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         { "scrlock", &lock_class_mtx_spin },
 408 #ifdef __i386__
 409         { "cy", &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         { "rm_spinlock", &lock_class_mtx_spin },
 426         { "turnstile chain", &lock_class_mtx_spin },
 427         { "turnstile lock", &lock_class_mtx_spin },
 428         { "sched lock", &lock_class_mtx_spin },
 429         { "td_contested", &lock_class_mtx_spin },
 430         { "callout", &lock_class_mtx_spin },
 431         { "entropy harvest mutex", &lock_class_mtx_spin },
 432         { "syscons video lock", &lock_class_mtx_spin },
 433         { "time lock", &lock_class_mtx_spin },
 434 #ifdef SMP
 435         { "smp rendezvous", &lock_class_mtx_spin },
 436 #endif
 437 #ifdef __powerpc__
 438         { "tlb0", &lock_class_mtx_spin },
 439 #endif
 440         /*
 441          * leaf locks
 442          */
 443         { "icu", &lock_class_mtx_spin },
 444 #if defined(SMP) && defined(__sparc64__)
 445         { "ipi", &lock_class_mtx_spin },
 446 #endif
 447 #ifdef __i386__
 448         { "allpmaps", &lock_class_mtx_spin },
 449         { "descriptor tables", &lock_class_mtx_spin },
 450 #endif
 451         { "clk", &lock_class_mtx_spin },
 452         { "mprof lock", &lock_class_mtx_spin },
 453         { "zombie lock", &lock_class_mtx_spin },
 454         { "ALD Queue", &lock_class_mtx_spin },
 455 #ifdef __ia64__
 456         { "MCA spin lock", &lock_class_mtx_spin },
 457 #endif
 458 #if defined(__i386__) || defined(__amd64__)
 459         { "pcicfg", &lock_class_mtx_spin },
 460         { "NDIS thread lock", &lock_class_mtx_spin },
 461 #endif
 462         { "tw_osl_io_lock", &lock_class_mtx_spin },
 463         { "tw_osl_q_lock", &lock_class_mtx_spin },
 464         { "tw_cl_io_lock", &lock_class_mtx_spin },
 465         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 466         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 467 #ifdef  HWPMC_HOOKS
 468         { "pmc-leaf", &lock_class_mtx_spin },
 469 #endif
 470         { "blocked lock", &lock_class_mtx_spin },
 471         { NULL, NULL },
 472         { NULL, NULL }
 473 };
 474
 475 #ifdef BLESSING
 476 /*
 477  * Pairs of locks which have been blessed
 478  * Don't complain about order problems with blessed locks
 479  */
 480 static struct witness_blessed blessed_list[] = {
 481 };
 482 static int blessed_count =
 483         sizeof(blessed_list) / sizeof(struct witness_blessed);
 484 #endif
 485
 486 /*
 487  * List of locks initialized prior to witness being initialized whose
 488  * enrollment is currently deferred.
 489  */
 490 STAILQ_HEAD(, lock_object) pending_locks =
 491     STAILQ_HEAD_INITIALIZER(pending_locks);
 492
 493 /*
 494  * This global is set to 0 once it becomes safe to use the witness code.
 495  */
 496 static int witness_cold = 1;
 497
 498 /*
 499  * This global is set to 1 once the static lock orders have been enrolled
 500  * so that a warning can be issued for any spin locks enrolled later.
 501  */
 502 static int witness_spin_warn = 0;
 503
 504 /*
 505  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 506  * assume that the early boot is single-threaded at least until after this
 507  * routine is completed.
 508  */
 509 static void
 510 witness_initialize(void *dummy __unused)
 511 {
 512         struct lock_object *lock;
 513         struct witness_order_list_entry *order;
 514         struct witness *w, *w1;
 515         int i;
 516
 517         /*
 518          * We have to release Giant before initializing its witness
 519          * structure so that WITNESS doesn't get confused.
 520          */
 521         mtx_unlock(&Giant);
 522         mtx_assert(&Giant, MA_NOTOWNED);
 523
 524         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 525         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 526             MTX_NOWITNESS | MTX_NOPROFILE);
 527         for (i = 0; i < WITNESS_COUNT; i++)
 528                 witness_free(&w_data[i]);
 529         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
 530                 witness_child_free(&w_childdata[i]);
 531         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 532                 witness_lock_list_free(&w_locklistdata[i]);
 533
 534         /* First add in all the specified order lists. */
 535         for (order = order_lists; order->w_name != NULL; order++) {
 536                 w = enroll(order->w_name, order->w_class);
 537                 if (w == NULL)
 538                         continue;
 539                 w->w_file = "order list";
 540                 for (order++; order->w_name != NULL; order++) {
 541                         w1 = enroll(order->w_name, order->w_class);
 542                         if (w1 == NULL)
 543                                 continue;
 544                         w1->w_file = "order list";
 545                         if (!itismychild(w, w1))
 546                                 panic("Not enough memory for static orders!");
 547                         w = w1;
 548                 }
 549         }
 550         witness_spin_warn = 1;
 551
 552         /* Iterate through all locks and add them to witness. */
 553         while (!STAILQ_EMPTY(&pending_locks)) {
 554                 lock = STAILQ_FIRST(&pending_locks);
 555                 STAILQ_REMOVE_HEAD(&pending_locks, lo_list);
 556                 KASSERT(lock->lo_flags & LO_WITNESS,
 557                     ("%s: lock %s is on pending list but not LO_WITNESS",
 558                     __func__, lock->lo_name));
 559                 lock->lo_witness = enroll(lock->lo_type, LOCK_CLASS(lock));
 560         }
 561
 562         /* Mark the witness code as being ready for use. */
 563         witness_cold = 0;
 564
 565         mtx_lock(&Giant);
 566 }
 567 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
 568
 569 static int
 570 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
 571 {
 572         int error, value;
 573
 574         value = witness_watch;
 575         error = sysctl_handle_int(oidp, &value, 0, req);
 576         if (error != 0 || req->newptr == NULL)
 577                 return (error);
 578         if (value == witness_watch)
 579                 return (0);
 580         if (value != 0)
 581                 return (EINVAL);
 582         witness_watch = 0;
 583         return (0);
 584 }
 585
 586 void
 587 witness_init(struct lock_object *lock)
 588 {
 589         struct lock_class *class;
 590
 591         /* Various sanity checks. */
 592         class = LOCK_CLASS(lock);
 593         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 594             (class->lc_flags & LC_RECURSABLE) == 0)
 595                 panic("%s: lock (%s) %s can not be recursable", __func__,
 596                     class->lc_name, lock->lo_name);
 597         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 598             (class->lc_flags & LC_SLEEPABLE) == 0)
 599                 panic("%s: lock (%s) %s can not be sleepable", __func__,
 600                     class->lc_name, lock->lo_name);
 601         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 602             (class->lc_flags & LC_UPGRADABLE) == 0)
 603                 panic("%s: lock (%s) %s can not be upgradable", __func__,
 604                     class->lc_name, lock->lo_name);
 605
 606         /*
 607          * If we shouldn't watch this lock, then just clear lo_witness.
 608          * Otherwise, if witness_cold is set, then it is too early to
 609          * enroll this lock, so defer it to witness_initialize() by adding
 610          * it to the pending_locks list.  If it is not too early, then enroll
 611          * the lock now.
 612          */
 613         if (witness_watch == 0 || panicstr != NULL ||
 614             (lock->lo_flags & LO_WITNESS) == 0)
 615                 lock->lo_witness = NULL;
 616         else if (witness_cold) {
 617                 STAILQ_INSERT_TAIL(&pending_locks, lock, lo_list);
 618                 lock->lo_flags |= LO_ENROLLPEND;
 619         } else
 620                 lock->lo_witness = enroll(lock->lo_type, class);
 621 }
 622
 623 void
 624 witness_destroy(struct lock_object *lock)
 625 {
 626         struct lock_class *class;
 627         struct witness *w;
 628
 629         class = LOCK_CLASS(lock);
 630         if (witness_cold)
 631                 panic("lock (%s) %s destroyed while witness_cold",
 632                     class->lc_name, lock->lo_name);
 633
 634         /* XXX: need to verify that no one holds the lock */
 635         if ((lock->lo_flags & (LO_WITNESS | LO_ENROLLPEND)) == LO_WITNESS &&
 636             lock->lo_witness != NULL) {
 637                 w = lock->lo_witness;
 638                 mtx_lock_spin(&w_mtx);
 639                 MPASS(w->w_refcount > 0);
 640                 w->w_refcount--;
 641
 642                 /*
 643                  * Lock is already released if we have an allocation failure
 644                  * and depart() fails.
 645                  */
 646                 if (w->w_refcount != 0 || depart(w))
 647                         mtx_unlock_spin(&w_mtx);
 648         }
 649
 650         /*
 651          * If this lock is destroyed before witness is up and running,
 652          * remove it from the pending list.
 653          */
 654         if (lock->lo_flags & LO_ENROLLPEND) {
 655                 STAILQ_REMOVE(&pending_locks, lock, lock_object, lo_list);
 656                 lock->lo_flags &= ~LO_ENROLLPEND;
 657         }
 658 }
 659
 660 #ifdef DDB
 661 static void
 662 witness_levelall (void)
 663 {
 664         struct witness_list *list;
 665         struct witness *w, *w1;
 666
 667         /*
 668          * First clear all levels.
 669          */
 670         STAILQ_FOREACH(w, &w_all, w_list) {
 671                 w->w_level = 0;
 672         }
 673
 674         /*
 675          * Look for locks with no parent and level all their descendants.
 676          */
 677         STAILQ_FOREACH(w, &w_all, w_list) {
 678                 /*
 679                  * This is just an optimization, technically we could get
 680                  * away just walking the all list each time.
 681                  */
 682                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
 683                         list = &w_sleep;
 684                 else
 685                         list = &w_spin;
 686                 STAILQ_FOREACH(w1, list, w_typelist) {
 687                         if (isitmychild(w1, w))
 688                                 goto skip;
 689                 }
 690                 witness_leveldescendents(w, 0);
 691         skip:
 692                 ;       /* silence GCC 3.x */
 693         }
 694 }
 695
 696 static void
 697 witness_leveldescendents(struct witness *parent, int level)
 698 {
 699         struct witness_child_list_entry *wcl;
 700         int i;
 701
 702         if (parent->w_level < level)
 703                 parent->w_level = level;
 704         level++;
 705         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 706                 for (i = 0; i < wcl->wcl_count; i++)
 707                         witness_leveldescendents(wcl->wcl_children[i], level);
 708 }
 709
 710 static void
 711 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
 712                            struct witness *parent, int indent)
 713 {
 714         struct witness_child_list_entry *wcl;
 715         int i, level;
 716
 717         level = parent->w_level;
 718         prnt("%-2d", level);
 719         for (i = 0; i < indent; i++)
 720                 prnt(" ");
 721         if (parent->w_refcount > 0)
 722                 prnt("%s", parent->w_name);
 723         else
 724                 prnt("(dead)");
 725         if (parent->w_displayed) {
 726                 prnt(" -- (already displayed)\n");
 727                 return;
 728         }
 729         parent->w_displayed = 1;
 730         if (parent->w_refcount > 0) {
 731                 if (parent->w_file != NULL)
 732                         prnt(" -- last acquired @ %s:%d", parent->w_file,
 733                             parent->w_line);
 734         }
 735         prnt("\n");
 736         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 737                 for (i = 0; i < wcl->wcl_count; i++)
 738                             witness_displaydescendants(prnt,
 739                                 wcl->wcl_children[i], indent + 1);
 740 }
 741
 742 static void
 743 witness_display_list(void(*prnt)(const char *fmt, ...),
 744                      struct witness_list *list)
 745 {
 746         struct witness *w;
 747
 748         STAILQ_FOREACH(w, list, w_typelist) {
 749                 if (w->w_file == NULL || w->w_level > 0)
 750                         continue;
 751                 /*
 752                  * This lock has no anscestors, display its descendants.
 753                  */
 754                 witness_displaydescendants(prnt, w, 0);
 755         }
 756 }
 757
 758 static void
 759 witness_display(void(*prnt)(const char *fmt, ...))
 760 {
 761         struct witness *w;
 762
 763         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 764         witness_levelall();
 765
 766         /* Clear all the displayed flags. */
 767         STAILQ_FOREACH(w, &w_all, w_list) {
 768                 w->w_displayed = 0;
 769         }
 770
 771         /*
 772          * First, handle sleep locks which have been acquired at least
 773          * once.
 774          */
 775         prnt("Sleep locks:\n");
 776         witness_display_list(prnt, &w_sleep);
 777
 778         /*
 779          * Now do spin locks which have been acquired at least once.
 780          */
 781         prnt("\nSpin locks:\n");
 782         witness_display_list(prnt, &w_spin);
 783
 784         /*
 785          * Finally, any locks which have not been acquired yet.
 786          */
 787         prnt("\nLocks which were never acquired:\n");
 788         STAILQ_FOREACH(w, &w_all, w_list) {
 789                 if (w->w_file != NULL || w->w_refcount == 0)
 790                         continue;
 791                 prnt("%s\n", w->w_name);
 792         }
 793 }
 794 #endif /* DDB */
 795
 796 /* Trim useless garbage from filenames. */
 797 static const char *
 798 fixup_filename(const char *file)
 799 {
 800
 801         if (file == NULL)
 802                 return (NULL);
 803         while (strncmp(file, "../", 3) == 0)
 804                 file += 3;
 805         return (file);
 806 }
 807
 808 int
 809 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
 810 {
 811
 812         if (witness_watch == 0 || panicstr != NULL)
 813                 return (0);
 814
 815         /* Require locks that witness knows about. */
 816         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
 817             lock2->lo_witness == NULL)
 818                 return (EINVAL);
 819
 820         MPASS(!mtx_owned(&w_mtx));
 821         mtx_lock_spin(&w_mtx);
 822
 823         /*
 824          * If we already have either an explicit or implied lock order that
 825          * is the other way around, then return an error.
 826          */
 827         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
 828                 mtx_unlock_spin(&w_mtx);
 829                 return (EDOOFUS);
 830         }
 831
 832         /* Try to add the new order. */
 833         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 834             lock2->lo_type, lock1->lo_type);
 835         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
 836                 return (ENOMEM);
 837         mtx_unlock_spin(&w_mtx);
 838         return (0);
 839 }
 840
 841 void
 842 witness_checkorder(struct lock_object *lock, int flags, const char *file,
 843     int line)
 844 {
 845         struct lock_list_entry **lock_list, *lle;
 846         struct lock_instance *lock1, *lock2;
 847         struct lock_class *class;
 848         struct witness *w, *w1;
 849         struct thread *td;
 850         int i, j;
 851
 852         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 853             panicstr != NULL)
 854                 return;
 855
 856         /*
 857          * Try locks do not block if they fail to acquire the lock, thus
 858          * there is no danger of deadlocks or of switching while holding a
 859          * spin lock if we acquire a lock via a try operation.  This
 860          * function shouldn't even be called for try locks, so panic if
 861          * that happens.
 862          */
 863         if (flags & LOP_TRYLOCK)
 864                 panic("%s should not be called for try lock operations",
 865                     __func__);
 866
 867         w = lock->lo_witness;
 868         class = LOCK_CLASS(lock);
 869         td = curthread;
 870         file = fixup_filename(file);
 871
 872         if (class->lc_flags & LC_SLEEPLOCK) {
 873                 /*
 874                  * Since spin locks include a critical section, this check
 875                  * implicitly enforces a lock order of all sleep locks before
 876                  * all spin locks.
 877                  */
 878                 if (td->td_critnest != 0 && !kdb_active)
 879                         panic("blockable sleep lock (%s) %s @ %s:%d",
 880                             class->lc_name, lock->lo_name, file, line);
 881
 882                 /*
 883                  * If this is the first lock acquired then just return as
 884                  * no order checking is needed.
 885                  */
 886                 if (td->td_sleeplocks == NULL)
 887                         return;
 888                 lock_list = &td->td_sleeplocks;
 889         } else {
 890                 /*
 891                  * If this is the first lock, just return as no order
 892                  * checking is needed.  We check this in both if clauses
 893                  * here as unifying the check would require us to use a
 894                  * critical section to ensure we don't migrate while doing
 895                  * the check.  Note that if this is not the first lock, we
 896                  * are already in a critical section and are safe for the
 897                  * rest of the check.
 898                  */
 899                 if (PCPU_GET(spinlocks) == NULL)
 900                         return;
 901                 lock_list = PCPU_PTR(spinlocks);
 902         }
 903
 904         /*
 905          * Check to see if we are recursing on a lock we already own.  If
 906          * so, make sure that we don't mismatch exclusive and shared lock
 907          * acquires.
 908          */
 909         lock1 = find_instance(*lock_list, lock);
 910         if (lock1 != NULL) {
 911                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
 912                     (flags & LOP_EXCLUSIVE) == 0) {
 913                         printf("shared lock of (%s) %s @ %s:%d\n",
 914                             class->lc_name, lock->lo_name, file, line);
 915                         printf("while exclusively locked from %s:%d\n",
 916                             lock1->li_file, lock1->li_line);
 917                         panic("share->excl");
 918                 }
 919                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
 920                     (flags & LOP_EXCLUSIVE) != 0) {
 921                         printf("exclusive lock of (%s) %s @ %s:%d\n",
 922                             class->lc_name, lock->lo_name, file, line);
 923                         printf("while share locked from %s:%d\n",
 924                             lock1->li_file, lock1->li_line);
 925                         panic("excl->share");
 926                 }
 927                 return;
 928         }
 929
 930         /*
 931          * Try locks do not block if they fail to acquire the lock, thus
 932          * there is no danger of deadlocks or of switching while holding a
 933          * spin lock if we acquire a lock via a try operation.
 934          */
 935         if (flags & LOP_TRYLOCK)
 936                 return;
 937
 938         /*
 939          * Check for duplicate locks of the same type.  Note that we only
 940          * have to check for this on the last lock we just acquired.  Any
 941          * other cases will be caught as lock order violations.
 942          */
 943         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 944         w1 = lock1->li_lock->lo_witness;
 945         if (w1 == w) {
 946                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
 947                     (flags & LOP_DUPOK))
 948                         return;
 949                 w->w_same_squawked = 1;
 950                 printf("acquiring duplicate lock of same type: \"%s\"\n",
 951                         lock->lo_type);
 952                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
 953                     lock1->li_file, lock1->li_line);
 954                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
 955 #ifdef KDB
 956                 goto debugger;
 957 #else
 958                 return;
 959 #endif
 960         }
 961         MPASS(!mtx_owned(&w_mtx));
 962         mtx_lock_spin(&w_mtx);
 963         /*
 964          * If we know that the the lock we are acquiring comes after
 965          * the lock we most recently acquired in the lock order tree,
 966          * then there is no need for any further checks.
 967          */
 968         if (isitmychild(w1, w)) {
 969                 mtx_unlock_spin(&w_mtx);
 970                 return;
 971         }
 972         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
 973                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
 974
 975                         MPASS(j < WITNESS_COUNT);
 976                         lock1 = &lle->ll_children[i];
 977                         w1 = lock1->li_lock->lo_witness;
 978
 979                         /*
 980                          * If this lock doesn't undergo witness checking,
 981                          * then skip it.
 982                          */
 983                         if (w1 == NULL) {
 984                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
 985                                     ("lock missing witness structure"));
 986                                 continue;
 987                         }
 988                         /*
 989                          * If we are locking Giant and this is a sleepable
 990                          * lock, then skip it.
 991                          */
 992                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
 993                             lock == &Giant.lock_object)
 994                                 continue;
 995                         /*
 996                          * If we are locking a sleepable lock and this lock
 997                          * is Giant, then skip it.
 998                          */
 999                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1000                             lock1->li_lock == &Giant.lock_object)
1001                                 continue;
1002                         /*
1003                          * If we are locking a sleepable lock and this lock
1004                          * isn't sleepable, we want to treat it as a lock
1005                          * order violation to enfore a general lock order of
1006                          * sleepable locks before non-sleepable locks.
1007                          */
1008                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1009                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1010                                 goto reversal;
1011                         /*
1012                          * If we are locking Giant and this is a non-sleepable
1013                          * lock, then treat it as a reversal.
1014                          */
1015                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1016                             lock == &Giant.lock_object)
1017                                 goto reversal;
1018                         /*
1019                          * Check the lock order hierarchy for a reveresal.
1020                          */
1021                         if (!isitmydescendant(w, w1))
1022                                 continue;
1023                 reversal:
1024                         /*
1025                          * We have a lock order violation, check to see if it
1026                          * is allowed or has already been yelled about.
1027                          */
1028                         mtx_unlock_spin(&w_mtx);
1029 #ifdef BLESSING
1030                         /*
1031                          * If the lock order is blessed, just bail.  We don't
1032                          * look for other lock order violations though, which
1033                          * may be a bug.
1034                          */
1035                         if (blessed(w, w1))
1036                                 return;
1037 #endif
1038                         if (lock1->li_lock == &Giant.lock_object) {
1039                                 if (w1->w_Giant_squawked)
1040                                         return;
1041                                 else
1042                                         w1->w_Giant_squawked = 1;
1043                         } else {
1044                                 if (w1->w_other_squawked)
1045                                         return;
1046                                 else
1047                                         w1->w_other_squawked = 1;
1048                         }
1049                         /*
1050                          * Ok, yell about it.
1051                          */
1052                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1053                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1054                                 printf(
1055                 "lock order reversal: (sleepable after non-sleepable)\n");
1056                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1057                             && lock == &Giant.lock_object)
1058                                 printf(
1059                 "lock order reversal: (Giant after non-sleepable)\n");
1060                         else
1061                                 printf("lock order reversal:\n");
1062                         /*
1063                          * Try to locate an earlier lock with
1064                          * witness w in our list.
1065                          */
1066                         do {
1067                                 lock2 = &lle->ll_children[i];
1068                                 MPASS(lock2->li_lock != NULL);
1069                                 if (lock2->li_lock->lo_witness == w)
1070                                         break;
1071                                 if (i == 0 && lle->ll_next != NULL) {
1072                                         lle = lle->ll_next;
1073                                         i = lle->ll_count - 1;
1074                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1075                                 } else
1076                                         i--;
1077                         } while (i >= 0);
1078                         if (i < 0) {
1079                                 printf(" 1st %p %s (%s) @ %s:%d\n",
1080                                     lock1->li_lock, lock1->li_lock->lo_name,
1081                                     lock1->li_lock->lo_type, lock1->li_file,
1082                                     lock1->li_line);
1083                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1084                                     lock->lo_name, lock->lo_type, file, line);
1085                         } else {
1086                                 printf(" 1st %p %s (%s) @ %s:%d\n",
1087                                     lock2->li_lock, lock2->li_lock->lo_name,
1088                                     lock2->li_lock->lo_type, lock2->li_file,
1089                                     lock2->li_line);
1090                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
1091                                     lock1->li_lock, lock1->li_lock->lo_name,
1092                                     lock1->li_lock->lo_type, lock1->li_file,
1093                                     lock1->li_line);
1094                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1095                                     lock->lo_name, lock->lo_type, file, line);
1096                         }
1097 #ifdef KDB
1098                         goto debugger;
1099 #else
1100                         return;
1101 #endif
1102                 }
1103         }
1104         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
1105         /*
1106          * If requested, build a new lock order.  However, don't build a new
1107          * relationship between a sleepable lock and Giant if it is in the
1108          * wrong direction.  The correct lock order is that sleepable locks
1109          * always come before Giant.
1110          */
1111         if (flags & LOP_NEWORDER &&
1112             !(lock1->li_lock == &Giant.lock_object &&
1113             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1114                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1115                     lock->lo_type, lock1->li_lock->lo_type);
1116                 if (!itismychild(lock1->li_lock->lo_witness, w))
1117                         /* Witness is dead. */
1118                         return;
1119         }
1120         mtx_unlock_spin(&w_mtx);
1121         return;
1122
1123 #ifdef KDB
1124 debugger:
1125         if (witness_trace)
1126                 kdb_backtrace();
1127         if (witness_kdb)
1128                 kdb_enter(KDB_WHY_WITNESS, __func__);
1129 #endif
1130 }
1131
1132 void
1133 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1134 {
1135         struct lock_list_entry **lock_list, *lle;
1136         struct lock_instance *instance;
1137         struct witness *w;
1138         struct thread *td;
1139
1140         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1141             panicstr != NULL)
1142                 return;
1143         w = lock->lo_witness;
1144         td = curthread;
1145         file = fixup_filename(file);
1146
1147         /* Determine lock list for this lock. */
1148         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1149                 lock_list = &td->td_sleeplocks;
1150         else
1151                 lock_list = PCPU_PTR(spinlocks);
1152
1153         /* Check to see if we are recursing on a lock we already own. */
1154         instance = find_instance(*lock_list, lock);
1155         if (instance != NULL) {
1156                 instance->li_flags++;
1157                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1158                     td->td_proc->p_pid, lock->lo_name,
1159                     instance->li_flags & LI_RECURSEMASK);
1160                 instance->li_file = file;
1161                 instance->li_line = line;
1162                 return;
1163         }
1164
1165         /* Update per-witness last file and line acquire. */
1166         w->w_file = file;
1167         w->w_line = line;
1168
1169         /* Find the next open lock instance in the list and fill it. */
1170         lle = *lock_list;
1171         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1172                 lle = witness_lock_list_get();
1173                 if (lle == NULL)
1174                         return;
1175                 lle->ll_next = *lock_list;
1176                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1177                     td->td_proc->p_pid, lle);
1178                 *lock_list = lle;
1179         }
1180         instance = &lle->ll_children[lle->ll_count++];
1181         instance->li_lock = lock;
1182         instance->li_line = line;
1183         instance->li_file = file;
1184         if ((flags & LOP_EXCLUSIVE) != 0)
1185                 instance->li_flags = LI_EXCLUSIVE;
1186         else
1187                 instance->li_flags = 0;
1188         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1189             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1190 }
1191
1192 void
1193 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1194 {
1195         struct lock_instance *instance;
1196         struct lock_class *class;
1197
1198         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1199         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1200                 return;
1201         class = LOCK_CLASS(lock);
1202         file = fixup_filename(file);
1203         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1204                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1205                     class->lc_name, lock->lo_name, file, line);
1206         if ((flags & LOP_TRYLOCK) == 0)
1207                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
1208                     lock->lo_name, file, line);
1209         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1210                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1211                     class->lc_name, lock->lo_name, file, line);
1212         instance = find_instance(curthread->td_sleeplocks, lock);
1213         if (instance == NULL)
1214                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1215                     class->lc_name, lock->lo_name, file, line);
1216         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1217                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1218                     class->lc_name, lock->lo_name, file, line);
1219         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1220                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1221                     class->lc_name, lock->lo_name,
1222                     instance->li_flags & LI_RECURSEMASK, file, line);
1223         instance->li_flags |= LI_EXCLUSIVE;
1224 }
1225
1226 void
1227 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1228     int line)
1229 {
1230         struct lock_instance *instance;
1231         struct lock_class *class;
1232
1233         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1234         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1235                 return;
1236         class = LOCK_CLASS(lock);
1237         file = fixup_filename(file);
1238         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1239                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1240                     class->lc_name, lock->lo_name, file, line);
1241         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1242                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1243                     class->lc_name, lock->lo_name, file, line);
1244         instance = find_instance(curthread->td_sleeplocks, lock);
1245         if (instance == NULL)
1246                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1247                     class->lc_name, lock->lo_name, file, line);
1248         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1249                 panic("downgrade of shared lock (%s) %s @ %s:%d",
1250                     class->lc_name, lock->lo_name, file, line);
1251         if ((instance->li_flags & LI_RECURSEMASK) != 0)
1252                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1253                     class->lc_name, lock->lo_name,
1254                     instance->li_flags & LI_RECURSEMASK, file, line);
1255         instance->li_flags &= ~LI_EXCLUSIVE;
1256 }
1257
1258 void
1259 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1260 {
1261         struct lock_list_entry **lock_list, *lle;
1262         struct lock_instance *instance;
1263         struct lock_class *class;
1264         struct thread *td;
1265         register_t s;
1266         int i, j;
1267
1268         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
1269             panicstr != NULL)
1270                 return;
1271         td = curthread;
1272         class = LOCK_CLASS(lock);
1273         file = fixup_filename(file);
1274
1275         /* Find lock instance associated with this lock. */
1276         if (class->lc_flags & LC_SLEEPLOCK)
1277                 lock_list = &td->td_sleeplocks;
1278         else
1279                 lock_list = PCPU_PTR(spinlocks);
1280         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1281                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1282                         instance = &(*lock_list)->ll_children[i];
1283                         if (instance->li_lock == lock)
1284                                 goto found;
1285                 }
1286         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
1287             file, line);
1288 found:
1289
1290         /* First, check for shared/exclusive mismatches. */
1291         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
1292             (flags & LOP_EXCLUSIVE) == 0) {
1293                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1294                     lock->lo_name, file, line);
1295                 printf("while exclusively locked from %s:%d\n",
1296                     instance->li_file, instance->li_line);
1297                 panic("excl->ushare");
1298         }
1299         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
1300             (flags & LOP_EXCLUSIVE) != 0) {
1301                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1302                     lock->lo_name, file, line);
1303                 printf("while share locked from %s:%d\n", instance->li_file,
1304                     instance->li_line);
1305                 panic("share->uexcl");
1306         }
1307
1308         /* If we are recursed, unrecurse. */
1309         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1310                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1311                     td->td_proc->p_pid, instance->li_lock->lo_name,
1312                     instance->li_flags);
1313                 instance->li_flags--;
1314                 return;
1315         }
1316
1317         /* Otherwise, remove this item from the list. */
1318         s = intr_disable();
1319         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1320             td->td_proc->p_pid, instance->li_lock->lo_name,
1321             (*lock_list)->ll_count - 1);
1322         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1323                 (*lock_list)->ll_children[j] =
1324                     (*lock_list)->ll_children[j + 1];
1325         (*lock_list)->ll_count--;
1326         intr_restore(s);
1327
1328         /* If this lock list entry is now empty, free it. */
1329         if ((*lock_list)->ll_count == 0) {
1330                 lle = *lock_list;
1331                 *lock_list = lle->ll_next;
1332                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1333                     td->td_proc->p_pid, lle);
1334                 witness_lock_list_free(lle);
1335         }
1336 }
1337
1338 /*
1339  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1340  * exempt Giant and sleepable locks from the checks as well.  If any
1341  * non-exempt locks are held, then a supplied message is printed to the
1342  * console along with a list of the offending locks.  If indicated in the
1343  * flags then a failure results in a panic as well.
1344  */
1345 int
1346 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1347 {
1348         struct lock_list_entry *lle;
1349         struct lock_instance *lock1;
1350         struct thread *td;
1351         va_list ap;
1352         int i, n;
1353
1354         if (witness_cold || witness_watch == 0 || panicstr != NULL)
1355                 return (0);
1356         n = 0;
1357         td = curthread;
1358         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1359                 for (i = lle->ll_count - 1; i >= 0; i--) {
1360                         lock1 = &lle->ll_children[i];
1361                         if (lock1->li_lock == lock)
1362                                 continue;
1363                         if (flags & WARN_GIANTOK &&
1364                             lock1->li_lock == &Giant.lock_object)
1365                                 continue;
1366                         if (flags & WARN_SLEEPOK &&
1367                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1368                                 continue;
1369                         if (n == 0) {
1370                                 va_start(ap, fmt);
1371                                 vprintf(fmt, ap);
1372                                 va_end(ap);
1373                                 printf(" with the following");
1374                                 if (flags & WARN_SLEEPOK)
1375                                         printf(" non-sleepable");
1376                                 printf(" locks held:\n");
1377                         }
1378                         n++;
1379                         witness_list_lock(lock1);
1380                 }
1381         if (PCPU_GET(spinlocks) != NULL) {
1382                 /*
1383                  * Since we already hold a spinlock preemption is
1384                  * already blocked.
1385                  */
1386                 if (n == 0) {
1387                         va_start(ap, fmt);
1388                         vprintf(fmt, ap);
1389                         va_end(ap);
1390                         printf(" with the following");
1391                         if (flags & WARN_SLEEPOK)
1392                                 printf(" non-sleepable");
1393                         printf(" locks held:\n");
1394                 }
1395                 n += witness_list_locks(PCPU_PTR(spinlocks));
1396         }
1397         if (flags & WARN_PANIC && n)
1398                 panic("witness_warn");
1399 #ifdef KDB
1400         else if (witness_kdb && n)
1401                 kdb_enter(KDB_WHY_WITNESS, __func__);
1402         else if (witness_trace && n)
1403                 kdb_backtrace();
1404 #endif
1405         return (n);
1406 }
1407
1408 const char *
1409 witness_file(struct lock_object *lock)
1410 {
1411         struct witness *w;
1412
1413         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1414                 return ("?");
1415         w = lock->lo_witness;
1416         return (w->w_file);
1417 }
1418
1419 int
1420 witness_line(struct lock_object *lock)
1421 {
1422         struct witness *w;
1423
1424         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
1425                 return (0);
1426         w = lock->lo_witness;
1427         return (w->w_line);
1428 }
1429
1430 static struct witness *
1431 enroll(const char *description, struct lock_class *lock_class)
1432 {
1433         struct witness *w;
1434
1435         if (witness_watch == 0 || panicstr != NULL)
1436                 return (NULL);
1437         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
1438                 return (NULL);
1439         mtx_lock_spin(&w_mtx);
1440         STAILQ_FOREACH(w, &w_all, w_list) {
1441                 if (w->w_name == description || (w->w_refcount > 0 &&
1442                     strcmp(description, w->w_name) == 0)) {
1443                         w->w_refcount++;
1444                         mtx_unlock_spin(&w_mtx);
1445                         if (lock_class != w->w_class)
1446                                 panic(
1447                                 "lock (%s) %s does not match earlier (%s) lock",
1448                                     description, lock_class->lc_name,
1449                                     w->w_class->lc_name);
1450                         return (w);
1451                 }
1452         }
1453         if ((w = witness_get()) == NULL)
1454                 goto out;
1455         w->w_name = description;
1456         w->w_class = lock_class;
1457         w->w_refcount = 1;
1458         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1459         if (lock_class->lc_flags & LC_SPINLOCK) {
1460                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1461                 w_spin_cnt++;
1462         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1463                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1464                 w_sleep_cnt++;
1465         } else {
1466                 mtx_unlock_spin(&w_mtx);
1467                 panic("lock class %s is not sleep or spin",
1468                     lock_class->lc_name);
1469         }
1470         mtx_unlock_spin(&w_mtx);
1471 out:
1472         /*
1473          * We issue a warning for any spin locks not defined in the static
1474          * order list as a way to discourage their use (folks should really
1475          * be using non-spin mutexes most of the time).  However, several
1476          * 3rd part device drivers use spin locks because that is all they
1477          * have available on Windows and Linux and they think that normal
1478          * mutexes are insufficient.
1479          */
1480         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
1481                 printf("WITNESS: spin lock %s not in order list\n",
1482                     description);
1483         return (w);
1484 }
1485
1486 /* Don't let the door bang you on the way out... */
1487 static int
1488 depart(struct witness *w)
1489 {
1490         struct witness_child_list_entry *wcl, *nwcl;
1491         struct witness_list *list;
1492         struct witness *parent;
1493
1494         MPASS(w->w_refcount == 0);
1495         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1496                 list = &w_sleep;
1497                 w_sleep_cnt--;
1498         } else {
1499                 list = &w_spin;
1500                 w_spin_cnt--;
1501         }
1502         /*
1503          * First, we run through the entire tree looking for any
1504          * witnesses that the outgoing witness is a child of.  For
1505          * each parent that we find, we reparent all the direct
1506          * children of the outgoing witness to its parent.
1507          */
1508         STAILQ_FOREACH(parent, list, w_typelist) {
1509                 if (!isitmychild(parent, w))
1510                         continue;
1511                 removechild(parent, w);
1512         }
1513
1514         /*
1515          * Now we go through and free up the child list of the
1516          * outgoing witness.
1517          */
1518         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
1519                 nwcl = wcl->wcl_next;
1520                 w_child_cnt--;
1521                 witness_child_free(wcl);
1522         }
1523
1524         /*
1525          * Detach from various lists and free.
1526          */
1527         STAILQ_REMOVE(list, w, witness, w_typelist);
1528         STAILQ_REMOVE(&w_all, w, witness, w_list);
1529         witness_free(w);
1530
1531         return (1);
1532 }
1533
1534 /*
1535  * Add "child" as a direct child of "parent".  Returns false if
1536  * we fail due to out of memory.
1537  */
1538 static int
1539 insertchild(struct witness *parent, struct witness *child)
1540 {
1541         struct witness_child_list_entry **wcl;
1542
1543         MPASS(child != NULL && parent != NULL);
1544
1545         /*
1546          * Insert "child" after "parent"
1547          */
1548         wcl = &parent->w_children;
1549         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
1550                 wcl = &(*wcl)->wcl_next;
1551         if (*wcl == NULL) {
1552                 *wcl = witness_child_get();
1553                 if (*wcl == NULL)
1554                         return (0);
1555                 w_child_cnt++;
1556         }
1557         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
1558
1559         return (1);
1560 }
1561
1562
1563 static int
1564 itismychild(struct witness *parent, struct witness *child)
1565 {
1566         struct witness_list *list;
1567
1568         MPASS(child != NULL && parent != NULL);
1569         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
1570             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
1571                 panic(
1572                 "%s: parent (%s) and child (%s) are not the same lock type",
1573                     __func__, parent->w_class->lc_name,
1574                     child->w_class->lc_name);
1575
1576         if (!insertchild(parent, child))
1577                 return (0);
1578
1579         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
1580                 list = &w_sleep;
1581         else
1582                 list = &w_spin;
1583         return (1);
1584 }
1585
1586 static void
1587 removechild(struct witness *parent, struct witness *child)
1588 {
1589         struct witness_child_list_entry **wcl, *wcl1;
1590         int i;
1591
1592         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
1593                 for (i = 0; i < (*wcl)->wcl_count; i++)
1594                         if ((*wcl)->wcl_children[i] == child)
1595                                 goto found;
1596         return;
1597 found:
1598         (*wcl)->wcl_count--;
1599         if ((*wcl)->wcl_count > i)
1600                 (*wcl)->wcl_children[i] =
1601                     (*wcl)->wcl_children[(*wcl)->wcl_count];
1602         MPASS((*wcl)->wcl_children[i] != NULL);
1603         if ((*wcl)->wcl_count != 0)
1604                 return;
1605         wcl1 = *wcl;
1606         *wcl = wcl1->wcl_next;
1607         w_child_cnt--;
1608         witness_child_free(wcl1);
1609 }
1610
1611 static int
1612 isitmychild(struct witness *parent, struct witness *child)
1613 {
1614         struct witness_child_list_entry *wcl;
1615         int i;
1616
1617         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1618                 for (i = 0; i < wcl->wcl_count; i++) {
1619                         if (wcl->wcl_children[i] == child)
1620                                 return (1);
1621                 }
1622         }
1623         return (0);
1624 }
1625
1626 static int
1627 isitmydescendant(struct witness *parent, struct witness *child)
1628 {
1629         struct witness_child_list_entry *wcl;
1630         int i, j;
1631
1632         if (isitmychild(parent, child))
1633                 return (1);
1634         j = 0;
1635         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
1636                 MPASS(j < 1000);
1637                 for (i = 0; i < wcl->wcl_count; i++) {
1638                         if (isitmydescendant(wcl->wcl_children[i], child))
1639                                 return (1);
1640                 }
1641                 j++;
1642         }
1643         return (0);
1644 }
1645
1646 #ifdef BLESSING
1647 static int
1648 blessed(struct witness *w1, struct witness *w2)
1649 {
1650         int i;
1651         struct witness_blessed *b;
1652
1653         for (i = 0; i < blessed_count; i++) {
1654                 b = &blessed_list[i];
1655                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1656                         if (strcmp(w2->w_name, b->b_lock2) == 0)
1657                                 return (1);
1658                         continue;
1659                 }
1660                 if (strcmp(w1->w_name, b->b_lock2) == 0)
1661                         if (strcmp(w2->w_name, b->b_lock1) == 0)
1662                                 return (1);
1663         }
1664         return (0);
1665 }
1666 #endif
1667
1668 static struct witness *
1669 witness_get(void)
1670 {
1671         struct witness *w;
1672
1673         if (witness_watch == 0) {
1674                 mtx_unlock_spin(&w_mtx);
1675                 return (NULL);
1676         }
1677         if (STAILQ_EMPTY(&w_free)) {
1678                 witness_watch = 0;
1679                 mtx_unlock_spin(&w_mtx);
1680                 printf("%s: witness exhausted\n", __func__);
1681                 return (NULL);
1682         }
1683         w = STAILQ_FIRST(&w_free);
1684         STAILQ_REMOVE_HEAD(&w_free, w_list);
1685         w_free_cnt--;
1686         bzero(w, sizeof(*w));
1687         return (w);
1688 }
1689
1690 static void
1691 witness_free(struct witness *w)
1692 {
1693
1694         STAILQ_INSERT_HEAD(&w_free, w, w_list);
1695         w_free_cnt++;
1696 }
1697
1698 static struct witness_child_list_entry *
1699 witness_child_get(void)
1700 {
1701         struct witness_child_list_entry *wcl;
1702
1703         if (witness_watch == 0) {
1704                 mtx_unlock_spin(&w_mtx);
1705                 return (NULL);
1706         }
1707         wcl = w_child_free;
1708         if (wcl == NULL) {
1709                 witness_watch = 0;
1710                 mtx_unlock_spin(&w_mtx);
1711                 printf("%s: witness exhausted\n", __func__);
1712                 return (NULL);
1713         }
1714         w_child_free = wcl->wcl_next;
1715         w_child_free_cnt--;
1716         bzero(wcl, sizeof(*wcl));
1717         return (wcl);
1718 }
1719
1720 static void
1721 witness_child_free(struct witness_child_list_entry *wcl)
1722 {
1723
1724         wcl->wcl_next = w_child_free;
1725         w_child_free = wcl;
1726         w_child_free_cnt++;
1727 }
1728
1729 static struct lock_list_entry *
1730 witness_lock_list_get(void)
1731 {
1732         struct lock_list_entry *lle;
1733
1734         if (witness_watch == 0)
1735                 return (NULL);
1736         mtx_lock_spin(&w_mtx);
1737         lle = w_lock_list_free;
1738         if (lle == NULL) {
1739                 witness_watch = 0;
1740                 mtx_unlock_spin(&w_mtx);
1741                 printf("%s: witness exhausted\n", __func__);
1742                 return (NULL);
1743         }
1744         w_lock_list_free = lle->ll_next;
1745         mtx_unlock_spin(&w_mtx);
1746         bzero(lle, sizeof(*lle));
1747         return (lle);
1748 }
1749
1750 static void
1751 witness_lock_list_free(struct lock_list_entry *lle)
1752 {
1753
1754         mtx_lock_spin(&w_mtx);
1755         lle->ll_next = w_lock_list_free;
1756         w_lock_list_free = lle;
1757         mtx_unlock_spin(&w_mtx);
1758 }
1759
1760 static struct lock_instance *
1761 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
1762 {
1763         struct lock_list_entry *lle;
1764         struct lock_instance *instance;
1765         int i;
1766
1767         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
1768                 for (i = lle->ll_count - 1; i >= 0; i--) {
1769                         instance = &lle->ll_children[i];
1770                         if (instance->li_lock == lock)
1771                                 return (instance);
1772                 }
1773         return (NULL);
1774 }
1775
1776 static void
1777 witness_list_lock(struct lock_instance *instance)
1778 {
1779         struct lock_object *lock;
1780
1781         lock = instance->li_lock;
1782         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
1783             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
1784         if (lock->lo_type != lock->lo_name)
1785                 printf(" (%s)", lock->lo_type);
1786         printf(" r = %d (%p) locked @ %s:%d\n",
1787             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
1788             instance->li_line);
1789 }
1790
1791 #ifdef DDB
1792 static int
1793 witness_thread_has_locks(struct thread *td)
1794 {
1795
1796         return (td->td_sleeplocks != NULL);
1797 }
1798
1799 static int
1800 witness_proc_has_locks(struct proc *p)
1801 {
1802         struct thread *td;
1803
1804         FOREACH_THREAD_IN_PROC(p, td) {
1805                 if (witness_thread_has_locks(td))
1806                         return (1);
1807         }
1808         return (0);
1809 }
1810 #endif
1811
1812 int
1813 witness_list_locks(struct lock_list_entry **lock_list)
1814 {
1815         struct lock_list_entry *lle;
1816         int i, nheld;
1817
1818         nheld = 0;
1819         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
1820                 for (i = lle->ll_count - 1; i >= 0; i--) {
1821                         witness_list_lock(&lle->ll_children[i]);
1822                         nheld++;
1823                 }
1824         return (nheld);
1825 }
1826
1827 /*
1828  * This is a bit risky at best.  We call this function when we have timed
1829  * out acquiring a spin lock, and we assume that the other CPU is stuck
1830  * with this lock held.  So, we go groveling around in the other CPU's
1831  * per-cpu data to try to find the lock instance for this spin lock to
1832  * see when it was last acquired.
1833  */
1834 void
1835 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
1836 {
1837         struct lock_instance *instance;
1838         struct pcpu *pc;
1839
1840         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
1841                 return;
1842         pc = pcpu_find(owner->td_oncpu);
1843         instance = find_instance(pc->pc_spinlocks, lock);
1844         if (instance != NULL)
1845                 witness_list_lock(instance);
1846 }
1847
1848 void
1849 witness_save(struct lock_object *lock, const char **filep, int *linep)
1850 {
1851         struct lock_list_entry *lock_list;
1852         struct lock_instance *instance;
1853         struct lock_class *class;
1854
1855         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1856         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1857                 return;
1858         class = LOCK_CLASS(lock);
1859         if (class->lc_flags & LC_SLEEPLOCK)
1860                 lock_list = curthread->td_sleeplocks;
1861         else {
1862                 if (witness_skipspin)
1863                         return;
1864                 lock_list = PCPU_GET(spinlocks);
1865         }
1866         instance = find_instance(lock_list, lock);
1867         if (instance == NULL)
1868                 panic("%s: lock (%s) %s not locked", __func__,
1869                     class->lc_name, lock->lo_name);
1870         *filep = instance->li_file;
1871         *linep = instance->li_line;
1872 }
1873
1874 void
1875 witness_restore(struct lock_object *lock, const char *file, int line)
1876 {
1877         struct lock_list_entry *lock_list;
1878         struct lock_instance *instance;
1879         struct lock_class *class;
1880
1881         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1882         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1883                 return;
1884         class = LOCK_CLASS(lock);
1885         if (class->lc_flags & LC_SLEEPLOCK)
1886                 lock_list = curthread->td_sleeplocks;
1887         else {
1888                 if (witness_skipspin)
1889                         return;
1890                 lock_list = PCPU_GET(spinlocks);
1891         }
1892         instance = find_instance(lock_list, lock);
1893         if (instance == NULL)
1894                 panic("%s: lock (%s) %s not locked", __func__,
1895                     class->lc_name, lock->lo_name);
1896         lock->lo_witness->w_file = file;
1897         lock->lo_witness->w_line = line;
1898         instance->li_file = file;
1899         instance->li_line = line;
1900 }
1901
1902 void
1903 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
1904 {
1905 #ifdef INVARIANT_SUPPORT
1906         struct lock_instance *instance;
1907         struct lock_class *class;
1908
1909         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
1910                 return;
1911         class = LOCK_CLASS(lock);
1912         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
1913                 instance = find_instance(curthread->td_sleeplocks, lock);
1914         else if ((class->lc_flags & LC_SPINLOCK) != 0)
1915                 instance = find_instance(PCPU_GET(spinlocks), lock);
1916         else {
1917                 panic("Lock (%s) %s is not sleep or spin!",
1918                     class->lc_name, lock->lo_name);
1919         }
1920         file = fixup_filename(file);
1921         switch (flags) {
1922         case LA_UNLOCKED:
1923                 if (instance != NULL)
1924                         panic("Lock (%s) %s locked @ %s:%d.",
1925                             class->lc_name, lock->lo_name, file, line);
1926                 break;
1927         case LA_LOCKED:
1928         case LA_LOCKED | LA_RECURSED:
1929         case LA_LOCKED | LA_NOTRECURSED:
1930         case LA_SLOCKED:
1931         case LA_SLOCKED | LA_RECURSED:
1932         case LA_SLOCKED | LA_NOTRECURSED:
1933         case LA_XLOCKED:
1934         case LA_XLOCKED | LA_RECURSED:
1935         case LA_XLOCKED | LA_NOTRECURSED:
1936                 if (instance == NULL) {
1937                         panic("Lock (%s) %s not locked @ %s:%d.",
1938                             class->lc_name, lock->lo_name, file, line);
1939                         break;
1940                 }
1941                 if ((flags & LA_XLOCKED) != 0 &&
1942                     (instance->li_flags & LI_EXCLUSIVE) == 0)
1943                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
1944                             class->lc_name, lock->lo_name, file, line);
1945                 if ((flags & LA_SLOCKED) != 0 &&
1946                     (instance->li_flags & LI_EXCLUSIVE) != 0)
1947                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
1948                             class->lc_name, lock->lo_name, file, line);
1949                 if ((flags & LA_RECURSED) != 0 &&
1950                     (instance->li_flags & LI_RECURSEMASK) == 0)
1951                         panic("Lock (%s) %s not recursed @ %s:%d.",
1952                             class->lc_name, lock->lo_name, file, line);
1953                 if ((flags & LA_NOTRECURSED) != 0 &&
1954                     (instance->li_flags & LI_RECURSEMASK) != 0)
1955                         panic("Lock (%s) %s recursed @ %s:%d.",
1956                             class->lc_name, lock->lo_name, file, line);
1957                 break;
1958         default:
1959                 panic("Invalid lock assertion at %s:%d.", file, line);
1960
1961         }
1962 #endif  /* INVARIANT_SUPPORT */
1963 }
1964
1965 #ifdef DDB
1966 static void
1967 witness_list(struct thread *td)
1968 {
1969
1970         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
1971         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
1972
1973         if (witness_watch == 0)
1974                 return;
1975
1976         witness_list_locks(&td->td_sleeplocks);
1977
1978         /*
1979          * We only handle spinlocks if td == curthread.  This is somewhat broken
1980          * if td is currently executing on some other CPU and holds spin locks
1981          * as we won't display those locks.  If we had a MI way of getting
1982          * the per-cpu data for a given cpu then we could use
1983          * td->td_oncpu to get the list of spinlocks for this thread
1984          * and "fix" this.
1985          *
1986          * That still wouldn't really fix this unless we locked the scheduler
1987          * lock or stopped the other CPU to make sure it wasn't changing the
1988          * list out from under us.  It is probably best to just not try to
1989          * handle threads on other CPU's for now.
1990          */
1991         if (td == curthread && PCPU_GET(spinlocks) != NULL)
1992                 witness_list_locks(PCPU_PTR(spinlocks));
1993 }
1994
1995 DB_SHOW_COMMAND(locks, db_witness_list)
1996 {
1997         struct thread *td;
1998
1999         if (have_addr)
2000                 td = db_lookup_thread(addr, TRUE);
2001         else
2002                 td = kdb_thread;
2003         witness_list(td);
2004 }
2005
2006 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
2007 {
2008         struct thread *td;
2009         struct proc *p;
2010
2011         /*
2012          * It would be nice to list only threads and processes that actually
2013          * held sleep locks, but that information is currently not exported
2014          * by WITNESS.
2015          */
2016         FOREACH_PROC_IN_SYSTEM(p) {
2017                 if (!witness_proc_has_locks(p))
2018                         continue;
2019                 FOREACH_THREAD_IN_PROC(p, td) {
2020                         if (!witness_thread_has_locks(td))
2021                                 continue;
2022                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2023                             td->td_name, td, td->td_tid);
2024                         witness_list(td);
2025                 }
2026         }
2027 }
2028
2029 DB_SHOW_COMMAND(witness, db_witness_display)
2030 {
2031
2032         witness_display(db_printf);
2033 }
2034 #endif