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         { "descriptor tables", &lock_class_mtx_spin },
 410 #endif
 411 #ifdef __sparc64__
 412         { "pcib_mtx", &lock_class_mtx_spin },
 413         { "rtc_mtx", &lock_class_mtx_spin },
 414 #endif
 415         { "scc_hwmtx", &lock_class_mtx_spin },
 416         { "uart_hwmtx", &lock_class_mtx_spin },
 417         { "fast_taskqueue", &lock_class_mtx_spin },
 418         { "intr table", &lock_class_mtx_spin },
 419 #ifdef  HWPMC_HOOKS
 420         { "pmc-per-proc", &lock_class_mtx_spin },
 421 #endif
 422         { "process slock", &lock_class_mtx_spin },
 423         { "sleepq chain", &lock_class_mtx_spin },
 424         { "umtx lock", &lock_class_mtx_spin },
 425         { "turnstile chain", &lock_class_mtx_spin },
 426         { "turnstile lock", &lock_class_mtx_spin },
 427         { "sched lock", &lock_class_mtx_spin },
 428         { "td_contested", &lock_class_mtx_spin },
 429         { "callout", &lock_class_mtx_spin },
 430         { "entropy harvest mutex", &lock_class_mtx_spin },
 431         { "syscons video lock", &lock_class_mtx_spin },
 432         { "time lock", &lock_class_mtx_spin },
 433         /*
 434          * leaf locks
 435          */
 436         { "allpmaps", &lock_class_mtx_spin },
 437         { "icu", &lock_class_mtx_spin },
 438 #ifdef SMP
 439         { "smp rendezvous", &lock_class_mtx_spin },
 440 #if defined(__i386__) || defined(__amd64__)
 441         { "tlb", &lock_class_mtx_spin },
 442 #endif
 443 #ifdef __sparc64__
 444         { "ipi", &lock_class_mtx_spin },
 445 #endif
 446 #endif
 447         { "clk", &lock_class_mtx_spin },
 448         { "mutex profiling lock", &lock_class_mtx_spin },
 449         { "kse lock", &lock_class_mtx_spin },
 450         { "zombie lock", &lock_class_mtx_spin },
 451         { "ALD Queue", &lock_class_mtx_spin },
 452 #ifdef __ia64__
 453         { "MCA spin lock", &lock_class_mtx_spin },
 454 #endif
 455 #if defined(__i386__) || defined(__amd64__)
 456         { "pcicfg", &lock_class_mtx_spin },
 457         { "NDIS thread lock", &lock_class_mtx_spin },
 458 #endif
 459         { "tw_osl_io_lock", &lock_class_mtx_spin },
 460         { "tw_osl_q_lock", &lock_class_mtx_spin },
 461         { "tw_cl_io_lock", &lock_class_mtx_spin },
 462         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 463         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 464 #ifdef  HWPMC_HOOKS
 465         { "pmc-leaf", &lock_class_mtx_spin },
 466 #endif
 467         { "blocked lock", &lock_class_mtx_spin },
 468         { NULL, NULL },
 469         { NULL, NULL }
 470 };
 471
 472 #ifdef BLESSING
 473 /*
 474  * Pairs of locks which have been blessed
 475  * Don't complain about order problems with blessed locks
 476  */
 477 static struct witness_blessed blessed_list[] = {
 478 };
 479 static int blessed_count =
 480         sizeof(blessed_list) / sizeof(struct witness_blessed);
 481 #endif
 482
 483 /*
 484  * List of locks initialized prior to witness being initialized whose
 485  * enrollment is currently deferred.
 486  */
 487 STAILQ_HEAD(, lock_object) pending_locks =
 488     STAILQ_HEAD_INITIALIZER(pending_locks);
 489
 490 /*
 491  * This global is set to 0 once it becomes safe to use the witness code.
 492  */
 493 static int witness_cold = 1;
 494
 495 /*
 496  * This global is set to 1 once the static lock orders have been enrolled
 497  * so that a warning can be issued for any spin locks enrolled later.
 498  */
 499 static int witness_spin_warn = 0;
 500
 501 /*
 502  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 503  * assume that the early boot is single-threaded at least until after this
 504  * routine is completed.
 505  */
 506 static void
 507 witness_initialize(void *dummy __unused)
 508 {
 509         struct lock_object *lock;
 510         struct witness_order_list_entry *order;
 511         struct witness *w, *w1;
 512         int i;
 513
 514         /*
 515          * We have to release Giant before initializing its witness
 516          * structure so that WITNESS doesn't get confused.
 517          */
 518         mtx_unlock(&Giant);
 519         mtx_assert(&Giant, MA_NOTOWNED);
 520
 521         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 522         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 523             MTX_NOWITNESS | MTX_NOPROFILE);
 524         for (i = 0; i < WITNESS_COUNT; i++)
 525                 witness_free(&w_data[i]);
 526         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
 527                 witness_child_free(&w_childdata[i]);
 528         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 529                 witness_lock_list_free(&w_locklistdata[i]);
 530
 531         /* First add in all the specified order lists. */
 532         for (order = order_lists; order->w_name != NULL; order++) {
 533                 w = enroll(order->w_name, order->w_class);
 534                 if (w == NULL)
 535                         continue;
 536                 w->w_file = "order list";
 537                 for (order++; order->w_name != NULL; order++) {
 538                         w1 = enroll(order->w_name, order->w_class);
 539                         if (w1 == NULL)
 540                                 continue;
 541                         w1->w_file = "order list";
 542                         if (!itismychild(w, w1))
 543                                 panic("Not enough memory for static orders!");
 544                         w = w1;
 545                 }
 546         }
 547         witness_spin_warn = 1;
 548
 549         /* Iterate through all locks and add them to witness. */
 550         while (!STAILQ_EMPTY(&pending_locks)) {
 551                 lock = STAILQ_FIRST(&pending_locks);
 552                 STAILQ_REMOVE_HEAD(&pending_locks, lo_list);
 553                 KASSERT(lock->lo_flags & LO_WITNESS,
 554                     ("%s: lock %s is on pending list but not LO_WITNESS",
 555                     __func__, lock->lo_name));
 556                 lock->lo_witness = enroll(lock->lo_type, LOCK_CLASS(lock));
 557         }
 558
 559         /* Mark the witness code as being ready for use. */
 560         witness_cold = 0;
 561
 562         mtx_lock(&Giant);
 563 }
 564 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
 565
 566 static int
 567 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
 568 {
 569         int error, value;
 570
 571         value = witness_watch;
 572         error = sysctl_handle_int(oidp, &value, 0, req);
 573         if (error != 0 || req->newptr == NULL)
 574                 return (error);
 575         if (value == witness_watch)
 576                 return (0);
 577         if (value != 0)
 578                 return (EINVAL);
 579         witness_watch = 0;
 580         return (0);
 581 }
 582
 583 void
 584 witness_init(struct lock_object *lock)
 585 {
 586         struct lock_class *class;
 587
 588         /* Various sanity checks. */
 589         class = LOCK_CLASS(lock);
 590         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 591             (class->lc_flags & LC_RECURSABLE) == 0)
 592                 panic("%s: lock (%s) %s can not be recursable", __func__,
 593                     class->lc_name, lock->lo_name);
 594         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 595             (class->lc_flags & LC_SLEEPABLE) == 0)
 596                 panic("%s: lock (%s) %s can not be sleepable", __func__,
 597                     class->lc_name, lock->lo_name);
 598         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 599             (class->lc_flags & LC_UPGRADABLE) == 0)
 600                 panic("%s: lock (%s) %s can not be upgradable", __func__,
 601                     class->lc_name, lock->lo_name);
 602
 603         /*
 604          * If we shouldn't watch this lock, then just clear lo_witness.
 605          * Otherwise, if witness_cold is set, then it is too early to
 606          * enroll this lock, so defer it to witness_initialize() by adding
 607          * it to the pending_locks list.  If it is not too early, then enroll
 608          * the lock now.
 609          */
 610         if (witness_watch == 0 || panicstr != NULL ||
 611             (lock->lo_flags & LO_WITNESS) == 0)
 612                 lock->lo_witness = NULL;
 613         else if (witness_cold) {
 614                 STAILQ_INSERT_TAIL(&pending_locks, lock, lo_list);
 615                 lock->lo_flags |= LO_ENROLLPEND;
 616         } else
 617                 lock->lo_witness = enroll(lock->lo_type, class);
 618 }
 619
 620 void
 621 witness_destroy(struct lock_object *lock)
 622 {
 623         struct lock_class *class;
 624         struct witness *w;
 625
 626         class = LOCK_CLASS(lock);
 627         if (witness_cold)
 628                 panic("lock (%s) %s destroyed while witness_cold",
 629                     class->lc_name, lock->lo_name);
 630
 631         /* XXX: need to verify that no one holds the lock */
 632         if ((lock->lo_flags & (LO_WITNESS | LO_ENROLLPEND)) == LO_WITNESS &&
 633             lock->lo_witness != NULL) {
 634                 w = lock->lo_witness;
 635                 mtx_lock_spin(&w_mtx);
 636                 MPASS(w->w_refcount > 0);
 637                 w->w_refcount--;
 638
 639                 /*
 640                  * Lock is already released if we have an allocation failure
 641                  * and depart() fails.
 642                  */
 643                 if (w->w_refcount != 0 || depart(w))
 644                         mtx_unlock_spin(&w_mtx);
 645         }
 646
 647         /*
 648          * If this lock is destroyed before witness is up and running,
 649          * remove it from the pending list.
 650          */
 651         if (lock->lo_flags & LO_ENROLLPEND) {
 652                 STAILQ_REMOVE(&pending_locks, lock, lock_object, lo_list);
 653                 lock->lo_flags &= ~LO_ENROLLPEND;
 654         }
 655 }
 656
 657 #ifdef DDB
 658 static void
 659 witness_levelall (void)
 660 {
 661         struct witness_list *list;
 662         struct witness *w, *w1;
 663
 664         /*
 665          * First clear all levels.
 666          */
 667         STAILQ_FOREACH(w, &w_all, w_list) {
 668                 w->w_level = 0;
 669         }
 670
 671         /*
 672          * Look for locks with no parent and level all their descendants.
 673          */
 674         STAILQ_FOREACH(w, &w_all, w_list) {
 675                 /*
 676                  * This is just an optimization, technically we could get
 677                  * away just walking the all list each time.
 678                  */
 679                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
 680                         list = &w_sleep;
 681                 else
 682                         list = &w_spin;
 683                 STAILQ_FOREACH(w1, list, w_typelist) {
 684                         if (isitmychild(w1, w))
 685                                 goto skip;
 686                 }
 687                 witness_leveldescendents(w, 0);
 688         skip:
 689                 ;       /* silence GCC 3.x */
 690         }
 691 }
 692
 693 static void
 694 witness_leveldescendents(struct witness *parent, int level)
 695 {
 696         struct witness_child_list_entry *wcl;
 697         int i;
 698
 699         if (parent->w_level < level)
 700                 parent->w_level = level;
 701         level++;
 702         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 703                 for (i = 0; i < wcl->wcl_count; i++)
 704                         witness_leveldescendents(wcl->wcl_children[i], level);
 705 }
 706
 707 static void
 708 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
 709                            struct witness *parent, int indent)
 710 {
 711         struct witness_child_list_entry *wcl;
 712         int i, level;
 713
 714         level = parent->w_level;
 715         prnt("%-2d", level);
 716         for (i = 0; i < indent; i++)
 717                 prnt(" ");
 718         if (parent->w_refcount > 0)
 719                 prnt("%s", parent->w_name);
 720         else
 721                 prnt("(dead)");
 722         if (parent->w_displayed) {
 723                 prnt(" -- (already displayed)\n");
 724                 return;
 725         }
 726         parent->w_displayed = 1;
 727         if (parent->w_refcount > 0) {
 728                 if (parent->w_file != NULL)
 729                         prnt(" -- last acquired @ %s:%d", parent->w_file,
 730                             parent->w_line);
 731         }
 732         prnt("\n");
 733         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 734                 for (i = 0; i < wcl->wcl_count; i++)
 735                             witness_displaydescendants(prnt,
 736                                 wcl->wcl_children[i], indent + 1);
 737 }
 738
 739 static void
 740 witness_display_list(void(*prnt)(const char *fmt, ...),
 741                      struct witness_list *list)
 742 {
 743         struct witness *w;
 744
 745         STAILQ_FOREACH(w, list, w_typelist) {
 746                 if (w->w_file == NULL || w->w_level > 0)
 747                         continue;
 748                 /*
 749                  * This lock has no anscestors, display its descendants.
 750                  */
 751                 witness_displaydescendants(prnt, w, 0);
 752         }
 753 }
 754
 755 static void
 756 witness_display(void(*prnt)(const char *fmt, ...))
 757 {
 758         struct witness *w;
 759
 760         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 761         witness_levelall();
 762
 763         /* Clear all the displayed flags. */
 764         STAILQ_FOREACH(w, &w_all, w_list) {
 765                 w->w_displayed = 0;
 766         }
 767
 768         /*
 769          * First, handle sleep locks which have been acquired at least
 770          * once.
 771          */
 772         prnt("Sleep locks:\n");
 773         witness_display_list(prnt, &w_sleep);
 774
 775         /*
 776          * Now do spin locks which have been acquired at least once.
 777          */
 778         prnt("\nSpin locks:\n");
 779         witness_display_list(prnt, &w_spin);
 780
 781         /*
 782          * Finally, any locks which have not been acquired yet.
 783          */
 784         prnt("\nLocks which were never acquired:\n");
 785         STAILQ_FOREACH(w, &w_all, w_list) {
 786                 if (w->w_file != NULL || w->w_refcount == 0)
 787                         continue;
 788                 prnt("%s\n", w->w_name);
 789         }
 790 }
 791 #endif /* DDB */
 792
 793 /* Trim useless garbage from filenames. */
 794 static const char *
 795 fixup_filename(const char *file)
 796 {
 797
 798         if (file == NULL)
 799                 return (NULL);
 800         while (strncmp(file, "../", 3) == 0)
 801                 file += 3;
 802         return (file);
 803 }
 804
 805 int
 806 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
 807 {
 808
 809         if (witness_watch == 0 || panicstr != NULL)
 810                 return (0);
 811
 812         /* Require locks that witness knows about. */
 813         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
 814             lock2->lo_witness == NULL)
 815                 return (EINVAL);
 816
 817         MPASS(!mtx_owned(&w_mtx));
 818         mtx_lock_spin(&w_mtx);
 819
 820         /*
 821          * If we already have either an explicit or implied lock order that
 822          * is the other way around, then return an error.
 823          */
 824         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
 825                 mtx_unlock_spin(&w_mtx);
 826                 return (EDOOFUS);
 827         }
 828
 829         /* Try to add the new order. */
 830         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 831             lock2->lo_type, lock1->lo_type);
 832         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
 833                 return (ENOMEM);
 834         mtx_unlock_spin(&w_mtx);
 835         return (0);
 836 }
 837
 838 void
 839 witness_checkorder(struct lock_object *lock, int flags, const char *file,
 840     int line)
 841 {
 842         struct lock_list_entry **lock_list, *lle;
 843         struct lock_instance *lock1, *lock2;
 844         struct lock_class *class;
 845         struct witness *w, *w1;
 846         struct thread *td;
 847         int i, j;
 848
 849         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 850             panicstr != NULL)
 851                 return;
 852
 853         /*
 854          * Try locks do not block if they fail to acquire the lock, thus
 855          * there is no danger of deadlocks or of switching while holding a
 856          * spin lock if we acquire a lock via a try operation.  This
 857          * function shouldn't even be called for try locks, so panic if
 858          * that happens.
 859          */
 860         if (flags & LOP_TRYLOCK)
 861                 panic("%s should not be called for try lock operations",
 862                     __func__);
 863
 864         w = lock->lo_witness;
 865         class = LOCK_CLASS(lock);
 866         td = curthread;
 867         file = fixup_filename(file);
 868
 869         if (class->lc_flags & LC_SLEEPLOCK) {
 870                 /*
 871                  * Since spin locks include a critical section, this check
 872                  * implicitly enforces a lock order of all sleep locks before
 873                  * all spin locks.
 874                  */
 875                 if (td->td_critnest != 0 && !kdb_active)
 876                         panic("blockable sleep lock (%s) %s @ %s:%d",
 877                             class->lc_name, lock->lo_name, file, line);
 878
 879                 /*
 880                  * If this is the first lock acquired then just return as
 881                  * no order checking is needed.
 882                  */
 883                 if (td->td_sleeplocks == NULL)
 884                         return;
 885                 lock_list = &td->td_sleeplocks;
 886         } else {
 887                 /*
 888                  * If this is the first lock, just return as no order
 889                  * checking is needed.  We check this in both if clauses
 890                  * here as unifying the check would require us to use a
 891                  * critical section to ensure we don't migrate while doing
 892                  * the check.  Note that if this is not the first lock, we
 893                  * are already in a critical section and are safe for the
 894                  * rest of the check.
 895                  */
 896                 if (PCPU_GET(spinlocks) == NULL)
 897                         return;
 898                 lock_list = PCPU_PTR(spinlocks);
 899         }
 900
 901         /*
 902          * Check to see if we are recursing on a lock we already own.  If
 903          * so, make sure that we don't mismatch exclusive and shared lock
 904          * acquires.
 905          */
 906         lock1 = find_instance(*lock_list, lock);
 907         if (lock1 != NULL) {
 908                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
 909                     (flags & LOP_EXCLUSIVE) == 0) {
 910                         printf("shared lock of (%s) %s @ %s:%d\n",
 911                             class->lc_name, lock->lo_name, file, line);
 912                         printf("while exclusively locked from %s:%d\n",
 913                             lock1->li_file, lock1->li_line);
 914                         panic("share->excl");
 915                 }
 916                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
 917                     (flags & LOP_EXCLUSIVE) != 0) {
 918                         printf("exclusive lock of (%s) %s @ %s:%d\n",
 919                             class->lc_name, lock->lo_name, file, line);
 920                         printf("while share locked from %s:%d\n",
 921                             lock1->li_file, lock1->li_line);
 922                         panic("excl->share");
 923                 }
 924                 return;
 925         }
 926
 927         /*
 928          * Try locks do not block if they fail to acquire the lock, thus
 929          * there is no danger of deadlocks or of switching while holding a
 930          * spin lock if we acquire a lock via a try operation.
 931          */
 932         if (flags & LOP_TRYLOCK)
 933                 return;
 934
 935         /*
 936          * Check for duplicate locks of the same type.  Note that we only
 937          * have to check for this on the last lock we just acquired.  Any
 938          * other cases will be caught as lock order violations.
 939          */
 940         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 941         w1 = lock1->li_lock->lo_witness;
 942         if (w1 == w) {
 943                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
 944                     (flags & LOP_DUPOK))
 945                         return;
 946                 w->w_same_squawked = 1;
 947                 printf("acquiring duplicate lock of same type: \"%s\"\n",
 948                         lock->lo_type);
 949                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
 950                     lock1->li_file, lock1->li_line);
 951                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
 952 #ifdef KDB
 953                 goto debugger;
 954 #else
 955                 return;
 956 #endif
 957         }
 958         MPASS(!mtx_owned(&w_mtx));
 959         mtx_lock_spin(&w_mtx);
 960         /*
 961          * If we know that the the lock we are acquiring comes after
 962          * the lock we most recently acquired in the lock order tree,
 963          * then there is no need for any further checks.
 964          */
 965         if (isitmychild(w1, w)) {
 966                 mtx_unlock_spin(&w_mtx);
 967                 return;
 968         }
 969         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
 970                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
 971
 972                         MPASS(j < WITNESS_COUNT);
 973                         lock1 = &lle->ll_children[i];
 974                         w1 = lock1->li_lock->lo_witness;
 975
 976                         /*
 977                          * If this lock doesn't undergo witness checking,
 978                          * then skip it.
 979                          */
 980                         if (w1 == NULL) {
 981                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
 982                                     ("lock missing witness structure"));
 983                                 continue;
 984                         }
 985                         /*
 986                          * If we are locking Giant and this is a sleepable
 987                          * lock, then skip it.
 988                          */
 989                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
 990                             lock == &Giant.lock_object)
 991                                 continue;
 992                         /*
 993                          * If we are locking a sleepable lock and this lock
 994                          * is Giant, then skip it.
 995                          */
 996                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 997                             lock1->li_lock == &Giant.lock_object)
 998                                 continue;
 999                         /*
1000                          * If we are locking a sleepable lock and this lock
1001                          * isn't sleepable, we want to treat it as a lock
1002                          * order violation to enfore a general lock order of
1003                          * sleepable locks before non-sleepable locks.
1004                          */
1005                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1006                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1007                                 goto reversal;
1008                         /*
1009                          * If we are locking Giant and this is a non-sleepable
1010                          * lock, then treat it as a reversal.
1011                          */
1012                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1013                             lock == &Giant.lock_object)
1014                                 goto reversal;
1015                         /*
1016                          * Check the lock order hierarchy for a reveresal.
1017                          */
1018                         if (!isitmydescendant(w, w1))
1019                                 continue;
1020                 reversal:
1021                         /*
1022                          * We have a lock order violation, check to see if it
1023                          * is allowed or has already been yelled about.
1024                          */
1025                         mtx_unlock_spin(&w_mtx);
1026 #ifdef BLESSING
1027                         /*
1028                          * If the lock order is blessed, just bail.  We don't
1029                          * look for other lock order violations though, which
1030                          * may be a bug.
1031                          */
1032                         if (blessed(w, w1))
1033                                 return;
1034 #endif
1035                         if (lock1->li_lock == &Giant.lock_object) {
1036                                 if (w1->w_Giant_squawked)
1037                                         return;
1038                                 else
1039                                         w1->w_Giant_squawked = 1;
1040                         } else {
1041                                 if (w1->w_other_squawked)
1042                                         return;
1043                                 else
1044                                         w1->w_other_squawked = 1;
1045                         }
1046                         /*
1047                          * Ok, yell about it.
1048                          */
1049                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1050                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1051                                 printf(
1052                 "lock order reversal: (sleepable after non-sleepable)\n");
1053                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1054                             && lock == &Giant.lock_object)
1055                                 printf(
1056                 "lock order reversal: (Giant after non-sleepable)\n");
1057                         else
1058                                 printf("lock order reversal:\n");
1059                         /*
1060                          * Try to locate an earlier lock with
1061                          * witness w in our list.
1062                          */
1063                         do {
1064                                 lock2 = &lle->ll_children[i];
1065                                 MPASS(lock2->li_lock != NULL);
1066                                 if (lock2->li_lock->lo_witness == w)
1067                                         break;
1068                                 if (i == 0 && lle->ll_next != NULL) {
1069                                         lle = lle->ll_next;
1070                                         i = lle->ll_count - 1;
1071                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1072                                 } else
1073                                         i--;
1074                         } while (i >= 0);
1075                         if (i < 0) {
1076                                 printf(" 1st %p %s (%s) @ %s:%d\n",
1077                                     lock1->li_lock, lock1->li_lock->lo_name,
1078                                     lock1->li_lock->lo_type, lock1->li_file,
1079                                     lock1->li_line);
1080                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1081                                     lock->lo_name, lock->lo_type, file, line);
1082                         } else {
1083                                 printf(" 1st %p %s (%s) @ %s:%d\n",
1084                                     lock2->li_lock, lock2->li_lock->lo_name,
1085                                     lock2->li_lock->lo_type, lock2->li_file,
1086                                     lock2->li_line);
1087                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
1088                                     lock1->li_lock, lock1->li_lock->lo_name,
1089                                     lock1->li_lock->lo_type, lock1->li_file,
1090                                     lock1->li_line);
1091                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1092                                     lock->lo_name, lock->lo_type, file, line);
1093                         }
1094 #ifdef KDB
1095                         goto debugger;
1096 #else
1097                         return;
1098 #endif
1099                 }
1100         }
1101         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
1102         /*
1103          * If requested, build a new lock order.  However, don't build a new
1104          * relationship between a sleepable lock and Giant if it is in the
1105          * wrong direction.  The correct lock order is that sleepable locks
1106          * always come before Giant.
1107          */
1108         if (flags & LOP_NEWORDER &&
1109             !(lock1->li_lock == &Giant.lock_object &&
1110             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1111                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1112                     lock->lo_type, lock1->li_lock->lo_type);
1113                 if (!itismychild(lock1->li_lock->lo_witness, w))
1114                         /* Witness is dead. */
1115                         return;
1116         }
1117         mtx_unlock_spin(&w_mtx);
1118         return;
1119
1120 #ifdef KDB
1121 debugger:
1122         if (witness_trace)
1123                 kdb_backtrace();
1124         if (witness_kdb)
1125                 kdb_enter(__func__);
1126 #endif
1127 }
1128
1129 void
1130 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1131 {
1132         struct lock_list_entry **lock_list, *lle;
1133         struct lock_instance *instance;
1134         struct witness *w;
1135         struct thread *td;
1136
1137         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1138             panicstr != NULL)
1139                 return;
1140         w = lock->lo_witness;
1141         td = curthread;
1142         file = fixup_filename(file);
1143
1144         /* Determine lock list for this lock. */
1145         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1146                 lock_list = &td->td_sleeplocks;
1147         else
1148                 lock_list = PCPU_PTR(spinlocks);
1149
1150         /* Check to see if we are recursing on a lock we already own. */
1151         instance = find_instance(*lock_list, lock);
1152         if (instance != NULL) {
1153                 instance->li_flags++;
1154                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1155                     td->td_proc->p_pid, lock->lo_name,
1156                     instance->li_flags & LI_RECURSEMASK);
1157                 instance->li_file = file;
1158                 instance->li_line = line;
1159                 return;
1160         }
1161
1162         /* Update per-witness last file and line acquire. */
1163         w->w_file = file;
1164         w->w_line = line;
1165
1166         /* Find the next open lock instance in the list and fill it. */
1167         lle = *lock_list;
1168         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1169                 lle = witness_lock_list_get();
1170                 if (lle == NULL)
1171                         return;
1172                 lle->ll_next = *lock_list;
1173                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1174                     td->td_proc->p_pid, lle);
1175                 *lock_list = lle;
1176         }
1177         instance = &lle->ll_children[lle->ll_count++];
1178         instance->li_lock = lock;
1179         instance->li_line = line;
1180         instance->li_file = file;
1181         if ((flags & LOP_EXCLUSIVE) != 0)
1182                 instance->li_flags = LI_EXCLUSIVE;
1183         else
1184                 instance->li_flags = 0;
1185         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1186             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1187 }
1188
1189 void
1190 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1191 {
1192         struct lock_instance *instance;
1193         struct lock_class *class;
1194
1195         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1196         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1197                 return;
1198         class = LOCK_CLASS(lock);
1199         file = fixup_filename(file);
1200         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1201                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1202                     class->lc_name, lock->lo_name, file, line);
1203         if ((flags & LOP_TRYLOCK) == 0)
1204                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1205                     lock->lo_name, file, line);
1206         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1207                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1208                     class->lc_name, lock->lo_name, file, line);
1209         instance = find_instance(curthread->td_sleeplocks, lock);
1210         if (instance == NULL)
1211                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1212                     class->lc_name, lock->lo_name, file, line);
1213         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1214                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1215                     class->lc_name, lock->lo_name, file, line);
1216         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1217                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1218                     class->lc_name, lock->lo_name,
1219                     instance->li_flags & LI_RECURSEMASK, file, line);
1220         instance->li_flags |= LI_EXCLUSIVE;
1221 }
1222
1223 void
1224 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1225     int line)
1226 {
1227         struct lock_instance *instance;
1228         struct lock_class *class;
1229
1230         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1231         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1232                 return;
1233         class = LOCK_CLASS(lock);
1234         file = fixup_filename(file);
1235         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1236                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1237                     class->lc_name, lock->lo_name, file, line);
1238         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1239                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1240                     class->lc_name, lock->lo_name, file, line);
1241         instance = find_instance(curthread->td_sleeplocks, lock);
1242         if (instance == NULL)
1243                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1244                     class->lc_name, lock->lo_name, file, line);
1245         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1246                 panic("downgrade of shared lock (%s) %s @ %s:%d",
1247                     class->lc_name, lock->lo_name, file, line);
1248         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1249                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1250                     class->lc_name, lock->lo_name,
1251                     instance->li_flags & LI_RECURSEMASK, file, line);
1252         instance->li_flags &= ~LI_EXCLUSIVE;
1253 }
1254
1255 void
1256 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1257 {
1258         struct lock_list_entry **lock_list, *lle;
1259         struct lock_instance *instance;
1260         struct lock_class *class;
1261         struct thread *td;
1262         register_t s;
1263         int i, j;
1264
1265         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1266             panicstr != NULL)
1267                 return;
1268         td = curthread;
1269         class = LOCK_CLASS(lock);
1270         file = fixup_filename(file);
1271
1272         /* Find lock instance associated with this lock. */
1273         if (class->lc_flags & LC_SLEEPLOCK)
1274                 lock_list = &td->td_sleeplocks;
1275         else
1276                 lock_list = PCPU_PTR(spinlocks);
1277         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1278                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1279                         instance = &(*lock_list)->ll_children[i];
1280                         if (instance->li_lock == lock)
1281                                 goto found;
1282                 }
1283         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1284             file, line);
1285 found:
1286
1287         /* First, check for shared/exclusive mismatches. */
1288         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1289             (flags & LOP_EXCLUSIVE) == 0) {
1290                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1291                     lock->lo_name, file, line);
1292                 printf("while exclusively locked from %s:%d\n",
1293                     instance->li_file, instance->li_line);
1294                 panic("excl->ushare");
1295         }
1296         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1297             (flags & LOP_EXCLUSIVE) != 0) {
1298                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1299                     lock->lo_name, file, line);
1300                 printf("while share locked from %s:%d\n", instance->li_file,
1301                     instance->li_line);
1302                 panic("share->uexcl");
1303         }
1304
1305         /* If we are recursed, unrecurse. */
1306         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1307                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1308                     td->td_proc->p_pid, instance->li_lock->lo_name,
1309                     instance->li_flags);
1310                 instance->li_flags--;
1311                 return;
1312         }
1313
1314         /* Otherwise, remove this item from the list. */
1315         s = intr_disable();
1316         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1317             td->td_proc->p_pid, instance->li_lock->lo_name,
1318             (*lock_list)->ll_count - 1);
1319         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1320                 (*lock_list)->ll_children[j] =
1321                     (*lock_list)->ll_children[j + 1];
1322         (*lock_list)->ll_count--;
1323         intr_restore(s);
1324
1325         /* If this lock list entry is now empty, free it. */
1326         if ((*lock_list)->ll_count == 0) {
1327                 lle = *lock_list;
1328                 *lock_list = lle->ll_next;
1329                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1330                     td->td_proc->p_pid, lle);
1331                 witness_lock_list_free(lle);
1332         }
1333 }
1334
1335 /*
1336  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1337  * exempt Giant and sleepable locks from the checks as well.  If any
1338  * non-exempt locks are held, then a supplied message is printed to the
1339  * console along with a list of the offending locks.  If indicated in the
1340  * flags then a failure results in a panic as well.
1341  */
1342 int
1343 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1344 {
1345         struct lock_list_entry *lle;
1346         struct lock_instance *lock1;
1347         struct thread *td;
1348         va_list ap;
1349         int i, n;
1350
1351         if (witness_cold || witness_watch == 0 || panicstr != NULL)
1352                 return (0);
1353         n = 0;
1354         td = curthread;
1355         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1356                 for (i = lle->ll_count - 1; i >= 0; i--) {
1357                         lock1 = &lle->ll_children[i];
1358                         if (lock1->li_lock == lock)
1359                                 continue;
1360                         if (flags & WARN_GIANTOK &&
1361                             lock1->li_lock == &Giant.lock_object)
1362                                 continue;
1363                         if (flags & WARN_SLEEPOK &&
1364                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1365                                 continue;
1366                         if (n == 0) {
1367                                 va_start(ap, fmt);
1368                                 vprintf(fmt, ap);
1369                                 va_end(ap);
1370                                 printf(" with the following");
1371                                 if (flags & WARN_SLEEPOK)
1372                                         printf(" non-sleepable");
1373                                 printf(" locks held:\n");
1374                         }
1375                         n++;
1376                         witness_list_lock(lock1);
1377                 }
1378         if (PCPU_GET(spinlocks) != NULL) {
1379                 /*
1380                  * Since we already hold a spinlock preemption is
1381                  * already blocked.
1382                  */
1383                 if (n == 0) {
1384                         va_start(ap, fmt);
1385                         vprintf(fmt, ap);
1386                         va_end(ap);
1387                         printf(" with the following");
1388                         if (flags & WARN_SLEEPOK)
1389                                 printf(" non-sleepable");
1390                         printf(" locks held:\n");
1391                 }
1392                 n += witness_list_locks(PCPU_PTR(spinlocks));
1393         }
1394         if (flags & WARN_PANIC && n)
1395                 panic("witness_warn");
1396 #ifdef KDB
1397         else if (witness_kdb && n)
1398                 kdb_enter(__func__);
1399         else if (witness_trace && n)
1400                 kdb_backtrace();
1401 #endif
1402         return (n);
1403 }
1404
1405 const char *
1406 witness_file(struct lock_object *lock)
1407 {
1408         struct witness *w;
1409
1410         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1411                 return ("?");
1412         w = lock->lo_witness;
1413         return (w->w_file);
1414 }
1415
1416 int
1417 witness_line(struct lock_object *lock)
1418 {
1419         struct witness *w;
1420
1421         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1422                 return (0);
1423         w = lock->lo_witness;
1424         return (w->w_line);
1425 }
1426
1427 static struct witness *
1428 enroll(const char *description, struct lock_class *lock_class)
1429 {
1430         struct witness *w;
1431
1432         if (witness_watch == 0 || panicstr != NULL)
1433                 return (NULL);
1434         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1435                 return (NULL);
1436         mtx_lock_spin(&w_mtx);
1437         STAILQ_FOREACH(w, &w_all, w_list) {
1438                 if (w->w_name == description || (w->w_refcount > 0 &&
1439                     strcmp(description, w->w_name) == 0)) {
1440                         w->w_refcount++;
1441                         mtx_unlock_spin(&w_mtx);
1442                         if (lock_class != w->w_class)
1443                                 panic(
1444                                 "lock (%s) %s does not match earlier (%s) lock",
1445                                     description, lock_class->lc_name,
1446                                     w->w_class->lc_name);
1447                         return (w);
1448                 }
1449         }
1450         if ((w = witness_get()) == NULL)
1451                 goto out;
1452         w->w_name = description;
1453         w->w_class = lock_class;
1454         w->w_refcount = 1;
1455         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1456         if (lock_class->lc_flags & LC_SPINLOCK) {
1457                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1458                 w_spin_cnt++;
1459         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1460                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1461                 w_sleep_cnt++;
1462         } else {
1463                 mtx_unlock_spin(&w_mtx);
1464                 panic("lock class %s is not sleep or spin",
1465                     lock_class->lc_name);
1466         }
1467         mtx_unlock_spin(&w_mtx);
1468 out:
1469         /*
1470          * We issue a warning for any spin locks not defined in the static
1471          * order list as a way to discourage their use (folks should really
1472          * be using non-spin mutexes most of the time).  However, several
1473          * 3rd part device drivers use spin locks because that is all they
1474          * have available on Windows and Linux and they think that normal
1475          * mutexes are insufficient.
1476          */
1477         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
1478                 printf("WITNESS: spin lock %s not in order list\n",
1479                     description);
1480         return (w);
1481 }
1482
1483 /* Don't let the door bang you on the way out... */
1484 static int
1485 depart(struct witness *w)
1486 {
1487         struct witness_child_list_entry *wcl, *nwcl;
1488         struct witness_list *list;
1489         struct witness *parent;
1490
1491         MPASS(w->w_refcount == 0);
1492         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1493                 list = &w_sleep;
1494                 w_sleep_cnt--;
1495         } else {
1496                 list = &w_spin;
1497                 w_spin_cnt--;
1498         }
1499         /*
1500          * First, we run through the entire tree looking for any
1501          * witnesses that the outgoing witness is a child of.  For
1502          * each parent that we find, we reparent all the direct
1503          * children of the outgoing witness to its parent.
1504          */
1505         STAILQ_FOREACH(parent, list, w_typelist) {
1506                 if (!isitmychild(parent, w))
1507                         continue;
1508                 removechild(parent, w);
1509         }
1510
1511         /*
1512          * Now we go through and free up the child list of the
1513          * outgoing witness.
1514          */
1515         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1516                 nwcl = wcl->wcl_next;
1517                 w_child_cnt--;
1518                 witness_child_free(wcl);
1519         }
1520
1521         /*
1522          * Detach from various lists and free.
1523          */
1524         STAILQ_REMOVE(list, w, witness, w_typelist);
1525         STAILQ_REMOVE(&w_all, w, witness, w_list);
1526         witness_free(w);
1527
1528         return (1);
1529 }
1530
1531 /*
1532  * Add "child" as a direct child of "parent".  Returns false if
1533  * we fail due to out of memory.
1534  */
1535 static int
1536 insertchild(struct witness *parent, struct witness *child)
1537 {
1538         struct witness_child_list_entry **wcl;
1539
1540         MPASS(child != NULL && parent != NULL);
1541
1542         /*
1543          * Insert "child" after "parent"
1544          */
1545         wcl = &parent->w_children;
1546         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1547                 wcl = &(*wcl)->wcl_next;
1548         if (*wcl == NULL) {
1549                 *wcl = witness_child_get();
1550                 if (*wcl == NULL)
1551                         return (0);
1552                 w_child_cnt++;
1553         }
1554         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1555
1556         return (1);
1557 }
1558
1559
1560 static int
1561 itismychild(struct witness *parent, struct witness *child)
1562 {
1563         struct witness_list *list;
1564
1565         MPASS(child != NULL && parent != NULL);
1566         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1567             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1568                 panic(
1569                 "%s: parent (%s) and child (%s) are not the same lock type",
1570                     __func__, parent->w_class->lc_name,
1571                     child->w_class->lc_name);
1572
1573         if (!insertchild(parent, child))
1574                 return (0);
1575
1576         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1577                 list = &w_sleep;
1578         else
1579                 list = &w_spin;
1580         return (1);
1581 }
1582
1583 static void
1584 removechild(struct witness *parent, struct witness *child)
1585 {
1586         struct witness_child_list_entry **wcl, *wcl1;
1587         int i;
1588
1589         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1590                 for (i = 0; i < (*wcl)->wcl_count; i++)
1591                         if ((*wcl)->wcl_children[i] == child)
1592                                 goto found;
1593         return;
1594 found:
1595         (*wcl)->wcl_count--;
1596         if ((*wcl)->wcl_count > i)
1597                 (*wcl)->wcl_children[i] =
1598                     (*wcl)->wcl_children[(*wcl)->wcl_count];
1599         MPASS((*wcl)->wcl_children[i] != NULL);
1600         if ((*wcl)->wcl_count != 0)
1601                 return;
1602         wcl1 = *wcl;
1603         *wcl = wcl1->wcl_next;
1604         w_child_cnt--;
1605         witness_child_free(wcl1);
1606 }
1607
1608 static int
1609 isitmychild(struct witness *parent, struct witness *child)
1610 {
1611         struct witness_child_list_entry *wcl;
1612         int i;
1613
1614         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1615                 for (i = 0; i < wcl->wcl_count; i++) {
1616                         if (wcl->wcl_children[i] == child)
1617                                 return (1);
1618                 }
1619         }
1620         return (0);
1621 }
1622
1623 static int
1624 isitmydescendant(struct witness *parent, struct witness *child)
1625 {
1626         struct witness_child_list_entry *wcl;
1627         int i, j;
1628
1629         if (isitmychild(parent, child))
1630                 return (1);
1631         j = 0;
1632         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1633                 MPASS(j < 1000);
1634                 for (i = 0; i < wcl->wcl_count; i++) {
1635                         if (isitmydescendant(wcl->wcl_children[i], child))
1636                                 return (1);
1637                 }
1638                 j++;
1639         }
1640         return (0);
1641 }
1642
1643 #ifdef BLESSING
1644 static int
1645 blessed(struct witness *w1, struct witness *w2)
1646 {
1647         int i;
1648         struct witness_blessed *b;
1649
1650         for (i = 0; i < blessed_count; i++) {
1651                 b = &blessed_list[i];
1652                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1653                         if (strcmp(w2->w_name, b->b_lock2) == 0)
1654                                 return (1);
1655                         continue;
1656                 }
1657                 if (strcmp(w1->w_name, b->b_lock2) == 0)
1658                         if (strcmp(w2->w_name, b->b_lock1) == 0)
1659                                 return (1);
1660         }
1661         return (0);
1662 }
1663 #endif
1664
1665 static struct witness *
1666 witness_get(void)
1667 {
1668         struct witness *w;
1669
1670         if (witness_watch == 0) {
1671                 mtx_unlock_spin(&w_mtx);
1672                 return (NULL);
1673         }
1674         if (STAILQ_EMPTY(&w_free)) {
1675                 witness_watch = 0;
1676                 mtx_unlock_spin(&w_mtx);
1677                 printf("%s: witness exhausted\n", __func__);
1678                 return (NULL);
1679         }
1680         w = STAILQ_FIRST(&w_free);
1681         STAILQ_REMOVE_HEAD(&w_free, w_list);
1682         w_free_cnt--;
1683         bzero(w, sizeof(*w));
1684         return (w);
1685 }
1686
1687 static void
1688 witness_free(struct witness *w)
1689 {
1690
1691         STAILQ_INSERT_HEAD(&w_free, w, w_list);
1692         w_free_cnt++;
1693 }
1694
1695 static struct witness_child_list_entry *
1696 witness_child_get(void)
1697 {
1698         struct witness_child_list_entry *wcl;
1699
1700         if (witness_watch == 0) {
1701                 mtx_unlock_spin(&w_mtx);
1702                 return (NULL);
1703         }
1704         wcl = w_child_free;
1705         if (wcl == NULL) {
1706                 witness_watch = 0;
1707                 mtx_unlock_spin(&w_mtx);
1708                 printf("%s: witness exhausted\n", __func__);
1709                 return (NULL);
1710         }
1711         w_child_free = wcl->wcl_next;
1712         w_child_free_cnt--;
1713         bzero(wcl, sizeof(*wcl));
1714         return (wcl);
1715 }
1716
1717 static void
1718 witness_child_free(struct witness_child_list_entry *wcl)
1719 {
1720
1721         wcl->wcl_next = w_child_free;
1722         w_child_free = wcl;
1723         w_child_free_cnt++;
1724 }
1725
1726 static struct lock_list_entry *
1727 witness_lock_list_get(void)
1728 {
1729         struct lock_list_entry *lle;
1730
1731         if (witness_watch == 0)
1732                 return (NULL);
1733         mtx_lock_spin(&w_mtx);
1734         lle = w_lock_list_free;
1735         if (lle == NULL) {
1736                 witness_watch = 0;
1737                 mtx_unlock_spin(&w_mtx);
1738                 printf("%s: witness exhausted\n", __func__);
1739                 return (NULL);
1740         }
1741         w_lock_list_free = lle->ll_next;
1742         mtx_unlock_spin(&w_mtx);
1743         bzero(lle, sizeof(*lle));
1744         return (lle);
1745 }
1746
1747 static void
1748 witness_lock_list_free(struct lock_list_entry *lle)
1749 {
1750
1751         mtx_lock_spin(&w_mtx);
1752         lle->ll_next = w_lock_list_free;
1753         w_lock_list_free = lle;
1754         mtx_unlock_spin(&w_mtx);
1755 }
1756
1757 static struct lock_instance *
1758 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1759 {
1760         struct lock_list_entry *lle;
1761         struct lock_instance *instance;
1762         int i;
1763
1764         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1765                 for (i = lle->ll_count - 1; i >= 0; i--) {
1766                         instance = &lle->ll_children[i];
1767                         if (instance->li_lock == lock)
1768                                 return (instance);
1769                 }
1770         return (NULL);
1771 }
1772
1773 static void
1774 witness_list_lock(struct lock_instance *instance)
1775 {
1776         struct lock_object *lock;
1777
1778         lock = instance->li_lock;
1779         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1780             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
1781         if (lock->lo_type != lock->lo_name)
1782                 printf(" (%s)", lock->lo_type);
1783         printf(" r = %d (%p) locked @ %s:%d\n",
1784             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1785             instance->li_line);
1786 }
1787
1788 #ifdef DDB
1789 static int
1790 witness_thread_has_locks(struct thread *td)
1791 {
1792
1793         return (td->td_sleeplocks != NULL);
1794 }
1795
1796 static int
1797 witness_proc_has_locks(struct proc *p)
1798 {
1799         struct thread *td;
1800
1801         FOREACH_THREAD_IN_PROC(p, td) {
1802                 if (witness_thread_has_locks(td))
1803                         return (1);
1804         }
1805         return (0);
1806 }
1807 #endif
1808
1809 int
1810 witness_list_locks(struct lock_list_entry **lock_list)
1811 {
1812         struct lock_list_entry *lle;
1813         int i, nheld;
1814
1815         nheld = 0;
1816         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1817                 for (i = lle->ll_count - 1; i >= 0; i--) {
1818                         witness_list_lock(&lle->ll_children[i]);
1819                         nheld++;
1820                 }
1821         return (nheld);
1822 }
1823
1824 /*
1825  * This is a bit risky at best.  We call this function when we have timed
1826  * out acquiring a spin lock, and we assume that the other CPU is stuck
1827  * with this lock held.  So, we go groveling around in the other CPU's
1828  * per-cpu data to try to find the lock instance for this spin lock to
1829  * see when it was last acquired.
1830  */
1831 void
1832 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1833 {
1834         struct lock_instance *instance;
1835         struct pcpu *pc;
1836
1837         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1838                 return;
1839         pc = pcpu_find(owner->td_oncpu);
1840         instance = find_instance(pc->pc_spinlocks, lock);
1841         if (instance != NULL)
1842                 witness_list_lock(instance);
1843 }
1844
1845 void
1846 witness_save(struct lock_object *lock, const char **filep, int *linep)
1847 {
1848         struct lock_list_entry *lock_list;
1849         struct lock_instance *instance;
1850         struct lock_class *class;
1851
1852         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1853         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1854                 return;
1855         class = LOCK_CLASS(lock);
1856         if (class->lc_flags & LC_SLEEPLOCK)
1857                 lock_list = curthread->td_sleeplocks;
1858         else {
1859                 if (witness_skipspin)
1860                         return;
1861                 lock_list = PCPU_GET(spinlocks);
1862         }
1863         instance = find_instance(lock_list, lock);
1864         if (instance == NULL)
1865                 panic("%s: lock (%s) %s not locked", __func__,
1866                     class->lc_name, lock->lo_name);
1867         *filep = instance->li_file;
1868         *linep = instance->li_line;
1869 }
1870
1871 void
1872 witness_restore(struct lock_object *lock, const char *file, int line)
1873 {
1874         struct lock_list_entry *lock_list;
1875         struct lock_instance *instance;
1876         struct lock_class *class;
1877
1878         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1879         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1880                 return;
1881         class = LOCK_CLASS(lock);
1882         if (class->lc_flags & LC_SLEEPLOCK)
1883                 lock_list = curthread->td_sleeplocks;
1884         else {
1885                 if (witness_skipspin)
1886                         return;
1887                 lock_list = PCPU_GET(spinlocks);
1888         }
1889         instance = find_instance(lock_list, lock);
1890         if (instance == NULL)
1891                 panic("%s: lock (%s) %s not locked", __func__,
1892                     class->lc_name, lock->lo_name);
1893         lock->lo_witness->w_file = file;
1894         lock->lo_witness->w_line = line;
1895         instance->li_file = file;
1896         instance->li_line = line;
1897 }
1898
1899 void
1900 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1901 {
1902 #ifdef INVARIANT_SUPPORT
1903         struct lock_instance *instance;
1904         struct lock_class *class;
1905
1906         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1907                 return;
1908         class = LOCK_CLASS(lock);
1909         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
1910                 instance = find_instance(curthread->td_sleeplocks, lock);
1911         else if ((class->lc_flags & LC_SPINLOCK) != 0)
1912                 instance = find_instance(PCPU_GET(spinlocks), lock);
1913         else {
1914                 panic("Lock (%s) %s is not sleep or spin!",
1915                     class->lc_name, lock->lo_name);
1916         }
1917         file = fixup_filename(file);
1918         switch (flags) {
1919         case LA_UNLOCKED:
1920                 if (instance != NULL)
1921                         panic("Lock (%s) %s locked @ %s:%d.",
1922                             class->lc_name, lock->lo_name, file, line);
1923                 break;
1924         case LA_LOCKED:
1925         case LA_LOCKED | LA_RECURSED:
1926         case LA_LOCKED | LA_NOTRECURSED:
1927         case LA_SLOCKED:
1928         case LA_SLOCKED | LA_RECURSED:
1929         case LA_SLOCKED | LA_NOTRECURSED:
1930         case LA_XLOCKED:
1931         case LA_XLOCKED | LA_RECURSED:
1932         case LA_XLOCKED | LA_NOTRECURSED:
1933                 if (instance == NULL) {
1934                         panic("Lock (%s) %s not locked @ %s:%d.",
1935                             class->lc_name, lock->lo_name, file, line);
1936                         break;
1937                 }
1938                 if ((flags & LA_XLOCKED) != 0 &&
1939                     (instance->li_flags & LI_EXCLUSIVE) == 0)
1940                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1941                             class->lc_name, lock->lo_name, file, line);
1942                 if ((flags & LA_SLOCKED) != 0 &&
1943                     (instance->li_flags & LI_EXCLUSIVE) != 0)
1944                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
1945                             class->lc_name, lock->lo_name, file, line);
1946                 if ((flags & LA_RECURSED) != 0 &&
1947                     (instance->li_flags & LI_RECURSEMASK) == 0)
1948                         panic("Lock (%s) %s not recursed @ %s:%d.",
1949                             class->lc_name, lock->lo_name, file, line);
1950                 if ((flags & LA_NOTRECURSED) != 0 &&
1951                     (instance->li_flags & LI_RECURSEMASK) != 0)
1952                         panic("Lock (%s) %s recursed @ %s:%d.",
1953                             class->lc_name, lock->lo_name, file, line);
1954                 break;
1955         default:
1956                 panic("Invalid lock assertion at %s:%d.", file, line);
1957
1958         }
1959 #endif  /* INVARIANT_SUPPORT */
1960 }
1961
1962 #ifdef DDB
1963 static void
1964 witness_list(struct thread *td)
1965 {
1966
1967         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1968         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1969
1970         if (witness_watch == 0)
1971                 return;
1972
1973         witness_list_locks(&td->td_sleeplocks);
1974
1975         /*
1976          * We only handle spinlocks if td == curthread.  This is somewhat broken
1977          * if td is currently executing on some other CPU and holds spin locks
1978          * as we won't display those locks.  If we had a MI way of getting
1979          * the per-cpu data for a given cpu then we could use
1980          * td->td_oncpu to get the list of spinlocks for this thread
1981          * and "fix" this.
1982          *
1983          * That still wouldn't really fix this unless we locked the scheduler
1984          * lock or stopped the other CPU to make sure it wasn't changing the
1985          * list out from under us.  It is probably best to just not try to
1986          * handle threads on other CPU's for now.
1987          */
1988         if (td == curthread && PCPU_GET(spinlocks) != NULL)
1989                 witness_list_locks(PCPU_PTR(spinlocks));
1990 }
1991
1992 DB_SHOW_COMMAND(locks, db_witness_list)
1993 {
1994         struct thread *td;
1995
1996         if (have_addr)
1997                 td = db_lookup_thread(addr, TRUE);
1998         else
1999                 td = kdb_thread;
2000         witness_list(td);
2001 }
2002
2003 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
2004 {
2005         struct thread *td;
2006         struct proc *p;
2007
2008         /*
2009          * It would be nice to list only threads and processes that actually
2010          * held sleep locks, but that information is currently not exported
2011          * by WITNESS.
2012          */
2013         FOREACH_PROC_IN_SYSTEM(p) {
2014                 if (!witness_proc_has_locks(p))
2015                         continue;
2016                 FOREACH_THREAD_IN_PROC(p, td) {
2017                         if (!witness_thread_has_locks(td))
2018                                 continue;
2019                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2020                             p->p_comm, td, td->td_tid);
2021                         witness_list(td);
2022                 }
2023         }
2024 }
2025
2026 DB_SHOW_COMMAND(witness, db_witness_display)
2027 {
2028
2029         witness_display(db_printf);
2030 }
2031 #endif