xref: /freebsd/sys/kern/kern_fail.c (revision ae477ca7da55f76d28859e1bd01cd1051e36f28f)
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 	if (name != NULL) {
483 		va_start(ap, fmt);
484 		vsnprintf(name, n + 1, fmt, ap);
485 		va_end(ap);
486 	}
487 	fp->fp_name = name;
488 	fp->fp_location = "";
489 	fp->fp_flags |= FAIL_POINT_DYNAMIC_NAME;
490 	fp->fp_pre_sleep_fn = NULL;
491 	fp->fp_pre_sleep_arg = NULL;
492 	fp->fp_post_sleep_fn = NULL;
493 	fp->fp_post_sleep_arg = NULL;
494 }
495 
496 void
497 fail_point_alloc_callout(struct fail_point *fp)
498 {
499 
500 	/**
501 	 * This assumes that calls to fail_point_use_timeout_path()
502 	 * will not race.
503 	 */
504 	if (fp->fp_callout != NULL)
505 		return;
506 	fp->fp_callout = fp_malloc(sizeof(*fp->fp_callout), M_WAITOK);
507 	callout_init(fp->fp_callout, CALLOUT_MPSAFE);
508 }
509 
510 /**
511  * Free the resources held by a fail_point, and wake any paused threads.
512  * Thou shalt not allow threads to hit this fail point after you enter this
513  * function, nor shall you call this multiple times for a given fp.
514  * @ingroup failpoint
515  */
516 void
517 fail_point_destroy(struct fail_point *fp)
518 {
519 
520 	fail_point_drain(fp, 0);
521 
522 	if ((fp->fp_flags & FAIL_POINT_DYNAMIC_NAME) != 0) {
523 		fp_free(__DECONST(void *, fp->fp_name));
524 		fp->fp_name = NULL;
525 	}
526 	fp->fp_flags = 0;
527 	if (fp->fp_callout) {
528 		fp_free(fp->fp_callout);
529 		fp->fp_callout = NULL;
530 	}
531 
532 	sx_xlock(&sx_fp_set);
533 	fail_point_garbage_collect();
534 	sx_xunlock(&sx_fp_set);
535 }
536 
537 /**
538  * This does the real work of evaluating a fail point. If the fail point tells
539  * us to return a value, this function returns 1 and fills in 'return_value'
540  * (return_value is allowed to be null). If the fail point tells us to panic,
541  * we never return. Otherwise we just return 0 after doing some work, which
542  * means "keep going".
543  */
544 enum fail_point_return_code
545 fail_point_eval_nontrivial(struct fail_point *fp, int *return_value)
546 {
547 	bool execute = false;
548 	struct fail_point_entry *ent;
549 	struct fail_point_setting *fp_setting;
550 	enum fail_point_return_code ret;
551 	int cont;
552 	int count;
553 	int msecs;
554 	int usecs;
555 
556 	ret = FAIL_POINT_RC_CONTINUE;
557 	cont = 0; /* don't continue by default */
558 
559 	fp_setting = fail_point_setting_get_ref(fp);
560 	if (fp_setting == NULL)
561 		goto abort;
562 
563 	TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
564 		if (ent->fe_stale)
565 			continue;
566 
567 		if (ent->fe_prob < PROB_MAX &&
568 		    ent->fe_prob < random() % PROB_MAX)
569 			continue;
570 
571 		if (ent->fe_pid != NO_PID && ent->fe_pid != curproc->p_pid)
572 			continue;
573 
574 		if (ent->fe_count != FE_COUNT_UNTRACKED) {
575 			count = ent->fe_count;
576 			while (count > 0) {
577 				if (atomic_cmpset_32(&ent->fe_count, count, count - 1)) {
578 					count--;
579 					execute = true;
580 					break;
581 				}
582 				count = ent->fe_count;
583 			}
584 			if (execute == false)
585 				/* We lost the race; consider the entry stale and bail now */
586 				continue;
587 			if (count == 0)
588 				ent->fe_stale = true;
589 		}
590 
591 		switch (ent->fe_type) {
592 		case FAIL_POINT_PANIC:
593 			panic("fail point %s panicking", fp->fp_name);
594 			/* NOTREACHED */
595 
596 		case FAIL_POINT_RETURN:
597 			if (return_value != NULL)
598 				*return_value = ent->fe_arg;
599 			ret = FAIL_POINT_RC_RETURN;
600 			break;
601 
602 		case FAIL_POINT_BREAK:
603 			printf("fail point %s breaking to debugger\n",
604 			        fp->fp_name);
605 			breakpoint();
606 			break;
607 
608 		case FAIL_POINT_PRINT:
609 			printf("fail point %s executing\n", fp->fp_name);
610 			cont = ent->fe_arg;
611 			break;
612 
613 		case FAIL_POINT_SLEEP:
614 			msecs = ent->fe_arg;
615 			if (msecs)
616 				fail_point_sleep(fp, msecs, &ret);
617 			break;
618 
619 		case FAIL_POINT_PAUSE:
620 			/**
621 			 * Pausing is inherently strange with multiple
622 			 * entries given our design.  That is because some
623 			 * entries could be unreachable, for instance in cases like:
624 			 * pause->return. We can never reach the return entry.
625 			 * The sysctl layer actually truncates all entries after
626 			 * a pause for this reason.
627 			 */
628 			mtx_lock_spin(&fp_setting->feq_mtx);
629 			fail_point_pause(fp, &ret, &fp_setting->feq_mtx);
630 			mtx_unlock_spin(&fp_setting->feq_mtx);
631 			break;
632 
633 		case FAIL_POINT_YIELD:
634 			kern_yield(PRI_UNCHANGED);
635 			break;
636 
637 		case FAIL_POINT_DELAY:
638 			usecs = ent->fe_arg;
639 			DELAY(usecs);
640 			break;
641 
642 		default:
643 			break;
644 		}
645 
646 		if (cont == 0)
647 			break;
648 	}
649 
650 	if (fail_point_is_off(fp))
651 		fail_point_eval_swap_out(fp, fp_setting);
652 
653 abort:
654 	fail_point_setting_release_ref(fp);
655 
656 	return (ret);
657 }
658 
659 /**
660  * Translate internal fail_point structure into human-readable text.
661  */
662 static void
663 fail_point_get(struct fail_point *fp, struct sbuf *sb,
664         bool verbose)
665 {
666 	struct fail_point_entry *ent;
667 	struct fail_point_setting *fp_setting;
668 	struct fail_point_entry *fp_entry_cpy;
669 	int cnt_sleeping;
670 	int idx;
671 	int printed_entry_count;
672 
673 	cnt_sleeping = 0;
674 	idx = 0;
675 	printed_entry_count = 0;
676 
677 	fp_entry_cpy = fp_malloc(sizeof(struct fail_point_entry) *
678 	        (FP_MAX_ENTRY_COUNT + 1), M_WAITOK);
679 
680 	fp_setting = fail_point_setting_get_ref(fp);
681 
682 	if (fp_setting != NULL) {
683 		TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
684 			if (ent->fe_stale)
685 				continue;
686 
687 			KASSERT(printed_entry_count < FP_MAX_ENTRY_COUNT,
688 			        ("FP entry list larger than allowed"));
689 
690 			fp_entry_cpy[printed_entry_count] = *ent;
691 			++printed_entry_count;
692 		}
693 	}
694 	fail_point_setting_release_ref(fp);
695 
696 	/* This is our equivalent of a NULL terminator */
697 	fp_entry_cpy[printed_entry_count].fe_type = FAIL_POINT_INVALID;
698 
699 	while (idx < printed_entry_count) {
700 		ent = &fp_entry_cpy[idx];
701 		++idx;
702 		if (ent->fe_prob < PROB_MAX) {
703 			int decimal = ent->fe_prob % (PROB_MAX / 100);
704 			int units = ent->fe_prob / (PROB_MAX / 100);
705 			sbuf_printf(sb, "%d", units);
706 			if (decimal) {
707 				int digits = PROB_DIGITS - 2;
708 				while (!(decimal % 10)) {
709 					digits--;
710 					decimal /= 10;
711 				}
712 				sbuf_printf(sb, ".%0*d", digits, decimal);
713 			}
714 			sbuf_printf(sb, "%%");
715 		}
716 		if (ent->fe_count >= 0)
717 			sbuf_printf(sb, "%d*", ent->fe_count);
718 		sbuf_printf(sb, "%s", fail_type_strings[ent->fe_type].name);
719 		if (ent->fe_arg)
720 			sbuf_printf(sb, "(%d)", ent->fe_arg);
721 		if (ent->fe_pid != NO_PID)
722 			sbuf_printf(sb, "[pid %d]", ent->fe_pid);
723 		if (TAILQ_NEXT(ent, fe_entries))
724 			sbuf_cat(sb, "->");
725 	}
726 	if (!printed_entry_count)
727 		sbuf_cat(sb, "off");
728 
729 	fp_free(fp_entry_cpy);
730 	if (verbose) {
731 #ifdef STACK
732 		/* Print number of sleeping threads. queue=0 is the argument
733 		 * used by msleep when sending our threads to sleep. */
734 		sbuf_cat(sb, "\nsleeping_thread_stacks = {\n");
735 		sleepq_sbuf_print_stacks(sb, FP_SLEEP_CHANNEL(fp), 0,
736 		        &cnt_sleeping);
737 
738 		sbuf_cat(sb, "},\n");
739 #endif
740 		sbuf_printf(sb, "sleeping_thread_count = %d,\n",
741 		        cnt_sleeping);
742 
743 #ifdef STACK
744 		sbuf_cat(sb, "paused_thread_stacks = {\n");
745 		sleepq_sbuf_print_stacks(sb, FP_PAUSE_CHANNEL(fp), 0,
746 		        &cnt_sleeping);
747 
748 		sbuf_cat(sb, "},\n");
749 #endif
750 		sbuf_printf(sb, "paused_thread_count = %d\n",
751 		        cnt_sleeping);
752 	}
753 }
754 
755 /**
756  * Set an internal fail_point structure from a human-readable failpoint string
757  * in a lock-safe manner.
758  */
759 static int
760 fail_point_set(struct fail_point *fp, char *buf)
761 {
762 	struct fail_point_entry *ent, *ent_next;
763 	struct fail_point_setting *entries;
764 	bool should_wake_paused;
765 	bool should_truncate;
766 	int error;
767 
768 	error = 0;
769 	should_wake_paused = false;
770 	should_truncate = false;
771 
772 	/* Parse new entries. */
773 	/**
774 	 * ref protects our new malloc'd stuff from being garbage collected
775 	 * before we link it.
776 	 */
777 	fail_point_setting_get_ref(fp);
778 	entries = fail_point_setting_new(fp);
779 	if (parse_fail_point(entries, buf) == NULL) {
780 		STAILQ_REMOVE(&fp_setting_garbage, entries,
781 		        fail_point_setting, fs_garbage_link);
782 		fail_point_setting_destroy(entries);
783 		error = EINVAL;
784 		goto end;
785 	}
786 
787 	/**
788 	 * Transfer the entries we are going to keep to a new list.
789 	 * Get rid of useless zero probability entries, and entries with hit
790 	 * count 0.
791 	 * If 'off' is present, and it has no hit count set, then all entries
792 	 *       after it are discarded since they are unreachable.
793 	 */
794 	TAILQ_FOREACH_SAFE(ent, &entries->fp_entry_queue, fe_entries, ent_next) {
795 		if (ent->fe_prob == 0 || ent->fe_count == 0) {
796 			printf("Discarding entry which cannot execute %s\n",
797 			        fail_type_strings[ent->fe_type].name);
798 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
799 			        fe_entries);
800 			fp_free(ent);
801 			continue;
802 		} else if (should_truncate) {
803 			printf("Discarding unreachable entry %s\n",
804 			        fail_type_strings[ent->fe_type].name);
805 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
806 			        fe_entries);
807 			fp_free(ent);
808 			continue;
809 		}
810 
811 		if (ent->fe_type == FAIL_POINT_OFF) {
812 			should_wake_paused = true;
813 			if (ent->fe_count == FE_COUNT_UNTRACKED) {
814 				should_truncate = true;
815 				TAILQ_REMOVE(&entries->fp_entry_queue, ent,
816 				        fe_entries);
817 				fp_free(ent);
818 			}
819 		} else if (ent->fe_type == FAIL_POINT_PAUSE) {
820 			should_truncate = true;
821 		} else if (ent->fe_type == FAIL_POINT_SLEEP && (fp->fp_flags &
822 		        FAIL_POINT_NONSLEEPABLE)) {
823 			/**
824 			 * If this fail point is annotated as being in a
825 			 * non-sleepable ctx, convert sleep to delay and
826 			 * convert the msec argument to usecs.
827 			 */
828 			printf("Sleep call request on fail point in "
829 			        "non-sleepable context; using delay instead "
830 			        "of sleep\n");
831 			ent->fe_type = FAIL_POINT_DELAY;
832 			ent->fe_arg *= 1000;
833 		}
834 	}
835 
836 	if (TAILQ_EMPTY(&entries->fp_entry_queue)) {
837 		entries = fail_point_swap_settings(fp, NULL);
838 		if (entries != NULL)
839 			wakeup(FP_PAUSE_CHANNEL(fp));
840 	} else {
841 		if (should_wake_paused)
842 			wakeup(FP_PAUSE_CHANNEL(fp));
843 		fail_point_swap_settings(fp, entries);
844 	}
845 
846 end:
847 #ifdef IWARNING
848 	if (error)
849 		IWARNING("Failed to set %s %s to %s",
850 		    fp->fp_name, fp->fp_location, buf);
851 	else
852 		INOTICE("Set %s %s to %s",
853 		    fp->fp_name, fp->fp_location, buf);
854 #endif /* IWARNING */
855 
856 	fail_point_setting_release_ref(fp);
857 	return (error);
858 }
859 
860 #define MAX_FAIL_POINT_BUF	1023
861 
862 /**
863  * Handle kernel failpoint set/get.
864  */
865 int
866 fail_point_sysctl(SYSCTL_HANDLER_ARGS)
867 {
868 	struct fail_point *fp;
869 	char *buf;
870 	struct sbuf sb, *sb_check;
871 	int error;
872 
873 	buf = NULL;
874 	error = 0;
875 	fp = arg1;
876 
877 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
878 	if (sb_check != &sb)
879 		return (ENOMEM);
880 
881 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
882 
883 	/* Setting */
884 	/**
885 	 * Lock protects any new entries from being garbage collected before we
886 	 * can link them to the fail point.
887 	 */
888 	sx_xlock(&sx_fp_set);
889 	if (req->newptr) {
890 		if (req->newlen > MAX_FAIL_POINT_BUF) {
891 			error = EINVAL;
892 			goto out;
893 		}
894 
895 		buf = fp_malloc(req->newlen + 1, M_WAITOK);
896 
897 		error = SYSCTL_IN(req, buf, req->newlen);
898 		if (error)
899 			goto out;
900 		buf[req->newlen] = '\0';
901 
902 		error = fail_point_set(fp, buf);
903 	}
904 
905 	fail_point_garbage_collect();
906 	sx_xunlock(&sx_fp_set);
907 
908 	/* Retrieving. */
909 	fail_point_get(fp, &sb, false);
910 
911 out:
912 	sbuf_finish(&sb);
913 	sbuf_delete(&sb);
914 
915 	if (buf)
916 		fp_free(buf);
917 
918 	return (error);
919 }
920 
921 int
922 fail_point_sysctl_status(SYSCTL_HANDLER_ARGS)
923 {
924 	struct fail_point *fp;
925 	struct sbuf sb, *sb_check;
926 
927 	fp = arg1;
928 
929 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
930 	if (sb_check != &sb)
931 		return (ENOMEM);
932 
933 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
934 
935 	/* Retrieving. */
936 	fail_point_get(fp, &sb, true);
937 
938 	sbuf_finish(&sb);
939 	sbuf_delete(&sb);
940 
941 	/**
942 	 * Lock protects any new entries from being garbage collected before we
943 	 * can link them to the fail point.
944 	 */
945 	sx_xlock(&sx_fp_set);
946 	fail_point_garbage_collect();
947 	sx_xunlock(&sx_fp_set);
948 
949 	return (0);
950 }
951 
952 int
953 fail_sysctl_drain_func(void *sysctl_args, const char *buf, int len)
954 {
955 	struct sysctl_req *sa;
956 	int error;
957 
958 	sa = sysctl_args;
959 
960 	error = SYSCTL_OUT(sa, buf, len);
961 
962 	if (error == ENOMEM)
963 		return (-1);
964 	else
965 		return (len);
966 }
967 
968 /**
969  * Internal helper function to translate a human-readable failpoint string
970  * into a internally-parsable fail_point structure.
971  */
972 static char *
973 parse_fail_point(struct fail_point_setting *ents, char *p)
974 {
975 	/*  <fail_point> ::
976 	 *      <term> ( "->" <term> )*
977 	 */
978 	uint8_t term_count;
979 
980 	term_count = 1;
981 
982 	p = parse_term(ents, p);
983 	if (p == NULL)
984 		return (NULL);
985 
986 	while (*p != '\0') {
987 		term_count++;
988 		if (p[0] != '-' || p[1] != '>' ||
989 		        (p = parse_term(ents, p+2)) == NULL ||
990 		        term_count > FP_MAX_ENTRY_COUNT)
991 			return (NULL);
992 	}
993 	return (p);
994 }
995 
996 /**
997  * Internal helper function to parse an individual term from a failpoint.
998  */
999 static char *
1000 parse_term(struct fail_point_setting *ents, char *p)
1001 {
1002 	struct fail_point_entry *ent;
1003 
1004 	ent = fail_point_entry_new(ents);
1005 
1006 	/*
1007 	 * <term> ::
1008 	 *     ( (<float> "%") | (<integer> "*" ) )*
1009 	 *     <type>
1010 	 *     [ "(" <integer> ")" ]
1011 	 *     [ "[pid " <integer> "]" ]
1012 	 */
1013 
1014 	/* ( (<float> "%") | (<integer> "*" ) )* */
1015 	while (isdigit(*p) || *p == '.') {
1016 		int units, decimal;
1017 
1018 		p = parse_number(&units, &decimal, p);
1019 		if (p == NULL)
1020 			return (NULL);
1021 
1022 		if (*p == '%') {
1023 			if (units > 100) /* prevent overflow early */
1024 				units = 100;
1025 			ent->fe_prob = units * (PROB_MAX / 100) + decimal;
1026 			if (ent->fe_prob > PROB_MAX)
1027 				ent->fe_prob = PROB_MAX;
1028 		} else if (*p == '*') {
1029 			if (!units || units < 0 || decimal)
1030 				return (NULL);
1031 			ent->fe_count = units;
1032 		} else
1033 			return (NULL);
1034 		p++;
1035 	}
1036 
1037 	/* <type> */
1038 	p = parse_type(ent, p);
1039 	if (p == NULL)
1040 		return (NULL);
1041 	if (*p == '\0')
1042 		return (p);
1043 
1044 	/* [ "(" <integer> ")" ] */
1045 	if (*p != '(')
1046 		return (p);
1047 	p++;
1048 	if (!isdigit(*p) && *p != '-')
1049 		return (NULL);
1050 	ent->fe_arg = strtol(p, &p, 0);
1051 	if (*p++ != ')')
1052 		return (NULL);
1053 
1054 	/* [ "[pid " <integer> "]" ] */
1055 #define PID_STRING "[pid "
1056 	if (strncmp(p, PID_STRING, sizeof(PID_STRING) - 1) != 0)
1057 		return (p);
1058 	p += sizeof(PID_STRING) - 1;
1059 	if (!isdigit(*p))
1060 		return (NULL);
1061 	ent->fe_pid = strtol(p, &p, 0);
1062 	if (*p++ != ']')
1063 		return (NULL);
1064 
1065 	return (p);
1066 }
1067 
1068 /**
1069  * Internal helper function to parse a numeric for a failpoint term.
1070  */
1071 static char *
1072 parse_number(int *out_units, int *out_decimal, char *p)
1073 {
1074 	char *old_p;
1075 
1076 	/**
1077 	 *  <number> ::
1078 	 *      <integer> [ "." <integer> ] |
1079 	 *      "." <integer>
1080 	 */
1081 
1082 	/* whole part */
1083 	old_p = p;
1084 	*out_units = strtol(p, &p, 10);
1085 	if (p == old_p && *p != '.')
1086 		return (NULL);
1087 
1088 	/* fractional part */
1089 	*out_decimal = 0;
1090 	if (*p == '.') {
1091 		int digits = 0;
1092 		p++;
1093 		while (isdigit(*p)) {
1094 			int digit = *p - '0';
1095 			if (digits < PROB_DIGITS - 2)
1096 				*out_decimal = *out_decimal * 10 + digit;
1097 			else if (digits == PROB_DIGITS - 2 && digit >= 5)
1098 				(*out_decimal)++;
1099 			digits++;
1100 			p++;
1101 		}
1102 		if (!digits) /* need at least one digit after '.' */
1103 			return (NULL);
1104 		while (digits++ < PROB_DIGITS - 2) /* add implicit zeros */
1105 			*out_decimal *= 10;
1106 	}
1107 
1108 	return (p); /* success */
1109 }
1110 
1111 /**
1112  * Internal helper function to parse an individual type for a failpoint term.
1113  */
1114 static char *
1115 parse_type(struct fail_point_entry *ent, char *beg)
1116 {
1117 	enum fail_point_t type;
1118 	int len;
1119 
1120 	for (type = FAIL_POINT_OFF; type < FAIL_POINT_NUMTYPES; type++) {
1121 		len = fail_type_strings[type].nmlen;
1122 		if (strncmp(fail_type_strings[type].name, beg, len) == 0) {
1123 			ent->fe_type = type;
1124 			return (beg + len);
1125 		}
1126 	}
1127 	return (NULL);
1128 }
1129 
1130 /* The fail point sysctl tree. */
1131 SYSCTL_NODE(_debug, OID_AUTO, fail_point, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1132     "fail points");
1133 
1134 /* Debugging/testing stuff for fail point */
1135 static int
1136 sysctl_test_fail_point(SYSCTL_HANDLER_ARGS)
1137 {
1138 
1139 	KFAIL_POINT_RETURN(DEBUG_FP, test_fail_point);
1140 	return (0);
1141 }
1142 SYSCTL_OID(_debug_fail_point, OID_AUTO, test_trigger_fail_point,
1143     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
1144     sysctl_test_fail_point, "A",
1145     "Trigger test fail points");
1146