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_hwpmc_hooks.h"
  89 #include "opt_witness.h"
  90
  91 #include <sys/param.h>
  92 #include <sys/bus.h>
  93 #include <sys/kdb.h>
  94 #include <sys/kernel.h>
  95 #include <sys/ktr.h>
  96 #include <sys/lock.h>
  97 #include <sys/malloc.h>
  98 #include <sys/mutex.h>
  99 #include <sys/priv.h>
 100 #include <sys/proc.h>
 101 #include <sys/sysctl.h>
 102 #include <sys/systm.h>
 103
 104 #include <ddb/ddb.h>
 105
 106 #include <machine/stdarg.h>
 107
 108 /* Note that these traces do not work with KTR_ALQ. */
 109 #if 0
 110 #define KTR_WITNESS     KTR_SUBSYS
 111 #else
 112 #define KTR_WITNESS     0
 113 #endif
 114
 115 /* Easier to stay with the old names. */
 116 #define lo_list         lo_witness_data.lod_list
 117 #define lo_witness      lo_witness_data.lod_witness
 118
 119 /* Define this to check for blessed mutexes */
 120 #undef BLESSING
 121
 122 #define WITNESS_COUNT 1024
 123 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
 124 /*
 125  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
 126  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
 127  * probably be safe for the most part, but it's still a SWAG.
 128  */
 129 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
 130
 131 #define WITNESS_NCHILDREN 6
 132
 133 struct witness_child_list_entry;
 134
 135 struct witness {
 136         const   char *w_name;
 137         struct  lock_class *w_class;
 138         STAILQ_ENTRY(witness) w_list;           /* List of all witnesses. */
 139         STAILQ_ENTRY(witness) w_typelist;       /* Witnesses of a type. */
 140         struct  witness_child_list_entry *w_children;   /* Great evilness... */
 141         const   char *w_file;
 142         int     w_line;
 143         u_int   w_level;
 144         u_int   w_refcount;
 145         u_char  w_Giant_squawked:1;
 146         u_char  w_other_squawked:1;
 147         u_char  w_same_squawked:1;
 148         u_char  w_displayed:1;
 149 };
 150
 151 struct witness_child_list_entry {
 152         struct  witness_child_list_entry *wcl_next;
 153         struct  witness *wcl_children[WITNESS_NCHILDREN];
 154         u_int   wcl_count;
 155 };
 156
 157 STAILQ_HEAD(witness_list, witness);
 158
 159 #ifdef BLESSING
 160 struct witness_blessed {
 161         const   char *b_lock1;
 162         const   char *b_lock2;
 163 };
 164 #endif
 165
 166 struct witness_order_list_entry {
 167         const   char *w_name;
 168         struct  lock_class *w_class;
 169 };
 170
 171 #ifdef BLESSING
 172 static int      blessed(struct witness *, struct witness *);
 173 #endif
 174 static int      depart(struct witness *w);
 175 static struct   witness *enroll(const char *description,
 176                                 struct lock_class *lock_class);
 177 static int      insertchild(struct witness *parent, struct witness *child);
 178 static int      isitmychild(struct witness *parent, struct witness *child);
 179 static int      isitmydescendant(struct witness *parent, struct witness *child);
 180 static int      itismychild(struct witness *parent, struct witness *child);
 181 static void     removechild(struct witness *parent, struct witness *child);
 182 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
 183 static const char *fixup_filename(const char *file);
 184 static struct   witness *witness_get(void);
 185 static void     witness_free(struct witness *m);
 186 static struct   witness_child_list_entry *witness_child_get(void);
 187 static void     witness_child_free(struct witness_child_list_entry *wcl);
 188 static struct   lock_list_entry *witness_lock_list_get(void);
 189 static void     witness_lock_list_free(struct lock_list_entry *lle);
 190 static struct   lock_instance *find_instance(struct lock_list_entry *lock_list,
 191                                              struct lock_object *lock);
 192 static void     witness_list_lock(struct lock_instance *instance);
 193 #ifdef DDB
 194 static void     witness_leveldescendents(struct witness *parent, int level);
 195 static void     witness_levelall(void);
 196 static void     witness_displaydescendants(void(*)(const char *fmt, ...),
 197                                            struct witness *, int indent);
 198 static void     witness_display_list(void(*prnt)(const char *fmt, ...),
 199                                      struct witness_list *list);
 200 static void     witness_display(void(*)(const char *fmt, ...));
 201 static void     witness_list(struct thread *td);
 202 #endif
 203
 204 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
 205
 206 /*
 207  * If set to 0, witness is disabled.  If set to a non-zero value, witness
 208  * performs full lock order checking for all locks.  At runtime, this
 209  * value may be set to 0 to turn off witness.  witness is not allowed be
 210  * turned on once it is turned off, however.
 211  */
 212 static int witness_watch = 1;
 213 TUNABLE_INT("debug.witness.watch", &witness_watch);
 214 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
 215     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
 216
 217 #ifdef KDB
 218 /*
 219  * When KDB is enabled and witness_kdb is set to 1, it will cause the system
 220  * to drop into kdebug() when:
 221  *      - a lock hierarchy violation occurs
 222  *      - locks are held when going to sleep.
 223  */
 224 #ifdef WITNESS_KDB
 225 int     witness_kdb = 1;
 226 #else
 227 int     witness_kdb = 0;
 228 #endif
 229 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
 230 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
 231
 232 /*
 233  * When KDB is enabled and witness_trace is set to 1, it will cause the system
 234  * to print a stack trace:
 235  *      - a lock hierarchy violation occurs
 236  *      - locks are held when going to sleep.
 237  */
 238 int     witness_trace = 1;
 239 TUNABLE_INT("debug.witness.trace", &witness_trace);
 240 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
 241 #endif /* KDB */
 242
 243 #ifdef WITNESS_SKIPSPIN
 244 int     witness_skipspin = 1;
 245 #else
 246 int     witness_skipspin = 0;
 247 #endif
 248 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
 249 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
 250     &witness_skipspin, 0, "");
 251
 252 static struct mtx w_mtx;
 253 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 254 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 255 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 256 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 257 static struct witness_child_list_entry *w_child_free = NULL;
 258 static struct lock_list_entry *w_lock_list_free = NULL;
 259
 260 static int w_free_cnt, w_spin_cnt, w_sleep_cnt, w_child_free_cnt, w_child_cnt;
 261 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
 262 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
 263 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
 264     "");
 265 SYSCTL_INT(_debug_witness, OID_AUTO, child_free_cnt, CTLFLAG_RD,
 266     &w_child_free_cnt, 0, "");
 267 SYSCTL_INT(_debug_witness, OID_AUTO, child_cnt, CTLFLAG_RD, &w_child_cnt, 0,
 268     "");
 269
 270 static struct witness w_data[WITNESS_COUNT];
 271 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
 272 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 273
 274 static struct witness_order_list_entry order_lists[] = {
 275         /*
 276          * sx locks
 277          */
 278         { "proctree", &lock_class_sx },
 279         { "allproc", &lock_class_sx },
 280         { "allprison", &lock_class_sx },
 281         { NULL, NULL },
 282         /*
 283          * Various mutexes
 284          */
 285         { "Giant", &lock_class_mtx_sleep },
 286         { "pipe mutex", &lock_class_mtx_sleep },
 287         { "sigio lock", &lock_class_mtx_sleep },
 288         { "process group", &lock_class_mtx_sleep },
 289         { "process lock", &lock_class_mtx_sleep },
 290         { "session", &lock_class_mtx_sleep },
 291         { "uidinfo hash", &lock_class_mtx_sleep },
 292         { "uidinfo struct", &lock_class_mtx_sleep },
 293 #ifdef  HWPMC_HOOKS
 294         { "pmc-sleep", &lock_class_mtx_sleep },
 295 #endif
 296         { NULL, NULL },
 297         /*
 298          * Sockets
 299          */
 300         { "accept", &lock_class_mtx_sleep },
 301         { "so_snd", &lock_class_mtx_sleep },
 302         { "so_rcv", &lock_class_mtx_sleep },
 303         { "sellck", &lock_class_mtx_sleep },
 304         { NULL, NULL },
 305         /*
 306          * Routing
 307          */
 308         { "so_rcv", &lock_class_mtx_sleep },
 309         { "radix node head", &lock_class_mtx_sleep },
 310         { "rtentry", &lock_class_mtx_sleep },
 311         { "ifaddr", &lock_class_mtx_sleep },
 312         { NULL, NULL },
 313         /*
 314          * Multicast - protocol locks before interface locks, after UDP locks.
 315          */
 316         { "udpinp", &lock_class_mtx_sleep },
 317         { "in_multi_mtx", &lock_class_mtx_sleep },
 318         { "igmp_mtx", &lock_class_mtx_sleep },
 319         { "if_addr_mtx", &lock_class_mtx_sleep },
 320         { NULL, NULL },
 321         /*
 322          * UNIX Domain Sockets
 323          */
 324         { "unp", &lock_class_mtx_sleep },
 325         { "so_snd", &lock_class_mtx_sleep },
 326         { NULL, NULL },
 327         /*
 328          * UDP/IP
 329          */
 330         { "udp", &lock_class_mtx_sleep },
 331         { "udpinp", &lock_class_mtx_sleep },
 332         { "so_snd", &lock_class_mtx_sleep },
 333         { NULL, NULL },
 334         /*
 335          * TCP/IP
 336          */
 337         { "tcp", &lock_class_mtx_sleep },
 338         { "tcpinp", &lock_class_mtx_sleep },
 339         { "so_snd", &lock_class_mtx_sleep },
 340         { NULL, NULL },
 341         /*
 342          * SLIP
 343          */
 344         { "slip_mtx", &lock_class_mtx_sleep },
 345         { "slip sc_mtx", &lock_class_mtx_sleep },
 346         { NULL, NULL },
 347         /*
 348          * netatalk
 349          */
 350         { "ddp_list_mtx", &lock_class_mtx_sleep },
 351         { "ddp_mtx", &lock_class_mtx_sleep },
 352         { NULL, NULL },
 353         /*
 354          * BPF
 355          */
 356         { "bpf global lock", &lock_class_mtx_sleep },
 357         { "bpf interface lock", &lock_class_mtx_sleep },
 358         { "bpf cdev lock", &lock_class_mtx_sleep },
 359         { NULL, NULL },
 360         /*
 361          * NFS server
 362          */
 363         { "nfsd_mtx", &lock_class_mtx_sleep },
 364         { "so_snd", &lock_class_mtx_sleep },
 365         { NULL, NULL },
 366
 367         /*
 368          * IEEE 802.11
 369          */
 370         { "802.11 com lock", &lock_class_mtx_sleep},
 371         { NULL, NULL },
 372         /*
 373          * Network drivers
 374          */
 375         { "network driver", &lock_class_mtx_sleep},
 376         { NULL, NULL },
 377
 378         /*
 379          * Netgraph
 380          */
 381         { "ng_node", &lock_class_mtx_sleep },
 382         { "ng_worklist", &lock_class_mtx_sleep },
 383         { NULL, NULL },
 384         /*
 385          * CDEV
 386          */
 387         { "system map", &lock_class_mtx_sleep },
 388         { "vm page queue mutex", &lock_class_mtx_sleep },
 389         { "vnode interlock", &lock_class_mtx_sleep },
 390         { "cdev", &lock_class_mtx_sleep },
 391         { NULL, NULL },
 392         /*
 393          * kqueue/VFS interaction
 394          */
 395         { "kqueue", &lock_class_mtx_sleep },
 396         { "struct mount mtx", &lock_class_mtx_sleep },
 397         { "vnode interlock", &lock_class_mtx_sleep },
 398         { NULL, NULL },
 399         /*
 400          * spin locks
 401          */
 402 #ifdef SMP
 403         { "ap boot", &lock_class_mtx_spin },
 404 #endif
 405         { "rm.mutex_mtx", &lock_class_mtx_spin },
 406         { "sio", &lock_class_mtx_spin },
 407 #ifdef __i386__
 408         { "cy", &lock_class_mtx_spin },
 409 #endif
 410 #ifdef __sparc64__
 411         { "pcib_mtx", &lock_class_mtx_spin },
 412         { "rtc_mtx", &lock_class_mtx_spin },
 413 #endif
 414         { "scc_hwmtx", &lock_class_mtx_spin },
 415         { "uart_hwmtx", &lock_class_mtx_spin },
 416         { "fast_taskqueue", &lock_class_mtx_spin },
 417         { "intr table", &lock_class_mtx_spin },
 418 #ifdef  HWPMC_HOOKS
 419         { "pmc-per-proc", &lock_class_mtx_spin },
 420 #endif
 421         { "process slock", &lock_class_mtx_spin },
 422         { "sleepq chain", &lock_class_mtx_spin },
 423         { "umtx lock", &lock_class_mtx_spin },
 424         { "turnstile chain", &lock_class_mtx_spin },
 425         { "turnstile lock", &lock_class_mtx_spin },
 426         { "sched lock", &lock_class_mtx_spin },
 427         { "td_contested", &lock_class_mtx_spin },
 428         { "callout", &lock_class_mtx_spin },
 429         { "entropy harvest mutex", &lock_class_mtx_spin },
 430         { "syscons video lock", &lock_class_mtx_spin },
 431         { "time lock", &lock_class_mtx_spin },
 432 #ifdef SMP
 433         { "smp rendezvous", &lock_class_mtx_spin },
 434 #endif
 435         /*
 436          * leaf locks
 437          */
 438         { "icu", &lock_class_mtx_spin },
 439 #if defined(SMP) && defined(__sparc64__)
 440         { "ipi", &lock_class_mtx_spin },
 441 #endif
 442 #ifdef __i386__
 443         { "allpmaps", &lock_class_mtx_spin },
 444         { "descriptor tables", &lock_class_mtx_spin },
 445 #endif
 446         { "clk", &lock_class_mtx_spin },
 447         { "mprof lock", &lock_class_mtx_spin },
 448         { "kse lock", &lock_class_mtx_spin },
 449         { "zombie lock", &lock_class_mtx_spin },
 450         { "ALD Queue", &lock_class_mtx_spin },
 451 #ifdef __ia64__
 452         { "MCA spin lock", &lock_class_mtx_spin },
 453 #endif
 454 #if defined(__i386__) || defined(__amd64__)
 455         { "pcicfg", &lock_class_mtx_spin },
 456         { "NDIS thread lock", &lock_class_mtx_spin },
 457 #endif
 458         { "tw_osl_io_lock", &lock_class_mtx_spin },
 459         { "tw_osl_q_lock", &lock_class_mtx_spin },
 460         { "tw_cl_io_lock", &lock_class_mtx_spin },
 461         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 462         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 463 #ifdef  HWPMC_HOOKS
 464         { "pmc-leaf", &lock_class_mtx_spin },
 465 #endif
 466         { "blocked lock", &lock_class_mtx_spin },
 467         { NULL, NULL },
 468         { NULL, NULL }
 469 };
 470
 471 #ifdef BLESSING
 472 /*
 473  * Pairs of locks which have been blessed
 474  * Don't complain about order problems with blessed locks
 475  */
 476 static struct witness_blessed blessed_list[] = {
 477 };
 478 static int blessed_count =
 479         sizeof(blessed_list) / sizeof(struct witness_blessed);
 480 #endif
 481
 482 /*
 483  * List of locks initialized prior to witness being initialized whose
 484  * enrollment is currently deferred.
 485  */
 486 STAILQ_HEAD(, lock_object) pending_locks =
 487     STAILQ_HEAD_INITIALIZER(pending_locks);
 488
 489 /*
 490  * This global is set to 0 once it becomes safe to use the witness code.
 491  */
 492 static int witness_cold = 1;
 493
 494 /*
 495  * This global is set to 1 once the static lock orders have been enrolled
 496  * so that a warning can be issued for any spin locks enrolled later.
 497  */
 498 static int witness_spin_warn = 0;
 499
 500 /*
 501  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 502  * assume that the early boot is single-threaded at least until after this
 503  * routine is completed.
 504  */
 505 static void
 506 witness_initialize(void *dummy __unused)
 507 {
 508         struct lock_object *lock;
 509         struct witness_order_list_entry *order;
 510         struct witness *w, *w1;
 511         int i;
 512
 513         /*
 514          * We have to release Giant before initializing its witness
 515          * structure so that WITNESS doesn't get confused.
 516          */
 517         mtx_unlock(&Giant);
 518         mtx_assert(&Giant, MA_NOTOWNED);
 519
 520         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 521         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 522             MTX_NOWITNESS | MTX_NOPROFILE);
 523         for (i = 0; i < WITNESS_COUNT; i++)
 524                 witness_free(&w_data[i]);
 525         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
 526                 witness_child_free(&w_childdata[i]);
 527         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 528                 witness_lock_list_free(&w_locklistdata[i]);
 529
 530         /* First add in all the specified order lists. */
 531         for (order = order_lists; order->w_name != NULL; order++) {
 532                 w = enroll(order->w_name, order->w_class);
 533                 if (w == NULL)
 534                         continue;
 535                 w->w_file = "order list";
 536                 for (order++; order->w_name != NULL; order++) {
 537                         w1 = enroll(order->w_name, order->w_class);
 538                         if (w1 == NULL)
 539                                 continue;
 540                         w1->w_file = "order list";
 541                         if (!itismychild(w, w1))
 542                                 panic("Not enough memory for static orders!");
 543                         w = w1;
 544                 }
 545         }
 546         witness_spin_warn = 1;
 547
 548         /* Iterate through all locks and add them to witness. */
 549         while (!STAILQ_EMPTY(&pending_locks)) {
 550                 lock = STAILQ_FIRST(&pending_locks);
 551                 STAILQ_REMOVE_HEAD(&pending_locks, lo_list);
 552                 KASSERT(lock->lo_flags & LO_WITNESS,
 553                     ("%s: lock %s is on pending list but not LO_WITNESS",
 554                     __func__, lock->lo_name));
 555                 lock->lo_witness = enroll(lock->lo_type, LOCK_CLASS(lock));
 556         }
 557
 558         /* Mark the witness code as being ready for use. */
 559         witness_cold = 0;
 560
 561         mtx_lock(&Giant);
 562 }
 563 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
 564
 565 static int
 566 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
 567 {
 568         int error, value;
 569
 570         value = witness_watch;
 571         error = sysctl_handle_int(oidp, &value, 0, req);
 572         if (error != 0 || req->newptr == NULL)
 573                 return (error);
 574         if (value == witness_watch)
 575                 return (0);
 576         if (value != 0)
 577                 return (EINVAL);
 578         witness_watch = 0;
 579         return (0);
 580 }
 581
 582 void
 583 witness_init(struct lock_object *lock)
 584 {
 585         struct lock_class *class;
 586
 587         /* Various sanity checks. */
 588         class = LOCK_CLASS(lock);
 589         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 590             (class->lc_flags & LC_RECURSABLE) == 0)
 591                 panic("%s: lock (%s) %s can not be recursable", __func__,
 592                     class->lc_name, lock->lo_name);
 593         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 594             (class->lc_flags & LC_SLEEPABLE) == 0)
 595                 panic("%s: lock (%s) %s can not be sleepable", __func__,
 596                     class->lc_name, lock->lo_name);
 597         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 598             (class->lc_flags & LC_UPGRADABLE) == 0)
 599                 panic("%s: lock (%s) %s can not be upgradable", __func__,
 600                     class->lc_name, lock->lo_name);
 601
 602         /*
 603          * If we shouldn't watch this lock, then just clear lo_witness.
 604          * Otherwise, if witness_cold is set, then it is too early to
 605          * enroll this lock, so defer it to witness_initialize() by adding
 606          * it to the pending_locks list.  If it is not too early, then enroll
 607          * the lock now.
 608          */
 609         if (witness_watch == 0 || panicstr != NULL ||
 610             (lock->lo_flags & LO_WITNESS) == 0)
 611                 lock->lo_witness = NULL;
 612         else if (witness_cold) {
 613                 STAILQ_INSERT_TAIL(&pending_locks, lock, lo_list);
 614                 lock->lo_flags |= LO_ENROLLPEND;
 615         } else
 616                 lock->lo_witness = enroll(lock->lo_type, class);
 617 }
 618
 619 void
 620 witness_destroy(struct lock_object *lock)
 621 {
 622         struct lock_class *class;
 623         struct witness *w;
 624
 625         class = LOCK_CLASS(lock);
 626         if (witness_cold)
 627                 panic("lock (%s) %s destroyed while witness_cold",
 628                     class->lc_name, lock->lo_name);
 629
 630         /* XXX: need to verify that no one holds the lock */
 631         if ((lock->lo_flags & (LO_WITNESS | LO_ENROLLPEND)) == LO_WITNESS &&
 632             lock->lo_witness != NULL) {
 633                 w = lock->lo_witness;
 634                 mtx_lock_spin(&w_mtx);
 635                 MPASS(w->w_refcount > 0);
 636                 w->w_refcount--;
 637
 638                 /*
 639                  * Lock is already released if we have an allocation failure
 640                  * and depart() fails.
 641                  */
 642                 if (w->w_refcount != 0 || depart(w))
 643                         mtx_unlock_spin(&w_mtx);
 644         }
 645
 646         /*
 647          * If this lock is destroyed before witness is up and running,
 648          * remove it from the pending list.
 649          */
 650         if (lock->lo_flags & LO_ENROLLPEND) {
 651                 STAILQ_REMOVE(&pending_locks, lock, lock_object, lo_list);
 652                 lock->lo_flags &= ~LO_ENROLLPEND;
 653         }
 654 }
 655
 656 #ifdef DDB
 657 static void
 658 witness_levelall (void)
 659 {
 660         struct witness_list *list;
 661         struct witness *w, *w1;
 662
 663         /*
 664          * First clear all levels.
 665          */
 666         STAILQ_FOREACH(w, &w_all, w_list) {
 667                 w->w_level = 0;
 668         }
 669
 670         /*
 671          * Look for locks with no parent and level all their descendants.
 672          */
 673         STAILQ_FOREACH(w, &w_all, w_list) {
 674                 /*
 675                  * This is just an optimization, technically we could get
 676                  * away just walking the all list each time.
 677                  */
 678                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
 679                         list = &w_sleep;
 680                 else
 681                         list = &w_spin;
 682                 STAILQ_FOREACH(w1, list, w_typelist) {
 683                         if (isitmychild(w1, w))
 684                                 goto skip;
 685                 }
 686                 witness_leveldescendents(w, 0);
 687         skip:
 688                 ;       /* silence GCC 3.x */
 689         }
 690 }
 691
 692 static void
 693 witness_leveldescendents(struct witness *parent, int level)
 694 {
 695         struct witness_child_list_entry *wcl;
 696         int i;
 697
 698         if (parent->w_level < level)
 699                 parent->w_level = level;
 700         level++;
 701         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 702                 for (i = 0; i < wcl->wcl_count; i++)
 703                         witness_leveldescendents(wcl->wcl_children[i], level);
 704 }
 705
 706 static void
 707 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
 708                            struct witness *parent, int indent)
 709 {
 710         struct witness_child_list_entry *wcl;
 711         int i, level;
 712
 713         level = parent->w_level;
 714         prnt("%-2d", level);
 715         for (i = 0; i < indent; i++)
 716                 prnt(" ");
 717         if (parent->w_refcount > 0)
 718                 prnt("%s", parent->w_name);
 719         else
 720                 prnt("(dead)");
 721         if (parent->w_displayed) {
 722                 prnt(" -- (already displayed)\n");
 723                 return;
 724         }
 725         parent->w_displayed = 1;
 726         if (parent->w_refcount > 0) {
 727                 if (parent->w_file != NULL)
 728                         prnt(" -- last acquired @ %s:%d", parent->w_file,
 729                             parent->w_line);
 730         }
 731         prnt("\n");
 732         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 733                 for (i = 0; i < wcl->wcl_count; i++)
 734                             witness_displaydescendants(prnt,
 735                                 wcl->wcl_children[i], indent + 1);
 736 }
 737
 738 static void
 739 witness_display_list(void(*prnt)(const char *fmt, ...),
 740                      struct witness_list *list)
 741 {
 742         struct witness *w;
 743
 744         STAILQ_FOREACH(w, list, w_typelist) {
 745                 if (w->w_file == NULL || w->w_level > 0)
 746                         continue;
 747                 /*
 748                  * This lock has no anscestors, display its descendants.
 749                  */
 750                 witness_displaydescendants(prnt, w, 0);
 751         }
 752 }
 753
 754 static void
 755 witness_display(void(*prnt)(const char *fmt, ...))
 756 {
 757         struct witness *w;
 758
 759         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 760         witness_levelall();
 761
 762         /* Clear all the displayed flags. */
 763         STAILQ_FOREACH(w, &w_all, w_list) {
 764                 w->w_displayed = 0;
 765         }
 766
 767         /*
 768          * First, handle sleep locks which have been acquired at least
 769          * once.
 770          */
 771         prnt("Sleep locks:\n");
 772         witness_display_list(prnt, &w_sleep);
 773
 774         /*
 775          * Now do spin locks which have been acquired at least once.
 776          */
 777         prnt("\nSpin locks:\n");
 778         witness_display_list(prnt, &w_spin);
 779
 780         /*
 781          * Finally, any locks which have not been acquired yet.
 782          */
 783         prnt("\nLocks which were never acquired:\n");
 784         STAILQ_FOREACH(w, &w_all, w_list) {
 785                 if (w->w_file != NULL || w->w_refcount == 0)
 786                         continue;
 787                 prnt("%s\n", w->w_name);
 788         }
 789 }
 790 #endif /* DDB */
 791
 792 /* Trim useless garbage from filenames. */
 793 static const char *
 794 fixup_filename(const char *file)
 795 {
 796
 797         if (file == NULL)
 798                 return (NULL);
 799         while (strncmp(file, "../", 3) == 0)
 800                 file += 3;
 801         return (file);
 802 }
 803
 804 int
 805 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
 806 {
 807
 808         if (witness_watch == 0 || panicstr != NULL)
 809                 return (0);
 810
 811         /* Require locks that witness knows about. */
 812         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
 813             lock2->lo_witness == NULL)
 814                 return (EINVAL);
 815
 816         MPASS(!mtx_owned(&w_mtx));
 817         mtx_lock_spin(&w_mtx);
 818
 819         /*
 820          * If we already have either an explicit or implied lock order that
 821          * is the other way around, then return an error.
 822          */
 823         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
 824                 mtx_unlock_spin(&w_mtx);
 825                 return (EDOOFUS);
 826         }
 827
 828         /* Try to add the new order. */
 829         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 830             lock2->lo_type, lock1->lo_type);
 831         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
 832                 return (ENOMEM);
 833         mtx_unlock_spin(&w_mtx);
 834         return (0);
 835 }
 836
 837 void
 838 witness_checkorder(struct lock_object *lock, int flags, const char *file,
 839     int line)
 840 {
 841         struct lock_list_entry **lock_list, *lle;
 842         struct lock_instance *lock1, *lock2;
 843         struct lock_class *class;
 844         struct witness *w, *w1;
 845         struct thread *td;
 846         int i, j;
 847
 848         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 849             panicstr != NULL)
 850                 return;
 851
 852         /*
 853          * Try locks do not block if they fail to acquire the lock, thus
 854          * there is no danger of deadlocks or of switching while holding a
 855          * spin lock if we acquire a lock via a try operation.  This
 856          * function shouldn't even be called for try locks, so panic if
 857          * that happens.
 858          */
 859         if (flags & LOP_TRYLOCK)
 860                 panic("%s should not be called for try lock operations",
 861                     __func__);
 862
 863         w = lock->lo_witness;
 864         class = LOCK_CLASS(lock);
 865         td = curthread;
 866         file = fixup_filename(file);
 867
 868         if (class->lc_flags & LC_SLEEPLOCK) {
 869                 /*
 870                  * Since spin locks include a critical section, this check
 871                  * implicitly enforces a lock order of all sleep locks before
 872                  * all spin locks.
 873                  */
 874                 if (td->td_critnest != 0 && !kdb_active)
 875                         panic("blockable sleep lock (%s) %s @ %s:%d",
 876                             class->lc_name, lock->lo_name, file, line);
 877
 878                 /*
 879                  * If this is the first lock acquired then just return as
 880                  * no order checking is needed.
 881                  */
 882                 if (td->td_sleeplocks == NULL)
 883                         return;
 884                 lock_list = &td->td_sleeplocks;
 885         } else {
 886                 /*
 887                  * If this is the first lock, just return as no order
 888                  * checking is needed.  We check this in both if clauses
 889                  * here as unifying the check would require us to use a
 890                  * critical section to ensure we don't migrate while doing
 891                  * the check.  Note that if this is not the first lock, we
 892                  * are already in a critical section and are safe for the
 893                  * rest of the check.
 894                  */
 895                 if (PCPU_GET(spinlocks) == NULL)
 896                         return;
 897                 lock_list = PCPU_PTR(spinlocks);
 898         }
 899
 900         /*
 901          * Check to see if we are recursing on a lock we already own.  If
 902          * so, make sure that we don't mismatch exclusive and shared lock
 903          * acquires.
 904          */
 905         lock1 = find_instance(*lock_list, lock);
 906         if (lock1 != NULL) {
 907                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
 908                     (flags & LOP_EXCLUSIVE) == 0) {
 909                         printf("shared lock of (%s) %s @ %s:%d\n",
 910                             class->lc_name, lock->lo_name, file, line);
 911                         printf("while exclusively locked from %s:%d\n",
 912                             lock1->li_file, lock1->li_line);
 913                         panic("share->excl");
 914                 }
 915                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
 916                     (flags & LOP_EXCLUSIVE) != 0) {
 917                         printf("exclusive lock of (%s) %s @ %s:%d\n",
 918                             class->lc_name, lock->lo_name, file, line);
 919                         printf("while share locked from %s:%d\n",
 920                             lock1->li_file, lock1->li_line);
 921                         panic("excl->share");
 922                 }
 923                 return;
 924         }
 925
 926         /*
 927          * Try locks do not block if they fail to acquire the lock, thus
 928          * there is no danger of deadlocks or of switching while holding a
 929          * spin lock if we acquire a lock via a try operation.
 930          */
 931         if (flags & LOP_TRYLOCK)
 932                 return;
 933
 934         /*
 935          * Check for duplicate locks of the same type.  Note that we only
 936          * have to check for this on the last lock we just acquired.  Any
 937          * other cases will be caught as lock order violations.
 938          */
 939         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 940         w1 = lock1->li_lock->lo_witness;
 941         if (w1 == w) {
 942                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
 943                     (flags & LOP_DUPOK))
 944                         return;
 945                 w->w_same_squawked = 1;
 946                 printf("acquiring duplicate lock of same type: \"%s\"\n",
 947                         lock->lo_type);
 948                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
 949                     lock1->li_file, lock1->li_line);
 950                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
 951 #ifdef KDB
 952                 goto debugger;
 953 #else
 954                 return;
 955 #endif
 956         }
 957         MPASS(!mtx_owned(&w_mtx));
 958         mtx_lock_spin(&w_mtx);
 959         /*
 960          * If we know that the the lock we are acquiring comes after
 961          * the lock we most recently acquired in the lock order tree,
 962          * then there is no need for any further checks.
 963          */
 964         if (isitmychild(w1, w)) {
 965                 mtx_unlock_spin(&w_mtx);
 966                 return;
 967         }
 968         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
 969                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
 970
 971                         MPASS(j < WITNESS_COUNT);
 972                         lock1 = &lle->ll_children[i];
 973                         w1 = lock1->li_lock->lo_witness;
 974
 975                         /*
 976                          * If this lock doesn't undergo witness checking,
 977                          * then skip it.
 978                          */
 979                         if (w1 == NULL) {
 980                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
 981                                     ("lock missing witness structure"));
 982                                 continue;
 983                         }
 984                         /*
 985                          * If we are locking Giant and this is a sleepable
 986                          * lock, then skip it.
 987                          */
 988                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
 989                             lock == &Giant.lock_object)
 990                                 continue;
 991                         /*
 992                          * If we are locking a sleepable lock and this lock
 993                          * is Giant, then skip it.
 994                          */
 995                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 996                             lock1->li_lock == &Giant.lock_object)
 997                                 continue;
 998                         /*
 999                          * If we are locking a sleepable lock and this lock
1000                          * isn't sleepable, we want to treat it as a lock
1001                          * order violation to enfore a general lock order of
1002                          * sleepable locks before non-sleepable locks.
1003                          */
1004                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1005                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1006                                 goto reversal;
1007                         /*
1008                          * If we are locking Giant and this is a non-sleepable
1009                          * lock, then treat it as a reversal.
1010                          */
1011                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1012                             lock == &Giant.lock_object)
1013                                 goto reversal;
1014                         /*
1015                          * Check the lock order hierarchy for a reveresal.
1016                          */
1017                         if (!isitmydescendant(w, w1))
1018                                 continue;
1019                 reversal:
1020                         /*
1021                          * We have a lock order violation, check to see if it
1022                          * is allowed or has already been yelled about.
1023                          */
1024                         mtx_unlock_spin(&w_mtx);
1025 #ifdef BLESSING
1026                         /*
1027                          * If the lock order is blessed, just bail.  We don't
1028                          * look for other lock order violations though, which
1029                          * may be a bug.
1030                          */
1031                         if (blessed(w, w1))
1032                                 return;
1033 #endif
1034                         if (lock1->li_lock == &Giant.lock_object) {
1035                                 if (w1->w_Giant_squawked)
1036                                         return;
1037                                 else
1038                                         w1->w_Giant_squawked = 1;
1039                         } else {
1040                                 if (w1->w_other_squawked)
1041                                         return;
1042                                 else
1043                                         w1->w_other_squawked = 1;
1044                         }
1045                         /*
1046                          * Ok, yell about it.
1047                          */
1048                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1049                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1050                                 printf(
1051                 "lock order reversal: (sleepable after non-sleepable)\n");
1052                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1053                             && lock == &Giant.lock_object)
1054                                 printf(
1055                 "lock order reversal: (Giant after non-sleepable)\n");
1056                         else
1057                                 printf("lock order reversal:\n");
1058                         /*
1059                          * Try to locate an earlier lock with
1060                          * witness w in our list.
1061                          */
1062                         do {
1063                                 lock2 = &lle->ll_children[i];
1064                                 MPASS(lock2->li_lock != NULL);
1065                                 if (lock2->li_lock->lo_witness == w)
1066                                         break;
1067                                 if (i == 0 && lle->ll_next != NULL) {
1068                                         lle = lle->ll_next;
1069                                         i = lle->ll_count - 1;
1070                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1071                                 } else
1072                                         i--;
1073                         } while (i >= 0);
1074                         if (i < 0) {
1075                                 printf(" 1st %p %s (%s) @ %s:%d\n",
1076                                     lock1->li_lock, lock1->li_lock->lo_name,
1077                                     lock1->li_lock->lo_type, lock1->li_file,
1078                                     lock1->li_line);
1079                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1080                                     lock->lo_name, lock->lo_type, file, line);
1081                         } else {
1082                                 printf(" 1st %p %s (%s) @ %s:%d\n",
1083                                     lock2->li_lock, lock2->li_lock->lo_name,
1084                                     lock2->li_lock->lo_type, lock2->li_file,
1085                                     lock2->li_line);
1086                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
1087                                     lock1->li_lock, lock1->li_lock->lo_name,
1088                                     lock1->li_lock->lo_type, lock1->li_file,
1089                                     lock1->li_line);
1090                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1091                                     lock->lo_name, lock->lo_type, file, line);
1092                         }
1093 #ifdef KDB
1094                         goto debugger;
1095 #else
1096                         return;
1097 #endif
1098                 }
1099         }
1100         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
1101         /*
1102          * If requested, build a new lock order.  However, don't build a new
1103          * relationship between a sleepable lock and Giant if it is in the
1104          * wrong direction.  The correct lock order is that sleepable locks
1105          * always come before Giant.
1106          */
1107         if (flags & LOP_NEWORDER &&
1108             !(lock1->li_lock == &Giant.lock_object &&
1109             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1110                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1111                     lock->lo_type, lock1->li_lock->lo_type);
1112                 if (!itismychild(lock1->li_lock->lo_witness, w))
1113                         /* Witness is dead. */
1114                         return;
1115         }
1116         mtx_unlock_spin(&w_mtx);
1117         return;
1118
1119 #ifdef KDB
1120 debugger:
1121         if (witness_trace)
1122                 kdb_backtrace();
1123         if (witness_kdb)
1124                 kdb_enter(__func__);
1125 #endif
1126 }
1127
1128 void
1129 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1130 {
1131         struct lock_list_entry **lock_list, *lle;
1132         struct lock_instance *instance;
1133         struct witness *w;
1134         struct thread *td;
1135
1136         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1137             panicstr != NULL)
1138                 return;
1139         w = lock->lo_witness;
1140         td = curthread;
1141         file = fixup_filename(file);
1142
1143         /* Determine lock list for this lock. */
1144         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1145                 lock_list = &td->td_sleeplocks;
1146         else
1147                 lock_list = PCPU_PTR(spinlocks);
1148
1149         /* Check to see if we are recursing on a lock we already own. */
1150         instance = find_instance(*lock_list, lock);
1151         if (instance != NULL) {
1152                 instance->li_flags++;
1153                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1154                     td->td_proc->p_pid, lock->lo_name,
1155                     instance->li_flags & LI_RECURSEMASK);
1156                 instance->li_file = file;
1157                 instance->li_line = line;
1158                 return;
1159         }
1160
1161         /* Update per-witness last file and line acquire. */
1162         w->w_file = file;
1163         w->w_line = line;
1164
1165         /* Find the next open lock instance in the list and fill it. */
1166         lle = *lock_list;
1167         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1168                 lle = witness_lock_list_get();
1169                 if (lle == NULL)
1170                         return;
1171                 lle->ll_next = *lock_list;
1172                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1173                     td->td_proc->p_pid, lle);
1174                 *lock_list = lle;
1175         }
1176         instance = &lle->ll_children[lle->ll_count++];
1177         instance->li_lock = lock;
1178         instance->li_line = line;
1179         instance->li_file = file;
1180         if ((flags & LOP_EXCLUSIVE) != 0)
1181                 instance->li_flags = LI_EXCLUSIVE;
1182         else
1183                 instance->li_flags = 0;
1184         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1185             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1186 }
1187
1188 void
1189 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1190 {
1191         struct lock_instance *instance;
1192         struct lock_class *class;
1193
1194         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1195         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1196                 return;
1197         class = LOCK_CLASS(lock);
1198         file = fixup_filename(file);
1199         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1200                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1201                     class->lc_name, lock->lo_name, file, line);
1202         if ((flags & LOP_TRYLOCK) == 0)
1203                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1204                     lock->lo_name, file, line);
1205         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1206                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1207                     class->lc_name, lock->lo_name, file, line);
1208         instance = find_instance(curthread->td_sleeplocks, lock);
1209         if (instance == NULL)
1210                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1211                     class->lc_name, lock->lo_name, file, line);
1212         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1213                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1214                     class->lc_name, lock->lo_name, file, line);
1215         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1216                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1217                     class->lc_name, lock->lo_name,
1218                     instance->li_flags & LI_RECURSEMASK, file, line);
1219         instance->li_flags |= LI_EXCLUSIVE;
1220 }
1221
1222 void
1223 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1224     int line)
1225 {
1226         struct lock_instance *instance;
1227         struct lock_class *class;
1228
1229         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1230         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1231                 return;
1232         class = LOCK_CLASS(lock);
1233         file = fixup_filename(file);
1234         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1235                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1236                     class->lc_name, lock->lo_name, file, line);
1237         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1238                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1239                     class->lc_name, lock->lo_name, file, line);
1240         instance = find_instance(curthread->td_sleeplocks, lock);
1241         if (instance == NULL)
1242                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1243                     class->lc_name, lock->lo_name, file, line);
1244         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1245                 panic("downgrade of shared lock (%s) %s @ %s:%d",
1246                     class->lc_name, lock->lo_name, file, line);
1247         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1248                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1249                     class->lc_name, lock->lo_name,
1250                     instance->li_flags & LI_RECURSEMASK, file, line);
1251         instance->li_flags &= ~LI_EXCLUSIVE;
1252 }
1253
1254 void
1255 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1256 {
1257         struct lock_list_entry **lock_list, *lle;
1258         struct lock_instance *instance;
1259         struct lock_class *class;
1260         struct thread *td;
1261         register_t s;
1262         int i, j;
1263
1264         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1265             panicstr != NULL)
1266                 return;
1267         td = curthread;
1268         class = LOCK_CLASS(lock);
1269         file = fixup_filename(file);
1270
1271         /* Find lock instance associated with this lock. */
1272         if (class->lc_flags & LC_SLEEPLOCK)
1273                 lock_list = &td->td_sleeplocks;
1274         else
1275                 lock_list = PCPU_PTR(spinlocks);
1276         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1277                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1278                         instance = &(*lock_list)->ll_children[i];
1279                         if (instance->li_lock == lock)
1280                                 goto found;
1281                 }
1282         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1283             file, line);
1284 found:
1285
1286         /* First, check for shared/exclusive mismatches. */
1287         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1288             (flags & LOP_EXCLUSIVE) == 0) {
1289                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1290                     lock->lo_name, file, line);
1291                 printf("while exclusively locked from %s:%d\n",
1292                     instance->li_file, instance->li_line);
1293                 panic("excl->ushare");
1294         }
1295         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1296             (flags & LOP_EXCLUSIVE) != 0) {
1297                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1298                     lock->lo_name, file, line);
1299                 printf("while share locked from %s:%d\n", instance->li_file,
1300                     instance->li_line);
1301                 panic("share->uexcl");
1302         }
1303
1304         /* If we are recursed, unrecurse. */
1305         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1306                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1307                     td->td_proc->p_pid, instance->li_lock->lo_name,
1308                     instance->li_flags);
1309                 instance->li_flags--;
1310                 return;
1311         }
1312
1313         /* Otherwise, remove this item from the list. */
1314         s = intr_disable();
1315         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1316             td->td_proc->p_pid, instance->li_lock->lo_name,
1317             (*lock_list)->ll_count - 1);
1318         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1319                 (*lock_list)->ll_children[j] =
1320                     (*lock_list)->ll_children[j + 1];
1321         (*lock_list)->ll_count--;
1322         intr_restore(s);
1323
1324         /* If this lock list entry is now empty, free it. */
1325         if ((*lock_list)->ll_count == 0) {
1326                 lle = *lock_list;
1327                 *lock_list = lle->ll_next;
1328                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1329                     td->td_proc->p_pid, lle);
1330                 witness_lock_list_free(lle);
1331         }
1332 }
1333
1334 /*
1335  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1336  * exempt Giant and sleepable locks from the checks as well.  If any
1337  * non-exempt locks are held, then a supplied message is printed to the
1338  * console along with a list of the offending locks.  If indicated in the
1339  * flags then a failure results in a panic as well.
1340  */
1341 int
1342 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1343 {
1344         struct lock_list_entry *lle;
1345         struct lock_instance *lock1;
1346         struct thread *td;
1347         va_list ap;
1348         int i, n;
1349
1350         if (witness_cold || witness_watch == 0 || panicstr != NULL)
1351                 return (0);
1352         n = 0;
1353         td = curthread;
1354         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1355                 for (i = lle->ll_count - 1; i >= 0; i--) {
1356                         lock1 = &lle->ll_children[i];
1357                         if (lock1->li_lock == lock)
1358                                 continue;
1359                         if (flags & WARN_GIANTOK &&
1360                             lock1->li_lock == &Giant.lock_object)
1361                                 continue;
1362                         if (flags & WARN_SLEEPOK &&
1363                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1364                                 continue;
1365                         if (n == 0) {
1366                                 va_start(ap, fmt);
1367                                 vprintf(fmt, ap);
1368                                 va_end(ap);
1369                                 printf(" with the following");
1370                                 if (flags & WARN_SLEEPOK)
1371                                         printf(" non-sleepable");
1372                                 printf(" locks held:\n");
1373                         }
1374                         n++;
1375                         witness_list_lock(lock1);
1376                 }
1377         if (PCPU_GET(spinlocks) != NULL) {
1378                 /*
1379                  * Since we already hold a spinlock preemption is
1380                  * already blocked.
1381                  */
1382                 if (n == 0) {
1383                         va_start(ap, fmt);
1384                         vprintf(fmt, ap);
1385                         va_end(ap);
1386                         printf(" with the following");
1387                         if (flags & WARN_SLEEPOK)
1388                                 printf(" non-sleepable");
1389                         printf(" locks held:\n");
1390                 }
1391                 n += witness_list_locks(PCPU_PTR(spinlocks));
1392         }
1393         if (flags & WARN_PANIC && n)
1394                 panic("witness_warn");
1395 #ifdef KDB
1396         else if (witness_kdb && n)
1397                 kdb_enter(__func__);
1398         else if (witness_trace && n)
1399                 kdb_backtrace();
1400 #endif
1401         return (n);
1402 }
1403
1404 const char *
1405 witness_file(struct lock_object *lock)
1406 {
1407         struct witness *w;
1408
1409         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1410                 return ("?");
1411         w = lock->lo_witness;
1412         return (w->w_file);
1413 }
1414
1415 int
1416 witness_line(struct lock_object *lock)
1417 {
1418         struct witness *w;
1419
1420         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1421                 return (0);
1422         w = lock->lo_witness;
1423         return (w->w_line);
1424 }
1425
1426 static struct witness *
1427 enroll(const char *description, struct lock_class *lock_class)
1428 {
1429         struct witness *w;
1430
1431         if (witness_watch == 0 || panicstr != NULL)
1432                 return (NULL);
1433         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1434                 return (NULL);
1435         mtx_lock_spin(&w_mtx);
1436         STAILQ_FOREACH(w, &w_all, w_list) {
1437                 if (w->w_name == description || (w->w_refcount > 0 &&
1438                     strcmp(description, w->w_name) == 0)) {
1439                         w->w_refcount++;
1440                         mtx_unlock_spin(&w_mtx);
1441                         if (lock_class != w->w_class)
1442                                 panic(
1443                                 "lock (%s) %s does not match earlier (%s) lock",
1444                                     description, lock_class->lc_name,
1445                                     w->w_class->lc_name);
1446                         return (w);
1447                 }
1448         }
1449         if ((w = witness_get()) == NULL)
1450                 goto out;
1451         w->w_name = description;
1452         w->w_class = lock_class;
1453         w->w_refcount = 1;
1454         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1455         if (lock_class->lc_flags & LC_SPINLOCK) {
1456                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1457                 w_spin_cnt++;
1458         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1459                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1460                 w_sleep_cnt++;
1461         } else {
1462                 mtx_unlock_spin(&w_mtx);
1463                 panic("lock class %s is not sleep or spin",
1464                     lock_class->lc_name);
1465         }
1466         mtx_unlock_spin(&w_mtx);
1467 out:
1468         /*
1469          * We issue a warning for any spin locks not defined in the static
1470          * order list as a way to discourage their use (folks should really
1471          * be using non-spin mutexes most of the time).  However, several
1472          * 3rd part device drivers use spin locks because that is all they
1473          * have available on Windows and Linux and they think that normal
1474          * mutexes are insufficient.
1475          */
1476         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
1477                 printf("WITNESS: spin lock %s not in order list\n",
1478                     description);
1479         return (w);
1480 }
1481
1482 /* Don't let the door bang you on the way out... */
1483 static int
1484 depart(struct witness *w)
1485 {
1486         struct witness_child_list_entry *wcl, *nwcl;
1487         struct witness_list *list;
1488         struct witness *parent;
1489
1490         MPASS(w->w_refcount == 0);
1491         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1492                 list = &w_sleep;
1493                 w_sleep_cnt--;
1494         } else {
1495                 list = &w_spin;
1496                 w_spin_cnt--;
1497         }
1498         /*
1499          * First, we run through the entire tree looking for any
1500          * witnesses that the outgoing witness is a child of.  For
1501          * each parent that we find, we reparent all the direct
1502          * children of the outgoing witness to its parent.
1503          */
1504         STAILQ_FOREACH(parent, list, w_typelist) {
1505                 if (!isitmychild(parent, w))
1506                         continue;
1507                 removechild(parent, w);
1508         }
1509
1510         /*
1511          * Now we go through and free up the child list of the
1512          * outgoing witness.
1513          */
1514         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1515                 nwcl = wcl->wcl_next;
1516                 w_child_cnt--;
1517                 witness_child_free(wcl);
1518         }
1519
1520         /*
1521          * Detach from various lists and free.
1522          */
1523         STAILQ_REMOVE(list, w, witness, w_typelist);
1524         STAILQ_REMOVE(&w_all, w, witness, w_list);
1525         witness_free(w);
1526
1527         return (1);
1528 }
1529
1530 /*
1531  * Add "child" as a direct child of "parent".  Returns false if
1532  * we fail due to out of memory.
1533  */
1534 static int
1535 insertchild(struct witness *parent, struct witness *child)
1536 {
1537         struct witness_child_list_entry **wcl;
1538
1539         MPASS(child != NULL && parent != NULL);
1540
1541         /*
1542          * Insert "child" after "parent"
1543          */
1544         wcl = &parent->w_children;
1545         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1546                 wcl = &(*wcl)->wcl_next;
1547         if (*wcl == NULL) {
1548                 *wcl = witness_child_get();
1549                 if (*wcl == NULL)
1550                         return (0);
1551                 w_child_cnt++;
1552         }
1553         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1554
1555         return (1);
1556 }
1557
1558
1559 static int
1560 itismychild(struct witness *parent, struct witness *child)
1561 {
1562         struct witness_list *list;
1563
1564         MPASS(child != NULL && parent != NULL);
1565         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1566             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1567                 panic(
1568                 "%s: parent (%s) and child (%s) are not the same lock type",
1569                     __func__, parent->w_class->lc_name,
1570                     child->w_class->lc_name);
1571
1572         if (!insertchild(parent, child))
1573                 return (0);
1574
1575         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1576                 list = &w_sleep;
1577         else
1578                 list = &w_spin;
1579         return (1);
1580 }
1581
1582 static void
1583 removechild(struct witness *parent, struct witness *child)
1584 {
1585         struct witness_child_list_entry **wcl, *wcl1;
1586         int i;
1587
1588         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1589                 for (i = 0; i < (*wcl)->wcl_count; i++)
1590                         if ((*wcl)->wcl_children[i] == child)
1591                                 goto found;
1592         return;
1593 found:
1594         (*wcl)->wcl_count--;
1595         if ((*wcl)->wcl_count > i)
1596                 (*wcl)->wcl_children[i] =
1597                     (*wcl)->wcl_children[(*wcl)->wcl_count];
1598         MPASS((*wcl)->wcl_children[i] != NULL);
1599         if ((*wcl)->wcl_count != 0)
1600                 return;
1601         wcl1 = *wcl;
1602         *wcl = wcl1->wcl_next;
1603         w_child_cnt--;
1604         witness_child_free(wcl1);
1605 }
1606
1607 static int
1608 isitmychild(struct witness *parent, struct witness *child)
1609 {
1610         struct witness_child_list_entry *wcl;
1611         int i;
1612
1613         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1614                 for (i = 0; i < wcl->wcl_count; i++) {
1615                         if (wcl->wcl_children[i] == child)
1616                                 return (1);
1617                 }
1618         }
1619         return (0);
1620 }
1621
1622 static int
1623 isitmydescendant(struct witness *parent, struct witness *child)
1624 {
1625         struct witness_child_list_entry *wcl;
1626         int i, j;
1627
1628         if (isitmychild(parent, child))
1629                 return (1);
1630         j = 0;
1631         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1632                 MPASS(j < 1000);
1633                 for (i = 0; i < wcl->wcl_count; i++) {
1634                         if (isitmydescendant(wcl->wcl_children[i], child))
1635                                 return (1);
1636                 }
1637                 j++;
1638         }
1639         return (0);
1640 }
1641
1642 #ifdef BLESSING
1643 static int
1644 blessed(struct witness *w1, struct witness *w2)
1645 {
1646         int i;
1647         struct witness_blessed *b;
1648
1649         for (i = 0; i < blessed_count; i++) {
1650                 b = &blessed_list[i];
1651                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1652                         if (strcmp(w2->w_name, b->b_lock2) == 0)
1653                                 return (1);
1654                         continue;
1655                 }
1656                 if (strcmp(w1->w_name, b->b_lock2) == 0)
1657                         if (strcmp(w2->w_name, b->b_lock1) == 0)
1658                                 return (1);
1659         }
1660         return (0);
1661 }
1662 #endif
1663
1664 static struct witness *
1665 witness_get(void)
1666 {
1667         struct witness *w;
1668
1669         if (witness_watch == 0) {
1670                 mtx_unlock_spin(&w_mtx);
1671                 return (NULL);
1672         }
1673         if (STAILQ_EMPTY(&w_free)) {
1674                 witness_watch = 0;
1675                 mtx_unlock_spin(&w_mtx);
1676                 printf("%s: witness exhausted\n", __func__);
1677                 return (NULL);
1678         }
1679         w = STAILQ_FIRST(&w_free);
1680         STAILQ_REMOVE_HEAD(&w_free, w_list);
1681         w_free_cnt--;
1682         bzero(w, sizeof(*w));
1683         return (w);
1684 }
1685
1686 static void
1687 witness_free(struct witness *w)
1688 {
1689
1690         STAILQ_INSERT_HEAD(&w_free, w, w_list);
1691         w_free_cnt++;
1692 }
1693
1694 static struct witness_child_list_entry *
1695 witness_child_get(void)
1696 {
1697         struct witness_child_list_entry *wcl;
1698
1699         if (witness_watch == 0) {
1700                 mtx_unlock_spin(&w_mtx);
1701                 return (NULL);
1702         }
1703         wcl = w_child_free;
1704         if (wcl == NULL) {
1705                 witness_watch = 0;
1706                 mtx_unlock_spin(&w_mtx);
1707                 printf("%s: witness exhausted\n", __func__);
1708                 return (NULL);
1709         }
1710         w_child_free = wcl->wcl_next;
1711         w_child_free_cnt--;
1712         bzero(wcl, sizeof(*wcl));
1713         return (wcl);
1714 }
1715
1716 static void
1717 witness_child_free(struct witness_child_list_entry *wcl)
1718 {
1719
1720         wcl->wcl_next = w_child_free;
1721         w_child_free = wcl;
1722         w_child_free_cnt++;
1723 }
1724
1725 static struct lock_list_entry *
1726 witness_lock_list_get(void)
1727 {
1728         struct lock_list_entry *lle;
1729
1730         if (witness_watch == 0)
1731                 return (NULL);
1732         mtx_lock_spin(&w_mtx);
1733         lle = w_lock_list_free;
1734         if (lle == NULL) {
1735                 witness_watch = 0;
1736                 mtx_unlock_spin(&w_mtx);
1737                 printf("%s: witness exhausted\n", __func__);
1738                 return (NULL);
1739         }
1740         w_lock_list_free = lle->ll_next;
1741         mtx_unlock_spin(&w_mtx);
1742         bzero(lle, sizeof(*lle));
1743         return (lle);
1744 }
1745
1746 static void
1747 witness_lock_list_free(struct lock_list_entry *lle)
1748 {
1749
1750         mtx_lock_spin(&w_mtx);
1751         lle->ll_next = w_lock_list_free;
1752         w_lock_list_free = lle;
1753         mtx_unlock_spin(&w_mtx);
1754 }
1755
1756 static struct lock_instance *
1757 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1758 {
1759         struct lock_list_entry *lle;
1760         struct lock_instance *instance;
1761         int i;
1762
1763         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1764                 for (i = lle->ll_count - 1; i >= 0; i--) {
1765                         instance = &lle->ll_children[i];
1766                         if (instance->li_lock == lock)
1767                                 return (instance);
1768                 }
1769         return (NULL);
1770 }
1771
1772 static void
1773 witness_list_lock(struct lock_instance *instance)
1774 {
1775         struct lock_object *lock;
1776
1777         lock = instance->li_lock;
1778         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1779             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
1780         if (lock->lo_type != lock->lo_name)
1781                 printf(" (%s)", lock->lo_type);
1782         printf(" r = %d (%p) locked @ %s:%d\n",
1783             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1784             instance->li_line);
1785 }
1786
1787 #ifdef DDB
1788 static int
1789 witness_thread_has_locks(struct thread *td)
1790 {
1791
1792         return (td->td_sleeplocks != NULL);
1793 }
1794
1795 static int
1796 witness_proc_has_locks(struct proc *p)
1797 {
1798         struct thread *td;
1799
1800         FOREACH_THREAD_IN_PROC(p, td) {
1801                 if (witness_thread_has_locks(td))
1802                         return (1);
1803         }
1804         return (0);
1805 }
1806 #endif
1807
1808 int
1809 witness_list_locks(struct lock_list_entry **lock_list)
1810 {
1811         struct lock_list_entry *lle;
1812         int i, nheld;
1813
1814         nheld = 0;
1815         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1816                 for (i = lle->ll_count - 1; i >= 0; i--) {
1817                         witness_list_lock(&lle->ll_children[i]);
1818                         nheld++;
1819                 }
1820         return (nheld);
1821 }
1822
1823 /*
1824  * This is a bit risky at best.  We call this function when we have timed
1825  * out acquiring a spin lock, and we assume that the other CPU is stuck
1826  * with this lock held.  So, we go groveling around in the other CPU's
1827  * per-cpu data to try to find the lock instance for this spin lock to
1828  * see when it was last acquired.
1829  */
1830 void
1831 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1832 {
1833         struct lock_instance *instance;
1834         struct pcpu *pc;
1835
1836         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1837                 return;
1838         pc = pcpu_find(owner->td_oncpu);
1839         instance = find_instance(pc->pc_spinlocks, lock);
1840         if (instance != NULL)
1841                 witness_list_lock(instance);
1842 }
1843
1844 void
1845 witness_save(struct lock_object *lock, const char **filep, int *linep)
1846 {
1847         struct lock_list_entry *lock_list;
1848         struct lock_instance *instance;
1849         struct lock_class *class;
1850
1851         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1852         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1853                 return;
1854         class = LOCK_CLASS(lock);
1855         if (class->lc_flags & LC_SLEEPLOCK)
1856                 lock_list = curthread->td_sleeplocks;
1857         else {
1858                 if (witness_skipspin)
1859                         return;
1860                 lock_list = PCPU_GET(spinlocks);
1861         }
1862         instance = find_instance(lock_list, lock);
1863         if (instance == NULL)
1864                 panic("%s: lock (%s) %s not locked", __func__,
1865                     class->lc_name, lock->lo_name);
1866         *filep = instance->li_file;
1867         *linep = instance->li_line;
1868 }
1869
1870 void
1871 witness_restore(struct lock_object *lock, const char *file, int line)
1872 {
1873         struct lock_list_entry *lock_list;
1874         struct lock_instance *instance;
1875         struct lock_class *class;
1876
1877         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1878         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1879                 return;
1880         class = LOCK_CLASS(lock);
1881         if (class->lc_flags & LC_SLEEPLOCK)
1882                 lock_list = curthread->td_sleeplocks;
1883         else {
1884                 if (witness_skipspin)
1885                         return;
1886                 lock_list = PCPU_GET(spinlocks);
1887         }
1888         instance = find_instance(lock_list, lock);
1889         if (instance == NULL)
1890                 panic("%s: lock (%s) %s not locked", __func__,
1891                     class->lc_name, lock->lo_name);
1892         lock->lo_witness->w_file = file;
1893         lock->lo_witness->w_line = line;
1894         instance->li_file = file;
1895         instance->li_line = line;
1896 }
1897
1898 void
1899 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1900 {
1901 #ifdef INVARIANT_SUPPORT
1902         struct lock_instance *instance;
1903         struct lock_class *class;
1904
1905         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1906                 return;
1907         class = LOCK_CLASS(lock);
1908         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
1909                 instance = find_instance(curthread->td_sleeplocks, lock);
1910         else if ((class->lc_flags & LC_SPINLOCK) != 0)
1911                 instance = find_instance(PCPU_GET(spinlocks), lock);
1912         else {
1913                 panic("Lock (%s) %s is not sleep or spin!",
1914                     class->lc_name, lock->lo_name);
1915         }
1916         file = fixup_filename(file);
1917         switch (flags) {
1918         case LA_UNLOCKED:
1919                 if (instance != NULL)
1920                         panic("Lock (%s) %s locked @ %s:%d.",
1921                             class->lc_name, lock->lo_name, file, line);
1922                 break;
1923         case LA_LOCKED:
1924         case LA_LOCKED | LA_RECURSED:
1925         case LA_LOCKED | LA_NOTRECURSED:
1926         case LA_SLOCKED:
1927         case LA_SLOCKED | LA_RECURSED:
1928         case LA_SLOCKED | LA_NOTRECURSED:
1929         case LA_XLOCKED:
1930         case LA_XLOCKED | LA_RECURSED:
1931         case LA_XLOCKED | LA_NOTRECURSED:
1932                 if (instance == NULL) {
1933                         panic("Lock (%s) %s not locked @ %s:%d.",
1934                             class->lc_name, lock->lo_name, file, line);
1935                         break;
1936                 }
1937                 if ((flags & LA_XLOCKED) != 0 &&
1938                     (instance->li_flags & LI_EXCLUSIVE) == 0)
1939                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1940                             class->lc_name, lock->lo_name, file, line);
1941                 if ((flags & LA_SLOCKED) != 0 &&
1942                     (instance->li_flags & LI_EXCLUSIVE) != 0)
1943                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
1944                             class->lc_name, lock->lo_name, file, line);
1945                 if ((flags & LA_RECURSED) != 0 &&
1946                     (instance->li_flags & LI_RECURSEMASK) == 0)
1947                         panic("Lock (%s) %s not recursed @ %s:%d.",
1948                             class->lc_name, lock->lo_name, file, line);
1949                 if ((flags & LA_NOTRECURSED) != 0 &&
1950                     (instance->li_flags & LI_RECURSEMASK) != 0)
1951                         panic("Lock (%s) %s recursed @ %s:%d.",
1952                             class->lc_name, lock->lo_name, file, line);
1953                 break;
1954         default:
1955                 panic("Invalid lock assertion at %s:%d.", file, line);
1956
1957         }
1958 #endif  /* INVARIANT_SUPPORT */
1959 }
1960
1961 #ifdef DDB
1962 static void
1963 witness_list(struct thread *td)
1964 {
1965
1966         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1967         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1968
1969         if (witness_watch == 0)
1970                 return;
1971
1972         witness_list_locks(&td->td_sleeplocks);
1973
1974         /*
1975          * We only handle spinlocks if td == curthread.  This is somewhat broken
1976          * if td is currently executing on some other CPU and holds spin locks
1977          * as we won't display those locks.  If we had a MI way of getting
1978          * the per-cpu data for a given cpu then we could use
1979          * td->td_oncpu to get the list of spinlocks for this thread
1980          * and "fix" this.
1981          *
1982          * That still wouldn't really fix this unless we locked the scheduler
1983          * lock or stopped the other CPU to make sure it wasn't changing the
1984          * list out from under us.  It is probably best to just not try to
1985          * handle threads on other CPU's for now.
1986          */
1987         if (td == curthread && PCPU_GET(spinlocks) != NULL)
1988                 witness_list_locks(PCPU_PTR(spinlocks));
1989 }
1990
1991 DB_SHOW_COMMAND(locks, db_witness_list)
1992 {
1993         struct thread *td;
1994
1995         if (have_addr)
1996                 td = db_lookup_thread(addr, TRUE);
1997         else
1998                 td = kdb_thread;
1999         witness_list(td);
2000 }
2001
2002 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
2003 {
2004         struct thread *td;
2005         struct proc *p;
2006
2007         /*
2008          * It would be nice to list only threads and processes that actually
2009          * held sleep locks, but that information is currently not exported
2010          * by WITNESS.
2011          */
2012         FOREACH_PROC_IN_SYSTEM(p) {
2013                 if (!witness_proc_has_locks(p))
2014                         continue;
2015                 FOREACH_THREAD_IN_PROC(p, td) {
2016                         if (!witness_thread_has_locks(td))
2017                                 continue;
2018                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2019                             p->p_comm, td, td->td_tid);
2020                         witness_list(td);
2021                 }
2022         }
2023 }
2024
2025 DB_SHOW_COMMAND(witness, db_witness_display)
2026 {
2027
2028         witness_display(db_printf);
2029 }
2030 #endif