sys/kern/kern_fail.c

   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         if (fp->fp_callout)
 407                 callout_drain(fp->fp_callout);
 408         fail_point_swap_settings(fp, entries);
 409 }
 410
 411 static inline void
 412 fail_point_pause(struct fail_point *fp, enum fail_point_return_code *pret,
 413         struct mtx *mtx_sleep)
 414 {
 415
 416         if (fp->fp_pre_sleep_fn)
 417                 fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
 418
 419         msleep_spin(FP_PAUSE_CHANNEL(fp), mtx_sleep, "failpt", 0);
 420
 421         if (fp->fp_post_sleep_fn)
 422                 fp->fp_post_sleep_fn(fp->fp_post_sleep_arg);
 423 }
 424
 425 static inline void
 426 fail_point_sleep(struct fail_point *fp, int msecs,
 427         enum fail_point_return_code *pret)
 428 {
 429         int timo;
 430
 431         /* Convert from millisecs to ticks, rounding up */
 432         timo = howmany((int64_t)msecs * hz, 1000L);
 433
 434         if (timo > 0) {
 435                 if (!(fp->fp_flags & FAIL_POINT_USE_TIMEOUT_PATH)) {
 436                         if (fp->fp_pre_sleep_fn)
 437                                 fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
 438
 439                         tsleep(FP_SLEEP_CHANNEL(fp), PWAIT, "failpt", timo);
 440
 441                         if (fp->fp_post_sleep_fn)
 442                                 fp->fp_post_sleep_fn(fp->fp_post_sleep_arg);
 443                 } else {
 444                         if (fp->fp_pre_sleep_fn)
 445                                 fp->fp_pre_sleep_fn(fp->fp_pre_sleep_arg);
 446
 447                         callout_reset(fp->fp_callout, timo,
 448                             fp->fp_post_sleep_fn, fp->fp_post_sleep_arg);
 449                         *pret = FAIL_POINT_RC_QUEUED;
 450                 }
 451         }
 452 }
 453
 454 static char *parse_fail_point(struct fail_point_setting *, char *);
 455 static char *parse_term(struct fail_point_setting *, char *);
 456 static char *parse_number(int *out_units, int *out_decimal, char *);
 457 static char *parse_type(struct fail_point_entry *, char *);
 458
 459 /**
 460  * Initialize a fail_point.  The name is formed in a printf-like fashion
 461  * from "fmt" and subsequent arguments.  This function is generally used
 462  * for custom failpoints located at odd places in the sysctl tree, and is
 463  * not explicitly needed for standard in-line-declared failpoints.
 464  *
 465  * @ingroup failpoint
 466  */
 467 void
 468 fail_point_init(struct fail_point *fp, const char *fmt, ...)
 469 {
 470         va_list ap;
 471         char *name;
 472         int n;
 473
 474         fp->fp_setting = NULL;
 475         fp->fp_flags = 0;
 476
 477         /* Figure out the size of the name. */
 478         va_start(ap, fmt);
 479         n = vsnprintf(NULL, 0, fmt, ap);
 480         va_end(ap);
 481
 482         /* Allocate the name and fill it in. */
 483         name = fp_malloc(n + 1, M_WAITOK);
 484         if (name != NULL) {
 485                 va_start(ap, fmt);
 486                 vsnprintf(name, n + 1, fmt, ap);
 487                 va_end(ap);
 488         }
 489         fp->fp_name = name;
 490         fp->fp_location = "";
 491         fp->fp_flags |= FAIL_POINT_DYNAMIC_NAME;
 492         fp->fp_pre_sleep_fn = NULL;
 493         fp->fp_pre_sleep_arg = NULL;
 494         fp->fp_post_sleep_fn = NULL;
 495         fp->fp_post_sleep_arg = NULL;
 496 }
 497
 498 void
 499 fail_point_alloc_callout(struct fail_point *fp)
 500 {
 501
 502         /**
 503          * This assumes that calls to fail_point_use_timeout_path()
 504          * will not race.
 505          */
 506         if (fp->fp_callout != NULL)
 507                 return;
 508         fp->fp_callout = fp_malloc(sizeof(*fp->fp_callout), M_WAITOK);
 509         callout_init(fp->fp_callout, CALLOUT_MPSAFE);
 510 }
 511
 512 /**
 513  * Free the resources held by a fail_point, and wake any paused threads.
 514  * Thou shalt not allow threads to hit this fail point after you enter this
 515  * function, nor shall you call this multiple times for a given fp.
 516  * @ingroup failpoint
 517  */
 518 void
 519 fail_point_destroy(struct fail_point *fp)
 520 {
 521
 522         fail_point_drain(fp, 0);
 523
 524         if ((fp->fp_flags & FAIL_POINT_DYNAMIC_NAME) != 0) {
 525                 fp_free(__DECONST(void *, fp->fp_name));
 526                 fp->fp_name = NULL;
 527         }
 528         fp->fp_flags = 0;
 529         if (fp->fp_callout) {
 530                 fp_free(fp->fp_callout);
 531                 fp->fp_callout = NULL;
 532         }
 533
 534         sx_xlock(&sx_fp_set);
 535         fail_point_garbage_collect();
 536         sx_xunlock(&sx_fp_set);
 537 }
 538
 539 /**
 540  * This does the real work of evaluating a fail point. If the fail point tells
 541  * us to return a value, this function returns 1 and fills in 'return_value'
 542  * (return_value is allowed to be null). If the fail point tells us to panic,
 543  * we never return. Otherwise we just return 0 after doing some work, which
 544  * means "keep going".
 545  */
 546 enum fail_point_return_code
 547 fail_point_eval_nontrivial(struct fail_point *fp, int *return_value)
 548 {
 549         bool execute = false;
 550         struct fail_point_entry *ent;
 551         struct fail_point_setting *fp_setting;
 552         enum fail_point_return_code ret;
 553         int cont;
 554         int count;
 555         int msecs;
 556         int usecs;
 557
 558         ret = FAIL_POINT_RC_CONTINUE;
 559         cont = 0; /* don't continue by default */
 560
 561         fp_setting = fail_point_setting_get_ref(fp);
 562         if (fp_setting == NULL)
 563                 goto abort;
 564
 565         TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
 566
 567                 if (ent->fe_stale)
 568                         continue;
 569
 570                 if (ent->fe_prob < PROB_MAX &&
 571                     ent->fe_prob < random() % PROB_MAX)
 572                         continue;
 573
 574                 if (ent->fe_pid != NO_PID && ent->fe_pid != curproc->p_pid)
 575                         continue;
 576
 577                 if (ent->fe_count != FE_COUNT_UNTRACKED) {
 578                         count = ent->fe_count;
 579                         while (count > 0) {
 580                                 if (atomic_cmpset_32(&ent->fe_count, count, count - 1)) {
 581                                         count--;
 582                                         execute = true;
 583                                         break;
 584                                 }
 585                                 count = ent->fe_count;
 586                         }
 587                         if (execute == false)
 588                                 /* We lost the race; consider the entry stale and bail now */
 589                                 continue;
 590                         if (count == 0)
 591                                 ent->fe_stale = true;
 592                 }
 593
 594                 switch (ent->fe_type) {
 595                 case FAIL_POINT_PANIC:
 596                         panic("fail point %s panicking", fp->fp_name);
 597                         /* NOTREACHED */
 598
 599                 case FAIL_POINT_RETURN:
 600                         if (return_value != NULL)
 601                                 *return_value = ent->fe_arg;
 602                         ret = FAIL_POINT_RC_RETURN;
 603                         break;
 604
 605                 case FAIL_POINT_BREAK:
 606                         printf("fail point %s breaking to debugger\n",
 607                                 fp->fp_name);
 608                         breakpoint();
 609                         break;
 610
 611                 case FAIL_POINT_PRINT:
 612                         printf("fail point %s executing\n", fp->fp_name);
 613                         cont = ent->fe_arg;
 614                         break;
 615
 616                 case FAIL_POINT_SLEEP:
 617                         msecs = ent->fe_arg;
 618                         if (msecs)
 619                                 fail_point_sleep(fp, msecs, &ret);
 620                         break;
 621
 622                 case FAIL_POINT_PAUSE:
 623                         /**
 624                          * Pausing is inherently strange with multiple
 625                          * entries given our design.  That is because some
 626                          * entries could be unreachable, for instance in cases like:
 627                          * pause->return. We can never reach the return entry.
 628                          * The sysctl layer actually truncates all entries after
 629                          * a pause for this reason.
 630                          */
 631                         mtx_lock_spin(&fp_setting->feq_mtx);
 632                         fail_point_pause(fp, &ret, &fp_setting->feq_mtx);
 633                         mtx_unlock_spin(&fp_setting->feq_mtx);
 634                         break;
 635
 636                 case FAIL_POINT_YIELD:
 637                         kern_yield(PRI_UNCHANGED);
 638                         break;
 639
 640                 case FAIL_POINT_DELAY:
 641                         usecs = ent->fe_arg;
 642                         DELAY(usecs);
 643                         break;
 644
 645                 default:
 646                         break;
 647                 }
 648
 649                 if (cont == 0)
 650                         break;
 651         }
 652
 653         if (fail_point_is_off(fp))
 654                 fail_point_eval_swap_out(fp, fp_setting);
 655
 656 abort:
 657         fail_point_setting_release_ref(fp);
 658
 659         return (ret);
 660 }
 661
 662 /**
 663  * Translate internal fail_point structure into human-readable text.
 664  */
 665 static void
 666 fail_point_get(struct fail_point *fp, struct sbuf *sb,
 667         bool verbose)
 668 {
 669         struct fail_point_entry *ent;
 670         struct fail_point_setting *fp_setting;
 671         struct fail_point_entry *fp_entry_cpy;
 672         int cnt_sleeping;
 673         int idx;
 674         int printed_entry_count;
 675
 676         cnt_sleeping = 0;
 677         idx = 0;
 678         printed_entry_count = 0;
 679
 680         fp_entry_cpy = fp_malloc(sizeof(struct fail_point_entry) *
 681                 (FP_MAX_ENTRY_COUNT + 1), M_WAITOK);
 682
 683         fp_setting = fail_point_setting_get_ref(fp);
 684
 685         if (fp_setting != NULL) {
 686                 TAILQ_FOREACH(ent, &fp_setting->fp_entry_queue, fe_entries) {
 687                         if (ent->fe_stale)
 688                                 continue;
 689
 690                         KASSERT(printed_entry_count < FP_MAX_ENTRY_COUNT,
 691                                 ("FP entry list larger than allowed"));
 692
 693                         fp_entry_cpy[printed_entry_count] = *ent;
 694                         ++printed_entry_count;
 695                 }
 696         }
 697         fail_point_setting_release_ref(fp);
 698
 699         /* This is our equivalent of a NULL terminator */
 700         fp_entry_cpy[printed_entry_count].fe_type = FAIL_POINT_INVALID;
 701
 702         while (idx < printed_entry_count) {
 703                 ent = &fp_entry_cpy[idx];
 704                 ++idx;
 705                 if (ent->fe_prob < PROB_MAX) {
 706                         int decimal = ent->fe_prob % (PROB_MAX / 100);
 707                         int units = ent->fe_prob / (PROB_MAX / 100);
 708                         sbuf_printf(sb, "%d", units);
 709                         if (decimal) {
 710                                 int digits = PROB_DIGITS - 2;
 711                                 while (!(decimal % 10)) {
 712                                         digits--;
 713                                         decimal /= 10;
 714                                 }
 715                                 sbuf_printf(sb, ".%0*d", digits, decimal);
 716                         }
 717                         sbuf_printf(sb, "%%");
 718                 }
 719                 if (ent->fe_count >= 0)
 720                         sbuf_printf(sb, "%d*", ent->fe_count);
 721                 sbuf_printf(sb, "%s", fail_type_strings[ent->fe_type].name);
 722                 if (ent->fe_arg)
 723                         sbuf_printf(sb, "(%d)", ent->fe_arg);
 724                 if (ent->fe_pid != NO_PID)
 725                         sbuf_printf(sb, "[pid %d]", ent->fe_pid);
 726                 if (TAILQ_NEXT(ent, fe_entries))
 727                         sbuf_printf(sb, "->");
 728         }
 729         if (!printed_entry_count)
 730                 sbuf_printf(sb, "off");
 731
 732         fp_free(fp_entry_cpy);
 733         if (verbose) {
 734 #ifdef STACK
 735                 /* Print number of sleeping threads. queue=0 is the argument
 736                  * used by msleep when sending our threads to sleep. */
 737                 sbuf_printf(sb, "\nsleeping_thread_stacks = {\n");
 738                 sleepq_sbuf_print_stacks(sb, FP_SLEEP_CHANNEL(fp), 0,
 739                         &cnt_sleeping);
 740
 741                 sbuf_printf(sb, "},\n");
 742 #endif
 743                 sbuf_printf(sb, "sleeping_thread_count = %d,\n",
 744                         cnt_sleeping);
 745
 746 #ifdef STACK
 747                 sbuf_printf(sb, "paused_thread_stacks = {\n");
 748                 sleepq_sbuf_print_stacks(sb, FP_PAUSE_CHANNEL(fp), 0,
 749                         &cnt_sleeping);
 750
 751                 sbuf_printf(sb, "},\n");
 752 #endif
 753                 sbuf_printf(sb, "paused_thread_count = %d\n",
 754                         cnt_sleeping);
 755         }
 756 }
 757
 758 /**
 759  * Set an internal fail_point structure from a human-readable failpoint string
 760  * in a lock-safe manner.
 761  */
 762 static int
 763 fail_point_set(struct fail_point *fp, char *buf)
 764 {
 765         struct fail_point_entry *ent, *ent_next;
 766         struct fail_point_setting *entries;
 767         bool should_wake_paused;
 768         bool should_truncate;
 769         int error;
 770
 771         error = 0;
 772         should_wake_paused = false;
 773         should_truncate = false;
 774
 775         /* Parse new entries. */
 776         /**
 777          * ref protects our new malloc'd stuff from being garbage collected
 778          * before we link it.
 779          */
 780         fail_point_setting_get_ref(fp);
 781         entries = fail_point_setting_new(fp);
 782         if (parse_fail_point(entries, buf) == NULL) {
 783                 STAILQ_REMOVE(&fp_setting_garbage, entries,
 784                         fail_point_setting, fs_garbage_link);
 785                 fail_point_setting_destroy(entries);
 786                 error = EINVAL;
 787                 goto end;
 788         }
 789
 790         /**
 791          * Transfer the entries we are going to keep to a new list.
 792          * Get rid of useless zero probability entries, and entries with hit
 793          * count 0.
 794          * If 'off' is present, and it has no hit count set, then all entries
 795          *       after it are discarded since they are unreachable.
 796          */
 797         TAILQ_FOREACH_SAFE(ent, &entries->fp_entry_queue, fe_entries, ent_next) {
 798                 if (ent->fe_prob == 0 || ent->fe_count == 0) {
 799                         printf("Discarding entry which cannot execute %s\n",
 800                                 fail_type_strings[ent->fe_type].name);
 801                         TAILQ_REMOVE(&entries->fp_entry_queue, ent,
 802                                 fe_entries);
 803                         fp_free(ent);
 804                         continue;
 805                 } else if (should_truncate) {
 806                         printf("Discarding unreachable entry %s\n",
 807                                 fail_type_strings[ent->fe_type].name);
 808                         TAILQ_REMOVE(&entries->fp_entry_queue, ent,
 809                                 fe_entries);
 810                         fp_free(ent);
 811                         continue;
 812                 }
 813
 814                 if (ent->fe_type == FAIL_POINT_OFF) {
 815                         should_wake_paused = true;
 816                         if (ent->fe_count == FE_COUNT_UNTRACKED) {
 817                                 should_truncate = true;
 818                                 TAILQ_REMOVE(&entries->fp_entry_queue, ent,
 819                                         fe_entries);
 820                                 fp_free(ent);
 821                         }
 822                 } else if (ent->fe_type == FAIL_POINT_PAUSE) {
 823                         should_truncate = true;
 824                 } else if (ent->fe_type == FAIL_POINT_SLEEP && (fp->fp_flags &
 825                         FAIL_POINT_NONSLEEPABLE)) {
 826                         /**
 827                          * If this fail point is annotated as being in a
 828                          * non-sleepable ctx, convert sleep to delay and
 829                          * convert the msec argument to usecs.
 830                          */
 831                         printf("Sleep call request on fail point in "
 832                                 "non-sleepable context; using delay instead "
 833                                 "of sleep\n");
 834                         ent->fe_type = FAIL_POINT_DELAY;
 835                         ent->fe_arg *= 1000;
 836                 }
 837         }
 838
 839         if (TAILQ_EMPTY(&entries->fp_entry_queue)) {
 840                 entries = fail_point_swap_settings(fp, NULL);
 841                 if (entries != NULL)
 842                         wakeup(FP_PAUSE_CHANNEL(fp));
 843         } else {
 844                 if (should_wake_paused)
 845                         wakeup(FP_PAUSE_CHANNEL(fp));
 846                 fail_point_swap_settings(fp, entries);
 847         }
 848
 849 end:
 850 #ifdef IWARNING
 851         if (error)
 852                 IWARNING("Failed to set %s %s to %s",
 853                     fp->fp_name, fp->fp_location, buf);
 854         else
 855                 INOTICE("Set %s %s to %s",
 856                     fp->fp_name, fp->fp_location, buf);
 857 #endif /* IWARNING */
 858
 859         fail_point_setting_release_ref(fp);
 860         return (error);
 861 }
 862
 863 #define MAX_FAIL_POINT_BUF      1023
 864
 865 /**
 866  * Handle kernel failpoint set/get.
 867  */
 868 int
 869 fail_point_sysctl(SYSCTL_HANDLER_ARGS)
 870 {
 871         struct fail_point *fp;
 872         char *buf;
 873         struct sbuf sb, *sb_check;
 874         int error;
 875
 876         buf = NULL;
 877         error = 0;
 878         fp = arg1;
 879
 880         sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
 881         if (sb_check != &sb)
 882                 return (ENOMEM);
 883
 884         sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
 885
 886         /* Setting */
 887         /**
 888          * Lock protects any new entries from being garbage collected before we
 889          * can link them to the fail point.
 890          */
 891         sx_xlock(&sx_fp_set);
 892         if (req->newptr) {
 893                 if (req->newlen > MAX_FAIL_POINT_BUF) {
 894                         error = EINVAL;
 895                         goto out;
 896                 }
 897
 898                 buf = fp_malloc(req->newlen + 1, M_WAITOK);
 899
 900                 error = SYSCTL_IN(req, buf, req->newlen);
 901                 if (error)
 902                         goto out;
 903                 buf[req->newlen] = '\0';
 904
 905                 error = fail_point_set(fp, buf);
 906         }
 907
 908         fail_point_garbage_collect();
 909         sx_xunlock(&sx_fp_set);
 910
 911         /* Retrieving. */
 912         fail_point_get(fp, &sb, false);
 913
 914 out:
 915         sbuf_finish(&sb);
 916         sbuf_delete(&sb);
 917
 918         if (buf)
 919                 fp_free(buf);
 920
 921         return (error);
 922 }
 923
 924 int
 925 fail_point_sysctl_status(SYSCTL_HANDLER_ARGS)
 926 {
 927         struct fail_point *fp;
 928         struct sbuf sb, *sb_check;
 929
 930         fp = arg1;
 931
 932         sb_check = sbuf_new(&sb, NULL, 1024, SBUF_AUTOEXTEND);
 933         if (sb_check != &sb)
 934                 return (ENOMEM);
 935
 936         sbuf_set_drain(&sb, (sbuf_drain_func *)fail_sysctl_drain_func, req);
 937
 938         /* Retrieving. */
 939         fail_point_get(fp, &sb, true);
 940
 941         sbuf_finish(&sb);
 942         sbuf_delete(&sb);
 943
 944         /**
 945          * Lock protects any new entries from being garbage collected before we
 946          * can link them to the fail point.
 947          */
 948         sx_xlock(&sx_fp_set);
 949         fail_point_garbage_collect();
 950         sx_xunlock(&sx_fp_set);
 951
 952         return (0);
 953 }
 954
 955 int
 956 fail_sysctl_drain_func(void *sysctl_args, const char *buf, int len)
 957 {
 958         struct sysctl_req *sa;
 959         int error;
 960
 961         sa = sysctl_args;
 962
 963         error = SYSCTL_OUT(sa, buf, len);
 964
 965         if (error == ENOMEM)
 966                 return (-1);
 967         else
 968                 return (len);
 969 }
 970
 971 /**
 972  * Internal helper function to translate a human-readable failpoint string
 973  * into a internally-parsable fail_point structure.
 974  */
 975 static char *
 976 parse_fail_point(struct fail_point_setting *ents, char *p)
 977 {
 978         /*  <fail_point> ::
 979          *      <term> ( "->" <term> )*
 980          */
 981         uint8_t term_count;
 982
 983         term_count = 1;
 984
 985         p = parse_term(ents, p);
 986         if (p == NULL)
 987                 return (NULL);
 988
 989         while (*p != '\0') {
 990                 term_count++;
 991                 if (p[0] != '-' || p[1] != '>' ||
 992                         (p = parse_term(ents, p+2)) == NULL ||
 993                         term_count > FP_MAX_ENTRY_COUNT)
 994                         return (NULL);
 995         }
 996         return (p);
 997 }
 998
 999 /**
1000  * Internal helper function to parse an individual term from a failpoint.
1001  */
1002 static char *
1003 parse_term(struct fail_point_setting *ents, char *p)
1004 {
1005         struct fail_point_entry *ent;
1006
1007         ent = fail_point_entry_new(ents);
1008
1009         /*
1010          * <term> ::
1011          *     ( (<float> "%") | (<integer> "*" ) )*
1012          *     <type>
1013          *     [ "(" <integer> ")" ]
1014          *     [ "[pid " <integer> "]" ]
1015          */
1016
1017         /* ( (<float> "%") | (<integer> "*" ) )* */
1018         while (isdigit(*p) || *p == '.') {
1019                 int units, decimal;
1020
1021                 p = parse_number(&units, &decimal, p);
1022                 if (p == NULL)
1023                         return (NULL);
1024
1025                 if (*p == '%') {
1026                         if (units > 100) /* prevent overflow early */
1027                                 units = 100;
1028                         ent->fe_prob = units * (PROB_MAX / 100) + decimal;
1029                         if (ent->fe_prob > PROB_MAX)
1030                                 ent->fe_prob = PROB_MAX;
1031                 } else if (*p == '*') {
1032                         if (!units || units < 0 || decimal)
1033                                 return (NULL);
1034                         ent->fe_count = units;
1035                 } else
1036                         return (NULL);
1037                 p++;
1038         }
1039
1040         /* <type> */
1041         p = parse_type(ent, p);
1042         if (p == NULL)
1043                 return (NULL);
1044         if (*p == '\0')
1045                 return (p);
1046
1047         /* [ "(" <integer> ")" ] */
1048         if (*p != '(')
1049                 return (p);
1050         p++;
1051         if (!isdigit(*p) && *p != '-')
1052                 return (NULL);
1053         ent->fe_arg = strtol(p, &p, 0);
1054         if (*p++ != ')')
1055                 return (NULL);
1056
1057         /* [ "[pid " <integer> "]" ] */
1058 #define PID_STRING "[pid "
1059         if (strncmp(p, PID_STRING, sizeof(PID_STRING) - 1) != 0)
1060                 return (p);
1061         p += sizeof(PID_STRING) - 1;
1062         if (!isdigit(*p))
1063                 return (NULL);
1064         ent->fe_pid = strtol(p, &p, 0);
1065         if (*p++ != ']')
1066                 return (NULL);
1067
1068         return (p);
1069 }
1070
1071 /**
1072  * Internal helper function to parse a numeric for a failpoint term.
1073  */
1074 static char *
1075 parse_number(int *out_units, int *out_decimal, char *p)
1076 {
1077         char *old_p;
1078
1079         /**
1080          *  <number> ::
1081          *      <integer> [ "." <integer> ] |
1082          *      "." <integer>
1083          */
1084
1085         /* whole part */
1086         old_p = p;
1087         *out_units = strtol(p, &p, 10);
1088         if (p == old_p && *p != '.')
1089                 return (NULL);
1090
1091         /* fractional part */
1092         *out_decimal = 0;
1093         if (*p == '.') {
1094                 int digits = 0;
1095                 p++;
1096                 while (isdigit(*p)) {
1097                         int digit = *p - '0';
1098                         if (digits < PROB_DIGITS - 2)
1099                                 *out_decimal = *out_decimal * 10 + digit;
1100                         else if (digits == PROB_DIGITS - 2 && digit >= 5)
1101                                 (*out_decimal)++;
1102                         digits++;
1103                         p++;
1104                 }
1105                 if (!digits) /* need at least one digit after '.' */
1106                         return (NULL);
1107                 while (digits++ < PROB_DIGITS - 2) /* add implicit zeros */
1108                         *out_decimal *= 10;
1109         }
1110
1111         return (p); /* success */
1112 }
1113
1114 /**
1115  * Internal helper function to parse an individual type for a failpoint term.
1116  */
1117 static char *
1118 parse_type(struct fail_point_entry *ent, char *beg)
1119 {
1120         enum fail_point_t type;
1121         int len;
1122
1123         for (type = FAIL_POINT_OFF; type < FAIL_POINT_NUMTYPES; type++) {
1124                 len = fail_type_strings[type].nmlen;
1125                 if (strncmp(fail_type_strings[type].name, beg, len) == 0) {
1126                         ent->fe_type = type;
1127                         return (beg + len);
1128                 }
1129         }
1130         return (NULL);
1131 }
1132
1133 /* The fail point sysctl tree. */
1134 SYSCTL_NODE(_debug, OID_AUTO, fail_point, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1135     "fail points");
1136
1137 /* Debugging/testing stuff for fail point */
1138 static int
1139 sysctl_test_fail_point(SYSCTL_HANDLER_ARGS)
1140 {
1141
1142         KFAIL_POINT_RETURN(DEBUG_FP, test_fail_point);
1143         return (0);
1144 }
1145 SYSCTL_OID(_debug_fail_point, OID_AUTO, test_trigger_fail_point,
1146     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0,
1147     sysctl_test_fail_point, "A",
1148     "Trigger test fail points");