sys/kern/subr_witness.c

   1 /*-
   2  * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved.
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  * 1. Redistributions of source code must retain the above copyright
   8  *    notice, this list of conditions and the following disclaimer.
   9  * 2. Redistributions in binary form must reproduce the above copyright
  10  *    notice, this list of conditions and the following disclaimer in the
  11  *    documentation and/or other materials provided with the distribution.
  12  * 3. Berkeley Software Design Inc's name may not be used to endorse or
  13  *    promote products derived from this software without specific prior
  14  *    written permission.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  *
  28  *      from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
  29  *      and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
  30  */
  31
  32 /*
  33  * Implementation of the `witness' lock verifier.  Originally implemented for
  34  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
  35  * classes in FreeBSD.
  36  */
  37
  38 /*
  39  *      Main Entry: witness
  40  *      Pronunciation: 'wit-n&s
  41  *      Function: noun
  42  *      Etymology: Middle English witnesse, from Old English witnes knowledge,
  43  *          testimony, witness, from 2wit
  44  *      Date: before 12th century
  45  *      1 : attestation of a fact or event : TESTIMONY
  46  *      2 : one that gives evidence; specifically : one who testifies in
  47  *          a cause or before a judicial tribunal
  48  *      3 : one asked to be present at a transaction so as to be able to
  49  *          testify to its having taken place
  50  *      4 : one who has personal knowledge of something
  51  *      5 a : something serving as evidence or proof : SIGN
  52  *        b : public affirmation by word or example of usually
  53  *            religious faith or conviction <the heroic witness to divine
  54  *            life -- Pilot>
  55  *      6 capitalized : a member of the Jehovah's Witnesses
  56  */
  57
  58 /*
  59  * Special rules concerning Giant and lock orders:
  60  *
  61  * 1) Giant must be acquired before any other mutexes.  Stated another way,
  62  *    no other mutex may be held when Giant is acquired.
  63  *
  64  * 2) Giant must be released when blocking on a sleepable lock.
  65  *
  66  * This rule is less obvious, but is a result of Giant providing the same
  67  * semantics as spl().  Basically, when a thread sleeps, it must release
  68  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
  69  * 2).
  70  *
  71  * 3) Giant may be acquired before or after sleepable locks.
  72  *
  73  * This rule is also not quite as obvious.  Giant may be acquired after
  74  * a sleepable lock because it is a non-sleepable lock and non-sleepable
  75  * locks may always be acquired while holding a sleepable lock.  The second
  76  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
  77  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
  78  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
  79  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
  80  * execute.  Thus, acquiring Giant both before and after a sleepable lock
  81  * will not result in a lock order reversal.
  82  */
  83
  84 #include <sys/cdefs.h>
  85 __FBSDID("$FreeBSD$");
  86
  87 #include "opt_ddb.h"
  88 #include "opt_witness.h"
  89
  90 #include <sys/param.h>
  91 #include <sys/bus.h>
  92 #include <sys/kdb.h>
  93 #include <sys/kernel.h>
  94 #include <sys/ktr.h>
  95 #include <sys/lock.h>
  96 #include <sys/malloc.h>
  97 #include <sys/mutex.h>
  98 #include <sys/proc.h>
  99 #include <sys/sysctl.h>
 100 #include <sys/systm.h>
 101
 102 #include <ddb/ddb.h>
 103
 104 #include <machine/stdarg.h>
 105
 106 /* Define this to check for blessed mutexes */
 107 #undef BLESSING
 108
 109 #define WITNESS_COUNT 1024
 110 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
 111 /*
 112  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
 113  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
 114  * probably be safe for the most part, but it's still a SWAG.
 115  */
 116 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
 117
 118 #define WITNESS_NCHILDREN 6
 119
 120 struct witness_child_list_entry;
 121
 122 struct witness {
 123         const   char *w_name;
 124         struct  lock_class *w_class;
 125         STAILQ_ENTRY(witness) w_list;           /* List of all witnesses. */
 126         STAILQ_ENTRY(witness) w_typelist;       /* Witnesses of a type. */
 127         struct  witness_child_list_entry *w_children;   /* Great evilness... */
 128         const   char *w_file;
 129         int     w_line;
 130         u_int   w_level;
 131         u_int   w_refcount;
 132         u_char  w_Giant_squawked:1;
 133         u_char  w_other_squawked:1;
 134         u_char  w_same_squawked:1;
 135         u_char  w_displayed:1;
 136 };
 137
 138 struct witness_child_list_entry {
 139         struct  witness_child_list_entry *wcl_next;
 140         struct  witness *wcl_children[WITNESS_NCHILDREN];
 141         u_int   wcl_count;
 142 };
 143
 144 STAILQ_HEAD(witness_list, witness);
 145
 146 #ifdef BLESSING
 147 struct witness_blessed {
 148         const   char *b_lock1;
 149         const   char *b_lock2;
 150 };
 151 #endif
 152
 153 struct witness_order_list_entry {
 154         const   char *w_name;
 155         struct  lock_class *w_class;
 156 };
 157
 158 #ifdef BLESSING
 159 static int      blessed(struct witness *, struct witness *);
 160 #endif
 161 static int      depart(struct witness *w);
 162 static struct   witness *enroll(const char *description,
 163                                 struct lock_class *lock_class);
 164 static int      insertchild(struct witness *parent, struct witness *child);
 165 static int      isitmychild(struct witness *parent, struct witness *child);
 166 static int      isitmydescendant(struct witness *parent, struct witness *child);
 167 static int      itismychild(struct witness *parent, struct witness *child);
 168 static int      rebalancetree(struct witness_list *list);
 169 static void     removechild(struct witness *parent, struct witness *child);
 170 static int      reparentchildren(struct witness *newparent,
 171                     struct witness *oldparent);
 172 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
 173 static void     witness_displaydescendants(void(*)(const char *fmt, ...),
 174                                            struct witness *, int indent);
 175 static const char *fixup_filename(const char *file);
 176 static void     witness_leveldescendents(struct witness *parent, int level);
 177 static void     witness_levelall(void);
 178 static struct   witness *witness_get(void);
 179 static void     witness_free(struct witness *m);
 180 static struct   witness_child_list_entry *witness_child_get(void);
 181 static void     witness_child_free(struct witness_child_list_entry *wcl);
 182 static struct   lock_list_entry *witness_lock_list_get(void);
 183 static void     witness_lock_list_free(struct lock_list_entry *lle);
 184 static struct   lock_instance *find_instance(struct lock_list_entry *lock_list,
 185                                              struct lock_object *lock);
 186 static void     witness_list_lock(struct lock_instance *instance);
 187 #ifdef DDB
 188 static void     witness_list(struct thread *td);
 189 static void     witness_display_list(void(*prnt)(const char *fmt, ...),
 190                                      struct witness_list *list);
 191 static void     witness_display(void(*)(const char *fmt, ...));
 192 #endif
 193
 194 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
 195
 196 /*
 197  * If set to 0, witness is disabled.  If set to 1, witness performs full lock
 198  * order checking for all locks.  If set to 2 or higher, then witness skips
 199  * the full lock order check if the lock being acquired is at a higher level
 200  * (i.e. farther down in the tree) than the current lock.  This last mode is
 201  * somewhat experimental and not considered fully safe.  At runtime, this
 202  * value may be set to 0 to turn off witness.  witness is not allowed be
 203  * turned on once it is turned off, however.
 204  */
 205 static int witness_watch = 1;
 206 TUNABLE_INT("debug.witness.watch", &witness_watch);
 207 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
 208     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
 209
 210 #ifdef KDB
 211 /*
 212  * When KDB is enabled and witness_kdb is set to 1, it will cause the system
 213  * to drop into kdebug() when:
 214  *      - a lock heirarchy violation occurs
 215  *      - locks are held when going to sleep.
 216  */
 217 #ifdef WITNESS_KDB
 218 int     witness_kdb = 1;
 219 #else
 220 int     witness_kdb = 0;
 221 #endif
 222 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
 223 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
 224
 225 /*
 226  * When KDB is enabled and witness_trace is set to 1, it will cause the system
 227  * to print a stack trace:
 228  *      - a lock heirarchy violation occurs
 229  *      - locks are held when going to sleep.
 230  */
 231 int     witness_trace = 1;
 232 TUNABLE_INT("debug.witness.trace", &witness_trace);
 233 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
 234 #endif /* KDB */
 235
 236 #ifdef WITNESS_SKIPSPIN
 237 int     witness_skipspin = 1;
 238 #else
 239 int     witness_skipspin = 0;
 240 #endif
 241 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
 242 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
 243     &witness_skipspin, 0, "");
 244
 245 static struct mtx w_mtx;
 246 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 247 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 248 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 249 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 250 static struct witness_child_list_entry *w_child_free = NULL;
 251 static struct lock_list_entry *w_lock_list_free = NULL;
 252
 253 static struct witness w_data[WITNESS_COUNT];
 254 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
 255 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 256
 257 static struct witness_order_list_entry order_lists[] = {
 258         { "proctree", &lock_class_sx },
 259         { "allproc", &lock_class_sx },
 260         { "Giant", &lock_class_mtx_sleep },
 261         { "filedesc structure", &lock_class_mtx_sleep },
 262         { "pipe mutex", &lock_class_mtx_sleep },
 263         { "sigio lock", &lock_class_mtx_sleep },
 264         { "process group", &lock_class_mtx_sleep },
 265         { "process lock", &lock_class_mtx_sleep },
 266         { "session", &lock_class_mtx_sleep },
 267         { "uidinfo hash", &lock_class_mtx_sleep },
 268         { "uidinfo struct", &lock_class_mtx_sleep },
 269         { "allprison", &lock_class_mtx_sleep },
 270         { NULL, NULL },
 271         /*
 272          * Sockets
 273          */
 274         { "filedesc structure", &lock_class_mtx_sleep },
 275         { "accept", &lock_class_mtx_sleep },
 276         { "so_snd", &lock_class_mtx_sleep },
 277         { "so_rcv", &lock_class_mtx_sleep },
 278         { "sellck", &lock_class_mtx_sleep },
 279         { NULL, NULL },
 280         /*
 281          * Routing
 282          */
 283         { "so_rcv", &lock_class_mtx_sleep },
 284         { "radix node head", &lock_class_mtx_sleep },
 285         { "rtentry", &lock_class_mtx_sleep },
 286         { "ifaddr", &lock_class_mtx_sleep },
 287         { NULL, NULL },
 288         /*
 289          * UNIX Domain Sockets
 290          */
 291         { "unp", &lock_class_mtx_sleep },
 292         { "so_snd", &lock_class_mtx_sleep },
 293         { NULL, NULL },
 294         /*
 295          * UDP/IP
 296          */
 297         { "udp", &lock_class_mtx_sleep },
 298         { "udpinp", &lock_class_mtx_sleep },
 299         { "so_snd", &lock_class_mtx_sleep },
 300         { NULL, NULL },
 301         /*
 302          * TCP/IP
 303          */
 304         { "tcp", &lock_class_mtx_sleep },
 305         { "tcpinp", &lock_class_mtx_sleep },
 306         { "so_snd", &lock_class_mtx_sleep },
 307         { NULL, NULL },
 308         /*
 309          * SLIP
 310          */
 311         { "slip_mtx", &lock_class_mtx_sleep },
 312         { "slip sc_mtx", &lock_class_mtx_sleep },
 313         { NULL, NULL },
 314         /*
 315          * netatalk
 316          */
 317         { "ddp_list_mtx", &lock_class_mtx_sleep },
 318         { "ddp_mtx", &lock_class_mtx_sleep },
 319         { NULL, NULL },
 320         /*
 321          * BPF
 322          */
 323         { "bpf global lock", &lock_class_mtx_sleep },
 324         { "bpf interface lock", &lock_class_mtx_sleep },
 325         { "bpf cdev lock", &lock_class_mtx_sleep },
 326         { NULL, NULL },
 327         /*
 328          * NFS server
 329          */
 330         { "nfsd_mtx", &lock_class_mtx_sleep },
 331         { "so_snd", &lock_class_mtx_sleep },
 332         { NULL, NULL },
 333         /*
 334          * CDEV
 335          */
 336         { "system map", &lock_class_mtx_sleep },
 337         { "vm page queue mutex", &lock_class_mtx_sleep },
 338         { "vnode interlock", &lock_class_mtx_sleep },
 339         { "cdev", &lock_class_mtx_sleep },
 340         { NULL, NULL },
 341         /*
 342          * spin locks
 343          */
 344 #ifdef SMP
 345         { "ap boot", &lock_class_mtx_spin },
 346 #endif
 347         { "sio", &lock_class_mtx_spin },
 348 #ifdef __i386__
 349         { "cy", &lock_class_mtx_spin },
 350 #endif
 351         { "uart_hwmtx", &lock_class_mtx_spin },
 352         { "sabtty", &lock_class_mtx_spin },
 353         { "zstty", &lock_class_mtx_spin },
 354         { "ng_node", &lock_class_mtx_spin },
 355         { "ng_worklist", &lock_class_mtx_spin },
 356         { "taskqueue_fast", &lock_class_mtx_spin },
 357         { "intr table", &lock_class_mtx_spin },
 358         { "ithread table lock", &lock_class_mtx_spin },
 359         { "sleepq chain", &lock_class_mtx_spin },
 360         { "sched lock", &lock_class_mtx_spin },
 361         { "turnstile chain", &lock_class_mtx_spin },
 362         { "td_contested", &lock_class_mtx_spin },
 363         { "callout", &lock_class_mtx_spin },
 364         { "entropy harvest mutex", &lock_class_mtx_spin },
 365         /*
 366          * leaf locks
 367          */
 368         { "allpmaps", &lock_class_mtx_spin },
 369         { "vm page queue free mutex", &lock_class_mtx_spin },
 370         { "icu", &lock_class_mtx_spin },
 371 #ifdef SMP
 372         { "smp rendezvous", &lock_class_mtx_spin },
 373 #if defined(__i386__) || defined(__amd64__)
 374         { "tlb", &lock_class_mtx_spin },
 375 #endif
 376 #ifdef __sparc64__
 377         { "ipi", &lock_class_mtx_spin },
 378         { "rtc_mtx", &lock_class_mtx_spin },
 379 #endif
 380 #endif
 381         { "clk", &lock_class_mtx_spin },
 382         { "mutex profiling lock", &lock_class_mtx_spin },
 383         { "kse zombie lock", &lock_class_mtx_spin },
 384         { "ALD Queue", &lock_class_mtx_spin },
 385 #ifdef __ia64__
 386         { "MCA spin lock", &lock_class_mtx_spin },
 387 #endif
 388 #if defined(__i386__) || defined(__amd64__)
 389         { "pcicfg", &lock_class_mtx_spin },
 390         { "NDIS thread lock", &lock_class_mtx_spin },
 391 #endif
 392         { "tw_osl_io_lock", &lock_class_mtx_spin },
 393         { "tw_osl_q_lock", &lock_class_mtx_spin },
 394         { "tw_cl_io_lock", &lock_class_mtx_spin },
 395         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 396         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 397         { NULL, NULL },
 398         { NULL, NULL }
 399 };
 400
 401 #ifdef BLESSING
 402 /*
 403  * Pairs of locks which have been blessed
 404  * Don't complain about order problems with blessed locks
 405  */
 406 static struct witness_blessed blessed_list[] = {
 407 };
 408 static int blessed_count =
 409         sizeof(blessed_list) / sizeof(struct witness_blessed);
 410 #endif
 411
 412 /*
 413  * List of all locks in the system.
 414  */
 415 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
 416
 417 static struct mtx all_mtx = {
 418         { &lock_class_mtx_sleep,        /* mtx_object.lo_class */
 419           "All locks list",             /* mtx_object.lo_name */
 420           "All locks list",             /* mtx_object.lo_type */
 421           LO_INITIALIZED,               /* mtx_object.lo_flags */
 422           { NULL, NULL },               /* mtx_object.lo_list */
 423           NULL },                       /* mtx_object.lo_witness */
 424         MTX_UNOWNED, 0                  /* mtx_lock, mtx_recurse */
 425 };
 426
 427 /*
 428  * This global is set to 0 once it becomes safe to use the witness code.
 429  */
 430 static int witness_cold = 1;
 431
 432 /*
 433  * Global variables for book keeping.
 434  */
 435 static int lock_cur_cnt;
 436 static int lock_max_cnt;
 437
 438 /*
 439  * The WITNESS-enabled diagnostic code.
 440  */
 441 static void
 442 witness_initialize(void *dummy __unused)
 443 {
 444         struct lock_object *lock;
 445         struct witness_order_list_entry *order;
 446         struct witness *w, *w1;
 447         int i;
 448
 449         /*
 450          * We have to release Giant before initializing its witness
 451          * structure so that WITNESS doesn't get confused.
 452          */
 453         mtx_unlock(&Giant);
 454         mtx_assert(&Giant, MA_NOTOWNED);
 455
 456         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 457         TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
 458         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 459             MTX_NOWITNESS);
 460         for (i = 0; i < WITNESS_COUNT; i++)
 461                 witness_free(&w_data[i]);
 462         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
 463                 witness_child_free(&w_childdata[i]);
 464         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 465                 witness_lock_list_free(&w_locklistdata[i]);
 466
 467         /* First add in all the specified order lists. */
 468         for (order = order_lists; order->w_name != NULL; order++) {
 469                 w = enroll(order->w_name, order->w_class);
 470                 if (w == NULL)
 471                         continue;
 472                 w->w_file = "order list";
 473                 for (order++; order->w_name != NULL; order++) {
 474                         w1 = enroll(order->w_name, order->w_class);
 475                         if (w1 == NULL)
 476                                 continue;
 477                         w1->w_file = "order list";
 478                         if (!itismychild(w, w1))
 479                                 panic("Not enough memory for static orders!");
 480                         w = w1;
 481                 }
 482         }
 483
 484         /* Iterate through all locks and add them to witness. */
 485         mtx_lock(&all_mtx);
 486         TAILQ_FOREACH(lock, &all_locks, lo_list) {
 487                 if (lock->lo_flags & LO_WITNESS)
 488                         lock->lo_witness = enroll(lock->lo_type,
 489                             lock->lo_class);
 490                 else
 491                         lock->lo_witness = NULL;
 492         }
 493         mtx_unlock(&all_mtx);
 494
 495         /* Mark the witness code as being ready for use. */
 496         atomic_store_rel_int(&witness_cold, 0);
 497
 498         mtx_lock(&Giant);
 499 }
 500 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
 501
 502 static int
 503 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
 504 {
 505         int error, value;
 506
 507         value = witness_watch;
 508         error = sysctl_handle_int(oidp, &value, 0, req);
 509         if (error != 0 || req->newptr == NULL)
 510                 return (error);
 511         error = suser(req->td);
 512         if (error != 0)
 513                 return (error);
 514         if (value == witness_watch)
 515                 return (0);
 516         if (value != 0)
 517                 return (EINVAL);
 518         witness_watch = 0;
 519         return (0);
 520 }
 521
 522 void
 523 witness_init(struct lock_object *lock)
 524 {
 525         struct lock_class *class;
 526
 527         class = lock->lo_class;
 528         if (lock->lo_flags & LO_INITIALIZED)
 529                 panic("%s: lock (%s) %s is already initialized", __func__,
 530                     class->lc_name, lock->lo_name);
 531         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 532             (class->lc_flags & LC_RECURSABLE) == 0)
 533                 panic("%s: lock (%s) %s can not be recursable", __func__,
 534                     class->lc_name, lock->lo_name);
 535         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 536             (class->lc_flags & LC_SLEEPABLE) == 0)
 537                 panic("%s: lock (%s) %s can not be sleepable", __func__,
 538                     class->lc_name, lock->lo_name);
 539         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 540             (class->lc_flags & LC_UPGRADABLE) == 0)
 541                 panic("%s: lock (%s) %s can not be upgradable", __func__,
 542                     class->lc_name, lock->lo_name);
 543
 544         mtx_lock(&all_mtx);
 545         TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
 546         lock->lo_flags |= LO_INITIALIZED;
 547         lock_cur_cnt++;
 548         if (lock_cur_cnt > lock_max_cnt)
 549                 lock_max_cnt = lock_cur_cnt;
 550         mtx_unlock(&all_mtx);
 551         if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
 552             (lock->lo_flags & LO_WITNESS) != 0)
 553                 lock->lo_witness = enroll(lock->lo_type, class);
 554         else
 555                 lock->lo_witness = NULL;
 556 }
 557
 558 void
 559 witness_destroy(struct lock_object *lock)
 560 {
 561         struct witness *w;
 562
 563         if (witness_cold)
 564                 panic("lock (%s) %s destroyed while witness_cold",
 565                     lock->lo_class->lc_name, lock->lo_name);
 566         if ((lock->lo_flags & LO_INITIALIZED) == 0)
 567                 panic("%s: lock (%s) %s is not initialized", __func__,
 568                     lock->lo_class->lc_name, lock->lo_name);
 569
 570         /* XXX: need to verify that no one holds the lock */
 571         w = lock->lo_witness;
 572         if (w != NULL) {
 573                 mtx_lock_spin(&w_mtx);
 574                 MPASS(w->w_refcount > 0);
 575                 w->w_refcount--;
 576
 577                 /*
 578                  * Lock is already released if we have an allocation failure
 579                  * and depart() fails.
 580                  */
 581                 if (w->w_refcount != 0 || depart(w))
 582                         mtx_unlock_spin(&w_mtx);
 583         }
 584
 585         mtx_lock(&all_mtx);
 586         lock_cur_cnt--;
 587         TAILQ_REMOVE(&all_locks, lock, lo_list);
 588         lock->lo_flags &= ~LO_INITIALIZED;
 589         mtx_unlock(&all_mtx);
 590 }
 591
 592 #ifdef DDB
 593 static void
 594 witness_display_list(void(*prnt)(const char *fmt, ...),
 595                      struct witness_list *list)
 596 {
 597         struct witness *w;
 598
 599         STAILQ_FOREACH(w, list, w_typelist) {
 600                 if (w->w_file == NULL || w->w_level > 0)
 601                         continue;
 602                 /*
 603                  * This lock has no anscestors, display its descendants.
 604                  */
 605                 witness_displaydescendants(prnt, w, 0);
 606         }
 607 }
 608
 609 static void
 610 witness_display(void(*prnt)(const char *fmt, ...))
 611 {
 612         struct witness *w;
 613
 614         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 615         witness_levelall();
 616
 617         /* Clear all the displayed flags. */
 618         STAILQ_FOREACH(w, &w_all, w_list) {
 619                 w->w_displayed = 0;
 620         }
 621
 622         /*
 623          * First, handle sleep locks which have been acquired at least
 624          * once.
 625          */
 626         prnt("Sleep locks:\n");
 627         witness_display_list(prnt, &w_sleep);
 628
 629         /*
 630          * Now do spin locks which have been acquired at least once.
 631          */
 632         prnt("\nSpin locks:\n");
 633         witness_display_list(prnt, &w_spin);
 634
 635         /*
 636          * Finally, any locks which have not been acquired yet.
 637          */
 638         prnt("\nLocks which were never acquired:\n");
 639         STAILQ_FOREACH(w, &w_all, w_list) {
 640                 if (w->w_file != NULL || w->w_refcount == 0)
 641                         continue;
 642                 prnt("%s\n", w->w_name);
 643         }
 644 }
 645 #endif /* DDB */
 646
 647 /* Trim useless garbage from filenames. */
 648 static const char *
 649 fixup_filename(const char *file)
 650 {
 651
 652         if (file == NULL)
 653                 return (NULL);
 654         while (strncmp(file, "../", 3) == 0)
 655                 file += 3;
 656         return (file);
 657 }
 658
 659 int
 660 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
 661 {
 662
 663         if (witness_watch == 0 || panicstr != NULL)
 664                 return (0);
 665
 666         /* Require locks that witness knows about. */
 667         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
 668             lock2->lo_witness == NULL)
 669                 return (EINVAL);
 670
 671         MPASS(!mtx_owned(&w_mtx));
 672         mtx_lock_spin(&w_mtx);
 673
 674         /*
 675          * If we already have either an explicit or implied lock order that
 676          * is the other way around, then return an error.
 677          */
 678         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
 679                 mtx_unlock_spin(&w_mtx);
 680                 return (EDOOFUS);
 681         }
 682
 683         /* Try to add the new order. */
 684         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 685             lock2->lo_type, lock1->lo_type);
 686         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
 687                 return (ENOMEM);
 688         mtx_unlock_spin(&w_mtx);
 689         return (0);
 690 }
 691
 692 void
 693 witness_checkorder(struct lock_object *lock, int flags, const char *file,
 694     int line)
 695 {
 696         struct lock_list_entry **lock_list, *lle;
 697         struct lock_instance *lock1, *lock2;
 698         struct lock_class *class;
 699         struct witness *w, *w1;
 700         struct thread *td;
 701         int i, j;
 702
 703         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 704             panicstr != NULL)
 705                 return;
 706
 707         /*
 708          * Try locks do not block if they fail to acquire the lock, thus
 709          * there is no danger of deadlocks or of switching while holding a
 710          * spin lock if we acquire a lock via a try operation.  This
 711          * function shouldn't even be called for try locks, so panic if
 712          * that happens.
 713          */
 714         if (flags & LOP_TRYLOCK)
 715                 panic("%s should not be called for try lock operations",
 716                     __func__);
 717
 718         w = lock->lo_witness;
 719         class = lock->lo_class;
 720         td = curthread;
 721         file = fixup_filename(file);
 722
 723         if (class->lc_flags & LC_SLEEPLOCK) {
 724                 /*
 725                  * Since spin locks include a critical section, this check
 726                  * implicitly enforces a lock order of all sleep locks before
 727                  * all spin locks.
 728                  */
 729                 if (td->td_critnest != 0 && !kdb_active)
 730                         panic("blockable sleep lock (%s) %s @ %s:%d",
 731                             class->lc_name, lock->lo_name, file, line);
 732
 733                 /*
 734                  * If this is the first lock acquired then just return as
 735                  * no order checking is needed.
 736                  */
 737                 if (td->td_sleeplocks == NULL)
 738                         return;
 739                 lock_list = &td->td_sleeplocks;
 740         } else {
 741                 /*
 742                  * If this is the first lock, just return as no order
 743                  * checking is needed.  We check this in both if clauses
 744                  * here as unifying the check would require us to use a
 745                  * critical section to ensure we don't migrate while doing
 746                  * the check.  Note that if this is not the first lock, we
 747                  * are already in a critical section and are safe for the
 748                  * rest of the check.
 749                  */
 750                 if (PCPU_GET(spinlocks) == NULL)
 751                         return;
 752                 lock_list = PCPU_PTR(spinlocks);
 753         }
 754
 755         /*
 756          * Check to see if we are recursing on a lock we already own.  If
 757          * so, make sure that we don't mismatch exclusive and shared lock
 758          * acquires.
 759          */
 760         lock1 = find_instance(*lock_list, lock);
 761         if (lock1 != NULL) {
 762                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
 763                     (flags & LOP_EXCLUSIVE) == 0) {
 764                         printf("shared lock of (%s) %s @ %s:%d\n",
 765                             class->lc_name, lock->lo_name, file, line);
 766                         printf("while exclusively locked from %s:%d\n",
 767                             lock1->li_file, lock1->li_line);
 768                         panic("share->excl");
 769                 }
 770                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
 771                     (flags & LOP_EXCLUSIVE) != 0) {
 772                         printf("exclusive lock of (%s) %s @ %s:%d\n",
 773                             class->lc_name, lock->lo_name, file, line);
 774                         printf("while share locked from %s:%d\n",
 775                             lock1->li_file, lock1->li_line);
 776                         panic("excl->share");
 777                 }
 778                 return;
 779         }
 780
 781         /*
 782          * Try locks do not block if they fail to acquire the lock, thus
 783          * there is no danger of deadlocks or of switching while holding a
 784          * spin lock if we acquire a lock via a try operation.
 785          */
 786         if (flags & LOP_TRYLOCK)
 787                 return;
 788
 789         /*
 790          * Check for duplicate locks of the same type.  Note that we only
 791          * have to check for this on the last lock we just acquired.  Any
 792          * other cases will be caught as lock order violations.
 793          */
 794         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 795         w1 = lock1->li_lock->lo_witness;
 796         if (w1 == w) {
 797                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
 798                     (flags & LOP_DUPOK))
 799                         return;
 800                 w->w_same_squawked = 1;
 801                 printf("acquiring duplicate lock of same type: \"%s\"\n",
 802                         lock->lo_type);
 803                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
 804                     lock1->li_file, lock1->li_line);
 805                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
 806 #ifdef KDB
 807                 goto debugger;
 808 #else
 809                 return;
 810 #endif
 811         }
 812         MPASS(!mtx_owned(&w_mtx));
 813         mtx_lock_spin(&w_mtx);
 814         /*
 815          * If we have a known higher number just say ok
 816          */
 817         if (witness_watch > 1 && w->w_level > w1->w_level) {
 818                 mtx_unlock_spin(&w_mtx);
 819                 return;
 820         }
 821         /*
 822          * If we know that the the lock we are acquiring comes after
 823          * the lock we most recently acquired in the lock order tree,
 824          * then there is no need for any further checks.
 825          */
 826         if (isitmydescendant(w1, w)) {
 827                 mtx_unlock_spin(&w_mtx);
 828                 return;
 829         }
 830         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
 831                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
 832
 833                         MPASS(j < WITNESS_COUNT);
 834                         lock1 = &lle->ll_children[i];
 835                         w1 = lock1->li_lock->lo_witness;
 836
 837                         /*
 838                          * If this lock doesn't undergo witness checking,
 839                          * then skip it.
 840                          */
 841                         if (w1 == NULL) {
 842                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
 843                                     ("lock missing witness structure"));
 844                                 continue;
 845                         }
 846                         /*
 847                          * If we are locking Giant and this is a sleepable
 848                          * lock, then skip it.
 849                          */
 850                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
 851                             lock == &Giant.mtx_object)
 852                                 continue;
 853                         /*
 854                          * If we are locking a sleepable lock and this lock
 855                          * is Giant, then skip it.
 856                          */
 857                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 858                             lock1->li_lock == &Giant.mtx_object)
 859                                 continue;
 860                         /*
 861                          * If we are locking a sleepable lock and this lock
 862                          * isn't sleepable, we want to treat it as a lock
 863                          * order violation to enfore a general lock order of
 864                          * sleepable locks before non-sleepable locks.
 865                          */
 866                         if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 867                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
 868                             /*
 869                              * Check the lock order hierarchy for a reveresal.
 870                              */
 871                             if (!isitmydescendant(w, w1))
 872                                 continue;
 873                         /*
 874                          * We have a lock order violation, check to see if it
 875                          * is allowed or has already been yelled about.
 876                          */
 877                         mtx_unlock_spin(&w_mtx);
 878 #ifdef BLESSING
 879                         /*
 880                          * If the lock order is blessed, just bail.  We don't
 881                          * look for other lock order violations though, which
 882                          * may be a bug.
 883                          */
 884                         if (blessed(w, w1))
 885                                 return;
 886 #endif
 887                         if (lock1->li_lock == &Giant.mtx_object) {
 888                                 if (w1->w_Giant_squawked)
 889                                         return;
 890                                 else
 891                                         w1->w_Giant_squawked = 1;
 892                         } else {
 893                                 if (w1->w_other_squawked)
 894                                         return;
 895                                 else
 896                                         w1->w_other_squawked = 1;
 897                         }
 898                         /*
 899                          * Ok, yell about it.
 900                          */
 901                         printf("lock order reversal\n");
 902                         /*
 903                          * Try to locate an earlier lock with
 904                          * witness w in our list.
 905                          */
 906                         do {
 907                                 lock2 = &lle->ll_children[i];
 908                                 MPASS(lock2->li_lock != NULL);
 909                                 if (lock2->li_lock->lo_witness == w)
 910                                         break;
 911                                 if (i == 0 && lle->ll_next != NULL) {
 912                                         lle = lle->ll_next;
 913                                         i = lle->ll_count - 1;
 914                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
 915                                 } else
 916                                         i--;
 917                         } while (i >= 0);
 918                         if (i < 0) {
 919                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 920                                     lock1->li_lock, lock1->li_lock->lo_name,
 921                                     lock1->li_lock->lo_type, lock1->li_file,
 922                                     lock1->li_line);
 923                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
 924                                     lock->lo_name, lock->lo_type, file, line);
 925                         } else {
 926                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 927                                     lock2->li_lock, lock2->li_lock->lo_name,
 928                                     lock2->li_lock->lo_type, lock2->li_file,
 929                                     lock2->li_line);
 930                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
 931                                     lock1->li_lock, lock1->li_lock->lo_name,
 932                                     lock1->li_lock->lo_type, lock1->li_file,
 933                                     lock1->li_line);
 934                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
 935                                     lock->lo_name, lock->lo_type, file, line);
 936                         }
 937 #ifdef KDB
 938                         goto debugger;
 939 #else
 940                         return;
 941 #endif
 942                 }
 943         }
 944         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 945         /*
 946          * If requested, build a new lock order.  However, don't build a new
 947          * relationship between a sleepable lock and Giant if it is in the
 948          * wrong direction.  The correct lock order is that sleepable locks
 949          * always come before Giant.
 950          */
 951         if (flags & LOP_NEWORDER &&
 952             !(lock1->li_lock == &Giant.mtx_object &&
 953             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
 954                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 955                     lock->lo_type, lock1->li_lock->lo_type);
 956                 if (!itismychild(lock1->li_lock->lo_witness, w))
 957                         /* Witness is dead. */
 958                         return;
 959         }
 960         mtx_unlock_spin(&w_mtx);
 961         return;
 962
 963 #ifdef KDB
 964 debugger:
 965         if (witness_trace)
 966                 kdb_backtrace();
 967         if (witness_kdb)
 968                 kdb_enter(__func__);
 969 #endif
 970 }
 971
 972 void
 973 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
 974 {
 975         struct lock_list_entry **lock_list, *lle;
 976         struct lock_instance *instance;
 977         struct witness *w;
 978         struct thread *td;
 979
 980         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 981             panicstr != NULL)
 982                 return;
 983         w = lock->lo_witness;
 984         td = curthread;
 985         file = fixup_filename(file);
 986
 987         /* Determine lock list for this lock. */
 988         if (lock->lo_class->lc_flags & LC_SLEEPLOCK)
 989                 lock_list = &td->td_sleeplocks;
 990         else
 991                 lock_list = PCPU_PTR(spinlocks);
 992
 993         /* Check to see if we are recursing on a lock we already own. */
 994         instance = find_instance(*lock_list, lock);
 995         if (instance != NULL) {
 996                 instance->li_flags++;
 997                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
 998                     td->td_proc->p_pid, lock->lo_name,
 999                     instance->li_flags & LI_RECURSEMASK);
1000                 instance->li_file = file;
1001                 instance->li_line = line;
1002                 return;
1003         }
1004
1005         /* Update per-witness last file and line acquire. */
1006         w->w_file = file;
1007         w->w_line = line;
1008
1009         /* Find the next open lock instance in the list and fill it. */
1010         lle = *lock_list;
1011         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1012                 lle = witness_lock_list_get();
1013                 if (lle == NULL)
1014                         return;
1015                 lle->ll_next = *lock_list;
1016                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1017                     td->td_proc->p_pid, lle);
1018                 *lock_list = lle;
1019         }
1020         instance = &lle->ll_children[lle->ll_count++];
1021         instance->li_lock = lock;
1022         instance->li_line = line;
1023         instance->li_file = file;
1024         if ((flags & LOP_EXCLUSIVE) != 0)
1025                 instance->li_flags = LI_EXCLUSIVE;
1026         else
1027                 instance->li_flags = 0;
1028         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1029             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1030 }
1031
1032 void
1033 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1034 {
1035         struct lock_instance *instance;
1036         struct lock_class *class;
1037
1038         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1039         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1040                 return;
1041         class = lock->lo_class;
1042         file = fixup_filename(file);
1043         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1044                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1045                     class->lc_name, lock->lo_name, file, line);
1046         if ((flags & LOP_TRYLOCK) == 0)
1047                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1048                     lock->lo_name, file, line);
1049         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1050                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1051                     class->lc_name, lock->lo_name, file, line);
1052         instance = find_instance(curthread->td_sleeplocks, lock);
1053         if (instance == NULL)
1054                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1055                     class->lc_name, lock->lo_name, file, line);
1056         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1057                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1058                     class->lc_name, lock->lo_name, file, line);
1059         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1060                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1061                     class->lc_name, lock->lo_name,
1062                     instance->li_flags & LI_RECURSEMASK, file, line);
1063         instance->li_flags |= LI_EXCLUSIVE;
1064 }
1065
1066 void
1067 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1068     int line)
1069 {
1070         struct lock_instance *instance;
1071         struct lock_class *class;
1072
1073         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1074         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1075                 return;
1076         class = lock->lo_class;
1077         file = fixup_filename(file);
1078         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1079                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1080                     class->lc_name, lock->lo_name, file, line);
1081         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1082                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1083                     class->lc_name, lock->lo_name, file, line);
1084         instance = find_instance(curthread->td_sleeplocks, lock);
1085         if (instance == NULL)
1086                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1087                     class->lc_name, lock->lo_name, file, line);
1088         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1089                 panic("downgrade of shared lock (%s) %s @ %s:%d",
1090                     class->lc_name, lock->lo_name, file, line);
1091         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1092                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1093                     class->lc_name, lock->lo_name,
1094                     instance->li_flags & LI_RECURSEMASK, file, line);
1095         instance->li_flags &= ~LI_EXCLUSIVE;
1096 }
1097
1098 void
1099 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1100 {
1101         struct lock_list_entry **lock_list, *lle;
1102         struct lock_instance *instance;
1103         struct lock_class *class;
1104         struct thread *td;
1105         register_t s;
1106         int i, j;
1107
1108         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1109             panicstr != NULL)
1110                 return;
1111         td = curthread;
1112         class = lock->lo_class;
1113         file = fixup_filename(file);
1114
1115         /* Find lock instance associated with this lock. */
1116         if (class->lc_flags & LC_SLEEPLOCK)
1117                 lock_list = &td->td_sleeplocks;
1118         else
1119                 lock_list = PCPU_PTR(spinlocks);
1120         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1121                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1122                         instance = &(*lock_list)->ll_children[i];
1123                         if (instance->li_lock == lock)
1124                                 goto found;
1125                 }
1126         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1127             file, line);
1128 found:
1129
1130         /* First, check for shared/exclusive mismatches. */
1131         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1132             (flags & LOP_EXCLUSIVE) == 0) {
1133                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1134                     lock->lo_name, file, line);
1135                 printf("while exclusively locked from %s:%d\n",
1136                     instance->li_file, instance->li_line);
1137                 panic("excl->ushare");
1138         }
1139         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1140             (flags & LOP_EXCLUSIVE) != 0) {
1141                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1142                     lock->lo_name, file, line);
1143                 printf("while share locked from %s:%d\n", instance->li_file,
1144                     instance->li_line);
1145                 panic("share->uexcl");
1146         }
1147
1148         /* If we are recursed, unrecurse. */
1149         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1150                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1151                     td->td_proc->p_pid, instance->li_lock->lo_name,
1152                     instance->li_flags);
1153                 instance->li_flags--;
1154                 return;
1155         }
1156
1157         /* Otherwise, remove this item from the list. */
1158         s = intr_disable();
1159         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1160             td->td_proc->p_pid, instance->li_lock->lo_name,
1161             (*lock_list)->ll_count - 1);
1162         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1163                 (*lock_list)->ll_children[j] =
1164                     (*lock_list)->ll_children[j + 1];
1165         (*lock_list)->ll_count--;
1166         intr_restore(s);
1167
1168         /* If this lock list entry is now empty, free it. */
1169         if ((*lock_list)->ll_count == 0) {
1170                 lle = *lock_list;
1171                 *lock_list = lle->ll_next;
1172                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1173                     td->td_proc->p_pid, lle);
1174                 witness_lock_list_free(lle);
1175         }
1176 }
1177
1178 /*
1179  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1180  * exempt Giant and sleepable locks from the checks as well.  If any
1181  * non-exempt locks are held, then a supplied message is printed to the
1182  * console along with a list of the offending locks.  If indicated in the
1183  * flags then a failure results in a panic as well.
1184  */
1185 int
1186 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1187 {
1188         struct lock_list_entry *lle;
1189         struct lock_instance *lock1;
1190         struct thread *td;
1191         va_list ap;
1192         int i, n;
1193
1194         if (witness_cold || witness_watch == 0 || panicstr != NULL)
1195                 return (0);
1196         n = 0;
1197         td = curthread;
1198         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1199                 for (i = lle->ll_count - 1; i >= 0; i--) {
1200                         lock1 = &lle->ll_children[i];
1201                         if (lock1->li_lock == lock)
1202                                 continue;
1203                         if (flags & WARN_GIANTOK &&
1204                             lock1->li_lock == &Giant.mtx_object)
1205                                 continue;
1206                         if (flags & WARN_SLEEPOK &&
1207                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1208                                 continue;
1209                         if (n == 0) {
1210                                 va_start(ap, fmt);
1211                                 vprintf(fmt, ap);
1212                                 va_end(ap);
1213                                 printf(" with the following");
1214                                 if (flags & WARN_SLEEPOK)
1215                                         printf(" non-sleepable");
1216                                 printf(" locks held:\n");
1217                         }
1218                         n++;
1219                         witness_list_lock(lock1);
1220                 }
1221         if (PCPU_GET(spinlocks) != NULL) {
1222                 /*
1223                  * Since we already hold a spinlock preemption is
1224                  * already blocked.
1225                  */
1226                 if (n == 0) {
1227                         va_start(ap, fmt);
1228                         vprintf(fmt, ap);
1229                         va_end(ap);
1230                         printf(" with the following");
1231                         if (flags & WARN_SLEEPOK)
1232                                 printf(" non-sleepable");
1233                         printf(" locks held:\n");
1234                 }
1235                 n += witness_list_locks(PCPU_PTR(spinlocks));
1236         }
1237         if (flags & WARN_PANIC && n)
1238                 panic("witness_warn");
1239 #ifdef KDB
1240         else if (witness_kdb && n)
1241                 kdb_enter(__func__);
1242         else if (witness_trace && n)
1243                 kdb_backtrace();
1244 #endif
1245         return (n);
1246 }
1247
1248 const char *
1249 witness_file(struct lock_object *lock)
1250 {
1251         struct witness *w;
1252
1253         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1254                 return ("?");
1255         w = lock->lo_witness;
1256         return (w->w_file);
1257 }
1258
1259 int
1260 witness_line(struct lock_object *lock)
1261 {
1262         struct witness *w;
1263
1264         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1265                 return (0);
1266         w = lock->lo_witness;
1267         return (w->w_line);
1268 }
1269
1270 static struct witness *
1271 enroll(const char *description, struct lock_class *lock_class)
1272 {
1273         struct witness *w;
1274
1275         if (witness_watch == 0 || panicstr != NULL)
1276                 return (NULL);
1277         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1278                 return (NULL);
1279         mtx_lock_spin(&w_mtx);
1280         STAILQ_FOREACH(w, &w_all, w_list) {
1281                 if (w->w_name == description || (w->w_refcount > 0 &&
1282                     strcmp(description, w->w_name) == 0)) {
1283                         w->w_refcount++;
1284                         mtx_unlock_spin(&w_mtx);
1285                         if (lock_class != w->w_class)
1286                                 panic(
1287                                 "lock (%s) %s does not match earlier (%s) lock",
1288                                     description, lock_class->lc_name,
1289                                     w->w_class->lc_name);
1290                         return (w);
1291                 }
1292         }
1293         /*
1294          * This isn't quite right, as witness_cold is still 0 while we
1295          * enroll all the locks initialized before witness_initialize().
1296          */
1297         if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
1298                 mtx_unlock_spin(&w_mtx);
1299                 panic("spin lock %s not in order list", description);
1300         }
1301         if ((w = witness_get()) == NULL)
1302                 return (NULL);
1303         w->w_name = description;
1304         w->w_class = lock_class;
1305         w->w_refcount = 1;
1306         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1307         if (lock_class->lc_flags & LC_SPINLOCK)
1308                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1309         else if (lock_class->lc_flags & LC_SLEEPLOCK)
1310                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1311         else {
1312                 mtx_unlock_spin(&w_mtx);
1313                 panic("lock class %s is not sleep or spin",
1314                     lock_class->lc_name);
1315         }
1316         mtx_unlock_spin(&w_mtx);
1317         return (w);
1318 }
1319
1320 /* Don't let the door bang you on the way out... */
1321 static int
1322 depart(struct witness *w)
1323 {
1324         struct witness_child_list_entry *wcl, *nwcl;
1325         struct witness_list *list;
1326         struct witness *parent;
1327
1328         MPASS(w->w_refcount == 0);
1329         if (w->w_class->lc_flags & LC_SLEEPLOCK)
1330                 list = &w_sleep;
1331         else
1332                 list = &w_spin;
1333         /*
1334          * First, we run through the entire tree looking for any
1335          * witnesses that the outgoing witness is a child of.  For
1336          * each parent that we find, we reparent all the direct
1337          * children of the outgoing witness to its parent.
1338          */
1339         STAILQ_FOREACH(parent, list, w_typelist) {
1340                 if (!isitmychild(parent, w))
1341                         continue;
1342                 removechild(parent, w);
1343                 if (!reparentchildren(parent, w))
1344                         return (0);
1345         }
1346
1347         /*
1348          * Now we go through and free up the child list of the
1349          * outgoing witness.
1350          */
1351         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1352                 nwcl = wcl->wcl_next;
1353                 witness_child_free(wcl);
1354         }
1355
1356         /*
1357          * Detach from various lists and free.
1358          */
1359         STAILQ_REMOVE(list, w, witness, w_typelist);
1360         STAILQ_REMOVE(&w_all, w, witness, w_list);
1361         witness_free(w);
1362
1363         /* Finally, fixup the tree. */
1364         return (rebalancetree(list));
1365 }
1366
1367 /*
1368  * Prune an entire lock order tree.  We look for cases where a lock
1369  * is now both a descendant and a direct child of a given lock.  In
1370  * that case, we want to remove the direct child link from the tree.
1371  *
1372  * Returns false if insertchild() fails.
1373  */
1374 static int
1375 rebalancetree(struct witness_list *list)
1376 {
1377         struct witness *child, *parent;
1378
1379         STAILQ_FOREACH(child, list, w_typelist) {
1380                 STAILQ_FOREACH(parent, list, w_typelist) {
1381                         if (!isitmychild(parent, child))
1382                                 continue;
1383                         removechild(parent, child);
1384                         if (isitmydescendant(parent, child))
1385                                 continue;
1386                         if (!insertchild(parent, child))
1387                                 return (0);
1388                 }
1389         }
1390         witness_levelall();
1391         return (1);
1392 }
1393
1394 /*
1395  * Add "child" as a direct child of "parent".  Returns false if
1396  * we fail due to out of memory.
1397  */
1398 static int
1399 insertchild(struct witness *parent, struct witness *child)
1400 {
1401         struct witness_child_list_entry **wcl;
1402
1403         MPASS(child != NULL && parent != NULL);
1404
1405         /*
1406          * Insert "child" after "parent"
1407          */
1408         wcl = &parent->w_children;
1409         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1410                 wcl = &(*wcl)->wcl_next;
1411         if (*wcl == NULL) {
1412                 *wcl = witness_child_get();
1413                 if (*wcl == NULL)
1414                         return (0);
1415         }
1416         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1417
1418         return (1);
1419 }
1420
1421 /*
1422  * Make all the direct descendants of oldparent be direct descendants
1423  * of newparent.
1424  */
1425 static int
1426 reparentchildren(struct witness *newparent, struct witness *oldparent)
1427 {
1428         struct witness_child_list_entry *wcl;
1429         int i;
1430
1431         /* Avoid making a witness a child of itself. */
1432         MPASS(!isitmychild(oldparent, newparent));
1433
1434         for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1435                 for (i = 0; i < wcl->wcl_count; i++)
1436                         if (!insertchild(newparent, wcl->wcl_children[i]))
1437                                 return (0);
1438         return (1);
1439 }
1440
1441 static int
1442 itismychild(struct witness *parent, struct witness *child)
1443 {
1444         struct witness_list *list;
1445
1446         MPASS(child != NULL && parent != NULL);
1447         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1448             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1449                 panic(
1450                 "%s: parent (%s) and child (%s) are not the same lock type",
1451                     __func__, parent->w_class->lc_name,
1452                     child->w_class->lc_name);
1453
1454         if (!insertchild(parent, child))
1455                 return (0);
1456
1457         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1458                 list = &w_sleep;
1459         else
1460                 list = &w_spin;
1461         return (rebalancetree(list));
1462 }
1463
1464 static void
1465 removechild(struct witness *parent, struct witness *child)
1466 {
1467         struct witness_child_list_entry **wcl, *wcl1;
1468         int i;
1469
1470         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1471                 for (i = 0; i < (*wcl)->wcl_count; i++)
1472                         if ((*wcl)->wcl_children[i] == child)
1473                                 goto found;
1474         return;
1475 found:
1476         (*wcl)->wcl_count--;
1477         if ((*wcl)->wcl_count > i)
1478                 (*wcl)->wcl_children[i] =
1479                     (*wcl)->wcl_children[(*wcl)->wcl_count];
1480         MPASS((*wcl)->wcl_children[i] != NULL);
1481         if ((*wcl)->wcl_count != 0)
1482                 return;
1483         wcl1 = *wcl;
1484         *wcl = wcl1->wcl_next;
1485         witness_child_free(wcl1);
1486 }
1487
1488 static int
1489 isitmychild(struct witness *parent, struct witness *child)
1490 {
1491         struct witness_child_list_entry *wcl;
1492         int i;
1493
1494         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1495                 for (i = 0; i < wcl->wcl_count; i++) {
1496                         if (wcl->wcl_children[i] == child)
1497                                 return (1);
1498                 }
1499         }
1500         return (0);
1501 }
1502
1503 static int
1504 isitmydescendant(struct witness *parent, struct witness *child)
1505 {
1506         struct witness_child_list_entry *wcl;
1507         int i, j;
1508
1509         if (isitmychild(parent, child))
1510                 return (1);
1511         j = 0;
1512         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1513                 MPASS(j < 1000);
1514                 for (i = 0; i < wcl->wcl_count; i++) {
1515                         if (isitmydescendant(wcl->wcl_children[i], child))
1516                                 return (1);
1517                 }
1518                 j++;
1519         }
1520         return (0);
1521 }
1522
1523 static void
1524 witness_levelall (void)
1525 {
1526         struct witness_list *list;
1527         struct witness *w, *w1;
1528
1529         /*
1530          * First clear all levels.
1531          */
1532         STAILQ_FOREACH(w, &w_all, w_list) {
1533                 w->w_level = 0;
1534         }
1535
1536         /*
1537          * Look for locks with no parent and level all their descendants.
1538          */
1539         STAILQ_FOREACH(w, &w_all, w_list) {
1540                 /*
1541                  * This is just an optimization, technically we could get
1542                  * away just walking the all list each time.
1543                  */
1544                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
1545                         list = &w_sleep;
1546                 else
1547                         list = &w_spin;
1548                 STAILQ_FOREACH(w1, list, w_typelist) {
1549                         if (isitmychild(w1, w))
1550                                 goto skip;
1551                 }
1552                 witness_leveldescendents(w, 0);
1553         skip:
1554                 ;       /* silence GCC 3.x */
1555         }
1556 }
1557
1558 static void
1559 witness_leveldescendents(struct witness *parent, int level)
1560 {
1561         struct witness_child_list_entry *wcl;
1562         int i;
1563
1564         if (parent->w_level < level)
1565                 parent->w_level = level;
1566         level++;
1567         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1568                 for (i = 0; i < wcl->wcl_count; i++)
1569                         witness_leveldescendents(wcl->wcl_children[i], level);
1570 }
1571
1572 static void
1573 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1574                            struct witness *parent, int indent)
1575 {
1576         struct witness_child_list_entry *wcl;
1577         int i, level;
1578
1579         level = parent->w_level;
1580         prnt("%-2d", level);
1581         for (i = 0; i < indent; i++)
1582                 prnt(" ");
1583         if (parent->w_refcount > 0)
1584                 prnt("%s", parent->w_name);
1585         else
1586                 prnt("(dead)");
1587         if (parent->w_displayed) {
1588                 prnt(" -- (already displayed)\n");
1589                 return;
1590         }
1591         parent->w_displayed = 1;
1592         if (parent->w_refcount > 0) {
1593                 if (parent->w_file != NULL)
1594                         prnt(" -- last acquired @ %s:%d", parent->w_file,
1595                             parent->w_line);
1596         }
1597         prnt("\n");
1598         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1599                 for (i = 0; i < wcl->wcl_count; i++)
1600                             witness_displaydescendants(prnt,
1601                                 wcl->wcl_children[i], indent + 1);
1602 }
1603
1604 #ifdef BLESSING
1605 static int
1606 blessed(struct witness *w1, struct witness *w2)
1607 {
1608         int i;
1609         struct witness_blessed *b;
1610
1611         for (i = 0; i < blessed_count; i++) {
1612                 b = &blessed_list[i];
1613                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1614                         if (strcmp(w2->w_name, b->b_lock2) == 0)
1615                                 return (1);
1616                         continue;
1617                 }
1618                 if (strcmp(w1->w_name, b->b_lock2) == 0)
1619                         if (strcmp(w2->w_name, b->b_lock1) == 0)
1620                                 return (1);
1621         }
1622         return (0);
1623 }
1624 #endif
1625
1626 static struct witness *
1627 witness_get(void)
1628 {
1629         struct witness *w;
1630
1631         if (witness_watch == 0) {
1632                 mtx_unlock_spin(&w_mtx);
1633                 return (NULL);
1634         }
1635         if (STAILQ_EMPTY(&w_free)) {
1636                 witness_watch = 0;
1637                 mtx_unlock_spin(&w_mtx);
1638                 printf("%s: witness exhausted\n", __func__);
1639                 return (NULL);
1640         }
1641         w = STAILQ_FIRST(&w_free);
1642         STAILQ_REMOVE_HEAD(&w_free, w_list);
1643         bzero(w, sizeof(*w));
1644         return (w);
1645 }
1646
1647 static void
1648 witness_free(struct witness *w)
1649 {
1650
1651         STAILQ_INSERT_HEAD(&w_free, w, w_list);
1652 }
1653
1654 static struct witness_child_list_entry *
1655 witness_child_get(void)
1656 {
1657         struct witness_child_list_entry *wcl;
1658
1659         if (witness_watch == 0) {
1660                 mtx_unlock_spin(&w_mtx);
1661                 return (NULL);
1662         }
1663         wcl = w_child_free;
1664         if (wcl == NULL) {
1665                 witness_watch = 0;
1666                 mtx_unlock_spin(&w_mtx);
1667                 printf("%s: witness exhausted\n", __func__);
1668                 return (NULL);
1669         }
1670         w_child_free = wcl->wcl_next;
1671         bzero(wcl, sizeof(*wcl));
1672         return (wcl);
1673 }
1674
1675 static void
1676 witness_child_free(struct witness_child_list_entry *wcl)
1677 {
1678
1679         wcl->wcl_next = w_child_free;
1680         w_child_free = wcl;
1681 }
1682
1683 static struct lock_list_entry *
1684 witness_lock_list_get(void)
1685 {
1686         struct lock_list_entry *lle;
1687
1688         if (witness_watch == 0)
1689                 return (NULL);
1690         mtx_lock_spin(&w_mtx);
1691         lle = w_lock_list_free;
1692         if (lle == NULL) {
1693                 witness_watch = 0;
1694                 mtx_unlock_spin(&w_mtx);
1695                 printf("%s: witness exhausted\n", __func__);
1696                 return (NULL);
1697         }
1698         w_lock_list_free = lle->ll_next;
1699         mtx_unlock_spin(&w_mtx);
1700         bzero(lle, sizeof(*lle));
1701         return (lle);
1702 }
1703
1704 static void
1705 witness_lock_list_free(struct lock_list_entry *lle)
1706 {
1707
1708         mtx_lock_spin(&w_mtx);
1709         lle->ll_next = w_lock_list_free;
1710         w_lock_list_free = lle;
1711         mtx_unlock_spin(&w_mtx);
1712 }
1713
1714 static struct lock_instance *
1715 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1716 {
1717         struct lock_list_entry *lle;
1718         struct lock_instance *instance;
1719         int i;
1720
1721         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1722                 for (i = lle->ll_count - 1; i >= 0; i--) {
1723                         instance = &lle->ll_children[i];
1724                         if (instance->li_lock == lock)
1725                                 return (instance);
1726                 }
1727         return (NULL);
1728 }
1729
1730 static void
1731 witness_list_lock(struct lock_instance *instance)
1732 {
1733         struct lock_object *lock;
1734
1735         lock = instance->li_lock;
1736         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1737             "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
1738         if (lock->lo_type != lock->lo_name)
1739                 printf(" (%s)", lock->lo_type);
1740         printf(" r = %d (%p) locked @ %s:%d\n",
1741             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1742             instance->li_line);
1743 }
1744
1745 #ifdef DDB
1746 static int
1747 witness_thread_has_locks(struct thread *td)
1748 {
1749
1750         return (td->td_sleeplocks != NULL);
1751 }
1752
1753 static int
1754 witness_proc_has_locks(struct proc *p)
1755 {
1756         struct thread *td;
1757
1758         FOREACH_THREAD_IN_PROC(p, td) {
1759                 if (witness_thread_has_locks(td))
1760                         return (1);
1761         }
1762         return (0);
1763 }
1764 #endif
1765
1766 int
1767 witness_list_locks(struct lock_list_entry **lock_list)
1768 {
1769         struct lock_list_entry *lle;
1770         int i, nheld;
1771
1772         nheld = 0;
1773         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1774                 for (i = lle->ll_count - 1; i >= 0; i--) {
1775                         witness_list_lock(&lle->ll_children[i]);
1776                         nheld++;
1777                 }
1778         return (nheld);
1779 }
1780
1781 /*
1782  * This is a bit risky at best.  We call this function when we have timed
1783  * out acquiring a spin lock, and we assume that the other CPU is stuck
1784  * with this lock held.  So, we go groveling around in the other CPU's
1785  * per-cpu data to try to find the lock instance for this spin lock to
1786  * see when it was last acquired.
1787  */
1788 void
1789 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1790 {
1791         struct lock_instance *instance;
1792         struct pcpu *pc;
1793
1794         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1795                 return;
1796         pc = pcpu_find(owner->td_oncpu);
1797         instance = find_instance(pc->pc_spinlocks, lock);
1798         if (instance != NULL)
1799                 witness_list_lock(instance);
1800 }
1801
1802 void
1803 witness_save(struct lock_object *lock, const char **filep, int *linep)
1804 {
1805         struct lock_instance *instance;
1806
1807         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1808         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1809                 return;
1810         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1811                 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1812                     lock->lo_class->lc_name, lock->lo_name);
1813         instance = find_instance(curthread->td_sleeplocks, lock);
1814         if (instance == NULL)
1815                 panic("%s: lock (%s) %s not locked", __func__,
1816                     lock->lo_class->lc_name, lock->lo_name);
1817         *filep = instance->li_file;
1818         *linep = instance->li_line;
1819 }
1820
1821 void
1822 witness_restore(struct lock_object *lock, const char *file, int line)
1823 {
1824         struct lock_instance *instance;
1825
1826         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1827         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1828                 return;
1829         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1830                 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1831                     lock->lo_class->lc_name, lock->lo_name);
1832         instance = find_instance(curthread->td_sleeplocks, lock);
1833         if (instance == NULL)
1834                 panic("%s: lock (%s) %s not locked", __func__,
1835                     lock->lo_class->lc_name, lock->lo_name);
1836         lock->lo_witness->w_file = file;
1837         lock->lo_witness->w_line = line;
1838         instance->li_file = file;
1839         instance->li_line = line;
1840 }
1841
1842 void
1843 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1844 {
1845 #ifdef INVARIANT_SUPPORT
1846         struct lock_instance *instance;
1847
1848         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1849                 return;
1850         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1851                 instance = find_instance(curthread->td_sleeplocks, lock);
1852         else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1853                 instance = find_instance(PCPU_GET(spinlocks), lock);
1854         else {
1855                 panic("Lock (%s) %s is not sleep or spin!",
1856                     lock->lo_class->lc_name, lock->lo_name);
1857         }
1858         file = fixup_filename(file);
1859         switch (flags) {
1860         case LA_UNLOCKED:
1861                 if (instance != NULL)
1862                         panic("Lock (%s) %s locked @ %s:%d.",
1863                             lock->lo_class->lc_name, lock->lo_name, file, line);
1864                 break;
1865         case LA_LOCKED:
1866         case LA_LOCKED | LA_RECURSED:
1867         case LA_LOCKED | LA_NOTRECURSED:
1868         case LA_SLOCKED:
1869         case LA_SLOCKED | LA_RECURSED:
1870         case LA_SLOCKED | LA_NOTRECURSED:
1871         case LA_XLOCKED:
1872         case LA_XLOCKED | LA_RECURSED:
1873         case LA_XLOCKED | LA_NOTRECURSED:
1874                 if (instance == NULL) {
1875                         panic("Lock (%s) %s not locked @ %s:%d.",
1876                             lock->lo_class->lc_name, lock->lo_name, file, line);
1877                         break;
1878                 }
1879                 if ((flags & LA_XLOCKED) != 0 &&
1880                     (instance->li_flags & LI_EXCLUSIVE) == 0)
1881                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1882                             lock->lo_class->lc_name, lock->lo_name, file, line);
1883                 if ((flags & LA_SLOCKED) != 0 &&
1884                     (instance->li_flags & LI_EXCLUSIVE) != 0)
1885                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
1886                             lock->lo_class->lc_name, lock->lo_name, file, line);
1887                 if ((flags & LA_RECURSED) != 0 &&
1888                     (instance->li_flags & LI_RECURSEMASK) == 0)
1889                         panic("Lock (%s) %s not recursed @ %s:%d.",
1890                             lock->lo_class->lc_name, lock->lo_name, file, line);
1891                 if ((flags & LA_NOTRECURSED) != 0 &&
1892                     (instance->li_flags & LI_RECURSEMASK) != 0)
1893                         panic("Lock (%s) %s recursed @ %s:%d.",
1894                             lock->lo_class->lc_name, lock->lo_name, file, line);
1895                 break;
1896         default:
1897                 panic("Invalid lock assertion at %s:%d.", file, line);
1898
1899         }
1900 #endif  /* INVARIANT_SUPPORT */
1901 }
1902
1903 #ifdef DDB
1904 static void
1905 witness_list(struct thread *td)
1906 {
1907
1908         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1909         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1910
1911         if (witness_watch == 0)
1912                 return;
1913
1914         witness_list_locks(&td->td_sleeplocks);
1915
1916         /*
1917          * We only handle spinlocks if td == curthread.  This is somewhat broken
1918          * if td is currently executing on some other CPU and holds spin locks
1919          * as we won't display those locks.  If we had a MI way of getting
1920          * the per-cpu data for a given cpu then we could use
1921          * td->td_oncpu to get the list of spinlocks for this thread
1922          * and "fix" this.
1923          *
1924          * That still wouldn't really fix this unless we locked sched_lock
1925          * or stopped the other CPU to make sure it wasn't changing the list
1926          * out from under us.  It is probably best to just not try to handle
1927          * threads on other CPU's for now.
1928          */
1929         if (td == curthread && PCPU_GET(spinlocks) != NULL)
1930                 witness_list_locks(PCPU_PTR(spinlocks));
1931 }
1932
1933 DB_SHOW_COMMAND(locks, db_witness_list)
1934 {
1935         struct thread *td;
1936         pid_t pid;
1937         struct proc *p;
1938
1939         if (have_addr) {
1940                 pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1941                     ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1942                     ((addr >> 16) % 16) * 10000;
1943                 /* sx_slock(&allproc_lock); */
1944                 FOREACH_PROC_IN_SYSTEM(p) {
1945                         if (p->p_pid == pid)
1946                                 break;
1947                 }
1948                 /* sx_sunlock(&allproc_lock); */
1949                 if (p == NULL) {
1950                         db_printf("pid %d not found\n", pid);
1951                         return;
1952                 }
1953                 FOREACH_THREAD_IN_PROC(p, td) {
1954                         witness_list(td);
1955                 }
1956         } else {
1957                 td = curthread;
1958                 witness_list(td);
1959         }
1960 }
1961
1962 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
1963 {
1964         struct thread *td;
1965         struct proc *p;
1966
1967         /*
1968          * It would be nice to list only threads and processes that actually
1969          * held sleep locks, but that information is currently not exported
1970          * by WITNESS.
1971          */
1972         FOREACH_PROC_IN_SYSTEM(p) {
1973                 if (!witness_proc_has_locks(p))
1974                         continue;
1975                 FOREACH_THREAD_IN_PROC(p, td) {
1976                         if (!witness_thread_has_locks(td))
1977                                 continue;
1978                         printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
1979                             p->p_comm, td, td->td_tid);
1980                         witness_list(td);
1981                 }
1982         }
1983 }
1984
1985 DB_SHOW_COMMAND(witness, db_witness_display)
1986 {
1987
1988         witness_display(db_printf);
1989 }
1990 #endif