xref: /freebsd/sys/kern/kern_fail.c (revision 87b759f0fa1f7554d50ce640c40138512bbded44)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2009 Isilon Inc http://www.isilon.com/
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 /**
28  * @file
29  *
30  * fail(9) Facility.
31  *
32  * @ingroup failpoint_private
33  */
34 /**
35  * @defgroup failpoint fail(9) Facility
36  *
37  * Failpoints allow for injecting fake errors into running code on the fly,
38  * without modifying code or recompiling with flags.  Failpoints are always
39  * present, and are very efficient when disabled.  Failpoints are described
40  * in man fail(9).
41  */
42 /**
43  * @defgroup failpoint_private Private fail(9) Implementation functions
44  *
45  * Private implementations for the actual failpoint code.
46  *
47  * @ingroup failpoint
48  */
49 /**
50  * @addtogroup failpoint_private
51  * @{
52  */
53 
54 #include <sys/cdefs.h>
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 	if (fp->fp_callout)
405 		callout_drain(fp->fp_callout);
406 	fail_point_swap_settings(fp, entries);
407 }
408 
409 static inline void
410 fail_point_pause(struct fail_point *fp, enum fail_point_return_code *pret,
411         struct mtx *mtx_sleep)
412 {
413 
414 	if (fp->fp_pre_sleep_fn)
415 		fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
416 
417 	msleep_spin(FP_PAUSE_CHANNEL(fp), mtx_sleep, "failpt", 0);
418 
419 	if (fp->fp_post_sleep_fn)
420 		fp->fp_post_sleep_fn(fp->fp_post_sleep_arg);
421 }
422 
423 static inline void
424 fail_point_sleep(struct fail_point *fp, int msecs,
425         enum fail_point_return_code *pret)
426 {
427 	int timo;
428 
429 	/* Convert from millisecs to ticks, rounding up */
430 	timo = howmany((int64_t)msecs * hz, 1000L);
431 
432 	if (timo > 0) {
433 		if (!(fp->fp_flags & FAIL_POINT_USE_TIMEOUT_PATH)) {
434 			if (fp->fp_pre_sleep_fn)
435 				fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
436 
437 			tsleep(FP_SLEEP_CHANNEL(fp), PWAIT, "failpt", timo);
438 
439 			if (fp->fp_post_sleep_fn)
440 				fp->fp_post_sleep_fn(fp->fp_post_sleep_arg);
441 		} else {
442 			if (fp->fp_pre_sleep_fn)
443 				fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
444 
445 			callout_reset(fp->fp_callout, timo,
446 			    fp->fp_post_sleep_fn, fp->fp_post_sleep_arg);
447 			*pret = FAIL_POINT_RC_QUEUED;
448 		}
449 	}
450 }
451 
452 static char *parse_fail_point(struct fail_point_setting *, char *);
453 static char *parse_term(struct fail_point_setting *, char *);
454 static char *parse_number(int *out_units, int *out_decimal, char *);
455 static char *parse_type(struct fail_point_entry *, char *);
456 
457 /**
458  * Initialize a fail_point.  The name is formed in a printf-like fashion
459  * from "fmt" and subsequent arguments.  This function is generally used
460  * for custom failpoints located at odd places in the sysctl tree, and is
461  * not explicitly needed for standard in-line-declared failpoints.
462  *
463  * @ingroup failpoint
464  */
465 void
466 fail_point_init(struct fail_point *fp, const char *fmt, ...)
467 {
468 	va_list ap;
469 	char *name;
470 	int n;
471 
472 	fp->fp_setting = NULL;
473 	fp->fp_flags = 0;
474 
475 	/* Figure out the size of the name. */
476 	va_start(ap, fmt);
477 	n = vsnprintf(NULL, 0, fmt, ap);
478 	va_end(ap);
479 
480 	/* Allocate the name and fill it in. */
481 	name = fp_malloc(n + 1, M_WAITOK);
482 	va_start(ap, fmt);
483 	vsnprintf(name, n + 1, fmt, ap);
484 	va_end(ap);
485 
486 	fp->fp_name = name;
487 	fp->fp_location = "";
488 	fp->fp_flags |= FAIL_POINT_DYNAMIC_NAME;
489 	fp->fp_pre_sleep_fn = NULL;
490 	fp->fp_pre_sleep_arg = NULL;
491 	fp->fp_post_sleep_fn = NULL;
492 	fp->fp_post_sleep_arg = NULL;
493 }
494 
495 void
496 fail_point_alloc_callout(struct fail_point *fp)
497 {
498 
499 	/**
500 	 * This assumes that calls to fail_point_use_timeout_path()
501 	 * will not race.
502 	 */
503 	if (fp->fp_callout != NULL)
504 		return;
505 	fp->fp_callout = fp_malloc(sizeof(*fp->fp_callout), M_WAITOK);
506 	callout_init(fp->fp_callout, CALLOUT_MPSAFE);
507 }
508 
509 /**
510  * Free the resources held by a fail_point, and wake any paused threads.
511  * Thou shalt not allow threads to hit this fail point after you enter this
512  * function, nor shall you call this multiple times for a given fp.
513  * @ingroup failpoint
514  */
515 void
516 fail_point_destroy(struct fail_point *fp)
517 {
518 
519 	fail_point_drain(fp, 0);
520 
521 	if ((fp->fp_flags & FAIL_POINT_DYNAMIC_NAME) != 0) {
522 		fp_free(__DECONST(void *, fp->fp_name));
523 		fp->fp_name = NULL;
524 	}
525 	fp->fp_flags = 0;
526 	if (fp->fp_callout) {
527 		fp_free(fp->fp_callout);
528 		fp->fp_callout = NULL;
529 	}
530 
531 	sx_xlock(&sx_fp_set);
532 	fail_point_garbage_collect();
533 	sx_xunlock(&sx_fp_set);
534 }
535 
536 /**
537  * This does the real work of evaluating a fail point. If the fail point tells
538  * us to return a value, this function returns 1 and fills in 'return_value'
539  * (return_value is allowed to be null). If the fail point tells us to panic,
540  * we never return. Otherwise we just return 0 after doing some work, which
541  * means "keep going".
542  */
543 enum fail_point_return_code
544 fail_point_eval_nontrivial(struct fail_point *fp, int *return_value)
545 {
546 	bool execute = false;
547 	struct fail_point_entry *ent;
548 	struct fail_point_setting *fp_setting;
549 	enum fail_point_return_code ret;
550 	int cont;
551 	int count;
552 	int msecs;
553 	int usecs;
554 
555 	ret = FAIL_POINT_RC_CONTINUE;
556 	cont = 0; /* don't continue by default */
557 
558 	fp_setting = fail_point_setting_get_ref(fp);
559 	if (fp_setting == NULL)
560 		goto abort;
561 
562 	TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
563 		if (ent->fe_stale)
564 			continue;
565 
566 		if (ent->fe_prob < PROB_MAX &&
567 		    ent->fe_prob < random() % PROB_MAX)
568 			continue;
569 
570 		if (ent->fe_pid != NO_PID && ent->fe_pid != curproc->p_pid)
571 			continue;
572 
573 		if (ent->fe_count != FE_COUNT_UNTRACKED) {
574 			count = ent->fe_count;
575 			while (count > 0) {
576 				if (atomic_cmpset_32(&ent->fe_count, count, count - 1)) {
577 					count--;
578 					execute = true;
579 					break;
580 				}
581 				count = ent->fe_count;
582 			}
583 			if (execute == false)
584 				/* We lost the race; consider the entry stale and bail now */
585 				continue;
586 			if (count == 0)
587 				ent->fe_stale = true;
588 		}
589 
590 		switch (ent->fe_type) {
591 		case FAIL_POINT_PANIC:
592 			panic("fail point %s panicking", fp->fp_name);
593 			/* NOTREACHED */
594 
595 		case FAIL_POINT_RETURN:
596 			if (return_value != NULL)
597 				*return_value = ent->fe_arg;
598 			ret = FAIL_POINT_RC_RETURN;
599 			break;
600 
601 		case FAIL_POINT_BREAK:
602 			printf("fail point %s breaking to debugger\n",
603 			        fp->fp_name);
604 			breakpoint();
605 			break;
606 
607 		case FAIL_POINT_PRINT:
608 			printf("fail point %s executing\n", fp->fp_name);
609 			cont = ent->fe_arg;
610 			break;
611 
612 		case FAIL_POINT_SLEEP:
613 			msecs = ent->fe_arg;
614 			if (msecs)
615 				fail_point_sleep(fp, msecs, &ret);
616 			break;
617 
618 		case FAIL_POINT_PAUSE:
619 			/**
620 			 * Pausing is inherently strange with multiple
621 			 * entries given our design.  That is because some
622 			 * entries could be unreachable, for instance in cases like:
623 			 * pause->return. We can never reach the return entry.
624 			 * The sysctl layer actually truncates all entries after
625 			 * a pause for this reason.
626 			 */
627 			mtx_lock_spin(&fp_setting->feq_mtx);
628 			fail_point_pause(fp, &ret, &fp_setting->feq_mtx);
629 			mtx_unlock_spin(&fp_setting->feq_mtx);
630 			break;
631 
632 		case FAIL_POINT_YIELD:
633 			kern_yield(PRI_UNCHANGED);
634 			break;
635 
636 		case FAIL_POINT_DELAY:
637 			usecs = ent->fe_arg;
638 			DELAY(usecs);
639 			break;
640 
641 		default:
642 			break;
643 		}
644 
645 		if (cont == 0)
646 			break;
647 	}
648 
649 	if (fail_point_is_off(fp))
650 		fail_point_eval_swap_out(fp, fp_setting);
651 
652 abort:
653 	fail_point_setting_release_ref(fp);
654 
655 	return (ret);
656 }
657 
658 /**
659  * Translate internal fail_point structure into human-readable text.
660  */
661 static void
662 fail_point_get(struct fail_point *fp, struct sbuf *sb,
663         bool verbose)
664 {
665 	struct fail_point_entry *ent;
666 	struct fail_point_setting *fp_setting;
667 	struct fail_point_entry *fp_entry_cpy;
668 	int cnt_sleeping;
669 	int idx;
670 	int printed_entry_count;
671 
672 	cnt_sleeping = 0;
673 	idx = 0;
674 	printed_entry_count = 0;
675 
676 	fp_entry_cpy = fp_malloc(sizeof(struct fail_point_entry) *
677 	        (FP_MAX_ENTRY_COUNT + 1), M_WAITOK);
678 
679 	fp_setting = fail_point_setting_get_ref(fp);
680 
681 	if (fp_setting != NULL) {
682 		TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
683 			if (ent->fe_stale)
684 				continue;
685 
686 			KASSERT(printed_entry_count < FP_MAX_ENTRY_COUNT,
687 			        ("FP entry list larger than allowed"));
688 
689 			fp_entry_cpy[printed_entry_count] = *ent;
690 			++printed_entry_count;
691 		}
692 	}
693 	fail_point_setting_release_ref(fp);
694 
695 	/* This is our equivalent of a NULL terminator */
696 	fp_entry_cpy[printed_entry_count].fe_type = FAIL_POINT_INVALID;
697 
698 	while (idx < printed_entry_count) {
699 		ent = &fp_entry_cpy[idx];
700 		++idx;
701 		if (ent->fe_prob < PROB_MAX) {
702 			int decimal = ent->fe_prob % (PROB_MAX / 100);
703 			int units = ent->fe_prob / (PROB_MAX / 100);
704 			sbuf_printf(sb, "%d", units);
705 			if (decimal) {
706 				int digits = PROB_DIGITS - 2;
707 				while (!(decimal % 10)) {
708 					digits--;
709 					decimal /= 10;
710 				}
711 				sbuf_printf(sb, ".%0*d", digits, decimal);
712 			}
713 			sbuf_printf(sb, "%%");
714 		}
715 		if (ent->fe_count >= 0)
716 			sbuf_printf(sb, "%d*", ent->fe_count);
717 		sbuf_printf(sb, "%s", fail_type_strings[ent->fe_type].name);
718 		if (ent->fe_arg)
719 			sbuf_printf(sb, "(%d)", ent->fe_arg);
720 		if (ent->fe_pid != NO_PID)
721 			sbuf_printf(sb, "[pid %d]", ent->fe_pid);
722 		if (TAILQ_NEXT(ent, fe_entries))
723 			sbuf_cat(sb, "->");
724 	}
725 	if (!printed_entry_count)
726 		sbuf_cat(sb, "off");
727 
728 	fp_free(fp_entry_cpy);
729 	if (verbose) {
730 #ifdef STACK
731 		/* Print number of sleeping threads. queue=0 is the argument
732 		 * used by msleep when sending our threads to sleep. */
733 		sbuf_cat(sb, "\nsleeping_thread_stacks = {\n");
734 		sleepq_sbuf_print_stacks(sb, FP_SLEEP_CHANNEL(fp), 0,
735 		        &cnt_sleeping);
736 
737 		sbuf_cat(sb, "},\n");
738 #endif
739 		sbuf_printf(sb, "sleeping_thread_count = %d,\n",
740 		        cnt_sleeping);
741 
742 #ifdef STACK
743 		sbuf_cat(sb, "paused_thread_stacks = {\n");
744 		sleepq_sbuf_print_stacks(sb, FP_PAUSE_CHANNEL(fp), 0,
745 		        &cnt_sleeping);
746 
747 		sbuf_cat(sb, "},\n");
748 #endif
749 		sbuf_printf(sb, "paused_thread_count = %d\n",
750 		        cnt_sleeping);
751 	}
752 }
753 
754 /**
755  * Set an internal fail_point structure from a human-readable failpoint string
756  * in a lock-safe manner.
757  */
758 static int
759 fail_point_set(struct fail_point *fp, char *buf)
760 {
761 	struct fail_point_entry *ent, *ent_next;
762 	struct fail_point_setting *entries;
763 	bool should_wake_paused;
764 	bool should_truncate;
765 	int error;
766 
767 	error = 0;
768 	should_wake_paused = false;
769 	should_truncate = false;
770 
771 	/* Parse new entries. */
772 	/**
773 	 * ref protects our new malloc'd stuff from being garbage collected
774 	 * before we link it.
775 	 */
776 	fail_point_setting_get_ref(fp);
777 	entries = fail_point_setting_new(fp);
778 	if (parse_fail_point(entries, buf) == NULL) {
779 		STAILQ_REMOVE(&fp_setting_garbage, entries,
780 		        fail_point_setting, fs_garbage_link);
781 		fail_point_setting_destroy(entries);
782 		error = EINVAL;
783 		goto end;
784 	}
785 
786 	/**
787 	 * Transfer the entries we are going to keep to a new list.
788 	 * Get rid of useless zero probability entries, and entries with hit
789 	 * count 0.
790 	 * If 'off' is present, and it has no hit count set, then all entries
791 	 *       after it are discarded since they are unreachable.
792 	 */
793 	TAILQ_FOREACH_SAFE(ent, &entries->fp_entry_queue, fe_entries, ent_next) {
794 		if (ent->fe_prob == 0 || ent->fe_count == 0) {
795 			printf("Discarding entry which cannot execute %s\n",
796 			        fail_type_strings[ent->fe_type].name);
797 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
798 			        fe_entries);
799 			fp_free(ent);
800 			continue;
801 		} else if (should_truncate) {
802 			printf("Discarding unreachable entry %s\n",
803 			        fail_type_strings[ent->fe_type].name);
804 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
805 			        fe_entries);
806 			fp_free(ent);
807 			continue;
808 		}
809 
810 		if (ent->fe_type == FAIL_POINT_OFF) {
811 			should_wake_paused = true;
812 			if (ent->fe_count == FE_COUNT_UNTRACKED) {
813 				should_truncate = true;
814 				TAILQ_REMOVE(&entries->fp_entry_queue, ent,
815 				        fe_entries);
816 				fp_free(ent);
817 			}
818 		} else if (ent->fe_type == FAIL_POINT_PAUSE) {
819 			should_truncate = true;
820 		} else if (ent->fe_type == FAIL_POINT_SLEEP && (fp->fp_flags &
821 		        FAIL_POINT_NONSLEEPABLE)) {
822 			/**
823 			 * If this fail point is annotated as being in a
824 			 * non-sleepable ctx, convert sleep to delay and
825 			 * convert the msec argument to usecs.
826 			 */
827 			printf("Sleep call request on fail point in "
828 			        "non-sleepable context; using delay instead "
829 			        "of sleep\n");
830 			ent->fe_type = FAIL_POINT_DELAY;
831 			ent->fe_arg *= 1000;
832 		}
833 	}
834 
835 	if (TAILQ_EMPTY(&entries->fp_entry_queue)) {
836 		entries = fail_point_swap_settings(fp, NULL);
837 		if (entries != NULL)
838 			wakeup(FP_PAUSE_CHANNEL(fp));
839 	} else {
840 		if (should_wake_paused)
841 			wakeup(FP_PAUSE_CHANNEL(fp));
842 		fail_point_swap_settings(fp, entries);
843 	}
844 
845 end:
846 #ifdef IWARNING
847 	if (error)
848 		IWARNING("Failed to set %s %s to %s",
849 		    fp->fp_name, fp->fp_location, buf);
850 	else
851 		INOTICE("Set %s %s to %s",
852 		    fp->fp_name, fp->fp_location, buf);
853 #endif /* IWARNING */
854 
855 	fail_point_setting_release_ref(fp);
856 	return (error);
857 }
858 
859 #define MAX_FAIL_POINT_BUF	1023
860 
861 /**
862  * Handle kernel failpoint set/get.
863  */
864 int
865 fail_point_sysctl(SYSCTL_HANDLER_ARGS)
866 {
867 	struct fail_point *fp;
868 	char *buf;
869 	struct sbuf sb, *sb_check;
870 	int error;
871 
872 	buf = NULL;
873 	error = 0;
874 	fp = arg1;
875 
876 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
877 	if (sb_check != &sb)
878 		return (ENOMEM);
879 
880 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
881 
882 	/* Setting */
883 	/**
884 	 * Lock protects any new entries from being garbage collected before we
885 	 * can link them to the fail point.
886 	 */
887 	sx_xlock(&sx_fp_set);
888 	if (req->newptr) {
889 		if (req->newlen > MAX_FAIL_POINT_BUF) {
890 			error = EINVAL;
891 			goto out;
892 		}
893 
894 		buf = fp_malloc(req->newlen + 1, M_WAITOK);
895 
896 		error = SYSCTL_IN(req, buf, req->newlen);
897 		if (error)
898 			goto out;
899 		buf[req->newlen] = '\0';
900 
901 		error = fail_point_set(fp, buf);
902 	}
903 
904 	fail_point_garbage_collect();
905 	sx_xunlock(&sx_fp_set);
906 
907 	/* Retrieving. */
908 	fail_point_get(fp, &sb, false);
909 
910 out:
911 	sbuf_finish(&sb);
912 	sbuf_delete(&sb);
913 
914 	if (buf)
915 		fp_free(buf);
916 
917 	return (error);
918 }
919 
920 int
921 fail_point_sysctl_status(SYSCTL_HANDLER_ARGS)
922 {
923 	struct fail_point *fp;
924 	struct sbuf sb, *sb_check;
925 
926 	fp = arg1;
927 
928 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
929 	if (sb_check != &sb)
930 		return (ENOMEM);
931 
932 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
933 
934 	/* Retrieving. */
935 	fail_point_get(fp, &sb, true);
936 
937 	sbuf_finish(&sb);
938 	sbuf_delete(&sb);
939 
940 	/**
941 	 * Lock protects any new entries from being garbage collected before we
942 	 * can link them to the fail point.
943 	 */
944 	sx_xlock(&sx_fp_set);
945 	fail_point_garbage_collect();
946 	sx_xunlock(&sx_fp_set);
947 
948 	return (0);
949 }
950 
951 int
952 fail_sysctl_drain_func(void *sysctl_args, const char *buf, int len)
953 {
954 	struct sysctl_req *sa;
955 	int error;
956 
957 	sa = sysctl_args;
958 
959 	error = SYSCTL_OUT(sa, buf, len);
960 
961 	if (error == ENOMEM)
962 		return (-1);
963 	else
964 		return (len);
965 }
966 
967 /**
968  * Internal helper function to translate a human-readable failpoint string
969  * into a internally-parsable fail_point structure.
970  */
971 static char *
972 parse_fail_point(struct fail_point_setting *ents, char *p)
973 {
974 	/*  <fail_point> ::
975 	 *      <term> ( "->" <term> )*
976 	 */
977 	uint8_t term_count;
978 
979 	term_count = 1;
980 
981 	p = parse_term(ents, p);
982 	if (p == NULL)
983 		return (NULL);
984 
985 	while (*p != '\0') {
986 		term_count++;
987 		if (p[0] != '-' || p[1] != '>' ||
988 		        (p = parse_term(ents, p+2)) == NULL ||
989 		        term_count > FP_MAX_ENTRY_COUNT)
990 			return (NULL);
991 	}
992 	return (p);
993 }
994 
995 /**
996  * Internal helper function to parse an individual term from a failpoint.
997  */
998 static char *
999 parse_term(struct fail_point_setting *ents, char *p)
1000 {
1001 	struct fail_point_entry *ent;
1002 
1003 	ent = fail_point_entry_new(ents);
1004 
1005 	/*
1006 	 * <term> ::
1007 	 *     ( (<float> "%") | (<integer> "*" ) )*
1008 	 *     <type>
1009 	 *     [ "(" <integer> ")" ]
1010 	 *     [ "[pid " <integer> "]" ]
1011 	 */
1012 
1013 	/* ( (<float> "%") | (<integer> "*" ) )* */
1014 	while (isdigit(*p) || *p == '.') {
1015 		int units, decimal;
1016 
1017 		p = parse_number(&units, &decimal, p);
1018 		if (p == NULL)
1019 			return (NULL);
1020 
1021 		if (*p == '%') {
1022 			if (units > 100) /* prevent overflow early */
1023 				units = 100;
1024 			ent->fe_prob = units * (PROB_MAX / 100) + decimal;
1025 			if (ent->fe_prob > PROB_MAX)
1026 				ent->fe_prob = PROB_MAX;
1027 		} else if (*p == '*') {
1028 			if (!units || units < 0 || decimal)
1029 				return (NULL);
1030 			ent->fe_count = units;
1031 		} else
1032 			return (NULL);
1033 		p++;
1034 	}
1035 
1036 	/* <type> */
1037 	p = parse_type(ent, p);
1038 	if (p == NULL)
1039 		return (NULL);
1040 	if (*p == '\0')
1041 		return (p);
1042 
1043 	/* [ "(" <integer> ")" ] */
1044 	if (*p != '(')
1045 		return (p);
1046 	p++;
1047 	if (!isdigit(*p) && *p != '-')
1048 		return (NULL);
1049 	ent->fe_arg = strtol(p, &p, 0);
1050 	if (*p++ != ')')
1051 		return (NULL);
1052 
1053 	/* [ "[pid " <integer> "]" ] */
1054 #define PID_STRING "[pid "
1055 	if (strncmp(p, PID_STRING, sizeof(PID_STRING) - 1) != 0)
1056 		return (p);
1057 	p += sizeof(PID_STRING) - 1;
1058 	if (!isdigit(*p))
1059 		return (NULL);
1060 	ent->fe_pid = strtol(p, &p, 0);
1061 	if (*p++ != ']')
1062 		return (NULL);
1063 
1064 	return (p);
1065 }
1066 
1067 /**
1068  * Internal helper function to parse a numeric for a failpoint term.
1069  */
1070 static char *
1071 parse_number(int *out_units, int *out_decimal, char *p)
1072 {
1073 	char *old_p;
1074 
1075 	/**
1076 	 *  <number> ::
1077 	 *      <integer> [ "." <integer> ] |
1078 	 *      "." <integer>
1079 	 */
1080 
1081 	/* whole part */
1082 	old_p = p;
1083 	*out_units = strtol(p, &p, 10);
1084 	if (p == old_p && *p != '.')
1085 		return (NULL);
1086 
1087 	/* fractional part */
1088 	*out_decimal = 0;
1089 	if (*p == '.') {
1090 		int digits = 0;
1091 		p++;
1092 		while (isdigit(*p)) {
1093 			int digit = *p - '0';
1094 			if (digits < PROB_DIGITS - 2)
1095 				*out_decimal = *out_decimal * 10 + digit;
1096 			else if (digits == PROB_DIGITS - 2 && digit >= 5)
1097 				(*out_decimal)++;
1098 			digits++;
1099 			p++;
1100 		}
1101 		if (!digits) /* need at least one digit after '.' */
1102 			return (NULL);
1103 		while (digits++ < PROB_DIGITS - 2) /* add implicit zeros */
1104 			*out_decimal *= 10;
1105 	}
1106 
1107 	return (p); /* success */
1108 }
1109 
1110 /**
1111  * Internal helper function to parse an individual type for a failpoint term.
1112  */
1113 static char *
1114 parse_type(struct fail_point_entry *ent, char *beg)
1115 {
1116 	enum fail_point_t type;
1117 	int len;
1118 
1119 	for (type = FAIL_POINT_OFF; type < FAIL_POINT_NUMTYPES; type++) {
1120 		len = fail_type_strings[type].nmlen;
1121 		if (strncmp(fail_type_strings[type].name, beg, len) == 0) {
1122 			ent->fe_type = type;
1123 			return (beg + len);
1124 		}
1125 	}
1126 	return (NULL);
1127 }
1128 
1129 /* The fail point sysctl tree. */
1130 SYSCTL_NODE(_debug, OID_AUTO, fail_point, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1131     "fail points");
1132 
1133 /* Debugging/testing stuff for fail point */
1134 static int
1135 sysctl_test_fail_point(SYSCTL_HANDLER_ARGS)
1136 {
1137 
1138 	KFAIL_POINT_RETURN(DEBUG_FP, test_fail_point);
1139 	return (0);
1140 }
1141 SYSCTL_OID(_debug_fail_point, OID_AUTO, test_trigger_fail_point,
1142     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
1143     sysctl_test_fail_point, "A",
1144     "Trigger test fail points");
1145