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          * spin locks
 335          */
 336 #ifdef SMP
 337         { "ap boot", &lock_class_mtx_spin },
 338 #endif
 339         { "sio", &lock_class_mtx_spin },
 340 #ifdef __i386__
 341         { "cy", &lock_class_mtx_spin },
 342 #endif
 343         { "uart_hwmtx", &lock_class_mtx_spin },
 344         { "sabtty", &lock_class_mtx_spin },
 345         { "zstty", &lock_class_mtx_spin },
 346         { "ng_node", &lock_class_mtx_spin },
 347         { "ng_worklist", &lock_class_mtx_spin },
 348         { "taskqueue_fast", &lock_class_mtx_spin },
 349         { "intr table", &lock_class_mtx_spin },
 350         { "ithread table lock", &lock_class_mtx_spin },
 351         { "sleepq chain", &lock_class_mtx_spin },
 352         { "sched lock", &lock_class_mtx_spin },
 353         { "turnstile chain", &lock_class_mtx_spin },
 354         { "td_contested", &lock_class_mtx_spin },
 355         { "callout", &lock_class_mtx_spin },
 356         { "entropy harvest mutex", &lock_class_mtx_spin },
 357         /*
 358          * leaf locks
 359          */
 360         { "allpmaps", &lock_class_mtx_spin },
 361         { "vm page queue free mutex", &lock_class_mtx_spin },
 362         { "icu", &lock_class_mtx_spin },
 363 #ifdef SMP
 364         { "smp rendezvous", &lock_class_mtx_spin },
 365 #if defined(__i386__) || defined(__amd64__)
 366         { "tlb", &lock_class_mtx_spin },
 367 #endif
 368 #ifdef __sparc64__
 369         { "ipi", &lock_class_mtx_spin },
 370 #endif
 371 #endif
 372         { "clk", &lock_class_mtx_spin },
 373         { "mutex profiling lock", &lock_class_mtx_spin },
 374         { "kse zombie lock", &lock_class_mtx_spin },
 375         { "ALD Queue", &lock_class_mtx_spin },
 376 #ifdef __ia64__
 377         { "MCA spin lock", &lock_class_mtx_spin },
 378 #endif
 379 #if defined(__i386__) || defined(__amd64__)
 380         { "pcicfg", &lock_class_mtx_spin },
 381         { "NDIS thread lock", &lock_class_mtx_spin },
 382 #endif
 383         { NULL, NULL },
 384         { NULL, NULL }
 385 };
 386
 387 #ifdef BLESSING
 388 /*
 389  * Pairs of locks which have been blessed
 390  * Don't complain about order problems with blessed locks
 391  */
 392 static struct witness_blessed blessed_list[] = {
 393 };
 394 static int blessed_count =
 395         sizeof(blessed_list) / sizeof(struct witness_blessed);
 396 #endif
 397
 398 /*
 399  * List of all locks in the system.
 400  */
 401 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
 402
 403 static struct mtx all_mtx = {
 404         { &lock_class_mtx_sleep,        /* mtx_object.lo_class */
 405           "All locks list",             /* mtx_object.lo_name */
 406           "All locks list",             /* mtx_object.lo_type */
 407           LO_INITIALIZED,               /* mtx_object.lo_flags */
 408           { NULL, NULL },               /* mtx_object.lo_list */
 409           NULL },                       /* mtx_object.lo_witness */
 410         MTX_UNOWNED, 0                  /* mtx_lock, mtx_recurse */
 411 };
 412
 413 /*
 414  * This global is set to 0 once it becomes safe to use the witness code.
 415  */
 416 static int witness_cold = 1;
 417
 418 /*
 419  * Global variables for book keeping.
 420  */
 421 static int lock_cur_cnt;
 422 static int lock_max_cnt;
 423
 424 /*
 425  * The WITNESS-enabled diagnostic code.
 426  */
 427 static void
 428 witness_initialize(void *dummy __unused)
 429 {
 430         struct lock_object *lock;
 431         struct witness_order_list_entry *order;
 432         struct witness *w, *w1;
 433         int i;
 434
 435         /*
 436          * We have to release Giant before initializing its witness
 437          * structure so that WITNESS doesn't get confused.
 438          */
 439         mtx_unlock(&Giant);
 440         mtx_assert(&Giant, MA_NOTOWNED);
 441
 442         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 443         TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
 444         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 445             MTX_NOWITNESS);
 446         for (i = 0; i < WITNESS_COUNT; i++)
 447                 witness_free(&w_data[i]);
 448         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
 449                 witness_child_free(&w_childdata[i]);
 450         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 451                 witness_lock_list_free(&w_locklistdata[i]);
 452
 453         /* First add in all the specified order lists. */
 454         for (order = order_lists; order->w_name != NULL; order++) {
 455                 w = enroll(order->w_name, order->w_class);
 456                 if (w == NULL)
 457                         continue;
 458                 w->w_file = "order list";
 459                 for (order++; order->w_name != NULL; order++) {
 460                         w1 = enroll(order->w_name, order->w_class);
 461                         if (w1 == NULL)
 462                                 continue;
 463                         w1->w_file = "order list";
 464                         if (!itismychild(w, w1))
 465                                 panic("Not enough memory for static orders!");
 466                         w = w1;
 467                 }
 468         }
 469
 470         /* Iterate through all locks and add them to witness. */
 471         mtx_lock(&all_mtx);
 472         TAILQ_FOREACH(lock, &all_locks, lo_list) {
 473                 if (lock->lo_flags & LO_WITNESS)
 474                         lock->lo_witness = enroll(lock->lo_type,
 475                             lock->lo_class);
 476                 else
 477                         lock->lo_witness = NULL;
 478         }
 479         mtx_unlock(&all_mtx);
 480
 481         /* Mark the witness code as being ready for use. */
 482         atomic_store_rel_int(&witness_cold, 0);
 483
 484         mtx_lock(&Giant);
 485 }
 486 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
 487
 488 static int
 489 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
 490 {
 491         int error, value;
 492
 493         value = witness_watch;
 494         error = sysctl_handle_int(oidp, &value, 0, req);
 495         if (error != 0 || req->newptr == NULL)
 496                 return (error);
 497         error = suser(req->td);
 498         if (error != 0)
 499                 return (error);
 500         if (value == witness_watch)
 501                 return (0);
 502         if (value != 0)
 503                 return (EINVAL);
 504         witness_watch = 0;
 505         return (0);
 506 }
 507
 508 void
 509 witness_init(struct lock_object *lock)
 510 {
 511         struct lock_class *class;
 512
 513         class = lock->lo_class;
 514         if (lock->lo_flags & LO_INITIALIZED)
 515                 panic("%s: lock (%s) %s is already initialized", __func__,
 516                     class->lc_name, lock->lo_name);
 517         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 518             (class->lc_flags & LC_RECURSABLE) == 0)
 519                 panic("%s: lock (%s) %s can not be recursable", __func__,
 520                     class->lc_name, lock->lo_name);
 521         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 522             (class->lc_flags & LC_SLEEPABLE) == 0)
 523                 panic("%s: lock (%s) %s can not be sleepable", __func__,
 524                     class->lc_name, lock->lo_name);
 525         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 526             (class->lc_flags & LC_UPGRADABLE) == 0)
 527                 panic("%s: lock (%s) %s can not be upgradable", __func__,
 528                     class->lc_name, lock->lo_name);
 529
 530         mtx_lock(&all_mtx);
 531         TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
 532         lock->lo_flags |= LO_INITIALIZED;
 533         lock_cur_cnt++;
 534         if (lock_cur_cnt > lock_max_cnt)
 535                 lock_max_cnt = lock_cur_cnt;
 536         mtx_unlock(&all_mtx);
 537         if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
 538             (lock->lo_flags & LO_WITNESS) != 0)
 539                 lock->lo_witness = enroll(lock->lo_type, class);
 540         else
 541                 lock->lo_witness = NULL;
 542 }
 543
 544 void
 545 witness_destroy(struct lock_object *lock)
 546 {
 547         struct witness *w;
 548
 549         if (witness_cold)
 550                 panic("lock (%s) %s destroyed while witness_cold",
 551                     lock->lo_class->lc_name, lock->lo_name);
 552         if ((lock->lo_flags & LO_INITIALIZED) == 0)
 553                 panic("%s: lock (%s) %s is not initialized", __func__,
 554                     lock->lo_class->lc_name, lock->lo_name);
 555
 556         /* XXX: need to verify that no one holds the lock */
 557         w = lock->lo_witness;
 558         if (w != NULL) {
 559                 mtx_lock_spin(&w_mtx);
 560                 MPASS(w->w_refcount > 0);
 561                 w->w_refcount--;
 562
 563                 /*
 564                  * Lock is already released if we have an allocation failure
 565                  * and depart() fails.
 566                  */
 567                 if (w->w_refcount != 0 || depart(w))
 568                         mtx_unlock_spin(&w_mtx);
 569         }
 570
 571         mtx_lock(&all_mtx);
 572         lock_cur_cnt--;
 573         TAILQ_REMOVE(&all_locks, lock, lo_list);
 574         lock->lo_flags &= ~LO_INITIALIZED;
 575         mtx_unlock(&all_mtx);
 576 }
 577
 578 #ifdef DDB
 579 static void
 580 witness_display_list(void(*prnt)(const char *fmt, ...),
 581                      struct witness_list *list)
 582 {
 583         struct witness *w;
 584
 585         STAILQ_FOREACH(w, list, w_typelist) {
 586                 if (w->w_file == NULL || w->w_level > 0)
 587                         continue;
 588                 /*
 589                  * This lock has no anscestors, display its descendants.
 590                  */
 591                 witness_displaydescendants(prnt, w, 0);
 592         }
 593 }
 594
 595 static void
 596 witness_display(void(*prnt)(const char *fmt, ...))
 597 {
 598         struct witness *w;
 599
 600         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 601         witness_levelall();
 602
 603         /* Clear all the displayed flags. */
 604         STAILQ_FOREACH(w, &w_all, w_list) {
 605                 w->w_displayed = 0;
 606         }
 607
 608         /*
 609          * First, handle sleep locks which have been acquired at least
 610          * once.
 611          */
 612         prnt("Sleep locks:\n");
 613         witness_display_list(prnt, &w_sleep);
 614
 615         /*
 616          * Now do spin locks which have been acquired at least once.
 617          */
 618         prnt("\nSpin locks:\n");
 619         witness_display_list(prnt, &w_spin);
 620
 621         /*
 622          * Finally, any locks which have not been acquired yet.
 623          */
 624         prnt("\nLocks which were never acquired:\n");
 625         STAILQ_FOREACH(w, &w_all, w_list) {
 626                 if (w->w_file != NULL || w->w_refcount == 0)
 627                         continue;
 628                 prnt("%s\n", w->w_name);
 629         }
 630 }
 631 #endif /* DDB */
 632
 633 /* Trim useless garbage from filenames. */
 634 static const char *
 635 fixup_filename(const char *file)
 636 {
 637
 638         if (file == NULL)
 639                 return (NULL);
 640         while (strncmp(file, "../", 3) == 0)
 641                 file += 3;
 642         return (file);
 643 }
 644
 645 int
 646 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
 647 {
 648
 649         if (witness_watch == 0 || panicstr != NULL)
 650                 return (0);
 651
 652         /* Require locks that witness knows about. */
 653         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
 654             lock2->lo_witness == NULL)
 655                 return (EINVAL);
 656
 657         MPASS(!mtx_owned(&w_mtx));
 658         mtx_lock_spin(&w_mtx);
 659
 660         /*
 661          * If we already have either an explicit or implied lock order that
 662          * is the other way around, then return an error.
 663          */
 664         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
 665                 mtx_unlock_spin(&w_mtx);
 666                 return (EDOOFUS);
 667         }
 668
 669         /* Try to add the new order. */
 670         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 671             lock2->lo_type, lock1->lo_type);
 672         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
 673                 return (ENOMEM);
 674         mtx_unlock_spin(&w_mtx);
 675         return (0);
 676 }
 677
 678 void
 679 witness_checkorder(struct lock_object *lock, int flags, const char *file,
 680     int line)
 681 {
 682         struct lock_list_entry **lock_list, *lle;
 683         struct lock_instance *lock1, *lock2;
 684         struct lock_class *class;
 685         struct witness *w, *w1;
 686         struct thread *td;
 687         int i, j;
 688
 689         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 690             panicstr != NULL)
 691                 return;
 692
 693         /*
 694          * Try locks do not block if they fail to acquire the lock, thus
 695          * there is no danger of deadlocks or of switching while holding a
 696          * spin lock if we acquire a lock via a try operation.  This
 697          * function shouldn't even be called for try locks, so panic if
 698          * that happens.
 699          */
 700         if (flags & LOP_TRYLOCK)
 701                 panic("%s should not be called for try lock operations",
 702                     __func__);
 703
 704         w = lock->lo_witness;
 705         class = lock->lo_class;
 706         td = curthread;
 707         file = fixup_filename(file);
 708
 709         if (class->lc_flags & LC_SLEEPLOCK) {
 710                 /*
 711                  * Since spin locks include a critical section, this check
 712                  * implicitly enforces a lock order of all sleep locks before
 713                  * all spin locks.
 714                  */
 715                 if (td->td_critnest != 0 && !kdb_active)
 716                         panic("blockable sleep lock (%s) %s @ %s:%d",
 717                             class->lc_name, lock->lo_name, file, line);
 718
 719                 /*
 720                  * If this is the first lock acquired then just return as
 721                  * no order checking is needed.
 722                  */
 723                 if (td->td_sleeplocks == NULL)
 724                         return;
 725                 lock_list = &td->td_sleeplocks;
 726         } else {
 727                 /*
 728                  * If this is the first lock, just return as no order
 729                  * checking is needed.  We check this in both if clauses
 730                  * here as unifying the check would require us to use a
 731                  * critical section to ensure we don't migrate while doing
 732                  * the check.  Note that if this is not the first lock, we
 733                  * are already in a critical section and are safe for the
 734                  * rest of the check.
 735                  */
 736                 if (PCPU_GET(spinlocks) == NULL)
 737                         return;
 738                 lock_list = PCPU_PTR(spinlocks);
 739         }
 740
 741         /*
 742          * Check to see if we are recursing on a lock we already own.  If
 743          * so, make sure that we don't mismatch exclusive and shared lock
 744          * acquires.
 745          */
 746         lock1 = find_instance(*lock_list, lock);
 747         if (lock1 != NULL) {
 748                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
 749                     (flags & LOP_EXCLUSIVE) == 0) {
 750                         printf("shared lock of (%s) %s @ %s:%d\n",
 751                             class->lc_name, lock->lo_name, file, line);
 752                         printf("while exclusively locked from %s:%d\n",
 753                             lock1->li_file, lock1->li_line);
 754                         panic("share->excl");
 755                 }
 756                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
 757                     (flags & LOP_EXCLUSIVE) != 0) {
 758                         printf("exclusive lock of (%s) %s @ %s:%d\n",
 759                             class->lc_name, lock->lo_name, file, line);
 760                         printf("while share locked from %s:%d\n",
 761                             lock1->li_file, lock1->li_line);
 762                         panic("excl->share");
 763                 }
 764                 return;
 765         }
 766
 767         /*
 768          * Try locks do not block if they fail to acquire the lock, thus
 769          * there is no danger of deadlocks or of switching while holding a
 770          * spin lock if we acquire a lock via a try operation.
 771          */
 772         if (flags & LOP_TRYLOCK)
 773                 return;
 774
 775         /*
 776          * Check for duplicate locks of the same type.  Note that we only
 777          * have to check for this on the last lock we just acquired.  Any
 778          * other cases will be caught as lock order violations.
 779          */
 780         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 781         w1 = lock1->li_lock->lo_witness;
 782         if (w1 == w) {
 783                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK))
 784                         return;
 785                 w->w_same_squawked = 1;
 786                 printf("acquiring duplicate lock of same type: \"%s\"\n",
 787                         lock->lo_type);
 788                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
 789                     lock1->li_file, lock1->li_line);
 790                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
 791 #ifdef KDB
 792                 goto debugger;
 793 #else
 794                 return;
 795 #endif
 796         }
 797         MPASS(!mtx_owned(&w_mtx));
 798         mtx_lock_spin(&w_mtx);
 799         /*
 800          * If we have a known higher number just say ok
 801          */
 802         if (witness_watch > 1 && w->w_level > w1->w_level) {
 803                 mtx_unlock_spin(&w_mtx);
 804                 return;
 805         }
 806         /*
 807          * If we know that the the lock we are acquiring comes after
 808          * the lock we most recently acquired in the lock order tree,
 809          * then there is no need for any further checks.
 810          */
 811         if (isitmydescendant(w1, w)) {
 812                 mtx_unlock_spin(&w_mtx);
 813                 return;
 814         }
 815         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
 816                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
 817
 818                         MPASS(j < WITNESS_COUNT);
 819                         lock1 = &lle->ll_children[i];
 820                         w1 = lock1->li_lock->lo_witness;
 821
 822                         /*
 823                          * If this lock doesn't undergo witness checking,
 824                          * then skip it.
 825                          */
 826                         if (w1 == NULL) {
 827                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
 828                                     ("lock missing witness structure"));
 829                                 continue;
 830                         }
 831                         /*
 832                          * If we are locking Giant and this is a sleepable
 833                          * lock, then skip it.
 834                          */
 835                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
 836                             lock == &Giant.mtx_object)
 837                                 continue;
 838                         /*
 839                          * If we are locking a sleepable lock and this lock
 840                          * is Giant, then skip it.
 841                          */
 842                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 843                             lock1->li_lock == &Giant.mtx_object)
 844                                 continue;
 845                         /*
 846                          * If we are locking a sleepable lock and this lock
 847                          * isn't sleepable, we want to treat it as a lock
 848                          * order violation to enfore a general lock order of
 849                          * sleepable locks before non-sleepable locks.
 850                          */
 851                         if (!((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 852                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
 853                             /*
 854                              * Check the lock order hierarchy for a reveresal.
 855                              */
 856                             if (!isitmydescendant(w, w1))
 857                                 continue;
 858                         /*
 859                          * We have a lock order violation, check to see if it
 860                          * is allowed or has already been yelled about.
 861                          */
 862                         mtx_unlock_spin(&w_mtx);
 863 #ifdef BLESSING
 864                         /*
 865                          * If the lock order is blessed, just bail.  We don't
 866                          * look for other lock order violations though, which
 867                          * may be a bug.
 868                          */
 869                         if (blessed(w, w1))
 870                                 return;
 871 #endif
 872                         if (lock1->li_lock == &Giant.mtx_object) {
 873                                 if (w1->w_Giant_squawked)
 874                                         return;
 875                                 else
 876                                         w1->w_Giant_squawked = 1;
 877                         } else {
 878                                 if (w1->w_other_squawked)
 879                                         return;
 880                                 else
 881                                         w1->w_other_squawked = 1;
 882                         }
 883                         /*
 884                          * Ok, yell about it.
 885                          */
 886                         printf("lock order reversal\n");
 887                         /*
 888                          * Try to locate an earlier lock with
 889                          * witness w in our list.
 890                          */
 891                         do {
 892                                 lock2 = &lle->ll_children[i];
 893                                 MPASS(lock2->li_lock != NULL);
 894                                 if (lock2->li_lock->lo_witness == w)
 895                                         break;
 896                                 if (i == 0 && lle->ll_next != NULL) {
 897                                         lle = lle->ll_next;
 898                                         i = lle->ll_count - 1;
 899                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
 900                                 } else
 901                                         i--;
 902                         } while (i >= 0);
 903                         if (i < 0) {
 904                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 905                                     lock1->li_lock, lock1->li_lock->lo_name,
 906                                     lock1->li_lock->lo_type, lock1->li_file,
 907                                     lock1->li_line);
 908                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
 909                                     lock->lo_name, lock->lo_type, file, line);
 910                         } else {
 911                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 912                                     lock2->li_lock, lock2->li_lock->lo_name,
 913                                     lock2->li_lock->lo_type, lock2->li_file,
 914                                     lock2->li_line);
 915                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
 916                                     lock1->li_lock, lock1->li_lock->lo_name,
 917                                     lock1->li_lock->lo_type, lock1->li_file,
 918                                     lock1->li_line);
 919                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
 920                                     lock->lo_name, lock->lo_type, file, line);
 921                         }
 922 #ifdef KDB
 923                         goto debugger;
 924 #else
 925                         return;
 926 #endif
 927                 }
 928         }
 929         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 930         /*
 931          * If requested, build a new lock order.  However, don't build a new
 932          * relationship between a sleepable lock and Giant if it is in the
 933          * wrong direction.  The correct lock order is that sleepable locks
 934          * always come before Giant.
 935          */
 936         if (flags & LOP_NEWORDER &&
 937             !(lock1->li_lock == &Giant.mtx_object &&
 938             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
 939                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 940                     lock->lo_type, lock1->li_lock->lo_type);
 941                 if (!itismychild(lock1->li_lock->lo_witness, w))
 942                         /* Witness is dead. */
 943                         return;
 944         }
 945         mtx_unlock_spin(&w_mtx);
 946         return;
 947
 948 #ifdef KDB
 949 debugger:
 950         if (witness_trace)
 951                 kdb_backtrace();
 952         if (witness_kdb)
 953                 kdb_enter(__func__);
 954 #endif
 955 }
 956
 957 void
 958 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
 959 {
 960         struct lock_list_entry **lock_list, *lle;
 961         struct lock_instance *instance;
 962         struct witness *w;
 963         struct thread *td;
 964
 965         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 966             panicstr != NULL)
 967                 return;
 968         w = lock->lo_witness;
 969         td = curthread;
 970         file = fixup_filename(file);
 971
 972         /* Determine lock list for this lock. */
 973         if (lock->lo_class->lc_flags & LC_SLEEPLOCK)
 974                 lock_list = &td->td_sleeplocks;
 975         else
 976                 lock_list = PCPU_PTR(spinlocks);
 977
 978         /* Check to see if we are recursing on a lock we already own. */
 979         instance = find_instance(*lock_list, lock);
 980         if (instance != NULL) {
 981                 instance->li_flags++;
 982                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
 983                     td->td_proc->p_pid, lock->lo_name,
 984                     instance->li_flags & LI_RECURSEMASK);
 985                 instance->li_file = file;
 986                 instance->li_line = line;
 987                 return;
 988         }
 989
 990         /* Update per-witness last file and line acquire. */
 991         w->w_file = file;
 992         w->w_line = line;
 993
 994         /* Find the next open lock instance in the list and fill it. */
 995         lle = *lock_list;
 996         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
 997                 lle = witness_lock_list_get();
 998                 if (lle == NULL)
 999                         return;
1000                 lle->ll_next = *lock_list;
1001                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1002                     td->td_proc->p_pid, lle);
1003                 *lock_list = lle;
1004         }
1005         instance = &lle->ll_children[lle->ll_count++];
1006         instance->li_lock = lock;
1007         instance->li_line = line;
1008         instance->li_file = file;
1009         if ((flags & LOP_EXCLUSIVE) != 0)
1010                 instance->li_flags = LI_EXCLUSIVE;
1011         else
1012                 instance->li_flags = 0;
1013         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1014             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1015 }
1016
1017 void
1018 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1019 {
1020         struct lock_instance *instance;
1021         struct lock_class *class;
1022
1023         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1024         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1025                 return;
1026         class = lock->lo_class;
1027         file = fixup_filename(file);
1028         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1029                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1030                     class->lc_name, lock->lo_name, file, line);
1031         if ((flags & LOP_TRYLOCK) == 0)
1032                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1033                     lock->lo_name, file, line);
1034         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1035                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1036                     class->lc_name, lock->lo_name, file, line);
1037         instance = find_instance(curthread->td_sleeplocks, lock);
1038         if (instance == NULL)
1039                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1040                     class->lc_name, lock->lo_name, file, line);
1041         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1042                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1043                     class->lc_name, lock->lo_name, file, line);
1044         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1045                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1046                     class->lc_name, lock->lo_name,
1047                     instance->li_flags & LI_RECURSEMASK, file, line);
1048         instance->li_flags |= LI_EXCLUSIVE;
1049 }
1050
1051 void
1052 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1053     int line)
1054 {
1055         struct lock_instance *instance;
1056         struct lock_class *class;
1057
1058         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1059         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1060                 return;
1061         class = lock->lo_class;
1062         file = fixup_filename(file);
1063         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1064                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1065                     class->lc_name, lock->lo_name, file, line);
1066         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1067                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1068                     class->lc_name, lock->lo_name, file, line);
1069         instance = find_instance(curthread->td_sleeplocks, lock);
1070         if (instance == NULL)
1071                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1072                     class->lc_name, lock->lo_name, file, line);
1073         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1074                 panic("downgrade of shared lock (%s) %s @ %s:%d",
1075                     class->lc_name, lock->lo_name, file, line);
1076         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1077                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1078                     class->lc_name, lock->lo_name,
1079                     instance->li_flags & LI_RECURSEMASK, file, line);
1080         instance->li_flags &= ~LI_EXCLUSIVE;
1081 }
1082
1083 void
1084 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1085 {
1086         struct lock_list_entry **lock_list, *lle;
1087         struct lock_instance *instance;
1088         struct lock_class *class;
1089         struct thread *td;
1090         register_t s;
1091         int i, j;
1092
1093         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1094             panicstr != NULL)
1095                 return;
1096         td = curthread;
1097         class = lock->lo_class;
1098         file = fixup_filename(file);
1099
1100         /* Find lock instance associated with this lock. */
1101         if (class->lc_flags & LC_SLEEPLOCK)
1102                 lock_list = &td->td_sleeplocks;
1103         else
1104                 lock_list = PCPU_PTR(spinlocks);
1105         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1106                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1107                         instance = &(*lock_list)->ll_children[i];
1108                         if (instance->li_lock == lock)
1109                                 goto found;
1110                 }
1111         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1112             file, line);
1113 found:
1114
1115         /* First, check for shared/exclusive mismatches. */
1116         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1117             (flags & LOP_EXCLUSIVE) == 0) {
1118                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1119                     lock->lo_name, file, line);
1120                 printf("while exclusively locked from %s:%d\n",
1121                     instance->li_file, instance->li_line);
1122                 panic("excl->ushare");
1123         }
1124         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1125             (flags & LOP_EXCLUSIVE) != 0) {
1126                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1127                     lock->lo_name, file, line);
1128                 printf("while share locked from %s:%d\n", instance->li_file,
1129                     instance->li_line);
1130                 panic("share->uexcl");
1131         }
1132
1133         /* If we are recursed, unrecurse. */
1134         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1135                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1136                     td->td_proc->p_pid, instance->li_lock->lo_name,
1137                     instance->li_flags);
1138                 instance->li_flags--;
1139                 return;
1140         }
1141
1142         /* Otherwise, remove this item from the list. */
1143         s = intr_disable();
1144         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1145             td->td_proc->p_pid, instance->li_lock->lo_name,
1146             (*lock_list)->ll_count - 1);
1147         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1148                 (*lock_list)->ll_children[j] =
1149                     (*lock_list)->ll_children[j + 1];
1150         (*lock_list)->ll_count--;
1151         intr_restore(s);
1152
1153         /* If this lock list entry is now empty, free it. */
1154         if ((*lock_list)->ll_count == 0) {
1155                 lle = *lock_list;
1156                 *lock_list = lle->ll_next;
1157                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1158                     td->td_proc->p_pid, lle);
1159                 witness_lock_list_free(lle);
1160         }
1161 }
1162
1163 /*
1164  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1165  * exempt Giant and sleepable locks from the checks as well.  If any
1166  * non-exempt locks are held, then a supplied message is printed to the
1167  * console along with a list of the offending locks.  If indicated in the
1168  * flags then a failure results in a panic as well.
1169  */
1170 int
1171 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1172 {
1173         struct lock_list_entry *lle;
1174         struct lock_instance *lock1;
1175         struct thread *td;
1176         va_list ap;
1177         int i, n;
1178
1179         if (witness_cold || witness_watch == 0 || panicstr != NULL)
1180                 return (0);
1181         n = 0;
1182         td = curthread;
1183         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1184                 for (i = lle->ll_count - 1; i >= 0; i--) {
1185                         lock1 = &lle->ll_children[i];
1186                         if (lock1->li_lock == lock)
1187                                 continue;
1188                         if (flags & WARN_GIANTOK &&
1189                             lock1->li_lock == &Giant.mtx_object)
1190                                 continue;
1191                         if (flags & WARN_SLEEPOK &&
1192                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1193                                 continue;
1194                         if (n == 0) {
1195                                 va_start(ap, fmt);
1196                                 vprintf(fmt, ap);
1197                                 va_end(ap);
1198                                 printf(" with the following");
1199                                 if (flags & WARN_SLEEPOK)
1200                                         printf(" non-sleepable");
1201                                 printf(" locks held:\n");
1202                         }
1203                         n++;
1204                         witness_list_lock(lock1);
1205                 }
1206         if (PCPU_GET(spinlocks) != NULL) {
1207                 /*
1208                  * Since we already hold a spinlock preemption is
1209                  * already blocked.
1210                  */
1211                 if (n == 0) {
1212                         va_start(ap, fmt);
1213                         vprintf(fmt, ap);
1214                         va_end(ap);
1215                         printf(" with the following");
1216                         if (flags & WARN_SLEEPOK)
1217                                 printf(" non-sleepable");
1218                         printf(" locks held:\n");
1219                 }
1220                 n += witness_list_locks(PCPU_PTR(spinlocks));
1221         }
1222         if (flags & WARN_PANIC && n)
1223                 panic("witness_warn");
1224 #ifdef KDB
1225         else if (witness_kdb && n)
1226                 kdb_enter(__func__);
1227         else if (witness_trace && n)
1228                 kdb_backtrace();
1229 #endif
1230         return (n);
1231 }
1232
1233 const char *
1234 witness_file(struct lock_object *lock)
1235 {
1236         struct witness *w;
1237
1238         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1239                 return ("?");
1240         w = lock->lo_witness;
1241         return (w->w_file);
1242 }
1243
1244 int
1245 witness_line(struct lock_object *lock)
1246 {
1247         struct witness *w;
1248
1249         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1250                 return (0);
1251         w = lock->lo_witness;
1252         return (w->w_line);
1253 }
1254
1255 static struct witness *
1256 enroll(const char *description, struct lock_class *lock_class)
1257 {
1258         struct witness *w;
1259
1260         if (witness_watch == 0 || panicstr != NULL)
1261                 return (NULL);
1262         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1263                 return (NULL);
1264         mtx_lock_spin(&w_mtx);
1265         STAILQ_FOREACH(w, &w_all, w_list) {
1266                 if (w->w_name == description || (w->w_refcount > 0 &&
1267                     strcmp(description, w->w_name) == 0)) {
1268                         w->w_refcount++;
1269                         mtx_unlock_spin(&w_mtx);
1270                         if (lock_class != w->w_class)
1271                                 panic(
1272                                 "lock (%s) %s does not match earlier (%s) lock",
1273                                     description, lock_class->lc_name,
1274                                     w->w_class->lc_name);
1275                         return (w);
1276                 }
1277         }
1278         /*
1279          * This isn't quite right, as witness_cold is still 0 while we
1280          * enroll all the locks initialized before witness_initialize().
1281          */
1282         if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) {
1283                 mtx_unlock_spin(&w_mtx);
1284                 panic("spin lock %s not in order list", description);
1285         }
1286         if ((w = witness_get()) == NULL)
1287                 return (NULL);
1288         w->w_name = description;
1289         w->w_class = lock_class;
1290         w->w_refcount = 1;
1291         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1292         if (lock_class->lc_flags & LC_SPINLOCK)
1293                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1294         else if (lock_class->lc_flags & LC_SLEEPLOCK)
1295                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1296         else {
1297                 mtx_unlock_spin(&w_mtx);
1298                 panic("lock class %s is not sleep or spin",
1299                     lock_class->lc_name);
1300         }
1301         mtx_unlock_spin(&w_mtx);
1302         return (w);
1303 }
1304
1305 /* Don't let the door bang you on the way out... */
1306 static int
1307 depart(struct witness *w)
1308 {
1309         struct witness_child_list_entry *wcl, *nwcl;
1310         struct witness_list *list;
1311         struct witness *parent;
1312
1313         MPASS(w->w_refcount == 0);
1314         if (w->w_class->lc_flags & LC_SLEEPLOCK)
1315                 list = &w_sleep;
1316         else
1317                 list = &w_spin;
1318         /*
1319          * First, we run through the entire tree looking for any
1320          * witnesses that the outgoing witness is a child of.  For
1321          * each parent that we find, we reparent all the direct
1322          * children of the outgoing witness to its parent.
1323          */
1324         STAILQ_FOREACH(parent, list, w_typelist) {
1325                 if (!isitmychild(parent, w))
1326                         continue;
1327                 removechild(parent, w);
1328                 if (!reparentchildren(parent, w))
1329                         return (0);
1330         }
1331
1332         /*
1333          * Now we go through and free up the child list of the
1334          * outgoing witness.
1335          */
1336         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1337                 nwcl = wcl->wcl_next;
1338                 witness_child_free(wcl);
1339         }
1340
1341         /*
1342          * Detach from various lists and free.
1343          */
1344         STAILQ_REMOVE(list, w, witness, w_typelist);
1345         STAILQ_REMOVE(&w_all, w, witness, w_list);
1346         witness_free(w);
1347
1348         /* Finally, fixup the tree. */
1349         return (rebalancetree(list));
1350 }
1351
1352 /*
1353  * Prune an entire lock order tree.  We look for cases where a lock
1354  * is now both a descendant and a direct child of a given lock.  In
1355  * that case, we want to remove the direct child link from the tree.
1356  *
1357  * Returns false if insertchild() fails.
1358  */
1359 static int
1360 rebalancetree(struct witness_list *list)
1361 {
1362         struct witness *child, *parent;
1363
1364         STAILQ_FOREACH(child, list, w_typelist) {
1365                 STAILQ_FOREACH(parent, list, w_typelist) {
1366                         if (!isitmychild(parent, child))
1367                                 continue;
1368                         removechild(parent, child);
1369                         if (isitmydescendant(parent, child))
1370                                 continue;
1371                         if (!insertchild(parent, child))
1372                                 return (0);
1373                 }
1374         }
1375         witness_levelall();
1376         return (1);
1377 }
1378
1379 /*
1380  * Add "child" as a direct child of "parent".  Returns false if
1381  * we fail due to out of memory.
1382  */
1383 static int
1384 insertchild(struct witness *parent, struct witness *child)
1385 {
1386         struct witness_child_list_entry **wcl;
1387
1388         MPASS(child != NULL && parent != NULL);
1389
1390         /*
1391          * Insert "child" after "parent"
1392          */
1393         wcl = &parent->w_children;
1394         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1395                 wcl = &(*wcl)->wcl_next;
1396         if (*wcl == NULL) {
1397                 *wcl = witness_child_get();
1398                 if (*wcl == NULL)
1399                         return (0);
1400         }
1401         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1402
1403         return (1);
1404 }
1405
1406 /*
1407  * Make all the direct descendants of oldparent be direct descendants
1408  * of newparent.
1409  */
1410 static int
1411 reparentchildren(struct witness *newparent, struct witness *oldparent)
1412 {
1413         struct witness_child_list_entry *wcl;
1414         int i;
1415
1416         /* Avoid making a witness a child of itself. */
1417         MPASS(!isitmychild(oldparent, newparent));
1418
1419         for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1420                 for (i = 0; i < wcl->wcl_count; i++)
1421                         if (!insertchild(newparent, wcl->wcl_children[i]))
1422                                 return (0);
1423         return (1);
1424 }
1425
1426 static int
1427 itismychild(struct witness *parent, struct witness *child)
1428 {
1429         struct witness_list *list;
1430
1431         MPASS(child != NULL && parent != NULL);
1432         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1433             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1434                 panic(
1435                 "%s: parent (%s) and child (%s) are not the same lock type",
1436                     __func__, parent->w_class->lc_name,
1437                     child->w_class->lc_name);
1438
1439         if (!insertchild(parent, child))
1440                 return (0);
1441
1442         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1443                 list = &w_sleep;
1444         else
1445                 list = &w_spin;
1446         return (rebalancetree(list));
1447 }
1448
1449 static void
1450 removechild(struct witness *parent, struct witness *child)
1451 {
1452         struct witness_child_list_entry **wcl, *wcl1;
1453         int i;
1454
1455         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1456                 for (i = 0; i < (*wcl)->wcl_count; i++)
1457                         if ((*wcl)->wcl_children[i] == child)
1458                                 goto found;
1459         return;
1460 found:
1461         (*wcl)->wcl_count--;
1462         if ((*wcl)->wcl_count > i)
1463                 (*wcl)->wcl_children[i] =
1464                     (*wcl)->wcl_children[(*wcl)->wcl_count];
1465         MPASS((*wcl)->wcl_children[i] != NULL);
1466         if ((*wcl)->wcl_count != 0)
1467                 return;
1468         wcl1 = *wcl;
1469         *wcl = wcl1->wcl_next;
1470         witness_child_free(wcl1);
1471 }
1472
1473 static int
1474 isitmychild(struct witness *parent, struct witness *child)
1475 {
1476         struct witness_child_list_entry *wcl;
1477         int i;
1478
1479         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1480                 for (i = 0; i < wcl->wcl_count; i++) {
1481                         if (wcl->wcl_children[i] == child)
1482                                 return (1);
1483                 }
1484         }
1485         return (0);
1486 }
1487
1488 static int
1489 isitmydescendant(struct witness *parent, struct witness *child)
1490 {
1491         struct witness_child_list_entry *wcl;
1492         int i, j;
1493
1494         if (isitmychild(parent, child))
1495                 return (1);
1496         j = 0;
1497         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1498                 MPASS(j < 1000);
1499                 for (i = 0; i < wcl->wcl_count; i++) {
1500                         if (isitmydescendant(wcl->wcl_children[i], child))
1501                                 return (1);
1502                 }
1503                 j++;
1504         }
1505         return (0);
1506 }
1507
1508 static void
1509 witness_levelall (void)
1510 {
1511         struct witness_list *list;
1512         struct witness *w, *w1;
1513
1514         /*
1515          * First clear all levels.
1516          */
1517         STAILQ_FOREACH(w, &w_all, w_list) {
1518                 w->w_level = 0;
1519         }
1520
1521         /*
1522          * Look for locks with no parent and level all their descendants.
1523          */
1524         STAILQ_FOREACH(w, &w_all, w_list) {
1525                 /*
1526                  * This is just an optimization, technically we could get
1527                  * away just walking the all list each time.
1528                  */
1529                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
1530                         list = &w_sleep;
1531                 else
1532                         list = &w_spin;
1533                 STAILQ_FOREACH(w1, list, w_typelist) {
1534                         if (isitmychild(w1, w))
1535                                 goto skip;
1536                 }
1537                 witness_leveldescendents(w, 0);
1538         skip:
1539                 ;       /* silence GCC 3.x */
1540         }
1541 }
1542
1543 static void
1544 witness_leveldescendents(struct witness *parent, int level)
1545 {
1546         struct witness_child_list_entry *wcl;
1547         int i;
1548
1549         if (parent->w_level < level)
1550                 parent->w_level = level;
1551         level++;
1552         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1553                 for (i = 0; i < wcl->wcl_count; i++)
1554                         witness_leveldescendents(wcl->wcl_children[i], level);
1555 }
1556
1557 static void
1558 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
1559                            struct witness *parent, int indent)
1560 {
1561         struct witness_child_list_entry *wcl;
1562         int i, level;
1563
1564         level = parent->w_level;
1565         prnt("%-2d", level);
1566         for (i = 0; i < indent; i++)
1567                 prnt(" ");
1568         if (parent->w_refcount > 0)
1569                 prnt("%s", parent->w_name);
1570         else
1571                 prnt("(dead)");
1572         if (parent->w_displayed) {
1573                 prnt(" -- (already displayed)\n");
1574                 return;
1575         }
1576         parent->w_displayed = 1;
1577         if (parent->w_refcount > 0) {
1578                 if (parent->w_file != NULL)
1579                         prnt(" -- last acquired @ %s:%d", parent->w_file,
1580                             parent->w_line);
1581         }
1582         prnt("\n");
1583         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
1584                 for (i = 0; i < wcl->wcl_count; i++)
1585                             witness_displaydescendants(prnt,
1586                                 wcl->wcl_children[i], indent + 1);
1587 }
1588
1589 #ifdef BLESSING
1590 static int
1591 blessed(struct witness *w1, struct witness *w2)
1592 {
1593         int i;
1594         struct witness_blessed *b;
1595
1596         for (i = 0; i < blessed_count; i++) {
1597                 b = &blessed_list[i];
1598                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1599                         if (strcmp(w2->w_name, b->b_lock2) == 0)
1600                                 return (1);
1601                         continue;
1602                 }
1603                 if (strcmp(w1->w_name, b->b_lock2) == 0)
1604                         if (strcmp(w2->w_name, b->b_lock1) == 0)
1605                                 return (1);
1606         }
1607         return (0);
1608 }
1609 #endif
1610
1611 static struct witness *
1612 witness_get(void)
1613 {
1614         struct witness *w;
1615
1616         if (witness_watch == 0) {
1617                 mtx_unlock_spin(&w_mtx);
1618                 return (NULL);
1619         }
1620         if (STAILQ_EMPTY(&w_free)) {
1621                 witness_watch = 0;
1622                 mtx_unlock_spin(&w_mtx);
1623                 printf("%s: witness exhausted\n", __func__);
1624                 return (NULL);
1625         }
1626         w = STAILQ_FIRST(&w_free);
1627         STAILQ_REMOVE_HEAD(&w_free, w_list);
1628         bzero(w, sizeof(*w));
1629         return (w);
1630 }
1631
1632 static void
1633 witness_free(struct witness *w)
1634 {
1635
1636         STAILQ_INSERT_HEAD(&w_free, w, w_list);
1637 }
1638
1639 static struct witness_child_list_entry *
1640 witness_child_get(void)
1641 {
1642         struct witness_child_list_entry *wcl;
1643
1644         if (witness_watch == 0) {
1645                 mtx_unlock_spin(&w_mtx);
1646                 return (NULL);
1647         }
1648         wcl = w_child_free;
1649         if (wcl == NULL) {
1650                 witness_watch = 0;
1651                 mtx_unlock_spin(&w_mtx);
1652                 printf("%s: witness exhausted\n", __func__);
1653                 return (NULL);
1654         }
1655         w_child_free = wcl->wcl_next;
1656         bzero(wcl, sizeof(*wcl));
1657         return (wcl);
1658 }
1659
1660 static void
1661 witness_child_free(struct witness_child_list_entry *wcl)
1662 {
1663
1664         wcl->wcl_next = w_child_free;
1665         w_child_free = wcl;
1666 }
1667
1668 static struct lock_list_entry *
1669 witness_lock_list_get(void)
1670 {
1671         struct lock_list_entry *lle;
1672
1673         if (witness_watch == 0)
1674                 return (NULL);
1675         mtx_lock_spin(&w_mtx);
1676         lle = w_lock_list_free;
1677         if (lle == NULL) {
1678                 witness_watch = 0;
1679                 mtx_unlock_spin(&w_mtx);
1680                 printf("%s: witness exhausted\n", __func__);
1681                 return (NULL);
1682         }
1683         w_lock_list_free = lle->ll_next;
1684         mtx_unlock_spin(&w_mtx);
1685         bzero(lle, sizeof(*lle));
1686         return (lle);
1687 }
1688
1689 static void
1690 witness_lock_list_free(struct lock_list_entry *lle)
1691 {
1692
1693         mtx_lock_spin(&w_mtx);
1694         lle->ll_next = w_lock_list_free;
1695         w_lock_list_free = lle;
1696         mtx_unlock_spin(&w_mtx);
1697 }
1698
1699 static struct lock_instance *
1700 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1701 {
1702         struct lock_list_entry *lle;
1703         struct lock_instance *instance;
1704         int i;
1705
1706         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1707                 for (i = lle->ll_count - 1; i >= 0; i--) {
1708                         instance = &lle->ll_children[i];
1709                         if (instance->li_lock == lock)
1710                                 return (instance);
1711                 }
1712         return (NULL);
1713 }
1714
1715 static void
1716 witness_list_lock(struct lock_instance *instance)
1717 {
1718         struct lock_object *lock;
1719
1720         lock = instance->li_lock;
1721         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1722             "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
1723         if (lock->lo_type != lock->lo_name)
1724                 printf(" (%s)", lock->lo_type);
1725         printf(" r = %d (%p) locked @ %s:%d\n",
1726             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1727             instance->li_line);
1728 }
1729
1730 #ifdef DDB
1731 static int
1732 witness_thread_has_locks(struct thread *td)
1733 {
1734
1735         return (td->td_sleeplocks != NULL);
1736 }
1737
1738 static int
1739 witness_proc_has_locks(struct proc *p)
1740 {
1741         struct thread *td;
1742
1743         FOREACH_THREAD_IN_PROC(p, td) {
1744                 if (witness_thread_has_locks(td))
1745                         return (1);
1746         }
1747         return (0);
1748 }
1749 #endif
1750
1751 int
1752 witness_list_locks(struct lock_list_entry **lock_list)
1753 {
1754         struct lock_list_entry *lle;
1755         int i, nheld;
1756
1757         nheld = 0;
1758         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1759                 for (i = lle->ll_count - 1; i >= 0; i--) {
1760                         witness_list_lock(&lle->ll_children[i]);
1761                         nheld++;
1762                 }
1763         return (nheld);
1764 }
1765
1766 /*
1767  * This is a bit risky at best.  We call this function when we have timed
1768  * out acquiring a spin lock, and we assume that the other CPU is stuck
1769  * with this lock held.  So, we go groveling around in the other CPU's
1770  * per-cpu data to try to find the lock instance for this spin lock to
1771  * see when it was last acquired.
1772  */
1773 void
1774 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1775 {
1776         struct lock_instance *instance;
1777         struct pcpu *pc;
1778
1779         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1780                 return;
1781         pc = pcpu_find(owner->td_oncpu);
1782         instance = find_instance(pc->pc_spinlocks, lock);
1783         if (instance != NULL)
1784                 witness_list_lock(instance);
1785 }
1786
1787 void
1788 witness_save(struct lock_object *lock, const char **filep, int *linep)
1789 {
1790         struct lock_instance *instance;
1791
1792         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1793         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1794                 return;
1795         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1796                 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1797                     lock->lo_class->lc_name, lock->lo_name);
1798         instance = find_instance(curthread->td_sleeplocks, lock);
1799         if (instance == NULL)
1800                 panic("%s: lock (%s) %s not locked", __func__,
1801                     lock->lo_class->lc_name, lock->lo_name);
1802         *filep = instance->li_file;
1803         *linep = instance->li_line;
1804 }
1805
1806 void
1807 witness_restore(struct lock_object *lock, const char *file, int line)
1808 {
1809         struct lock_instance *instance;
1810
1811         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1812         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1813                 return;
1814         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
1815                 panic("%s: lock (%s) %s is not a sleep lock", __func__,
1816                     lock->lo_class->lc_name, lock->lo_name);
1817         instance = find_instance(curthread->td_sleeplocks, lock);
1818         if (instance == NULL)
1819                 panic("%s: lock (%s) %s not locked", __func__,
1820                     lock->lo_class->lc_name, lock->lo_name);
1821         lock->lo_witness->w_file = file;
1822         lock->lo_witness->w_line = line;
1823         instance->li_file = file;
1824         instance->li_line = line;
1825 }
1826
1827 void
1828 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1829 {
1830 #ifdef INVARIANT_SUPPORT
1831         struct lock_instance *instance;
1832
1833         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1834                 return;
1835         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
1836                 instance = find_instance(curthread->td_sleeplocks, lock);
1837         else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
1838                 instance = find_instance(PCPU_GET(spinlocks), lock);
1839         else {
1840                 panic("Lock (%s) %s is not sleep or spin!",
1841                     lock->lo_class->lc_name, lock->lo_name);
1842         }
1843         file = fixup_filename(file);
1844         switch (flags) {
1845         case LA_UNLOCKED:
1846                 if (instance != NULL)
1847                         panic("Lock (%s) %s locked @ %s:%d.",
1848                             lock->lo_class->lc_name, lock->lo_name, file, line);
1849                 break;
1850         case LA_LOCKED:
1851         case LA_LOCKED | LA_RECURSED:
1852         case LA_LOCKED | LA_NOTRECURSED:
1853         case LA_SLOCKED:
1854         case LA_SLOCKED | LA_RECURSED:
1855         case LA_SLOCKED | LA_NOTRECURSED:
1856         case LA_XLOCKED:
1857         case LA_XLOCKED | LA_RECURSED:
1858         case LA_XLOCKED | LA_NOTRECURSED:
1859                 if (instance == NULL) {
1860                         panic("Lock (%s) %s not locked @ %s:%d.",
1861                             lock->lo_class->lc_name, lock->lo_name, file, line);
1862                         break;
1863                 }
1864                 if ((flags & LA_XLOCKED) != 0 &&
1865                     (instance->li_flags & LI_EXCLUSIVE) == 0)
1866                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1867                             lock->lo_class->lc_name, lock->lo_name, file, line);
1868                 if ((flags & LA_SLOCKED) != 0 &&
1869                     (instance->li_flags & LI_EXCLUSIVE) != 0)
1870                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
1871                             lock->lo_class->lc_name, lock->lo_name, file, line);
1872                 if ((flags & LA_RECURSED) != 0 &&
1873                     (instance->li_flags & LI_RECURSEMASK) == 0)
1874                         panic("Lock (%s) %s not recursed @ %s:%d.",
1875                             lock->lo_class->lc_name, lock->lo_name, file, line);
1876                 if ((flags & LA_NOTRECURSED) != 0 &&
1877                     (instance->li_flags & LI_RECURSEMASK) != 0)
1878                         panic("Lock (%s) %s recursed @ %s:%d.",
1879                             lock->lo_class->lc_name, lock->lo_name, file, line);
1880                 break;
1881         default:
1882                 panic("Invalid lock assertion at %s:%d.", file, line);
1883
1884         }
1885 #endif  /* INVARIANT_SUPPORT */
1886 }
1887
1888 #ifdef DDB
1889 static void
1890 witness_list(struct thread *td)
1891 {
1892
1893         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1894         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1895
1896         if (witness_watch == 0)
1897                 return;
1898
1899         witness_list_locks(&td->td_sleeplocks);
1900
1901         /*
1902          * We only handle spinlocks if td == curthread.  This is somewhat broken
1903          * if td is currently executing on some other CPU and holds spin locks
1904          * as we won't display those locks.  If we had a MI way of getting
1905          * the per-cpu data for a given cpu then we could use
1906          * td->td_oncpu to get the list of spinlocks for this thread
1907          * and "fix" this.
1908          *
1909          * That still wouldn't really fix this unless we locked sched_lock
1910          * or stopped the other CPU to make sure it wasn't changing the list
1911          * out from under us.  It is probably best to just not try to handle
1912          * threads on other CPU's for now.
1913          */
1914         if (td == curthread && PCPU_GET(spinlocks) != NULL)
1915                 witness_list_locks(PCPU_PTR(spinlocks));
1916 }
1917
1918 DB_SHOW_COMMAND(locks, db_witness_list)
1919 {
1920         struct thread *td;
1921         pid_t pid;
1922         struct proc *p;
1923
1924         if (have_addr) {
1925                 pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
1926                     ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
1927                     ((addr >> 16) % 16) * 10000;
1928                 /* sx_slock(&allproc_lock); */
1929                 FOREACH_PROC_IN_SYSTEM(p) {
1930                         if (p->p_pid == pid)
1931                                 break;
1932                 }
1933                 /* sx_sunlock(&allproc_lock); */
1934                 if (p == NULL) {
1935                         db_printf("pid %d not found\n", pid);
1936                         return;
1937                 }
1938                 FOREACH_THREAD_IN_PROC(p, td) {
1939                         witness_list(td);
1940                 }
1941         } else {
1942                 td = curthread;
1943                 witness_list(td);
1944         }
1945 }
1946
1947 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
1948 {
1949         struct thread *td;
1950         struct proc *p;
1951
1952         /*
1953          * It would be nice to list only threads and processes that actually
1954          * held sleep locks, but that information is currently not exported
1955          * by WITNESS.
1956          */
1957         FOREACH_PROC_IN_SYSTEM(p) {
1958                 if (!witness_proc_has_locks(p))
1959                         continue;
1960                 FOREACH_THREAD_IN_PROC(p, td) {
1961                         if (!witness_thread_has_locks(td))
1962                                 continue;
1963                         printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
1964                             p->p_comm, td, td->td_tid);
1965                         witness_list(td);
1966                 }
1967         }
1968 }
1969
1970 DB_SHOW_COMMAND(witness, db_witness_display)
1971 {
1972
1973         witness_display(db_printf);
1974 }
1975 #endif