
438 				fp->fp_post_sleep_fn(fp->fp_post_sleep_arg);
439 		} else {
440 			if (fp->fp_pre_sleep_fn)
441 				fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
442 
443 			timeout(fp->fp_post_sleep_fn, fp->fp_post_sleep_arg,
444 			    timo);
445 			*pret = FAIL_POINT_RC_QUEUED;
446 		}
447 	}
448 }
449 
450 static char *parse_fail_point(struct fail_point_setting *, char *);
451 static char *parse_term(struct fail_point_setting *, char *);
452 static char *parse_number(int *out_units, int *out_decimal, char *);
453 static char *parse_type(struct fail_point_entry *, char *);
454 
455 /**
456  * Initialize a fail_point.  The name is formed in a printf-like fashion
457  * from "fmt" and subsequent arguments.  This function is generally used
458  * for custom failpoints located at odd places in the sysctl tree, and is
459  * not explicitly needed for standard in-line-declared failpoints.
460  *
461  * @ingroup failpoint
462  */
463 void
464 fail_point_init(struct fail_point *fp, const char *fmt, ...)
465 {
466 	va_list ap;
467 	char *name;
468 	int n;
469 
470 	fp->fp_setting = NULL;
471 	fp->fp_flags = 0;
472 
473 	/* Figure out the size of the name. */
474 	va_start(ap, fmt);
475 	n = vsnprintf(NULL, 0, fmt, ap);
476 	va_end(ap);
477 
478 	/* Allocate the name and fill it in. */
479 	name = fp_malloc(n + 1, M_WAITOK);
480 	if (name != NULL) {
481 		va_start(ap, fmt);
482 		vsnprintf(name, n + 1, fmt, ap);
483 		va_end(ap);
484 	}
485 	fp->fp_name = name;
486 	fp->fp_location = "";
487 	fp->fp_flags |= FAIL_POINT_DYNAMIC_NAME;
488 	fp->fp_pre_sleep_fn = NULL;
489 	fp->fp_pre_sleep_arg = NULL;
490 	fp->fp_post_sleep_fn = NULL;
491 	fp->fp_post_sleep_arg = NULL;
492 }
493 
494 /**
495  * Free the resources held by a fail_point, and wake any paused threads.
496  * Thou shalt not allow threads to hit this fail point after you enter this
497  * function, nor shall you call this multiple times for a given fp.
498  * @ingroup failpoint
499  */
500 void
501 fail_point_destroy(struct fail_point *fp)
502 {
503 
504 	fail_point_drain(fp, 0);
505 
506 	if ((fp->fp_flags & FAIL_POINT_DYNAMIC_NAME) != 0) {
507 		fp_free(__DECONST(void *, fp->fp_name));
508 		fp->fp_name = NULL;
509 	}
510 	fp->fp_flags = 0;
511 
512 	sx_xlock(&sx_fp_set);
513 	fail_point_garbage_collect();
514 	sx_xunlock(&sx_fp_set);
515 }
516 
517 /**
518  * This does the real work of evaluating a fail point. If the fail point tells
519  * us to return a value, this function returns 1 and fills in 'return_value'
520  * (return_value is allowed to be null). If the fail point tells us to panic,
521  * we never return. Otherwise we just return 0 after doing some work, which
522  * means "keep going".
523  */
524 enum fail_point_return_code
525 fail_point_eval_nontrivial(struct fail_point *fp, int *return_value)
526 {
527 	bool execute = false;
528 	struct fail_point_entry *ent;
529 	struct fail_point_setting *fp_setting;
530 	enum fail_point_return_code ret;
531 	int cont;
532 	int count;
533 	int msecs;
534 	int usecs;
535 
536 	ret = FAIL_POINT_RC_CONTINUE;
537 	cont = 0; /* don't continue by default */
538 
539 	fp_setting = fail_point_setting_get_ref(fp);
540 	if (fp_setting == NULL)
541 		goto abort;
542 
543 	TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
544 
545 		if (ent->fe_stale)
546 			continue;
547 
548 		if (ent->fe_prob < PROB_MAX &&
549 		    ent->fe_prob < random() % PROB_MAX)
550 			continue;
551 
552 		if (ent->fe_pid != NO_PID && ent->fe_pid != curproc->p_pid)
553 			continue;
554 
555 		if (ent->fe_count != FE_COUNT_UNTRACKED) {
556 			count = ent->fe_count;
557 			while (count > 0) {
558 				if (atomic_cmpset_32(&ent->fe_count, count, count - 1)) {
559 					count--;
560 					execute = true;
561 					break;
562 				}
563 				count = ent->fe_count;
564 			}
565 			if (execute == false)
566 				/* We lost the race; consider the entry stale and bail now */
567 				continue;
568 			if (count == 0)
569 				ent->fe_stale = true;
570 		}
571 
572 		switch (ent->fe_type) {
573 		case FAIL_POINT_PANIC:
574 			panic("fail point %s panicking", fp->fp_name);
575 			/* NOTREACHED */
576 
577 		case FAIL_POINT_RETURN:
578 			if (return_value != NULL)
579 				*return_value = ent->fe_arg;
580 			ret = FAIL_POINT_RC_RETURN;
581 			break;
582 
583 		case FAIL_POINT_BREAK:
584 			printf("fail point %s breaking to debugger\n",
585 			        fp->fp_name);
586 			breakpoint();
587 			break;
588 
589 		case FAIL_POINT_PRINT:
590 			printf("fail point %s executing\n", fp->fp_name);
591 			cont = ent->fe_arg;
592 			break;
593 
594 		case FAIL_POINT_SLEEP:
595 			msecs = ent->fe_arg;
596 			if (msecs)
597 				fail_point_sleep(fp, msecs, &ret);
598 			break;
599 
600 		case FAIL_POINT_PAUSE:
601 			/**
602 			 * Pausing is inherently strange with multiple
603 			 * entries given our design.  That is because some
604 			 * entries could be unreachable, for instance in cases like:
605 			 * pause->return. We can never reach the return entry.
606 			 * The sysctl layer actually truncates all entries after
607 			 * a pause for this reason.
608 			 */
609 			mtx_lock_spin(&fp_setting->feq_mtx);
610 			fail_point_pause(fp, &ret, &fp_setting->feq_mtx);
611 			mtx_unlock_spin(&fp_setting->feq_mtx);
612 			break;
613 
614 		case FAIL_POINT_YIELD:
615 			kern_yield(PRI_UNCHANGED);
616 			break;
617 
618 		case FAIL_POINT_DELAY:
619 			usecs = ent->fe_arg;
620 			DELAY(usecs);
621 			break;
622 
623 		default:
624 			break;
625 		}
626 
627 		if (cont == 0)
628 			break;
629 	}
630 
631 	if (fail_point_is_off(fp))
632 		fail_point_eval_swap_out(fp, fp_setting);
633 
634 abort:
635 	fail_point_setting_release_ref(fp);
636 
637 	return (ret);
638 }
639 
640 /**
641  * Translate internal fail_point structure into human-readable text.
642  */
643 static void
644 fail_point_get(struct fail_point *fp, struct sbuf *sb,
645         bool verbose)
646 {
647 	struct fail_point_entry *ent;
648 	struct fail_point_setting *fp_setting;
649 	struct fail_point_entry *fp_entry_cpy;
650 	int cnt_sleeping;
651 	int idx;
652 	int printed_entry_count;
653 
654 	cnt_sleeping = 0;
655 	idx = 0;
656 	printed_entry_count = 0;
657 
658 	fp_entry_cpy = fp_malloc(sizeof(struct fail_point_entry) *
659 	        (FP_MAX_ENTRY_COUNT + 1), M_WAITOK);
660 
661 	fp_setting = fail_point_setting_get_ref(fp);
662 
663 	if (fp_setting != NULL) {
664 		TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
665 			if (ent->fe_stale)
666 				continue;
667 
668 			KASSERT(printed_entry_count < FP_MAX_ENTRY_COUNT,
669 			        ("FP entry list larger than allowed"));
670 
671 			fp_entry_cpy[printed_entry_count] = *ent;
672 			++printed_entry_count;
673 		}
674 	}
675 	fail_point_setting_release_ref(fp);
676 
677 	/* This is our equivalent of a NULL terminator */
678 	fp_entry_cpy[printed_entry_count].fe_type = FAIL_POINT_INVALID;
679 
680 	while (idx < printed_entry_count) {
681 		ent = &fp_entry_cpy[idx];
682 		++idx;
683 		if (ent->fe_prob < PROB_MAX) {
684 			int decimal = ent->fe_prob % (PROB_MAX / 100);
685 			int units = ent->fe_prob / (PROB_MAX / 100);
686 			sbuf_printf(sb, "%d", units);
687 			if (decimal) {
688 				int digits = PROB_DIGITS - 2;
689 				while (!(decimal % 10)) {
690 					digits--;
691 					decimal /= 10;
692 				}
693 				sbuf_printf(sb, ".%0*d", digits, decimal);
694 			}
695 			sbuf_printf(sb, "%%");
696 		}
697 		if (ent->fe_count >= 0)
698 			sbuf_printf(sb, "%d*", ent->fe_count);
699 		sbuf_printf(sb, "%s", fail_type_strings[ent->fe_type].name);
700 		if (ent->fe_arg)
701 			sbuf_printf(sb, "(%d)", ent->fe_arg);
702 		if (ent->fe_pid != NO_PID)
703 			sbuf_printf(sb, "[pid %d]", ent->fe_pid);
704 		if (TAILQ_NEXT(ent, fe_entries))
705 			sbuf_printf(sb, "->");
706 	}
707 	if (!printed_entry_count)
708 		sbuf_printf(sb, "off");
709 
710 	fp_free(fp_entry_cpy);
711 	if (verbose) {
712 #ifdef STACK
713 		/* Print number of sleeping threads. queue=0 is the argument
714 		 * used by msleep when sending our threads to sleep. */
715 		sbuf_printf(sb, "\nsleeping_thread_stacks = {\n");
716 		sleepq_sbuf_print_stacks(sb, FP_SLEEP_CHANNEL(fp), 0,
717 		        &cnt_sleeping);
718 
719 		sbuf_printf(sb, "},\n");
720 #endif
721 		sbuf_printf(sb, "sleeping_thread_count = %d,\n",
722 		        cnt_sleeping);
723 
724 #ifdef STACK
725 		sbuf_printf(sb, "paused_thread_stacks = {\n");
726 		sleepq_sbuf_print_stacks(sb, FP_PAUSE_CHANNEL(fp), 0,
727 		        &cnt_sleeping);
728 
729 		sbuf_printf(sb, "},\n");
730 #endif
731 		sbuf_printf(sb, "paused_thread_count = %d\n",
732 		        cnt_sleeping);
733 	}
734 }
735 
736 /**
737  * Set an internal fail_point structure from a human-readable failpoint string
738  * in a lock-safe manner.
739  */
740 static int
741 fail_point_set(struct fail_point *fp, char *buf)
742 {
743 	struct fail_point_entry *ent, *ent_next;
744 	struct fail_point_setting *entries;
745 	bool should_wake_paused;
746 	bool should_truncate;
747 	int error;
748 
749 	error = 0;
750 	should_wake_paused = false;
751 	should_truncate = false;
752 
753 	/* Parse new entries. */
754 	/**
755 	 * ref protects our new malloc'd stuff from being garbage collected
756 	 * before we link it.
757 	 */
758 	fail_point_setting_get_ref(fp);
759 	entries = fail_point_setting_new(fp);
760 	if (parse_fail_point(entries, buf) == NULL) {
761 		STAILQ_REMOVE(&fp_setting_garbage, entries,
762 		        fail_point_setting, fs_garbage_link);
763 		fail_point_setting_destroy(entries);
764 		error = EINVAL;
765 		goto end;
766 	}
767 
768 	/**
769 	 * Transfer the entries we are going to keep to a new list.
770 	 * Get rid of useless zero probability entries, and entries with hit
771 	 * count 0.
772 	 * If 'off' is present, and it has no hit count set, then all entries
773 	 *       after it are discarded since they are unreachable.
774 	 */
775 	TAILQ_FOREACH_SAFE(ent, &entries->fp_entry_queue, fe_entries, ent_next) {
776 		if (ent->fe_prob == 0 || ent->fe_count == 0) {
777 			printf("Discarding entry which cannot execute %s\n",
778 			        fail_type_strings[ent->fe_type].name);
779 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
780 			        fe_entries);
781 			fp_free(ent);
782 			continue;
783 		} else if (should_truncate) {
784 			printf("Discarding unreachable entry %s\n",
785 			        fail_type_strings[ent->fe_type].name);
786 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
787 			        fe_entries);
788 			fp_free(ent);
789 			continue;
790 		}
791 
792 		if (ent->fe_type == FAIL_POINT_OFF) {
793 			should_wake_paused = true;
794 			if (ent->fe_count == FE_COUNT_UNTRACKED) {
795 				should_truncate = true;
796 				TAILQ_REMOVE(&entries->fp_entry_queue, ent,
797 				        fe_entries);
798 				fp_free(ent);
799 			}
800 		} else if (ent->fe_type == FAIL_POINT_PAUSE) {
801 			should_truncate = true;
802 		} else if (ent->fe_type == FAIL_POINT_SLEEP && (fp->fp_flags &
803 		        FAIL_POINT_NONSLEEPABLE)) {
804 			/**
805 			 * If this fail point is annotated as being in a
806 			 * non-sleepable ctx, convert sleep to delay and
807 			 * convert the msec argument to usecs.
808 			 */
809 			printf("Sleep call request on fail point in "
810 			        "non-sleepable context; using delay instead "
811 			        "of sleep\n");
812 			ent->fe_type = FAIL_POINT_DELAY;
813 			ent->fe_arg *= 1000;
814 		}
815 	}
816 
817 	if (TAILQ_EMPTY(&entries->fp_entry_queue)) {
818 		entries = fail_point_swap_settings(fp, NULL);
819 		if (entries != NULL)
820 			wakeup(FP_PAUSE_CHANNEL(fp));
821 	} else {
822 		if (should_wake_paused)
823 			wakeup(FP_PAUSE_CHANNEL(fp));
824 		fail_point_swap_settings(fp, entries);
825 	}
826 
827 end:
828 #ifdef IWARNING
829 	if (error)
830 		IWARNING("Failed to set %s %s to %s",
831 		    fp->fp_name, fp->fp_location, buf);
832 	else
833 		INOTICE("Set %s %s to %s",
834 		    fp->fp_name, fp->fp_location, buf);
835 #endif /* IWARNING */
836 
837 	fail_point_setting_release_ref(fp);
838 	return (error);
839 }
840 
841 #define MAX_FAIL_POINT_BUF	1023
842 
843 /**
844  * Handle kernel failpoint set/get.
845  */
846 int
847 fail_point_sysctl(SYSCTL_HANDLER_ARGS)
848 {
849 	struct fail_point *fp;
850 	char *buf;
851 	struct sbuf sb, *sb_check;
852 	int error;
853 
854 	buf = NULL;
855 	error = 0;
856 	fp = arg1;
857 
858 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
859 	if (sb_check != &sb)
860 		return (ENOMEM);
861 
862 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
863 
864 	/* Setting */
865 	/**
866 	 * Lock protects any new entries from being garbage collected before we
867 	 * can link them to the fail point.
868 	 */
869 	sx_xlock(&sx_fp_set);
870 	if (req->newptr) {
871 		if (req->newlen > MAX_FAIL_POINT_BUF) {
872 			error = EINVAL;
873 			goto out;
874 		}
875 
876 		buf = fp_malloc(req->newlen + 1, M_WAITOK);
877 
878 		error = SYSCTL_IN(req, buf, req->newlen);
879 		if (error)
880 			goto out;
881 		buf[req->newlen] = '\0';
882 
883 		error = fail_point_set(fp, buf);
884 	}
885 
886 	fail_point_garbage_collect();
887 	sx_xunlock(&sx_fp_set);
888 
889 	/* Retrieving. */
890 	fail_point_get(fp, &sb, false);
891 
892 out:
893 	sbuf_finish(&sb);
894 	sbuf_delete(&sb);
895 
896 	if (buf)
897 		fp_free(buf);
898 
899 	return (error);
900 }
901 
902 int
903 fail_point_sysctl_status(SYSCTL_HANDLER_ARGS)
904 {
905 	struct fail_point *fp;
906 	struct sbuf sb, *sb_check;
907 
908 	fp = arg1;
909 
910 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
911 	if (sb_check != &sb)
912 		return (ENOMEM);
913 
914 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
915 
916 	/* Retrieving. */
917 	fail_point_get(fp, &sb, true);
918 
919 	sbuf_finish(&sb);
920 	sbuf_delete(&sb);
921 
922 	/**
923 	 * Lock protects any new entries from being garbage collected before we
924 	 * can link them to the fail point.
925 	 */
926 	sx_xlock(&sx_fp_set);
927 	fail_point_garbage_collect();
928 	sx_xunlock(&sx_fp_set);
929 
930 	return (0);
931 }
932 
933 int
934 fail_sysctl_drain_func(void *sysctl_args, const char *buf, int len)
935 {
936 	struct sysctl_req *sa;
937 	int error;
938 
939 	sa = sysctl_args;
940 
941 	error = SYSCTL_OUT(sa, buf, len);
942 
943 	if (error == ENOMEM)
944 		return (-1);
945 	else
946 		return (len);
947 }
948 
949 /**
950  * Internal helper function to translate a human-readable failpoint string
951  * into a internally-parsable fail_point structure.
952  */
953 static char *
954 parse_fail_point(struct fail_point_setting *ents, char *p)
955 {
956 	/*  <fail_point> ::
957 	 *      <term> ( "->" <term> )*
958 	 */
959 	uint8_t term_count;
960 
961 	term_count = 1;
962 
963 	p = parse_term(ents, p);
964 	if (p == NULL)
965 		return (NULL);
966 
967 	while (*p != '\0') {
968 		term_count++;
969 		if (p[0] != '-' || p[1] != '>' ||
970 		        (p = parse_term(ents, p+2)) == NULL ||
971 		        term_count > FP_MAX_ENTRY_COUNT)
972 			return (NULL);
973 	}
974 	return (p);
975 }
976 
977 /**
978  * Internal helper function to parse an individual term from a failpoint.
979  */
980 static char *
981 parse_term(struct fail_point_setting *ents, char *p)
982 {
983 	struct fail_point_entry *ent;
984 
985 	ent = fail_point_entry_new(ents);
986 
987 	/*
988 	 * <term> ::
989 	 *     ( (<float> "%") | (<integer> "*" ) )*
990 	 *     <type>
991 	 *     [ "(" <integer> ")" ]
992 	 *     [ "[pid " <integer> "]" ]
993 	 */
994 
995 	/* ( (<float> "%") | (<integer> "*" ) )* */
996 	while (isdigit(*p) || *p == '.') {
997 		int units, decimal;
998 
999 		p = parse_number(&units, &decimal, p);
1000 		if (p == NULL)
1001 			return (NULL);
1002 
1003 		if (*p == '%') {
1004 			if (units > 100) /* prevent overflow early */
1005 				units = 100;
1006 			ent->fe_prob = units * (PROB_MAX / 100) + decimal;
1007 			if (ent->fe_prob > PROB_MAX)
1008 				ent->fe_prob = PROB_MAX;
1009 		} else if (*p == '*') {
1010 			if (!units || units < 0 || decimal)
1011 				return (NULL);
1012 			ent->fe_count = units;
1013 		} else
1014 			return (NULL);
1015 		p++;
1016 	}
1017 
1018 	/* <type> */
1019 	p = parse_type(ent, p);
1020 	if (p == NULL)
1021 		return (NULL);
1022 	if (*p == '\0')
1023 		return (p);
1024 
1025 	/* [ "(" <integer> ")" ] */
1026 	if (*p != '(')
1027 		return (p);
1028 	p++;
1029 	if (!isdigit(*p) && *p != '-')
1030 		return (NULL);
1031 	ent->fe_arg = strtol(p, &p, 0);
1032 	if (*p++ != ')')
1033 		return (NULL);
1034 
1035 	/* [ "[pid " <integer> "]" ] */
1036 #define PID_STRING "[pid "
1037 	if (strncmp(p, PID_STRING, sizeof(PID_STRING) - 1) != 0)
1038 		return (p);
1039 	p += sizeof(PID_STRING) - 1;
1040 	if (!isdigit(*p))
1041 		return (NULL);
1042 	ent->fe_pid = strtol(p, &p, 0);
1043 	if (*p++ != ']')
1044 		return (NULL);
1045 
1046 	return (p);
1047 }
1048 
1049 /**
1050  * Internal helper function to parse a numeric for a failpoint term.
1051  */
1052 static char *
1053 parse_number(int *out_units, int *out_decimal, char *p)
1054 {
1055 	char *old_p;
1056 
1057 	/**
1058 	 *  <number> ::
1059 	 *      <integer> [ "." <integer> ] |
1060 	 *      "." <integer>
1061 	 */
1062 
1063 	/* whole part */
1064 	old_p = p;
1065 	*out_units = strtol(p, &p, 10);
1066 	if (p == old_p && *p != '.')
1067 		return (NULL);
1068 
1069 	/* fractional part */
1070 	*out_decimal = 0;
1071 	if (*p == '.') {
1072 		int digits = 0;
1073 		p++;
1074 		while (isdigit(*p)) {
1075 			int digit = *p - '0';
1076 			if (digits < PROB_DIGITS - 2)
1077 				*out_decimal = *out_decimal * 10 + digit;
1078 			else if (digits == PROB_DIGITS - 2 && digit >= 5)
1079 				(*out_decimal)++;
1080 			digits++;
1081 			p++;
1082 		}
1083 		if (!digits) /* need at least one digit after '.' */
1084 			return (NULL);
1085 		while (digits++ < PROB_DIGITS - 2) /* add implicit zeros */
1086 			*out_decimal *= 10;
1087 	}
1088 
1089 	return (p); /* success */
1090 }
1091 
1092 /**
1093  * Internal helper function to parse an individual type for a failpoint term.
1094  */
1095 static char *
1096 parse_type(struct fail_point_entry *ent, char *beg)
1097 {
1098 	enum fail_point_t type;
1099 	int len;
1100 
1101 	for (type = FAIL_POINT_OFF; type < FAIL_POINT_NUMTYPES; type++) {
1102 		len = fail_type_strings[type].nmlen;
1103 		if (strncmp(fail_type_strings[type].name, beg, len) == 0) {
1104 			ent->fe_type = type;
1105 			return (beg + len);
1106 		}
1107 	}
1108 	return (NULL);
1109 }
1110 
1111 /* The fail point sysctl tree. */
1112 SYSCTL_NODE(_debug, OID_AUTO, fail_point, CTLFLAG_RW, 0, "fail points");
1113 
1114 /* Debugging/testing stuff for fail point */
1115 static int
1116 sysctl_test_fail_point(SYSCTL_HANDLER_ARGS)
1117 {
1118 
1119 	KFAIL_POINT_RETURN(DEBUG_FP, test_fail_point);
1120 	return (0);
1121 }
1122 SYSCTL_OID(_debug_fail_point, OID_AUTO, test_trigger_fail_point,
1123         CTLTYPE_STRING | CTLFLAG_RD, NULL, 0, sysctl_test_fail_point, "A",
1124         "Trigger test fail points");
1125