xref: /freebsd/sys/kern/kern_fail.c (revision eb69d1f144a6fcc765d1b9d44a5ae8082353e70b)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
56 
57 #include "opt_stack.h"
58 
59 #include <sys/ctype.h>
60 #include <sys/errno.h>
61 #include <sys/fail.h>
62 #include <sys/kernel.h>
63 #include <sys/libkern.h>
64 #include <sys/limits.h>
65 #include <sys/lock.h>
66 #include <sys/malloc.h>
67 #include <sys/mutex.h>
68 #include <sys/proc.h>
69 #include <sys/sbuf.h>
70 #include <sys/sleepqueue.h>
71 #include <sys/sx.h>
72 #include <sys/sysctl.h>
73 #include <sys/types.h>
74 
75 #include <machine/atomic.h>
76 #include <machine/stdarg.h>
77 
78 #ifdef ILOG_DEFINE_FOR_FILE
79 ILOG_DEFINE_FOR_FILE(L_ISI_FAIL_POINT, L_ILOG, fail_point);
80 #endif
81 
82 static MALLOC_DEFINE(M_FAIL_POINT, "Fail Points", "fail points system");
83 #define fp_free(ptr) free(ptr, M_FAIL_POINT)
84 #define fp_malloc(size, flags) malloc((size), M_FAIL_POINT, (flags))
85 #define fs_free(ptr) fp_free(ptr)
86 #define fs_malloc() fp_malloc(sizeof(struct fail_point_setting), \
87     M_WAITOK | M_ZERO)
88 
89 /**
90  * These define the wchans that are used for sleeping, pausing respectively.
91  * They are chosen arbitrarily but need to be distinct to the failpoint and
92  * the sleep/pause distinction.
93  */
94 #define FP_SLEEP_CHANNEL(fp) (void*)(fp)
95 #define FP_PAUSE_CHANNEL(fp) __DEVOLATILE(void*, &fp->fp_setting)
96 
97 /**
98  * Don't allow more than this many entries in a fail point set by sysctl.
99  * The 99.99...% case is to have 1 entry.  I can't imagine having this many
100  * entries, so it should not limit us.  Saves on re-mallocs while holding
101  * a non-sleepable lock.
102  */
103 #define FP_MAX_ENTRY_COUNT 20
104 
105 /* Used to drain sbufs to the sysctl output */
106 int fail_sysctl_drain_func(void *, const char *, int);
107 
108 /* Head of tailq of struct fail_point_entry */
109 TAILQ_HEAD(fail_point_entry_queue, fail_point_entry);
110 
111 /**
112  * fp entries garbage list; outstanding entries are cleaned up in the
113  * garbage collector
114  */
115 STAILQ_HEAD(fail_point_setting_garbage, fail_point_setting);
116 static struct fail_point_setting_garbage fp_setting_garbage =
117         STAILQ_HEAD_INITIALIZER(fp_setting_garbage);
118 static struct mtx mtx_garbage_list;
119 MTX_SYSINIT(mtx_garbage_list, &mtx_garbage_list, "fail point garbage mtx",
120         MTX_SPIN);
121 
122 static struct sx sx_fp_set;
123 SX_SYSINIT(sx_fp_set, &sx_fp_set, "fail point set sx");
124 
125 /**
126  * Failpoint types.
127  * Don't change these without changing fail_type_strings in fail.c.
128  * @ingroup failpoint_private
129  */
130 enum fail_point_t {
131 	FAIL_POINT_OFF,		/**< don't fail */
132 	FAIL_POINT_PANIC,	/**< panic */
133 	FAIL_POINT_RETURN,	/**< return an errorcode */
134 	FAIL_POINT_BREAK,	/**< break into the debugger */
135 	FAIL_POINT_PRINT,	/**< print a message */
136 	FAIL_POINT_SLEEP,	/**< sleep for some msecs */
137 	FAIL_POINT_PAUSE,	/**< sleep until failpoint is set to off */
138 	FAIL_POINT_YIELD,	/**< yield the cpu */
139 	FAIL_POINT_DELAY,	/**< busy wait the cpu */
140 	FAIL_POINT_NUMTYPES,
141 	FAIL_POINT_INVALID = -1
142 };
143 
144 static struct {
145 	const char *name;
146 	int	nmlen;
147 } fail_type_strings[] = {
148 #define	FP_TYPE_NM_LEN(s)	{ s, sizeof(s) - 1 }
149 	[FAIL_POINT_OFF] =	FP_TYPE_NM_LEN("off"),
150 	[FAIL_POINT_PANIC] =	FP_TYPE_NM_LEN("panic"),
151 	[FAIL_POINT_RETURN] =	FP_TYPE_NM_LEN("return"),
152 	[FAIL_POINT_BREAK] =	FP_TYPE_NM_LEN("break"),
153 	[FAIL_POINT_PRINT] =	FP_TYPE_NM_LEN("print"),
154 	[FAIL_POINT_SLEEP] =	FP_TYPE_NM_LEN("sleep"),
155 	[FAIL_POINT_PAUSE] =	FP_TYPE_NM_LEN("pause"),
156 	[FAIL_POINT_YIELD] =	FP_TYPE_NM_LEN("yield"),
157 	[FAIL_POINT_DELAY] =	FP_TYPE_NM_LEN("delay"),
158 };
159 
160 #define FE_COUNT_UNTRACKED (INT_MIN)
161 
162 /**
163  * Internal structure tracking a single term of a complete failpoint.
164  * @ingroup failpoint_private
165  */
166 struct fail_point_entry {
167 	volatile bool	fe_stale;
168 	enum fail_point_t	fe_type;	/**< type of entry */
169 	int		fe_arg;		/**< argument to type (e.g. return value) */
170 	int		fe_prob;	/**< likelihood of firing in millionths */
171 	int32_t		fe_count;	/**< number of times to fire, -1 means infinite */
172 	pid_t		fe_pid;		/**< only fail for this process */
173 	struct fail_point	*fe_parent;	/**< backpointer to fp */
174 	TAILQ_ENTRY(fail_point_entry)	fe_entries; /**< next entry ptr */
175 };
176 
177 struct fail_point_setting {
178 	STAILQ_ENTRY(fail_point_setting) fs_garbage_link;
179 	struct fail_point_entry_queue fp_entry_queue;
180 	struct fail_point * fs_parent;
181 	struct mtx feq_mtx; /* Gives fail_point_pause something to do.  */
182 };
183 
184 /**
185  * Defines stating the equivalent of probablilty one (100%)
186  */
187 enum {
188 	PROB_MAX = 1000000,	/* probability between zero and this number */
189 	PROB_DIGITS = 6		/* number of zero's in above number */
190 };
191 
192 /* Get a ref on an fp's fp_setting */
193 static inline struct fail_point_setting *fail_point_setting_get_ref(
194         struct fail_point *fp);
195 /* Release a ref on an fp_setting */
196 static inline void fail_point_setting_release_ref(struct fail_point *fp);
197 /* Allocate and initialize a struct fail_point_setting */
198 static struct fail_point_setting *fail_point_setting_new(struct
199         fail_point *);
200 /* Free a struct fail_point_setting */
201 static void fail_point_setting_destroy(struct fail_point_setting *fp_setting);
202 /* Allocate and initialize a struct fail_point_entry */
203 static struct fail_point_entry *fail_point_entry_new(struct
204         fail_point_setting *);
205 /* Free a struct fail_point_entry */
206 static void fail_point_entry_destroy(struct fail_point_entry *fp_entry);
207 /* Append fp setting to garbage list */
208 static inline void fail_point_setting_garbage_append(
209         struct fail_point_setting *fp_setting);
210 /* Swap fp's setting with fp_setting_new */
211 static inline struct fail_point_setting *
212         fail_point_swap_settings(struct fail_point *fp,
213         struct fail_point_setting *fp_setting_new);
214 /* Free up any zero-ref setting in the garbage queue */
215 static void fail_point_garbage_collect(void);
216 /* If this fail point's setting are empty, then swap it out to NULL. */
217 static inline void fail_point_eval_swap_out(struct fail_point *fp,
218         struct fail_point_setting *fp_setting);
219 
220 bool
221 fail_point_is_off(struct fail_point *fp)
222 {
223 	bool return_val;
224 	struct fail_point_setting *fp_setting;
225 	struct fail_point_entry *ent;
226 
227 	return_val = true;
228 
229 	fp_setting = fail_point_setting_get_ref(fp);
230 	if (fp_setting != NULL) {
231 		TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue,
232 		    fe_entries) {
233 			if (!ent->fe_stale) {
234 				return_val = false;
235 				break;
236 			}
237 		}
238 	}
239 	fail_point_setting_release_ref(fp);
240 
241 	return (return_val);
242 }
243 
244 /* Allocate and initialize a struct fail_point_setting */
245 static struct fail_point_setting *
246 fail_point_setting_new(struct fail_point *fp)
247 {
248 	struct fail_point_setting *fs_new;
249 
250 	fs_new = fs_malloc();
251 	fs_new->fs_parent = fp;
252 	TAILQ_INIT(&fs_new->fp_entry_queue);
253 	mtx_init(&fs_new->feq_mtx, "fail point entries", NULL, MTX_SPIN);
254 
255 	fail_point_setting_garbage_append(fs_new);
256 
257 	return (fs_new);
258 }
259 
260 /* Free a struct fail_point_setting */
261 static void
262 fail_point_setting_destroy(struct fail_point_setting *fp_setting)
263 {
264 	struct fail_point_entry *ent;
265 
266 	while (!TAILQ_EMPTY(&fp_setting->fp_entry_queue)) {
267 		ent = TAILQ_FIRST(&fp_setting->fp_entry_queue);
268 		TAILQ_REMOVE(&fp_setting->fp_entry_queue, ent, fe_entries);
269 		fail_point_entry_destroy(ent);
270 	}
271 
272 	fs_free(fp_setting);
273 }
274 
275 /* Allocate and initialize a struct fail_point_entry */
276 static struct fail_point_entry *
277 fail_point_entry_new(struct fail_point_setting *fp_setting)
278 {
279 	struct fail_point_entry *fp_entry;
280 
281 	fp_entry = fp_malloc(sizeof(struct fail_point_entry),
282 	        M_WAITOK | M_ZERO);
283 	fp_entry->fe_parent = fp_setting->fs_parent;
284 	fp_entry->fe_prob = PROB_MAX;
285 	fp_entry->fe_pid = NO_PID;
286 	fp_entry->fe_count = FE_COUNT_UNTRACKED;
287 	TAILQ_INSERT_TAIL(&fp_setting->fp_entry_queue, fp_entry,
288 	        fe_entries);
289 
290 	return (fp_entry);
291 }
292 
293 /* Free a struct fail_point_entry */
294 static void
295 fail_point_entry_destroy(struct fail_point_entry *fp_entry)
296 {
297 
298 	fp_free(fp_entry);
299 }
300 
301 /* Get a ref on an fp's fp_setting */
302 static inline struct fail_point_setting *
303 fail_point_setting_get_ref(struct fail_point *fp)
304 {
305 	struct fail_point_setting *fp_setting;
306 
307 	/* Invariant: if we have a ref, our pointer to fp_setting is safe */
308 	atomic_add_acq_32(&fp->fp_ref_cnt, 1);
309 	fp_setting = fp->fp_setting;
310 
311 	return (fp_setting);
312 }
313 
314 /* Release a ref on an fp_setting */
315 static inline void
316 fail_point_setting_release_ref(struct fail_point *fp)
317 {
318 
319 	KASSERT(&fp->fp_ref_cnt > 0, ("Attempting to deref w/no refs"));
320 	atomic_subtract_rel_32(&fp->fp_ref_cnt, 1);
321 }
322 
323 /* Append fp entries to fp garbage list */
324 static inline void
325 fail_point_setting_garbage_append(struct fail_point_setting *fp_setting)
326 {
327 
328 	mtx_lock_spin(&mtx_garbage_list);
329 	STAILQ_INSERT_TAIL(&fp_setting_garbage, fp_setting,
330 	        fs_garbage_link);
331 	mtx_unlock_spin(&mtx_garbage_list);
332 }
333 
334 /* Swap fp's entries with fp_setting_new */
335 static struct fail_point_setting *
336 fail_point_swap_settings(struct fail_point *fp,
337         struct fail_point_setting *fp_setting_new)
338 {
339 	struct fail_point_setting *fp_setting_old;
340 
341 	fp_setting_old = fp->fp_setting;
342 	fp->fp_setting = fp_setting_new;
343 
344 	return (fp_setting_old);
345 }
346 
347 static inline void
348 fail_point_eval_swap_out(struct fail_point *fp,
349         struct fail_point_setting *fp_setting)
350 {
351 
352 	/* We may have already been swapped out and replaced; ignore. */
353 	if (fp->fp_setting == fp_setting)
354 		fail_point_swap_settings(fp, NULL);
355 }
356 
357 /* Free up any zero-ref entries in the garbage queue */
358 static void
359 fail_point_garbage_collect(void)
360 {
361 	struct fail_point_setting *fs_current, *fs_next;
362 	struct fail_point_setting_garbage fp_ents_free_list;
363 
364 	/**
365 	  * We will transfer the entries to free to fp_ents_free_list while holding
366 	  * the spin mutex, then free it after we drop the lock. This avoids
367 	  * triggering witness due to sleepable mutexes in the memory
368 	  * allocator.
369 	  */
370 	STAILQ_INIT(&fp_ents_free_list);
371 
372 	mtx_lock_spin(&mtx_garbage_list);
373 	STAILQ_FOREACH_SAFE(fs_current, &fp_setting_garbage, fs_garbage_link,
374 	    fs_next) {
375 		if (fs_current->fs_parent->fp_setting != fs_current &&
376 		        fs_current->fs_parent->fp_ref_cnt == 0) {
377 			STAILQ_REMOVE(&fp_setting_garbage, fs_current,
378 			        fail_point_setting, fs_garbage_link);
379 			STAILQ_INSERT_HEAD(&fp_ents_free_list, fs_current,
380 			        fs_garbage_link);
381 		}
382 	}
383 	mtx_unlock_spin(&mtx_garbage_list);
384 
385 	STAILQ_FOREACH_SAFE(fs_current, &fp_ents_free_list, fs_garbage_link,
386 	        fs_next)
387 		fail_point_setting_destroy(fs_current);
388 }
389 
390 /* Drain out all refs from this fail point */
391 static inline void
392 fail_point_drain(struct fail_point *fp, int expected_ref)
393 {
394 	struct fail_point_setting *entries;
395 
396 	entries = fail_point_swap_settings(fp, NULL);
397 	/**
398 	 * We have unpaused all threads; so we will wait no longer
399 	 * than the time taken for the longest remaining sleep, or
400 	 * the length of time of a long-running code block.
401 	 */
402 	while (fp->fp_ref_cnt > expected_ref) {
403 		wakeup(FP_PAUSE_CHANNEL(fp));
404 		tsleep(&fp, PWAIT, "fail_point_drain", hz / 100);
405 	}
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 			timeout(fp->fp_post_sleep_fn, fp->fp_post_sleep_arg,
446 			    timo);
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 /**
497  * Free the resources held by a fail_point, and wake any paused threads.
498  * Thou shalt not allow threads to hit this fail point after you enter this
499  * function, nor shall you call this multiple times for a given fp.
500  * @ingroup failpoint
501  */
502 void
503 fail_point_destroy(struct fail_point *fp)
504 {
505 
506 	fail_point_drain(fp, 0);
507 
508 	if ((fp->fp_flags & FAIL_POINT_DYNAMIC_NAME) != 0) {
509 		fp_free(__DECONST(void *, fp->fp_name));
510 		fp->fp_name = NULL;
511 	}
512 	fp->fp_flags = 0;
513 
514 	sx_xlock(&sx_fp_set);
515 	fail_point_garbage_collect();
516 	sx_xunlock(&sx_fp_set);
517 }
518 
519 /**
520  * This does the real work of evaluating a fail point. If the fail point tells
521  * us to return a value, this function returns 1 and fills in 'return_value'
522  * (return_value is allowed to be null). If the fail point tells us to panic,
523  * we never return. Otherwise we just return 0 after doing some work, which
524  * means "keep going".
525  */
526 enum fail_point_return_code
527 fail_point_eval_nontrivial(struct fail_point *fp, int *return_value)
528 {
529 	bool execute = false;
530 	struct fail_point_entry *ent;
531 	struct fail_point_setting *fp_setting;
532 	enum fail_point_return_code ret;
533 	int cont;
534 	int count;
535 	int msecs;
536 	int usecs;
537 
538 	ret = FAIL_POINT_RC_CONTINUE;
539 	cont = 0; /* don't continue by default */
540 
541 	fp_setting = fail_point_setting_get_ref(fp);
542 	if (fp_setting == NULL)
543 		goto abort;
544 
545 	TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
546 
547 		if (ent->fe_stale)
548 			continue;
549 
550 		if (ent->fe_prob < PROB_MAX &&
551 		    ent->fe_prob < random() % PROB_MAX)
552 			continue;
553 
554 		if (ent->fe_pid != NO_PID && ent->fe_pid != curproc->p_pid)
555 			continue;
556 
557 		if (ent->fe_count != FE_COUNT_UNTRACKED) {
558 			count = ent->fe_count;
559 			while (count > 0) {
560 				if (atomic_cmpset_32(&ent->fe_count, count, count - 1)) {
561 					count--;
562 					execute = true;
563 					break;
564 				}
565 				count = ent->fe_count;
566 			}
567 			if (execute == false)
568 				/* We lost the race; consider the entry stale and bail now */
569 				continue;
570 			if (count == 0)
571 				ent->fe_stale = true;
572 		}
573 
574 		switch (ent->fe_type) {
575 		case FAIL_POINT_PANIC:
576 			panic("fail point %s panicking", fp->fp_name);
577 			/* NOTREACHED */
578 
579 		case FAIL_POINT_RETURN:
580 			if (return_value != NULL)
581 				*return_value = ent->fe_arg;
582 			ret = FAIL_POINT_RC_RETURN;
583 			break;
584 
585 		case FAIL_POINT_BREAK:
586 			printf("fail point %s breaking to debugger\n",
587 			        fp->fp_name);
588 			breakpoint();
589 			break;
590 
591 		case FAIL_POINT_PRINT:
592 			printf("fail point %s executing\n", fp->fp_name);
593 			cont = ent->fe_arg;
594 			break;
595 
596 		case FAIL_POINT_SLEEP:
597 			msecs = ent->fe_arg;
598 			if (msecs)
599 				fail_point_sleep(fp, msecs, &ret);
600 			break;
601 
602 		case FAIL_POINT_PAUSE:
603 			/**
604 			 * Pausing is inherently strange with multiple
605 			 * entries given our design.  That is because some
606 			 * entries could be unreachable, for instance in cases like:
607 			 * pause->return. We can never reach the return entry.
608 			 * The sysctl layer actually truncates all entries after
609 			 * a pause for this reason.
610 			 */
611 			mtx_lock_spin(&fp_setting->feq_mtx);
612 			fail_point_pause(fp, &ret, &fp_setting->feq_mtx);
613 			mtx_unlock_spin(&fp_setting->feq_mtx);
614 			break;
615 
616 		case FAIL_POINT_YIELD:
617 			kern_yield(PRI_UNCHANGED);
618 			break;
619 
620 		case FAIL_POINT_DELAY:
621 			usecs = ent->fe_arg;
622 			DELAY(usecs);
623 			break;
624 
625 		default:
626 			break;
627 		}
628 
629 		if (cont == 0)
630 			break;
631 	}
632 
633 	if (fail_point_is_off(fp))
634 		fail_point_eval_swap_out(fp, fp_setting);
635 
636 abort:
637 	fail_point_setting_release_ref(fp);
638 
639 	return (ret);
640 }
641 
642 /**
643  * Translate internal fail_point structure into human-readable text.
644  */
645 static void
646 fail_point_get(struct fail_point *fp, struct sbuf *sb,
647         bool verbose)
648 {
649 	struct fail_point_entry *ent;
650 	struct fail_point_setting *fp_setting;
651 	struct fail_point_entry *fp_entry_cpy;
652 	int cnt_sleeping;
653 	int idx;
654 	int printed_entry_count;
655 
656 	cnt_sleeping = 0;
657 	idx = 0;
658 	printed_entry_count = 0;
659 
660 	fp_entry_cpy = fp_malloc(sizeof(struct fail_point_entry) *
661 	        (FP_MAX_ENTRY_COUNT + 1), M_WAITOK);
662 
663 	fp_setting = fail_point_setting_get_ref(fp);
664 
665 	if (fp_setting != NULL) {
666 		TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
667 			if (ent->fe_stale)
668 				continue;
669 
670 			KASSERT(printed_entry_count < FP_MAX_ENTRY_COUNT,
671 			        ("FP entry list larger than allowed"));
672 
673 			fp_entry_cpy[printed_entry_count] = *ent;
674 			++printed_entry_count;
675 		}
676 	}
677 	fail_point_setting_release_ref(fp);
678 
679 	/* This is our equivalent of a NULL terminator */
680 	fp_entry_cpy[printed_entry_count].fe_type = FAIL_POINT_INVALID;
681 
682 	while (idx < printed_entry_count) {
683 		ent = &fp_entry_cpy[idx];
684 		++idx;
685 		if (ent->fe_prob < PROB_MAX) {
686 			int decimal = ent->fe_prob % (PROB_MAX / 100);
687 			int units = ent->fe_prob / (PROB_MAX / 100);
688 			sbuf_printf(sb, "%d", units);
689 			if (decimal) {
690 				int digits = PROB_DIGITS - 2;
691 				while (!(decimal % 10)) {
692 					digits--;
693 					decimal /= 10;
694 				}
695 				sbuf_printf(sb, ".%0*d", digits, decimal);
696 			}
697 			sbuf_printf(sb, "%%");
698 		}
699 		if (ent->fe_count >= 0)
700 			sbuf_printf(sb, "%d*", ent->fe_count);
701 		sbuf_printf(sb, "%s", fail_type_strings[ent->fe_type].name);
702 		if (ent->fe_arg)
703 			sbuf_printf(sb, "(%d)", ent->fe_arg);
704 		if (ent->fe_pid != NO_PID)
705 			sbuf_printf(sb, "[pid %d]", ent->fe_pid);
706 		if (TAILQ_NEXT(ent, fe_entries))
707 			sbuf_printf(sb, "->");
708 	}
709 	if (!printed_entry_count)
710 		sbuf_printf(sb, "off");
711 
712 	fp_free(fp_entry_cpy);
713 	if (verbose) {
714 #ifdef STACK
715 		/* Print number of sleeping threads. queue=0 is the argument
716 		 * used by msleep when sending our threads to sleep. */
717 		sbuf_printf(sb, "\nsleeping_thread_stacks = {\n");
718 		sleepq_sbuf_print_stacks(sb, FP_SLEEP_CHANNEL(fp), 0,
719 		        &cnt_sleeping);
720 
721 		sbuf_printf(sb, "},\n");
722 #endif
723 		sbuf_printf(sb, "sleeping_thread_count = %d,\n",
724 		        cnt_sleeping);
725 
726 #ifdef STACK
727 		sbuf_printf(sb, "paused_thread_stacks = {\n");
728 		sleepq_sbuf_print_stacks(sb, FP_PAUSE_CHANNEL(fp), 0,
729 		        &cnt_sleeping);
730 
731 		sbuf_printf(sb, "},\n");
732 #endif
733 		sbuf_printf(sb, "paused_thread_count = %d\n",
734 		        cnt_sleeping);
735 	}
736 }
737 
738 /**
739  * Set an internal fail_point structure from a human-readable failpoint string
740  * in a lock-safe manner.
741  */
742 static int
743 fail_point_set(struct fail_point *fp, char *buf)
744 {
745 	struct fail_point_entry *ent, *ent_next;
746 	struct fail_point_setting *entries;
747 	bool should_wake_paused;
748 	bool should_truncate;
749 	int error;
750 
751 	error = 0;
752 	should_wake_paused = false;
753 	should_truncate = false;
754 
755 	/* Parse new entries. */
756 	/**
757 	 * ref protects our new malloc'd stuff from being garbage collected
758 	 * before we link it.
759 	 */
760 	fail_point_setting_get_ref(fp);
761 	entries = fail_point_setting_new(fp);
762 	if (parse_fail_point(entries, buf) == NULL) {
763 		STAILQ_REMOVE(&fp_setting_garbage, entries,
764 		        fail_point_setting, fs_garbage_link);
765 		fail_point_setting_destroy(entries);
766 		error = EINVAL;
767 		goto end;
768 	}
769 
770 	/**
771 	 * Transfer the entries we are going to keep to a new list.
772 	 * Get rid of useless zero probability entries, and entries with hit
773 	 * count 0.
774 	 * If 'off' is present, and it has no hit count set, then all entries
775 	 *       after it are discarded since they are unreachable.
776 	 */
777 	TAILQ_FOREACH_SAFE(ent, &entries->fp_entry_queue, fe_entries, ent_next) {
778 		if (ent->fe_prob == 0 || ent->fe_count == 0) {
779 			printf("Discarding entry which cannot execute %s\n",
780 			        fail_type_strings[ent->fe_type].name);
781 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
782 			        fe_entries);
783 			fp_free(ent);
784 			continue;
785 		} else if (should_truncate) {
786 			printf("Discarding unreachable entry %s\n",
787 			        fail_type_strings[ent->fe_type].name);
788 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
789 			        fe_entries);
790 			fp_free(ent);
791 			continue;
792 		}
793 
794 		if (ent->fe_type == FAIL_POINT_OFF) {
795 			should_wake_paused = true;
796 			if (ent->fe_count == FE_COUNT_UNTRACKED) {
797 				should_truncate = true;
798 				TAILQ_REMOVE(&entries->fp_entry_queue, ent,
799 				        fe_entries);
800 				fp_free(ent);
801 			}
802 		} else if (ent->fe_type == FAIL_POINT_PAUSE) {
803 			should_truncate = true;
804 		} else if (ent->fe_type == FAIL_POINT_SLEEP && (fp->fp_flags &
805 		        FAIL_POINT_NONSLEEPABLE)) {
806 			/**
807 			 * If this fail point is annotated as being in a
808 			 * non-sleepable ctx, convert sleep to delay and
809 			 * convert the msec argument to usecs.
810 			 */
811 			printf("Sleep call request on fail point in "
812 			        "non-sleepable context; using delay instead "
813 			        "of sleep\n");
814 			ent->fe_type = FAIL_POINT_DELAY;
815 			ent->fe_arg *= 1000;
816 		}
817 	}
818 
819 	if (TAILQ_EMPTY(&entries->fp_entry_queue)) {
820 		entries = fail_point_swap_settings(fp, NULL);
821 		if (entries != NULL)
822 			wakeup(FP_PAUSE_CHANNEL(fp));
823 	} else {
824 		if (should_wake_paused)
825 			wakeup(FP_PAUSE_CHANNEL(fp));
826 		fail_point_swap_settings(fp, entries);
827 	}
828 
829 end:
830 #ifdef IWARNING
831 	if (error)
832 		IWARNING("Failed to set %s %s to %s",
833 		    fp->fp_name, fp->fp_location, buf);
834 	else
835 		INOTICE("Set %s %s to %s",
836 		    fp->fp_name, fp->fp_location, buf);
837 #endif /* IWARNING */
838 
839 	fail_point_setting_release_ref(fp);
840 	return (error);
841 }
842 
843 #define MAX_FAIL_POINT_BUF	1023
844 
845 /**
846  * Handle kernel failpoint set/get.
847  */
848 int
849 fail_point_sysctl(SYSCTL_HANDLER_ARGS)
850 {
851 	struct fail_point *fp;
852 	char *buf;
853 	struct sbuf sb, *sb_check;
854 	int error;
855 
856 	buf = NULL;
857 	error = 0;
858 	fp = arg1;
859 
860 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
861 	if (sb_check != &sb)
862 		return (ENOMEM);
863 
864 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
865 
866 	/* Setting */
867 	/**
868 	 * Lock protects any new entries from being garbage collected before we
869 	 * can link them to the fail point.
870 	 */
871 	sx_xlock(&sx_fp_set);
872 	if (req->newptr) {
873 		if (req->newlen > MAX_FAIL_POINT_BUF) {
874 			error = EINVAL;
875 			goto out;
876 		}
877 
878 		buf = fp_malloc(req->newlen + 1, M_WAITOK);
879 
880 		error = SYSCTL_IN(req, buf, req->newlen);
881 		if (error)
882 			goto out;
883 		buf[req->newlen] = '\0';
884 
885 		error = fail_point_set(fp, buf);
886 	}
887 
888 	fail_point_garbage_collect();
889 	sx_xunlock(&sx_fp_set);
890 
891 	/* Retrieving. */
892 	fail_point_get(fp, &sb, false);
893 
894 out:
895 	sbuf_finish(&sb);
896 	sbuf_delete(&sb);
897 
898 	if (buf)
899 		fp_free(buf);
900 
901 	return (error);
902 }
903 
904 int
905 fail_point_sysctl_status(SYSCTL_HANDLER_ARGS)
906 {
907 	struct fail_point *fp;
908 	struct sbuf sb, *sb_check;
909 
910 	fp = arg1;
911 
912 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
913 	if (sb_check != &sb)
914 		return (ENOMEM);
915 
916 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
917 
918 	/* Retrieving. */
919 	fail_point_get(fp, &sb, true);
920 
921 	sbuf_finish(&sb);
922 	sbuf_delete(&sb);
923 
924 	/**
925 	 * Lock protects any new entries from being garbage collected before we
926 	 * can link them to the fail point.
927 	 */
928 	sx_xlock(&sx_fp_set);
929 	fail_point_garbage_collect();
930 	sx_xunlock(&sx_fp_set);
931 
932 	return (0);
933 }
934 
935 int
936 fail_sysctl_drain_func(void *sysctl_args, const char *buf, int len)
937 {
938 	struct sysctl_req *sa;
939 	int error;
940 
941 	sa = sysctl_args;
942 
943 	error = SYSCTL_OUT(sa, buf, len);
944 
945 	if (error == ENOMEM)
946 		return (-1);
947 	else
948 		return (len);
949 }
950 
951 /**
952  * Internal helper function to translate a human-readable failpoint string
953  * into a internally-parsable fail_point structure.
954  */
955 static char *
956 parse_fail_point(struct fail_point_setting *ents, char *p)
957 {
958 	/*  <fail_point> ::
959 	 *      <term> ( "->" <term> )*
960 	 */
961 	uint8_t term_count;
962 
963 	term_count = 1;
964 
965 	p = parse_term(ents, p);
966 	if (p == NULL)
967 		return (NULL);
968 
969 	while (*p != '\0') {
970 		term_count++;
971 		if (p[0] != '-' || p[1] != '>' ||
972 		        (p = parse_term(ents, p+2)) == NULL ||
973 		        term_count > FP_MAX_ENTRY_COUNT)
974 			return (NULL);
975 	}
976 	return (p);
977 }
978 
979 /**
980  * Internal helper function to parse an individual term from a failpoint.
981  */
982 static char *
983 parse_term(struct fail_point_setting *ents, char *p)
984 {
985 	struct fail_point_entry *ent;
986 
987 	ent = fail_point_entry_new(ents);
988 
989 	/*
990 	 * <term> ::
991 	 *     ( (<float> "%") | (<integer> "*" ) )*
992 	 *     <type>
993 	 *     [ "(" <integer> ")" ]
994 	 *     [ "[pid " <integer> "]" ]
995 	 */
996 
997 	/* ( (<float> "%") | (<integer> "*" ) )* */
998 	while (isdigit(*p) || *p == '.') {
999 		int units, decimal;
1000 
1001 		p = parse_number(&units, &decimal, p);
1002 		if (p == NULL)
1003 			return (NULL);
1004 
1005 		if (*p == '%') {
1006 			if (units > 100) /* prevent overflow early */
1007 				units = 100;
1008 			ent->fe_prob = units * (PROB_MAX / 100) + decimal;
1009 			if (ent->fe_prob > PROB_MAX)
1010 				ent->fe_prob = PROB_MAX;
1011 		} else if (*p == '*') {
1012 			if (!units || units < 0 || decimal)
1013 				return (NULL);
1014 			ent->fe_count = units;
1015 		} else
1016 			return (NULL);
1017 		p++;
1018 	}
1019 
1020 	/* <type> */
1021 	p = parse_type(ent, p);
1022 	if (p == NULL)
1023 		return (NULL);
1024 	if (*p == '\0')
1025 		return (p);
1026 
1027 	/* [ "(" <integer> ")" ] */
1028 	if (*p != '(')
1029 		return (p);
1030 	p++;
1031 	if (!isdigit(*p) && *p != '-')
1032 		return (NULL);
1033 	ent->fe_arg = strtol(p, &p, 0);
1034 	if (*p++ != ')')
1035 		return (NULL);
1036 
1037 	/* [ "[pid " <integer> "]" ] */
1038 #define PID_STRING "[pid "
1039 	if (strncmp(p, PID_STRING, sizeof(PID_STRING) - 1) != 0)
1040 		return (p);
1041 	p += sizeof(PID_STRING) - 1;
1042 	if (!isdigit(*p))
1043 		return (NULL);
1044 	ent->fe_pid = strtol(p, &p, 0);
1045 	if (*p++ != ']')
1046 		return (NULL);
1047 
1048 	return (p);
1049 }
1050 
1051 /**
1052  * Internal helper function to parse a numeric for a failpoint term.
1053  */
1054 static char *
1055 parse_number(int *out_units, int *out_decimal, char *p)
1056 {
1057 	char *old_p;
1058 
1059 	/**
1060 	 *  <number> ::
1061 	 *      <integer> [ "." <integer> ] |
1062 	 *      "." <integer>
1063 	 */
1064 
1065 	/* whole part */
1066 	old_p = p;
1067 	*out_units = strtol(p, &p, 10);
1068 	if (p == old_p && *p != '.')
1069 		return (NULL);
1070 
1071 	/* fractional part */
1072 	*out_decimal = 0;
1073 	if (*p == '.') {
1074 		int digits = 0;
1075 		p++;
1076 		while (isdigit(*p)) {
1077 			int digit = *p - '0';
1078 			if (digits < PROB_DIGITS - 2)
1079 				*out_decimal = *out_decimal * 10 + digit;
1080 			else if (digits == PROB_DIGITS - 2 && digit >= 5)
1081 				(*out_decimal)++;
1082 			digits++;
1083 			p++;
1084 		}
1085 		if (!digits) /* need at least one digit after '.' */
1086 			return (NULL);
1087 		while (digits++ < PROB_DIGITS - 2) /* add implicit zeros */
1088 			*out_decimal *= 10;
1089 	}
1090 
1091 	return (p); /* success */
1092 }
1093 
1094 /**
1095  * Internal helper function to parse an individual type for a failpoint term.
1096  */
1097 static char *
1098 parse_type(struct fail_point_entry *ent, char *beg)
1099 {
1100 	enum fail_point_t type;
1101 	int len;
1102 
1103 	for (type = FAIL_POINT_OFF; type < FAIL_POINT_NUMTYPES; type++) {
1104 		len = fail_type_strings[type].nmlen;
1105 		if (strncmp(fail_type_strings[type].name, beg, len) == 0) {
1106 			ent->fe_type = type;
1107 			return (beg + len);
1108 		}
1109 	}
1110 	return (NULL);
1111 }
1112 
1113 /* The fail point sysctl tree. */
1114 SYSCTL_NODE(_debug, OID_AUTO, fail_point, CTLFLAG_RW, 0, "fail points");
1115 
1116 /* Debugging/testing stuff for fail point */
1117 static int
1118 sysctl_test_fail_point(SYSCTL_HANDLER_ARGS)
1119 {
1120 
1121 	KFAIL_POINT_RETURN(DEBUG_FP, test_fail_point);
1122 	return (0);
1123 }
1124 SYSCTL_OID(_debug_fail_point, OID_AUTO, test_trigger_fail_point,
1125         CTLTYPE_STRING | CTLFLAG_RD, NULL, 0, sysctl_test_fail_point, "A",
1126         "Trigger test fail points");
1127