xref: /freebsd/sys/kern/kern_fail.c (revision 331613ddd8a516e8eaf841f293754fa47cb339aa)
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/stdarg.h>
70 #include <sys/sx.h>
71 #include <sys/sysctl.h>
72 #include <sys/types.h>
73 
74 #include <machine/atomic.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 };
180 
181 /**
182  * Defines stating the equivalent of probablilty one (100%)
183  */
184 enum {
185 	PROB_MAX = 1000000,	/* probability between zero and this number */
186 	PROB_DIGITS = 6		/* number of zero's in above number */
187 };
188 
189 /* Get a ref on an fp's fp_setting */
190 static inline struct fail_point_setting *fail_point_setting_get_ref(
191         struct fail_point *fp);
192 /* Release a ref on an fp_setting */
193 static inline void fail_point_setting_release_ref(struct fail_point *fp);
194 /* Allocate and initialize a struct fail_point_setting */
195 static struct fail_point_setting *fail_point_setting_new(struct
196         fail_point *);
197 /* Free a struct fail_point_setting */
198 static void fail_point_setting_destroy(struct fail_point_setting *fp_setting);
199 /* Allocate and initialize a struct fail_point_entry */
200 static struct fail_point_entry *fail_point_entry_new(struct
201         fail_point_setting *);
202 /* Free a struct fail_point_entry */
203 static void fail_point_entry_destroy(struct fail_point_entry *fp_entry);
204 /* Append fp setting to garbage list */
205 static inline void fail_point_setting_garbage_append(
206         struct fail_point_setting *fp_setting);
207 /* Swap fp's setting with fp_setting_new */
208 static inline struct fail_point_setting *
209         fail_point_swap_settings(struct fail_point *fp,
210         struct fail_point_setting *fp_setting_new);
211 /* Free up any zero-ref setting in the garbage queue */
212 static void fail_point_garbage_collect(void);
213 /* If this fail point's setting are empty, then swap it out to NULL. */
214 static inline void fail_point_eval_swap_out(struct fail_point *fp,
215         struct fail_point_setting *fp_setting);
216 
217 bool
fail_point_is_off(struct fail_point * fp)218 fail_point_is_off(struct fail_point *fp)
219 {
220 	bool return_val;
221 	struct fail_point_setting *fp_setting;
222 	struct fail_point_entry *ent;
223 
224 	return_val = true;
225 
226 	fp_setting = fail_point_setting_get_ref(fp);
227 	if (fp_setting != NULL) {
228 		TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue,
229 		    fe_entries) {
230 			if (!ent->fe_stale) {
231 				return_val = false;
232 				break;
233 			}
234 		}
235 	}
236 	fail_point_setting_release_ref(fp);
237 
238 	return (return_val);
239 }
240 
241 /* Allocate and initialize a struct fail_point_setting */
242 static struct fail_point_setting *
fail_point_setting_new(struct fail_point * fp)243 fail_point_setting_new(struct fail_point *fp)
244 {
245 	struct fail_point_setting *fs_new;
246 
247 	fs_new = fs_malloc();
248 	fs_new->fs_parent = fp;
249 	TAILQ_INIT(&fs_new->fp_entry_queue);
250 
251 	fail_point_setting_garbage_append(fs_new);
252 
253 	return (fs_new);
254 }
255 
256 /* Free a struct fail_point_setting */
257 static void
fail_point_setting_destroy(struct fail_point_setting * fp_setting)258 fail_point_setting_destroy(struct fail_point_setting *fp_setting)
259 {
260 	struct fail_point_entry *ent;
261 
262 	while (!TAILQ_EMPTY(&fp_setting->fp_entry_queue)) {
263 		ent = TAILQ_FIRST(&fp_setting->fp_entry_queue);
264 		TAILQ_REMOVE(&fp_setting->fp_entry_queue, ent, fe_entries);
265 		fail_point_entry_destroy(ent);
266 	}
267 
268 	fs_free(fp_setting);
269 }
270 
271 /* Allocate and initialize a struct fail_point_entry */
272 static struct fail_point_entry *
fail_point_entry_new(struct fail_point_setting * fp_setting)273 fail_point_entry_new(struct fail_point_setting *fp_setting)
274 {
275 	struct fail_point_entry *fp_entry;
276 
277 	fp_entry = fp_malloc(sizeof(struct fail_point_entry),
278 	        M_WAITOK | M_ZERO);
279 	fp_entry->fe_parent = fp_setting->fs_parent;
280 	fp_entry->fe_prob = PROB_MAX;
281 	fp_entry->fe_pid = NO_PID;
282 	fp_entry->fe_count = FE_COUNT_UNTRACKED;
283 	TAILQ_INSERT_TAIL(&fp_setting->fp_entry_queue, fp_entry,
284 	        fe_entries);
285 
286 	return (fp_entry);
287 }
288 
289 /* Free a struct fail_point_entry */
290 static void
fail_point_entry_destroy(struct fail_point_entry * fp_entry)291 fail_point_entry_destroy(struct fail_point_entry *fp_entry)
292 {
293 
294 	fp_free(fp_entry);
295 }
296 
297 /* Get a ref on an fp's fp_setting */
298 static inline struct fail_point_setting *
fail_point_setting_get_ref(struct fail_point * fp)299 fail_point_setting_get_ref(struct fail_point *fp)
300 {
301 	struct fail_point_setting *fp_setting;
302 
303 	/* Invariant: if we have a ref, our pointer to fp_setting is safe */
304 	atomic_add_acq_32(&fp->fp_ref_cnt, 1);
305 	fp_setting = fp->fp_setting;
306 
307 	return (fp_setting);
308 }
309 
310 /* Release a ref on an fp_setting */
311 static inline void
fail_point_setting_release_ref(struct fail_point * fp)312 fail_point_setting_release_ref(struct fail_point *fp)
313 {
314 
315 	KASSERT(&fp->fp_ref_cnt > 0, ("Attempting to deref w/no refs"));
316 	atomic_subtract_rel_32(&fp->fp_ref_cnt, 1);
317 }
318 
319 /* Append fp entries to fp garbage list */
320 static inline void
fail_point_setting_garbage_append(struct fail_point_setting * fp_setting)321 fail_point_setting_garbage_append(struct fail_point_setting *fp_setting)
322 {
323 
324 	mtx_lock_spin(&mtx_garbage_list);
325 	STAILQ_INSERT_TAIL(&fp_setting_garbage, fp_setting,
326 	        fs_garbage_link);
327 	mtx_unlock_spin(&mtx_garbage_list);
328 }
329 
330 /* Swap fp's entries with fp_setting_new */
331 static struct fail_point_setting *
fail_point_swap_settings(struct fail_point * fp,struct fail_point_setting * fp_setting_new)332 fail_point_swap_settings(struct fail_point *fp,
333         struct fail_point_setting *fp_setting_new)
334 {
335 	struct fail_point_setting *fp_setting_old;
336 
337 	fp_setting_old = fp->fp_setting;
338 	fp->fp_setting = fp_setting_new;
339 
340 	return (fp_setting_old);
341 }
342 
343 static inline void
fail_point_eval_swap_out(struct fail_point * fp,struct fail_point_setting * fp_setting)344 fail_point_eval_swap_out(struct fail_point *fp,
345         struct fail_point_setting *fp_setting)
346 {
347 
348 	/* We may have already been swapped out and replaced; ignore. */
349 	if (fp->fp_setting == fp_setting)
350 		fail_point_swap_settings(fp, NULL);
351 }
352 
353 /* Free up any zero-ref entries in the garbage queue */
354 static void
fail_point_garbage_collect(void)355 fail_point_garbage_collect(void)
356 {
357 	struct fail_point_setting *fs_current, *fs_next;
358 	struct fail_point_setting_garbage fp_ents_free_list;
359 
360 	/**
361 	  * We will transfer the entries to free to fp_ents_free_list while holding
362 	  * the spin mutex, then free it after we drop the lock. This avoids
363 	  * triggering witness due to sleepable mutexes in the memory
364 	  * allocator.
365 	  */
366 	STAILQ_INIT(&fp_ents_free_list);
367 
368 	mtx_lock_spin(&mtx_garbage_list);
369 	STAILQ_FOREACH_SAFE(fs_current, &fp_setting_garbage, fs_garbage_link,
370 	    fs_next) {
371 		if (fs_current->fs_parent->fp_setting != fs_current &&
372 		        fs_current->fs_parent->fp_ref_cnt == 0) {
373 			STAILQ_REMOVE(&fp_setting_garbage, fs_current,
374 			        fail_point_setting, fs_garbage_link);
375 			STAILQ_INSERT_HEAD(&fp_ents_free_list, fs_current,
376 			        fs_garbage_link);
377 		}
378 	}
379 	mtx_unlock_spin(&mtx_garbage_list);
380 
381 	STAILQ_FOREACH_SAFE(fs_current, &fp_ents_free_list, fs_garbage_link,
382 	        fs_next)
383 		fail_point_setting_destroy(fs_current);
384 }
385 
386 /* Drain out all refs from this fail point */
387 static inline void
fail_point_drain(struct fail_point * fp,int expected_ref)388 fail_point_drain(struct fail_point *fp, int expected_ref)
389 {
390 	struct fail_point_setting *entries;
391 
392 	entries = fail_point_swap_settings(fp, NULL);
393 	/**
394 	 * We have unpaused all threads; so we will wait no longer
395 	 * than the time taken for the longest remaining sleep, or
396 	 * the length of time of a long-running code block.
397 	 */
398 	while (fp->fp_ref_cnt > expected_ref) {
399 		wakeup(FP_PAUSE_CHANNEL(fp));
400 		tsleep(&fp, PWAIT, "fail_point_drain", hz / 100);
401 	}
402 	if (fp->fp_callout)
403 		callout_drain(fp->fp_callout);
404 	fail_point_swap_settings(fp, entries);
405 }
406 
407 static inline void
fail_point_pause(struct fail_point * fp,enum fail_point_return_code * pret)408 fail_point_pause(struct fail_point *fp, enum fail_point_return_code *pret)
409 {
410 
411 	if (fp->fp_pre_sleep_fn)
412 		fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
413 
414 	tsleep(FP_PAUSE_CHANNEL(fp), 0, "failpt", 0);
415 
416 	if (fp->fp_post_sleep_fn)
417 		fp->fp_post_sleep_fn(fp->fp_post_sleep_arg);
418 }
419 
420 static inline void
fail_point_sleep(struct fail_point * fp,int msecs,enum fail_point_return_code * pret)421 fail_point_sleep(struct fail_point *fp, int msecs,
422         enum fail_point_return_code *pret)
423 {
424 	int timo;
425 
426 	/* Convert from millisecs to ticks, rounding up */
427 	timo = howmany((int64_t)msecs * hz, 1000L);
428 
429 	if (timo > 0) {
430 		if (!(fp->fp_flags & FAIL_POINT_USE_TIMEOUT_PATH)) {
431 			if (fp->fp_pre_sleep_fn)
432 				fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
433 
434 			tsleep(FP_SLEEP_CHANNEL(fp), PWAIT, "failpt", timo);
435 
436 			if (fp->fp_post_sleep_fn)
437 				fp->fp_post_sleep_fn(fp->fp_post_sleep_arg);
438 		} else {
439 			if (fp->fp_pre_sleep_fn)
440 				fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
441 
442 			callout_reset(fp->fp_callout, timo,
443 			    fp->fp_post_sleep_fn, fp->fp_post_sleep_arg);
444 			*pret = FAIL_POINT_RC_QUEUED;
445 		}
446 	}
447 }
448 
449 static char *parse_fail_point(struct fail_point_setting *, char *);
450 static char *parse_term(struct fail_point_setting *, char *);
451 static char *parse_number(int *out_units, int *out_decimal, char *);
452 static char *parse_type(struct fail_point_entry *, char *);
453 
454 /**
455  * Initialize a fail_point.  The name is formed in a printf-like fashion
456  * from "fmt" and subsequent arguments.  This function is generally used
457  * for custom failpoints located at odd places in the sysctl tree, and is
458  * not explicitly needed for standard in-line-declared failpoints.
459  *
460  * @ingroup failpoint
461  */
462 void
fail_point_init(struct fail_point * fp,const char * fmt,...)463 fail_point_init(struct fail_point *fp, const char *fmt, ...)
464 {
465 	va_list ap;
466 	char *name;
467 	int n;
468 
469 	fp->fp_setting = NULL;
470 	fp->fp_flags = 0;
471 
472 	/* Figure out the size of the name. */
473 	va_start(ap, fmt);
474 	n = vsnprintf(NULL, 0, fmt, ap);
475 	va_end(ap);
476 
477 	/* Allocate the name and fill it in. */
478 	name = fp_malloc(n + 1, M_WAITOK);
479 	va_start(ap, fmt);
480 	vsnprintf(name, n + 1, fmt, ap);
481 	va_end(ap);
482 
483 	fp->fp_name = name;
484 	fp->fp_location = "";
485 	fp->fp_flags |= FAIL_POINT_DYNAMIC_NAME;
486 	fp->fp_pre_sleep_fn = NULL;
487 	fp->fp_pre_sleep_arg = NULL;
488 	fp->fp_post_sleep_fn = NULL;
489 	fp->fp_post_sleep_arg = NULL;
490 }
491 
492 void
fail_point_alloc_callout(struct fail_point * fp)493 fail_point_alloc_callout(struct fail_point *fp)
494 {
495 
496 	/**
497 	 * This assumes that calls to fail_point_use_timeout_path()
498 	 * will not race.
499 	 */
500 	if (fp->fp_callout != NULL)
501 		return;
502 	fp->fp_callout = fp_malloc(sizeof(*fp->fp_callout), M_WAITOK);
503 	callout_init(fp->fp_callout, CALLOUT_MPSAFE);
504 }
505 
506 /**
507  * Free the resources held by a fail_point, and wake any paused threads.
508  * Thou shalt not allow threads to hit this fail point after you enter this
509  * function, nor shall you call this multiple times for a given fp.
510  * @ingroup failpoint
511  */
512 void
fail_point_destroy(struct fail_point * fp)513 fail_point_destroy(struct fail_point *fp)
514 {
515 
516 	fail_point_drain(fp, 0);
517 
518 	if ((fp->fp_flags & FAIL_POINT_DYNAMIC_NAME) != 0) {
519 		fp_free(__DECONST(void *, fp->fp_name));
520 		fp->fp_name = NULL;
521 	}
522 	fp->fp_flags = 0;
523 	if (fp->fp_callout) {
524 		fp_free(fp->fp_callout);
525 		fp->fp_callout = NULL;
526 	}
527 
528 	sx_xlock(&sx_fp_set);
529 	fail_point_garbage_collect();
530 	sx_xunlock(&sx_fp_set);
531 }
532 
533 /**
534  * This does the real work of evaluating a fail point. If the fail point tells
535  * us to return a value, this function returns 1 and fills in 'return_value'
536  * (return_value is allowed to be null). If the fail point tells us to panic,
537  * we never return. Otherwise we just return 0 after doing some work, which
538  * means "keep going".
539  */
540 enum fail_point_return_code
fail_point_eval_nontrivial(struct fail_point * fp,int * return_value)541 fail_point_eval_nontrivial(struct fail_point *fp, int *return_value)
542 {
543 	bool execute = false;
544 	struct fail_point_entry *ent;
545 	struct fail_point_setting *fp_setting;
546 	enum fail_point_return_code ret;
547 	int cont;
548 	int count;
549 	int msecs;
550 	int usecs;
551 
552 	ret = FAIL_POINT_RC_CONTINUE;
553 	cont = 0; /* don't continue by default */
554 
555 	fp_setting = fail_point_setting_get_ref(fp);
556 	if (fp_setting == NULL)
557 		goto abort;
558 
559 	TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
560 		if (ent->fe_stale)
561 			continue;
562 
563 		if (ent->fe_prob < PROB_MAX &&
564 		    ent->fe_prob < random() % PROB_MAX)
565 			continue;
566 
567 		if (ent->fe_pid != NO_PID && ent->fe_pid != curproc->p_pid)
568 			continue;
569 
570 		if (ent->fe_count != FE_COUNT_UNTRACKED) {
571 			count = ent->fe_count;
572 			while (count > 0) {
573 				if (atomic_cmpset_32(&ent->fe_count, count, count - 1)) {
574 					count--;
575 					execute = true;
576 					break;
577 				}
578 				count = ent->fe_count;
579 			}
580 			if (execute == false)
581 				/* We lost the race; consider the entry stale and bail now */
582 				continue;
583 			if (count == 0)
584 				ent->fe_stale = true;
585 		}
586 
587 		switch (ent->fe_type) {
588 		case FAIL_POINT_PANIC:
589 			panic("fail point %s panicking", fp->fp_name);
590 			/* NOTREACHED */
591 
592 		case FAIL_POINT_RETURN:
593 			if (return_value != NULL)
594 				*return_value = ent->fe_arg;
595 			ret = FAIL_POINT_RC_RETURN;
596 			break;
597 
598 		case FAIL_POINT_BREAK:
599 			printf("fail point %s breaking to debugger\n",
600 			        fp->fp_name);
601 			breakpoint();
602 			break;
603 
604 		case FAIL_POINT_PRINT:
605 			printf("fail point %s executing\n", fp->fp_name);
606 			cont = ent->fe_arg;
607 			break;
608 
609 		case FAIL_POINT_SLEEP:
610 			msecs = ent->fe_arg;
611 			if (msecs)
612 				fail_point_sleep(fp, msecs, &ret);
613 			break;
614 
615 		case FAIL_POINT_PAUSE:
616 			/**
617 			 * Pausing is inherently strange with multiple
618 			 * entries given our design.  That is because some
619 			 * entries could be unreachable, for instance in cases like:
620 			 * pause->return. We can never reach the return entry.
621 			 * The sysctl layer actually truncates all entries after
622 			 * a pause for this reason.
623 			 */
624 			fail_point_pause(fp, &ret);
625 			break;
626 
627 		case FAIL_POINT_YIELD:
628 			kern_yield(PRI_UNCHANGED);
629 			break;
630 
631 		case FAIL_POINT_DELAY:
632 			usecs = ent->fe_arg;
633 			DELAY(usecs);
634 			break;
635 
636 		default:
637 			break;
638 		}
639 
640 		if (cont == 0)
641 			break;
642 	}
643 
644 	if (fail_point_is_off(fp))
645 		fail_point_eval_swap_out(fp, fp_setting);
646 
647 abort:
648 	fail_point_setting_release_ref(fp);
649 
650 	return (ret);
651 }
652 
653 /**
654  * Translate internal fail_point structure into human-readable text.
655  */
656 static void
fail_point_get(struct fail_point * fp,struct sbuf * sb,bool verbose)657 fail_point_get(struct fail_point *fp, struct sbuf *sb,
658         bool verbose)
659 {
660 	struct fail_point_entry *ent;
661 	struct fail_point_setting *fp_setting;
662 	struct fail_point_entry *fp_entry_cpy;
663 	int cnt_sleeping;
664 	int idx;
665 	int printed_entry_count;
666 
667 	cnt_sleeping = 0;
668 	idx = 0;
669 	printed_entry_count = 0;
670 
671 	fp_entry_cpy = fp_malloc(sizeof(struct fail_point_entry) *
672 	        (FP_MAX_ENTRY_COUNT + 1), M_WAITOK);
673 
674 	fp_setting = fail_point_setting_get_ref(fp);
675 
676 	if (fp_setting != NULL) {
677 		TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
678 			if (ent->fe_stale)
679 				continue;
680 
681 			KASSERT(printed_entry_count < FP_MAX_ENTRY_COUNT,
682 			        ("FP entry list larger than allowed"));
683 
684 			fp_entry_cpy[printed_entry_count] = *ent;
685 			++printed_entry_count;
686 		}
687 	}
688 	fail_point_setting_release_ref(fp);
689 
690 	/* This is our equivalent of a NULL terminator */
691 	fp_entry_cpy[printed_entry_count].fe_type = FAIL_POINT_INVALID;
692 
693 	while (idx < printed_entry_count) {
694 		ent = &fp_entry_cpy[idx];
695 		++idx;
696 		if (ent->fe_prob < PROB_MAX) {
697 			int decimal = ent->fe_prob % (PROB_MAX / 100);
698 			int units = ent->fe_prob / (PROB_MAX / 100);
699 			sbuf_printf(sb, "%d", units);
700 			if (decimal) {
701 				int digits = PROB_DIGITS - 2;
702 				while (!(decimal % 10)) {
703 					digits--;
704 					decimal /= 10;
705 				}
706 				sbuf_printf(sb, ".%0*d", digits, decimal);
707 			}
708 			sbuf_printf(sb, "%%");
709 		}
710 		if (ent->fe_count >= 0)
711 			sbuf_printf(sb, "%d*", ent->fe_count);
712 		sbuf_printf(sb, "%s", fail_type_strings[ent->fe_type].name);
713 		if (ent->fe_arg)
714 			sbuf_printf(sb, "(%d)", ent->fe_arg);
715 		if (ent->fe_pid != NO_PID)
716 			sbuf_printf(sb, "[pid %d]", ent->fe_pid);
717 		if (TAILQ_NEXT(ent, fe_entries))
718 			sbuf_cat(sb, "->");
719 	}
720 	if (!printed_entry_count)
721 		sbuf_cat(sb, "off");
722 
723 	fp_free(fp_entry_cpy);
724 	if (verbose) {
725 #ifdef STACK
726 		/* Print number of sleeping threads. queue=0 is the argument
727 		 * used by msleep when sending our threads to sleep. */
728 		sbuf_cat(sb, "\nsleeping_thread_stacks = {\n");
729 		sleepq_sbuf_print_stacks(sb, FP_SLEEP_CHANNEL(fp), 0,
730 		        &cnt_sleeping);
731 
732 		sbuf_cat(sb, "},\n");
733 #endif
734 		sbuf_printf(sb, "sleeping_thread_count = %d,\n",
735 		        cnt_sleeping);
736 
737 #ifdef STACK
738 		sbuf_cat(sb, "paused_thread_stacks = {\n");
739 		sleepq_sbuf_print_stacks(sb, FP_PAUSE_CHANNEL(fp), 0,
740 		        &cnt_sleeping);
741 
742 		sbuf_cat(sb, "},\n");
743 #endif
744 		sbuf_printf(sb, "paused_thread_count = %d\n",
745 		        cnt_sleeping);
746 	}
747 }
748 
749 /**
750  * Set an internal fail_point structure from a human-readable failpoint string
751  * in a lock-safe manner.
752  */
753 static int
fail_point_set(struct fail_point * fp,char * buf)754 fail_point_set(struct fail_point *fp, char *buf)
755 {
756 	struct fail_point_entry *ent, *ent_next;
757 	struct fail_point_setting *entries;
758 	bool should_wake_paused;
759 	bool should_truncate;
760 	int error;
761 
762 	error = 0;
763 	should_wake_paused = false;
764 	should_truncate = false;
765 
766 	/* Parse new entries. */
767 	/**
768 	 * ref protects our new malloc'd stuff from being garbage collected
769 	 * before we link it.
770 	 */
771 	fail_point_setting_get_ref(fp);
772 	entries = fail_point_setting_new(fp);
773 	if (parse_fail_point(entries, buf) == NULL) {
774 		STAILQ_REMOVE(&fp_setting_garbage, entries,
775 		        fail_point_setting, fs_garbage_link);
776 		fail_point_setting_destroy(entries);
777 		error = EINVAL;
778 		goto end;
779 	}
780 
781 	/**
782 	 * Transfer the entries we are going to keep to a new list.
783 	 * Get rid of useless zero probability entries, and entries with hit
784 	 * count 0.
785 	 * If 'off' is present, and it has no hit count set, then all entries
786 	 *       after it are discarded since they are unreachable.
787 	 */
788 	TAILQ_FOREACH_SAFE(ent, &entries->fp_entry_queue, fe_entries, ent_next) {
789 		if (ent->fe_prob == 0 || ent->fe_count == 0) {
790 			printf("Discarding entry which cannot execute %s\n",
791 			        fail_type_strings[ent->fe_type].name);
792 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
793 			        fe_entries);
794 			fp_free(ent);
795 			continue;
796 		} else if (should_truncate) {
797 			printf("Discarding unreachable entry %s\n",
798 			        fail_type_strings[ent->fe_type].name);
799 			TAILQ_REMOVE(&entries->fp_entry_queue, ent,
800 			        fe_entries);
801 			fp_free(ent);
802 			continue;
803 		}
804 
805 		if (ent->fe_type == FAIL_POINT_OFF) {
806 			should_wake_paused = true;
807 			if (ent->fe_count == FE_COUNT_UNTRACKED) {
808 				should_truncate = true;
809 				TAILQ_REMOVE(&entries->fp_entry_queue, ent,
810 				        fe_entries);
811 				fp_free(ent);
812 			}
813 		} else if (ent->fe_type == FAIL_POINT_PAUSE) {
814 			should_truncate = true;
815 		} else if (ent->fe_type == FAIL_POINT_SLEEP && (fp->fp_flags &
816 		        FAIL_POINT_NONSLEEPABLE)) {
817 			/**
818 			 * If this fail point is annotated as being in a
819 			 * non-sleepable ctx, convert sleep to delay and
820 			 * convert the msec argument to usecs.
821 			 */
822 			printf("Sleep call request on fail point in "
823 			        "non-sleepable context; using delay instead "
824 			        "of sleep\n");
825 			ent->fe_type = FAIL_POINT_DELAY;
826 			ent->fe_arg *= 1000;
827 		}
828 	}
829 
830 	if (TAILQ_EMPTY(&entries->fp_entry_queue)) {
831 		entries = fail_point_swap_settings(fp, NULL);
832 		if (entries != NULL)
833 			wakeup(FP_PAUSE_CHANNEL(fp));
834 	} else {
835 		if (should_wake_paused)
836 			wakeup(FP_PAUSE_CHANNEL(fp));
837 		fail_point_swap_settings(fp, entries);
838 	}
839 
840 end:
841 #ifdef IWARNING
842 	if (error)
843 		IWARNING("Failed to set %s %s to %s",
844 		    fp->fp_name, fp->fp_location, buf);
845 	else
846 		INOTICE("Set %s %s to %s",
847 		    fp->fp_name, fp->fp_location, buf);
848 #endif /* IWARNING */
849 
850 	fail_point_setting_release_ref(fp);
851 	return (error);
852 }
853 
854 #define MAX_FAIL_POINT_BUF	1023
855 
856 /**
857  * Handle kernel failpoint set/get.
858  */
859 int
fail_point_sysctl(SYSCTL_HANDLER_ARGS)860 fail_point_sysctl(SYSCTL_HANDLER_ARGS)
861 {
862 	struct fail_point *fp;
863 	char *buf;
864 	struct sbuf sb, *sb_check;
865 	int error;
866 
867 	buf = NULL;
868 	error = 0;
869 	fp = arg1;
870 
871 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
872 	if (sb_check != &sb)
873 		return (ENOMEM);
874 
875 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
876 
877 	/* Setting */
878 	/**
879 	 * Lock protects any new entries from being garbage collected before we
880 	 * can link them to the fail point.
881 	 */
882 	sx_xlock(&sx_fp_set);
883 	if (req->newptr) {
884 		if (req->newlen > MAX_FAIL_POINT_BUF) {
885 			error = EINVAL;
886 			goto out;
887 		}
888 
889 		buf = fp_malloc(req->newlen + 1, M_WAITOK);
890 
891 		error = SYSCTL_IN(req, buf, req->newlen);
892 		if (error)
893 			goto out;
894 		buf[req->newlen] = '\0';
895 
896 		error = fail_point_set(fp, buf);
897 	}
898 
899 	fail_point_garbage_collect();
900 	sx_xunlock(&sx_fp_set);
901 
902 	/* Retrieving. */
903 	fail_point_get(fp, &sb, false);
904 
905 out:
906 	sbuf_finish(&sb);
907 	sbuf_delete(&sb);
908 
909 	if (buf)
910 		fp_free(buf);
911 
912 	return (error);
913 }
914 
915 int
fail_point_sysctl_status(SYSCTL_HANDLER_ARGS)916 fail_point_sysctl_status(SYSCTL_HANDLER_ARGS)
917 {
918 	struct fail_point *fp;
919 	struct sbuf sb, *sb_check;
920 
921 	fp = arg1;
922 
923 	sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
924 	if (sb_check != &sb)
925 		return (ENOMEM);
926 
927 	sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
928 
929 	/* Retrieving. */
930 	fail_point_get(fp, &sb, true);
931 
932 	sbuf_finish(&sb);
933 	sbuf_delete(&sb);
934 
935 	/**
936 	 * Lock protects any new entries from being garbage collected before we
937 	 * can link them to the fail point.
938 	 */
939 	sx_xlock(&sx_fp_set);
940 	fail_point_garbage_collect();
941 	sx_xunlock(&sx_fp_set);
942 
943 	return (0);
944 }
945 
946 int
fail_sysctl_drain_func(void * sysctl_args,const char * buf,int len)947 fail_sysctl_drain_func(void *sysctl_args, const char *buf, int len)
948 {
949 	struct sysctl_req *sa;
950 	int error;
951 
952 	sa = sysctl_args;
953 
954 	error = SYSCTL_OUT(sa, buf, len);
955 
956 	if (error == ENOMEM)
957 		return (-1);
958 	else
959 		return (len);
960 }
961 
962 /**
963  * Internal helper function to translate a human-readable failpoint string
964  * into a internally-parsable fail_point structure.
965  */
966 static char *
parse_fail_point(struct fail_point_setting * ents,char * p)967 parse_fail_point(struct fail_point_setting *ents, char *p)
968 {
969 	/*  <fail_point> ::
970 	 *      <term> ( "->" <term> )*
971 	 */
972 	uint8_t term_count;
973 
974 	term_count = 1;
975 
976 	p = parse_term(ents, p);
977 	if (p == NULL)
978 		return (NULL);
979 
980 	while (*p != '\0') {
981 		term_count++;
982 		if (p[0] != '-' || p[1] != '>' ||
983 		        (p = parse_term(ents, p+2)) == NULL ||
984 		        term_count > FP_MAX_ENTRY_COUNT)
985 			return (NULL);
986 	}
987 	return (p);
988 }
989 
990 /**
991  * Internal helper function to parse an individual term from a failpoint.
992  */
993 static char *
parse_term(struct fail_point_setting * ents,char * p)994 parse_term(struct fail_point_setting *ents, char *p)
995 {
996 	struct fail_point_entry *ent;
997 
998 	ent = fail_point_entry_new(ents);
999 
1000 	/*
1001 	 * <term> ::
1002 	 *     ( (<float> "%") | (<integer> "*" ) )*
1003 	 *     <type>
1004 	 *     [ "(" <integer> ")" ]
1005 	 *     [ "[pid " <integer> "]" ]
1006 	 */
1007 
1008 	/* ( (<float> "%") | (<integer> "*" ) )* */
1009 	while (isdigit(*p) || *p == '.') {
1010 		int units, decimal;
1011 
1012 		p = parse_number(&units, &decimal, p);
1013 		if (p == NULL)
1014 			return (NULL);
1015 
1016 		if (*p == '%') {
1017 			if (units > 100) /* prevent overflow early */
1018 				units = 100;
1019 			ent->fe_prob = units * (PROB_MAX / 100) + decimal;
1020 			if (ent->fe_prob > PROB_MAX)
1021 				ent->fe_prob = PROB_MAX;
1022 		} else if (*p == '*') {
1023 			if (!units || units < 0 || decimal)
1024 				return (NULL);
1025 			ent->fe_count = units;
1026 		} else
1027 			return (NULL);
1028 		p++;
1029 	}
1030 
1031 	/* <type> */
1032 	p = parse_type(ent, p);
1033 	if (p == NULL)
1034 		return (NULL);
1035 	if (*p == '\0')
1036 		return (p);
1037 
1038 	/* [ "(" <integer> ")" ] */
1039 	if (*p != '(')
1040 		return (p);
1041 	p++;
1042 	if (!isdigit(*p) && *p != '-')
1043 		return (NULL);
1044 	ent->fe_arg = strtol(p, &p, 0);
1045 	if (*p++ != ')')
1046 		return (NULL);
1047 
1048 	/* [ "[pid " <integer> "]" ] */
1049 #define PID_STRING "[pid "
1050 	if (strncmp(p, PID_STRING, sizeof(PID_STRING) - 1) != 0)
1051 		return (p);
1052 	p += sizeof(PID_STRING) - 1;
1053 	if (!isdigit(*p))
1054 		return (NULL);
1055 	ent->fe_pid = strtol(p, &p, 0);
1056 	if (*p++ != ']')
1057 		return (NULL);
1058 
1059 	return (p);
1060 }
1061 
1062 /**
1063  * Internal helper function to parse a numeric for a failpoint term.
1064  */
1065 static char *
parse_number(int * out_units,int * out_decimal,char * p)1066 parse_number(int *out_units, int *out_decimal, char *p)
1067 {
1068 	char *old_p;
1069 
1070 	/**
1071 	 *  <number> ::
1072 	 *      <integer> [ "." <integer> ] |
1073 	 *      "." <integer>
1074 	 */
1075 
1076 	/* whole part */
1077 	old_p = p;
1078 	*out_units = strtol(p, &p, 10);
1079 	if (p == old_p && *p != '.')
1080 		return (NULL);
1081 
1082 	/* fractional part */
1083 	*out_decimal = 0;
1084 	if (*p == '.') {
1085 		int digits = 0;
1086 		p++;
1087 		while (isdigit(*p)) {
1088 			int digit = *p - '0';
1089 			if (digits < PROB_DIGITS - 2)
1090 				*out_decimal = *out_decimal * 10 + digit;
1091 			else if (digits == PROB_DIGITS - 2 && digit >= 5)
1092 				(*out_decimal)++;
1093 			digits++;
1094 			p++;
1095 		}
1096 		if (!digits) /* need at least one digit after '.' */
1097 			return (NULL);
1098 		while (digits++ < PROB_DIGITS - 2) /* add implicit zeros */
1099 			*out_decimal *= 10;
1100 	}
1101 
1102 	return (p); /* success */
1103 }
1104 
1105 /**
1106  * Internal helper function to parse an individual type for a failpoint term.
1107  */
1108 static char *
parse_type(struct fail_point_entry * ent,char * beg)1109 parse_type(struct fail_point_entry *ent, char *beg)
1110 {
1111 	enum fail_point_t type;
1112 	int len;
1113 
1114 	for (type = FAIL_POINT_OFF; type < FAIL_POINT_NUMTYPES; type++) {
1115 		len = fail_type_strings[type].nmlen;
1116 		if (strncmp(fail_type_strings[type].name, beg, len) == 0) {
1117 			ent->fe_type = type;
1118 			return (beg + len);
1119 		}
1120 	}
1121 	return (NULL);
1122 }
1123 
1124 /* The fail point sysctl tree. */
1125 SYSCTL_NODE(_debug, OID_AUTO, fail_point, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1126     "fail points");
1127 
1128 /* Debugging/testing stuff for fail point */
1129 static int
sysctl_test_fail_point(SYSCTL_HANDLER_ARGS)1130 sysctl_test_fail_point(SYSCTL_HANDLER_ARGS)
1131 {
1132 
1133 	KFAIL_POINT_RETURN(DEBUG_FP, test_fail_point);
1134 	return (0);
1135 }
1136 SYSCTL_OID(_debug_fail_point, OID_AUTO, test_trigger_fail_point,
1137     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
1138     sysctl_test_fail_point, "A",
1139     "Trigger test fail points");
1140