sys/kern/subr_witness.c

   1 /*-
   2  * Copyright (c) 2008 Isilon Systems, Inc.
   3  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
   4  * Copyright (c) 1998 Berkeley Software Design, Inc.
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Berkeley Software Design Inc's name may not be used to endorse or
  16  *    promote products derived from this software without specific prior
  17  *    written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
  20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  22  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
  23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  29  * SUCH DAMAGE.
  30  *
  31  *      from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
  32  *      and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
  33  */
  34
  35 /*
  36  * Implementation of the `witness' lock verifier.  Originally implemented for
  37  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
  38  * classes in FreeBSD.
  39  */
  40
  41 /*
  42  *      Main Entry: witness
  43  *      Pronunciation: 'wit-n&s
  44  *      Function: noun
  45  *      Etymology: Middle English witnesse, from Old English witnes knowledge,
  46  *          testimony, witness, from 2wit
  47  *      Date: before 12th century
  48  *      1 : attestation of a fact or event : TESTIMONY
  49  *      2 : one that gives evidence; specifically : one who testifies in
  50  *          a cause or before a judicial tribunal
  51  *      3 : one asked to be present at a transaction so as to be able to
  52  *          testify to its having taken place
  53  *      4 : one who has personal knowledge of something
  54  *      5 a : something serving as evidence or proof : SIGN
  55  *        b : public affirmation by word or example of usually
  56  *            religious faith or conviction <the heroic witness to divine
  57  *            life -- Pilot>
  58  *      6 capitalized : a member of the Jehovah's Witnesses
  59  */
  60
  61 /*
  62  * Special rules concerning Giant and lock orders:
  63  *
  64  * 1) Giant must be acquired before any other mutexes.  Stated another way,
  65  *    no other mutex may be held when Giant is acquired.
  66  *
  67  * 2) Giant must be released when blocking on a sleepable lock.
  68  *
  69  * This rule is less obvious, but is a result of Giant providing the same
  70  * semantics as spl().  Basically, when a thread sleeps, it must release
  71  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
  72  * 2).
  73  *
  74  * 3) Giant may be acquired before or after sleepable locks.
  75  *
  76  * This rule is also not quite as obvious.  Giant may be acquired after
  77  * a sleepable lock because it is a non-sleepable lock and non-sleepable
  78  * locks may always be acquired while holding a sleepable lock.  The second
  79  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
  80  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
  81  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
  82  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
  83  * execute.  Thus, acquiring Giant both before and after a sleepable lock
  84  * will not result in a lock order reversal.
  85  */
  86
  87 #include <sys/cdefs.h>
  88 __FBSDID("$FreeBSD$");
  89
  90 #include "opt_ddb.h"
  91 #include "opt_hwpmc_hooks.h"
  92 #include "opt_stack.h"
  93 #include "opt_witness.h"
  94
  95 #include <sys/param.h>
  96 #include <sys/bus.h>
  97 #include <sys/kdb.h>
  98 #include <sys/kernel.h>
  99 #include <sys/ktr.h>
 100 #include <sys/lock.h>
 101 #include <sys/malloc.h>
 102 #include <sys/mutex.h>
 103 #include <sys/priv.h>
 104 #include <sys/proc.h>
 105 #include <sys/sbuf.h>
 106 #include <sys/sched.h>
 107 #include <sys/stack.h>
 108 #include <sys/sysctl.h>
 109 #include <sys/systm.h>
 110
 111 #ifdef DDB
 112 #include <ddb/ddb.h>
 113 #endif
 114
 115 #include <machine/stdarg.h>
 116
 117 #if !defined(DDB) && !defined(STACK)
 118 #error "DDB or STACK options are required for WITNESS"
 119 #endif
 120
 121 /* Note that these traces do not work with KTR_ALQ. */
 122 #if 0
 123 #define KTR_WITNESS     KTR_SUBSYS
 124 #else
 125 #define KTR_WITNESS     0
 126 #endif
 127
 128 #define LI_RECURSEMASK  0x0000ffff      /* Recursion depth of lock instance. */
 129 #define LI_EXCLUSIVE    0x00010000      /* Exclusive lock instance. */
 130 #define LI_NORELEASE    0x00020000      /* Lock not allowed to be released. */
 131
 132 /* Define this to check for blessed mutexes */
 133 #undef BLESSING
 134
 135 #ifndef WITNESS_COUNT
 136 #define WITNESS_COUNT           1536
 137 #endif
 138 #define WITNESS_CHILDCOUNT      (WITNESS_COUNT * 4)
 139 #define WITNESS_HASH_SIZE       251     /* Prime, gives load factor < 2 */
 140 #define WITNESS_PENDLIST        (1024 + MAXCPU)
 141
 142 /* Allocate 256 KB of stack data space */
 143 #define WITNESS_LO_DATA_COUNT   2048
 144
 145 /* Prime, gives load factor of ~2 at full load */
 146 #define WITNESS_LO_HASH_SIZE    1021
 147
 148 /*
 149  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
 150  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
 151  * probably be safe for the most part, but it's still a SWAG.
 152  */
 153 #define LOCK_NCHILDREN  5
 154 #define LOCK_CHILDCOUNT 2048
 155
 156 #define MAX_W_NAME      64
 157
 158 #define BADSTACK_SBUF_SIZE      (256 * WITNESS_COUNT)
 159 #define FULLGRAPH_SBUF_SIZE     512
 160
 161 /*
 162  * These flags go in the witness relationship matrix and describe the
 163  * relationship between any two struct witness objects.
 164  */
 165 #define WITNESS_UNRELATED        0x00    /* No lock order relation. */
 166 #define WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
 167 #define WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
 168 #define WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
 169 #define WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
 170 #define WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
 171 #define WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
 172 #define WITNESS_RELATED_MASK                                            \
 173         (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
 174 #define WITNESS_REVERSAL         0x10    /* A lock order reversal has been
 175                                           * observed. */
 176 #define WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
 177 #define WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
 178 #define WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
 179
 180 /* Descendant to ancestor flags */
 181 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
 182
 183 /* Ancestor to descendant flags */
 184 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
 185
 186 #define WITNESS_INDEX_ASSERT(i)                                         \
 187         MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
 188
 189 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
 190
 191 /*
 192  * Lock instances.  A lock instance is the data associated with a lock while
 193  * it is held by witness.  For example, a lock instance will hold the
 194  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
 195  * are held in a per-cpu list while sleep locks are held in per-thread list.
 196  */
 197 struct lock_instance {
 198         struct lock_object      *li_lock;
 199         const char              *li_file;
 200         int                     li_line;
 201         u_int                   li_flags;
 202 };
 203
 204 /*
 205  * A simple list type used to build the list of locks held by a thread
 206  * or CPU.  We can't simply embed the list in struct lock_object since a
 207  * lock may be held by more than one thread if it is a shared lock.  Locks
 208  * are added to the head of the list, so we fill up each list entry from
 209  * "the back" logically.  To ease some of the arithmetic, we actually fill
 210  * in each list entry the normal way (children[0] then children[1], etc.) but
 211  * when we traverse the list we read children[count-1] as the first entry
 212  * down to children[0] as the final entry.
 213  */
 214 struct lock_list_entry {
 215         struct lock_list_entry  *ll_next;
 216         struct lock_instance    ll_children[LOCK_NCHILDREN];
 217         u_int                   ll_count;
 218 };
 219
 220 /*
 221  * The main witness structure. One of these per named lock type in the system
 222  * (for example, "vnode interlock").
 223  */
 224 struct witness {
 225         char                    w_name[MAX_W_NAME];
 226         uint32_t                w_index;  /* Index in the relationship matrix */
 227         struct lock_class       *w_class;
 228         STAILQ_ENTRY(witness)   w_list;         /* List of all witnesses. */
 229         STAILQ_ENTRY(witness)   w_typelist;     /* Witnesses of a type. */
 230         struct witness          *w_hash_next; /* Linked list in hash buckets. */
 231         const char              *w_file; /* File where last acquired */
 232         uint32_t                w_line; /* Line where last acquired */
 233         uint32_t                w_refcount;
 234         uint16_t                w_num_ancestors; /* direct/indirect
 235                                                   * ancestor count */
 236         uint16_t                w_num_descendants; /* direct/indirect
 237                                                     * descendant count */
 238         int16_t                 w_ddb_level;
 239         unsigned                w_displayed:1;
 240         unsigned                w_reversed:1;
 241 };
 242
 243 STAILQ_HEAD(witness_list, witness);
 244
 245 /*
 246  * The witness hash table. Keys are witness names (const char *), elements are
 247  * witness objects (struct witness *).
 248  */
 249 struct witness_hash {
 250         struct witness  *wh_array[WITNESS_HASH_SIZE];
 251         uint32_t        wh_size;
 252         uint32_t        wh_count;
 253 };
 254
 255 /*
 256  * Key type for the lock order data hash table.
 257  */
 258 struct witness_lock_order_key {
 259         uint16_t        from;
 260         uint16_t        to;
 261 };
 262
 263 struct witness_lock_order_data {
 264         struct stack                    wlod_stack;
 265         struct witness_lock_order_key   wlod_key;
 266         struct witness_lock_order_data  *wlod_next;
 267 };
 268
 269 /*
 270  * The witness lock order data hash table. Keys are witness index tuples
 271  * (struct witness_lock_order_key), elements are lock order data objects
 272  * (struct witness_lock_order_data).
 273  */
 274 struct witness_lock_order_hash {
 275         struct witness_lock_order_data  *wloh_array[WITNESS_LO_HASH_SIZE];
 276         u_int   wloh_size;
 277         u_int   wloh_count;
 278 };
 279
 280 #ifdef BLESSING
 281 struct witness_blessed {
 282         const char      *b_lock1;
 283         const char      *b_lock2;
 284 };
 285 #endif
 286
 287 struct witness_pendhelp {
 288         const char              *wh_type;
 289         struct lock_object      *wh_lock;
 290 };
 291
 292 struct witness_order_list_entry {
 293         const char              *w_name;
 294         struct lock_class       *w_class;
 295 };
 296
 297 /*
 298  * Returns 0 if one of the locks is a spin lock and the other is not.
 299  * Returns 1 otherwise.
 300  */
 301 static __inline int
 302 witness_lock_type_equal(struct witness *w1, struct witness *w2)
 303 {
 304
 305         return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
 306                 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
 307 }
 308
 309 static __inline int
 310 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
 311     const struct witness_lock_order_key *b)
 312 {
 313
 314         return (a->from == b->from && a->to == b->to);
 315 }
 316
 317 static int      _isitmyx(struct witness *w1, struct witness *w2, int rmask,
 318                     const char *fname);
 319 #ifdef KDB
 320 static void     _witness_debugger(int cond, const char *msg);
 321 #endif
 322 static void     adopt(struct witness *parent, struct witness *child);
 323 #ifdef BLESSING
 324 static int      blessed(struct witness *, struct witness *);
 325 #endif
 326 static void     depart(struct witness *w);
 327 static struct witness   *enroll(const char *description,
 328                             struct lock_class *lock_class);
 329 static struct lock_instance     *find_instance(struct lock_list_entry *list,
 330                                     const struct lock_object *lock);
 331 static int      isitmychild(struct witness *parent, struct witness *child);
 332 static int      isitmydescendant(struct witness *parent, struct witness *child);
 333 static void     itismychild(struct witness *parent, struct witness *child);
 334 static int      sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
 335 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
 336 static int      sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
 337 static void     witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
 338 #ifdef DDB
 339 static void     witness_ddb_compute_levels(void);
 340 static void     witness_ddb_display(int(*)(const char *fmt, ...));
 341 static void     witness_ddb_display_descendants(int(*)(const char *fmt, ...),
 342                     struct witness *, int indent);
 343 static void     witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
 344                     struct witness_list *list);
 345 static void     witness_ddb_level_descendants(struct witness *parent, int l);
 346 static void     witness_ddb_list(struct thread *td);
 347 #endif
 348 static void     witness_free(struct witness *m);
 349 static struct witness   *witness_get(void);
 350 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
 351 static struct witness   *witness_hash_get(const char *key);
 352 static void     witness_hash_put(struct witness *w);
 353 static void     witness_init_hash_tables(void);
 354 static void     witness_increment_graph_generation(void);
 355 static void     witness_lock_list_free(struct lock_list_entry *lle);
 356 static struct lock_list_entry   *witness_lock_list_get(void);
 357 static int      witness_lock_order_add(struct witness *parent,
 358                     struct witness *child);
 359 static int      witness_lock_order_check(struct witness *parent,
 360                     struct witness *child);
 361 static struct witness_lock_order_data   *witness_lock_order_get(
 362                                             struct witness *parent,
 363                                             struct witness *child);
 364 static void     witness_list_lock(struct lock_instance *instance,
 365                     int (*prnt)(const char *fmt, ...));
 366 static void     witness_setflag(struct lock_object *lock, int flag, int set);
 367
 368 #ifdef KDB
 369 #define witness_debugger(c)     _witness_debugger(c, __func__)
 370 #else
 371 #define witness_debugger(c)
 372 #endif
 373
 374 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
 375     "Witness Locking");
 376
 377 /*
 378  * If set to 0, lock order checking is disabled.  If set to -1,
 379  * witness is completely disabled.  Otherwise witness performs full
 380  * lock order checking for all locks.  At runtime, lock order checking
 381  * may be toggled.  However, witness cannot be reenabled once it is
 382  * completely disabled.
 383  */
 384 static int witness_watch = 1;
 385 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
 386     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
 387
 388 #ifdef KDB
 389 /*
 390  * When KDB is enabled and witness_kdb is 1, it will cause the system
 391  * to drop into kdebug() when:
 392  *      - a lock hierarchy violation occurs
 393  *      - locks are held when going to sleep.
 394  */
 395 #ifdef WITNESS_KDB
 396 int     witness_kdb = 1;
 397 #else
 398 int     witness_kdb = 0;
 399 #endif
 400 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
 401
 402 /*
 403  * When KDB is enabled and witness_trace is 1, it will cause the system
 404  * to print a stack trace:
 405  *      - a lock hierarchy violation occurs
 406  *      - locks are held when going to sleep.
 407  */
 408 int     witness_trace = 1;
 409 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
 410 #endif /* KDB */
 411
 412 #ifdef WITNESS_SKIPSPIN
 413 int     witness_skipspin = 1;
 414 #else
 415 int     witness_skipspin = 0;
 416 #endif
 417 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
 418
 419 /*
 420  * Call this to print out the relations between locks.
 421  */
 422 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
 423     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
 424
 425 /*
 426  * Call this to print out the witness faulty stacks.
 427  */
 428 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
 429     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
 430
 431 static struct mtx w_mtx;
 432
 433 /* w_list */
 434 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 435 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 436
 437 /* w_typelist */
 438 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 439 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 440
 441 /* lock list */
 442 static struct lock_list_entry *w_lock_list_free = NULL;
 443 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
 444 static u_int pending_cnt;
 445
 446 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
 447 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
 448 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
 449 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
 450     "");
 451
 452 static struct witness *w_data;
 453 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
 454 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 455 static struct witness_hash w_hash;      /* The witness hash table. */
 456
 457 /* The lock order data hash */
 458 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
 459 static struct witness_lock_order_data *w_lofree = NULL;
 460 static struct witness_lock_order_hash w_lohash;
 461 static int w_max_used_index = 0;
 462 static unsigned int w_generation = 0;
 463 static const char w_notrunning[] = "Witness not running\n";
 464 static const char w_stillcold[] = "Witness is still cold\n";
 465
 466
 467 static struct witness_order_list_entry order_lists[] = {
 468         /*
 469          * sx locks
 470          */
 471         { "proctree", &lock_class_sx },
 472         { "allproc", &lock_class_sx },
 473         { "allprison", &lock_class_sx },
 474         { NULL, NULL },
 475         /*
 476          * Various mutexes
 477          */
 478         { "Giant", &lock_class_mtx_sleep },
 479         { "pipe mutex", &lock_class_mtx_sleep },
 480         { "sigio lock", &lock_class_mtx_sleep },
 481         { "process group", &lock_class_mtx_sleep },
 482         { "process lock", &lock_class_mtx_sleep },
 483         { "session", &lock_class_mtx_sleep },
 484         { "uidinfo hash", &lock_class_rw },
 485 #ifdef  HWPMC_HOOKS
 486         { "pmc-sleep", &lock_class_mtx_sleep },
 487 #endif
 488         { "time lock", &lock_class_mtx_sleep },
 489         { NULL, NULL },
 490         /*
 491          * Sockets
 492          */
 493         { "accept", &lock_class_mtx_sleep },
 494         { "so_snd", &lock_class_mtx_sleep },
 495         { "so_rcv", &lock_class_mtx_sleep },
 496         { "sellck", &lock_class_mtx_sleep },
 497         { NULL, NULL },
 498         /*
 499          * Routing
 500          */
 501         { "so_rcv", &lock_class_mtx_sleep },
 502         { "radix node head", &lock_class_rw },
 503         { "rtentry", &lock_class_mtx_sleep },
 504         { "ifaddr", &lock_class_mtx_sleep },
 505         { NULL, NULL },
 506         /*
 507          * IPv4 multicast:
 508          * protocol locks before interface locks, after UDP locks.
 509          */
 510         { "udpinp", &lock_class_rw },
 511         { "in_multi_mtx", &lock_class_mtx_sleep },
 512         { "igmp_mtx", &lock_class_mtx_sleep },
 513         { "if_addr_lock", &lock_class_rw },
 514         { NULL, NULL },
 515         /*
 516          * IPv6 multicast:
 517          * protocol locks before interface locks, after UDP locks.
 518          */
 519         { "udpinp", &lock_class_rw },
 520         { "in6_multi_mtx", &lock_class_mtx_sleep },
 521         { "mld_mtx", &lock_class_mtx_sleep },
 522         { "if_addr_lock", &lock_class_rw },
 523         { NULL, NULL },
 524         /*
 525          * UNIX Domain Sockets
 526          */
 527         { "unp_global_rwlock", &lock_class_rw },
 528         { "unp_list_lock", &lock_class_mtx_sleep },
 529         { "unp", &lock_class_mtx_sleep },
 530         { "so_snd", &lock_class_mtx_sleep },
 531         { NULL, NULL },
 532         /*
 533          * UDP/IP
 534          */
 535         { "udp", &lock_class_rw },
 536         { "udpinp", &lock_class_rw },
 537         { "so_snd", &lock_class_mtx_sleep },
 538         { NULL, NULL },
 539         /*
 540          * TCP/IP
 541          */
 542         { "tcp", &lock_class_rw },
 543         { "tcpinp", &lock_class_rw },
 544         { "so_snd", &lock_class_mtx_sleep },
 545         { NULL, NULL },
 546         /*
 547          * BPF
 548          */
 549         { "bpf global lock", &lock_class_mtx_sleep },
 550         { "bpf interface lock", &lock_class_rw },
 551         { "bpf cdev lock", &lock_class_mtx_sleep },
 552         { NULL, NULL },
 553         /*
 554          * NFS server
 555          */
 556         { "nfsd_mtx", &lock_class_mtx_sleep },
 557         { "so_snd", &lock_class_mtx_sleep },
 558         { NULL, NULL },
 559
 560         /*
 561          * IEEE 802.11
 562          */
 563         { "802.11 com lock", &lock_class_mtx_sleep},
 564         { NULL, NULL },
 565         /*
 566          * Network drivers
 567          */
 568         { "network driver", &lock_class_mtx_sleep},
 569         { NULL, NULL },
 570
 571         /*
 572          * Netgraph
 573          */
 574         { "ng_node", &lock_class_mtx_sleep },
 575         { "ng_worklist", &lock_class_mtx_sleep },
 576         { NULL, NULL },
 577         /*
 578          * CDEV
 579          */
 580         { "vm map (system)", &lock_class_mtx_sleep },
 581         { "vm page queue", &lock_class_mtx_sleep },
 582         { "vnode interlock", &lock_class_mtx_sleep },
 583         { "cdev", &lock_class_mtx_sleep },
 584         { NULL, NULL },
 585         /*
 586          * VM
 587          */
 588         { "vm map (user)", &lock_class_sx },
 589         { "vm object", &lock_class_rw },
 590         { "vm page", &lock_class_mtx_sleep },
 591         { "vm page queue", &lock_class_mtx_sleep },
 592         { "pmap pv global", &lock_class_rw },
 593         { "pmap", &lock_class_mtx_sleep },
 594         { "pmap pv list", &lock_class_rw },
 595         { "vm page free queue", &lock_class_mtx_sleep },
 596         { NULL, NULL },
 597         /*
 598          * kqueue/VFS interaction
 599          */
 600         { "kqueue", &lock_class_mtx_sleep },
 601         { "struct mount mtx", &lock_class_mtx_sleep },
 602         { "vnode interlock", &lock_class_mtx_sleep },
 603         { NULL, NULL },
 604         /*
 605          * ZFS locking
 606          */
 607         { "dn->dn_mtx", &lock_class_sx },
 608         { "dr->dt.di.dr_mtx", &lock_class_sx },
 609         { "db->db_mtx", &lock_class_sx },
 610         { NULL, NULL },
 611         /*
 612          * spin locks
 613          */
 614 #ifdef SMP
 615         { "ap boot", &lock_class_mtx_spin },
 616 #endif
 617         { "rm.mutex_mtx", &lock_class_mtx_spin },
 618         { "sio", &lock_class_mtx_spin },
 619         { "scrlock", &lock_class_mtx_spin },
 620 #ifdef __i386__
 621         { "cy", &lock_class_mtx_spin },
 622 #endif
 623 #ifdef __sparc64__
 624         { "pcib_mtx", &lock_class_mtx_spin },
 625         { "rtc_mtx", &lock_class_mtx_spin },
 626 #endif
 627         { "scc_hwmtx", &lock_class_mtx_spin },
 628         { "uart_hwmtx", &lock_class_mtx_spin },
 629         { "fast_taskqueue", &lock_class_mtx_spin },
 630         { "intr table", &lock_class_mtx_spin },
 631 #ifdef  HWPMC_HOOKS
 632         { "pmc-per-proc", &lock_class_mtx_spin },
 633 #endif
 634         { "process slock", &lock_class_mtx_spin },
 635         { "sleepq chain", &lock_class_mtx_spin },
 636         { "umtx lock", &lock_class_mtx_spin },
 637         { "rm_spinlock", &lock_class_mtx_spin },
 638         { "turnstile chain", &lock_class_mtx_spin },
 639         { "turnstile lock", &lock_class_mtx_spin },
 640         { "sched lock", &lock_class_mtx_spin },
 641         { "td_contested", &lock_class_mtx_spin },
 642         { "callout", &lock_class_mtx_spin },
 643         { "entropy harvest mutex", &lock_class_mtx_spin },
 644         { "syscons video lock", &lock_class_mtx_spin },
 645 #ifdef SMP
 646         { "smp rendezvous", &lock_class_mtx_spin },
 647 #endif
 648 #ifdef __powerpc__
 649         { "tlb0", &lock_class_mtx_spin },
 650 #endif
 651         /*
 652          * leaf locks
 653          */
 654         { "intrcnt", &lock_class_mtx_spin },
 655         { "icu", &lock_class_mtx_spin },
 656 #ifdef __i386__
 657         { "allpmaps", &lock_class_mtx_spin },
 658         { "descriptor tables", &lock_class_mtx_spin },
 659 #endif
 660         { "clk", &lock_class_mtx_spin },
 661         { "cpuset", &lock_class_mtx_spin },
 662         { "mprof lock", &lock_class_mtx_spin },
 663         { "zombie lock", &lock_class_mtx_spin },
 664         { "ALD Queue", &lock_class_mtx_spin },
 665 #if defined(__i386__) || defined(__amd64__)
 666         { "pcicfg", &lock_class_mtx_spin },
 667         { "NDIS thread lock", &lock_class_mtx_spin },
 668 #endif
 669         { "tw_osl_io_lock", &lock_class_mtx_spin },
 670         { "tw_osl_q_lock", &lock_class_mtx_spin },
 671         { "tw_cl_io_lock", &lock_class_mtx_spin },
 672         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 673         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 674 #ifdef  HWPMC_HOOKS
 675         { "pmc-leaf", &lock_class_mtx_spin },
 676 #endif
 677         { "blocked lock", &lock_class_mtx_spin },
 678         { NULL, NULL },
 679         { NULL, NULL }
 680 };
 681
 682 #ifdef BLESSING
 683 /*
 684  * Pairs of locks which have been blessed
 685  * Don't complain about order problems with blessed locks
 686  */
 687 static struct witness_blessed blessed_list[] = {
 688 };
 689 static int blessed_count =
 690         sizeof(blessed_list) / sizeof(struct witness_blessed);
 691 #endif
 692
 693 /*
 694  * This global is set to 0 once it becomes safe to use the witness code.
 695  */
 696 static int witness_cold = 1;
 697
 698 /*
 699  * This global is set to 1 once the static lock orders have been enrolled
 700  * so that a warning can be issued for any spin locks enrolled later.
 701  */
 702 static int witness_spin_warn = 0;
 703
 704 /* Trim useless garbage from filenames. */
 705 static const char *
 706 fixup_filename(const char *file)
 707 {
 708
 709         if (file == NULL)
 710                 return (NULL);
 711         while (strncmp(file, "../", 3) == 0)
 712                 file += 3;
 713         return (file);
 714 }
 715
 716 /*
 717  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 718  * assume that the early boot is single-threaded at least until after this
 719  * routine is completed.
 720  */
 721 static void
 722 witness_initialize(void *dummy __unused)
 723 {
 724         struct lock_object *lock;
 725         struct witness_order_list_entry *order;
 726         struct witness *w, *w1;
 727         int i;
 728
 729         w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
 730             M_NOWAIT | M_ZERO);
 731
 732         /*
 733          * We have to release Giant before initializing its witness
 734          * structure so that WITNESS doesn't get confused.
 735          */
 736         mtx_unlock(&Giant);
 737         mtx_assert(&Giant, MA_NOTOWNED);
 738
 739         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 740         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 741             MTX_NOWITNESS | MTX_NOPROFILE);
 742         for (i = WITNESS_COUNT - 1; i >= 0; i--) {
 743                 w = &w_data[i];
 744                 memset(w, 0, sizeof(*w));
 745                 w_data[i].w_index = i;  /* Witness index never changes. */
 746                 witness_free(w);
 747         }
 748         KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
 749             ("%s: Invalid list of free witness objects", __func__));
 750
 751         /* Witness with index 0 is not used to aid in debugging. */
 752         STAILQ_REMOVE_HEAD(&w_free, w_list);
 753         w_free_cnt--;
 754
 755         memset(w_rmatrix, 0,
 756             (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
 757
 758         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 759                 witness_lock_list_free(&w_locklistdata[i]);
 760         witness_init_hash_tables();
 761
 762         /* First add in all the specified order lists. */
 763         for (order = order_lists; order->w_name != NULL; order++) {
 764                 w = enroll(order->w_name, order->w_class);
 765                 if (w == NULL)
 766                         continue;
 767                 w->w_file = "order list";
 768                 for (order++; order->w_name != NULL; order++) {
 769                         w1 = enroll(order->w_name, order->w_class);
 770                         if (w1 == NULL)
 771                                 continue;
 772                         w1->w_file = "order list";
 773                         itismychild(w, w1);
 774                         w = w1;
 775                 }
 776         }
 777         witness_spin_warn = 1;
 778
 779         /* Iterate through all locks and add them to witness. */
 780         for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
 781                 lock = pending_locks[i].wh_lock;
 782                 KASSERT(lock->lo_flags & LO_WITNESS,
 783                     ("%s: lock %s is on pending list but not LO_WITNESS",
 784                     __func__, lock->lo_name));
 785                 lock->lo_witness = enroll(pending_locks[i].wh_type,
 786                     LOCK_CLASS(lock));
 787         }
 788
 789         /* Mark the witness code as being ready for use. */
 790         witness_cold = 0;
 791
 792         mtx_lock(&Giant);
 793 }
 794 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
 795     NULL);
 796
 797 void
 798 witness_init(struct lock_object *lock, const char *type)
 799 {
 800         struct lock_class *class;
 801
 802         /* Various sanity checks. */
 803         class = LOCK_CLASS(lock);
 804         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 805             (class->lc_flags & LC_RECURSABLE) == 0)
 806                 kassert_panic("%s: lock (%s) %s can not be recursable",
 807                     __func__, class->lc_name, lock->lo_name);
 808         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 809             (class->lc_flags & LC_SLEEPABLE) == 0)
 810                 kassert_panic("%s: lock (%s) %s can not be sleepable",
 811                     __func__, class->lc_name, lock->lo_name);
 812         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 813             (class->lc_flags & LC_UPGRADABLE) == 0)
 814                 kassert_panic("%s: lock (%s) %s can not be upgradable",
 815                     __func__, class->lc_name, lock->lo_name);
 816
 817         /*
 818          * If we shouldn't watch this lock, then just clear lo_witness.
 819          * Otherwise, if witness_cold is set, then it is too early to
 820          * enroll this lock, so defer it to witness_initialize() by adding
 821          * it to the pending_locks list.  If it is not too early, then enroll
 822          * the lock now.
 823          */
 824         if (witness_watch < 1 || panicstr != NULL ||
 825             (lock->lo_flags & LO_WITNESS) == 0)
 826                 lock->lo_witness = NULL;
 827         else if (witness_cold) {
 828                 pending_locks[pending_cnt].wh_lock = lock;
 829                 pending_locks[pending_cnt++].wh_type = type;
 830                 if (pending_cnt > WITNESS_PENDLIST)
 831                         panic("%s: pending locks list is too small, "
 832                             "increase WITNESS_PENDLIST\n",
 833                             __func__);
 834         } else
 835                 lock->lo_witness = enroll(type, class);
 836 }
 837
 838 void
 839 witness_destroy(struct lock_object *lock)
 840 {
 841         struct lock_class *class;
 842         struct witness *w;
 843
 844         class = LOCK_CLASS(lock);
 845
 846         if (witness_cold)
 847                 panic("lock (%s) %s destroyed while witness_cold",
 848                     class->lc_name, lock->lo_name);
 849
 850         /* XXX: need to verify that no one holds the lock */
 851         if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
 852                 return;
 853         w = lock->lo_witness;
 854
 855         mtx_lock_spin(&w_mtx);
 856         MPASS(w->w_refcount > 0);
 857         w->w_refcount--;
 858
 859         if (w->w_refcount == 0)
 860                 depart(w);
 861         mtx_unlock_spin(&w_mtx);
 862 }
 863
 864 #ifdef DDB
 865 static void
 866 witness_ddb_compute_levels(void)
 867 {
 868         struct witness *w;
 869
 870         /*
 871          * First clear all levels.
 872          */
 873         STAILQ_FOREACH(w, &w_all, w_list)
 874                 w->w_ddb_level = -1;
 875
 876         /*
 877          * Look for locks with no parents and level all their descendants.
 878          */
 879         STAILQ_FOREACH(w, &w_all, w_list) {
 880
 881                 /* If the witness has ancestors (is not a root), skip it. */
 882                 if (w->w_num_ancestors > 0)
 883                         continue;
 884                 witness_ddb_level_descendants(w, 0);
 885         }
 886 }
 887
 888 static void
 889 witness_ddb_level_descendants(struct witness *w, int l)
 890 {
 891         int i;
 892
 893         if (w->w_ddb_level >= l)
 894                 return;
 895
 896         w->w_ddb_level = l;
 897         l++;
 898
 899         for (i = 1; i <= w_max_used_index; i++) {
 900                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 901                         witness_ddb_level_descendants(&w_data[i], l);
 902         }
 903 }
 904
 905 static void
 906 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
 907     struct witness *w, int indent)
 908 {
 909         int i;
 910
 911         for (i = 0; i < indent; i++)
 912                 prnt(" ");
 913         prnt("%s (type: %s, depth: %d, active refs: %d)",
 914              w->w_name, w->w_class->lc_name,
 915              w->w_ddb_level, w->w_refcount);
 916         if (w->w_displayed) {
 917                 prnt(" -- (already displayed)\n");
 918                 return;
 919         }
 920         w->w_displayed = 1;
 921         if (w->w_file != NULL && w->w_line != 0)
 922                 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
 923                     w->w_line);
 924         else
 925                 prnt(" -- never acquired\n");
 926         indent++;
 927         WITNESS_INDEX_ASSERT(w->w_index);
 928         for (i = 1; i <= w_max_used_index; i++) {
 929                 if (db_pager_quit)
 930                         return;
 931                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 932                         witness_ddb_display_descendants(prnt, &w_data[i],
 933                             indent);
 934         }
 935 }
 936
 937 static void
 938 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
 939     struct witness_list *list)
 940 {
 941         struct witness *w;
 942
 943         STAILQ_FOREACH(w, list, w_typelist) {
 944                 if (w->w_file == NULL || w->w_ddb_level > 0)
 945                         continue;
 946
 947                 /* This lock has no anscestors - display its descendants. */
 948                 witness_ddb_display_descendants(prnt, w, 0);
 949                 if (db_pager_quit)
 950                         return;
 951         }
 952 }
 953
 954 static void
 955 witness_ddb_display(int(*prnt)(const char *fmt, ...))
 956 {
 957         struct witness *w;
 958
 959         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
 960         witness_ddb_compute_levels();
 961
 962         /* Clear all the displayed flags. */
 963         STAILQ_FOREACH(w, &w_all, w_list)
 964                 w->w_displayed = 0;
 965
 966         /*
 967          * First, handle sleep locks which have been acquired at least
 968          * once.
 969          */
 970         prnt("Sleep locks:\n");
 971         witness_ddb_display_list(prnt, &w_sleep);
 972         if (db_pager_quit)
 973                 return;
 974
 975         /*
 976          * Now do spin locks which have been acquired at least once.
 977          */
 978         prnt("\nSpin locks:\n");
 979         witness_ddb_display_list(prnt, &w_spin);
 980         if (db_pager_quit)
 981                 return;
 982
 983         /*
 984          * Finally, any locks which have not been acquired yet.
 985          */
 986         prnt("\nLocks which were never acquired:\n");
 987         STAILQ_FOREACH(w, &w_all, w_list) {
 988                 if (w->w_file != NULL || w->w_refcount == 0)
 989                         continue;
 990                 prnt("%s (type: %s, depth: %d)\n", w->w_name,
 991                     w->w_class->lc_name, w->w_ddb_level);
 992                 if (db_pager_quit)
 993                         return;
 994         }
 995 }
 996 #endif /* DDB */
 997
 998 int
 999 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1000 {
1001
1002         if (witness_watch == -1 || panicstr != NULL)
1003                 return (0);
1004
1005         /* Require locks that witness knows about. */
1006         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1007             lock2->lo_witness == NULL)
1008                 return (EINVAL);
1009
1010         mtx_assert(&w_mtx, MA_NOTOWNED);
1011         mtx_lock_spin(&w_mtx);
1012
1013         /*
1014          * If we already have either an explicit or implied lock order that
1015          * is the other way around, then return an error.
1016          */
1017         if (witness_watch &&
1018             isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1019                 mtx_unlock_spin(&w_mtx);
1020                 return (EDOOFUS);
1021         }
1022
1023         /* Try to add the new order. */
1024         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1025             lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1026         itismychild(lock1->lo_witness, lock2->lo_witness);
1027         mtx_unlock_spin(&w_mtx);
1028         return (0);
1029 }
1030
1031 void
1032 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1033     int line, struct lock_object *interlock)
1034 {
1035         struct lock_list_entry *lock_list, *lle;
1036         struct lock_instance *lock1, *lock2, *plock;
1037         struct lock_class *class, *iclass;
1038         struct witness *w, *w1;
1039         struct thread *td;
1040         int i, j;
1041
1042         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1043             panicstr != NULL)
1044                 return;
1045
1046         w = lock->lo_witness;
1047         class = LOCK_CLASS(lock);
1048         td = curthread;
1049
1050         if (class->lc_flags & LC_SLEEPLOCK) {
1051
1052                 /*
1053                  * Since spin locks include a critical section, this check
1054                  * implicitly enforces a lock order of all sleep locks before
1055                  * all spin locks.
1056                  */
1057                 if (td->td_critnest != 0 && !kdb_active)
1058                         kassert_panic("acquiring blockable sleep lock with "
1059                             "spinlock or critical section held (%s) %s @ %s:%d",
1060                             class->lc_name, lock->lo_name,
1061                             fixup_filename(file), line);
1062
1063                 /*
1064                  * If this is the first lock acquired then just return as
1065                  * no order checking is needed.
1066                  */
1067                 lock_list = td->td_sleeplocks;
1068                 if (lock_list == NULL || lock_list->ll_count == 0)
1069                         return;
1070         } else {
1071
1072                 /*
1073                  * If this is the first lock, just return as no order
1074                  * checking is needed.  Avoid problems with thread
1075                  * migration pinning the thread while checking if
1076                  * spinlocks are held.  If at least one spinlock is held
1077                  * the thread is in a safe path and it is allowed to
1078                  * unpin it.
1079                  */
1080                 sched_pin();
1081                 lock_list = PCPU_GET(spinlocks);
1082                 if (lock_list == NULL || lock_list->ll_count == 0) {
1083                         sched_unpin();
1084                         return;
1085                 }
1086                 sched_unpin();
1087         }
1088
1089         /*
1090          * Check to see if we are recursing on a lock we already own.  If
1091          * so, make sure that we don't mismatch exclusive and shared lock
1092          * acquires.
1093          */
1094         lock1 = find_instance(lock_list, lock);
1095         if (lock1 != NULL) {
1096                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1097                     (flags & LOP_EXCLUSIVE) == 0) {
1098                         printf("shared lock of (%s) %s @ %s:%d\n",
1099                             class->lc_name, lock->lo_name,
1100                             fixup_filename(file), line);
1101                         printf("while exclusively locked from %s:%d\n",
1102                             fixup_filename(lock1->li_file), lock1->li_line);
1103                         kassert_panic("excl->share");
1104                 }
1105                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1106                     (flags & LOP_EXCLUSIVE) != 0) {
1107                         printf("exclusive lock of (%s) %s @ %s:%d\n",
1108                             class->lc_name, lock->lo_name,
1109                             fixup_filename(file), line);
1110                         printf("while share locked from %s:%d\n",
1111                             fixup_filename(lock1->li_file), lock1->li_line);
1112                         kassert_panic("share->excl");
1113                 }
1114                 return;
1115         }
1116
1117         /* Warn if the interlock is not locked exactly once. */
1118         if (interlock != NULL) {
1119                 iclass = LOCK_CLASS(interlock);
1120                 lock1 = find_instance(lock_list, interlock);
1121                 if (lock1 == NULL)
1122                         kassert_panic("interlock (%s) %s not locked @ %s:%d",
1123                             iclass->lc_name, interlock->lo_name,
1124                             fixup_filename(file), line);
1125                 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1126                         kassert_panic("interlock (%s) %s recursed @ %s:%d",
1127                             iclass->lc_name, interlock->lo_name,
1128                             fixup_filename(file), line);
1129         }
1130
1131         /*
1132          * Find the previously acquired lock, but ignore interlocks.
1133          */
1134         plock = &lock_list->ll_children[lock_list->ll_count - 1];
1135         if (interlock != NULL && plock->li_lock == interlock) {
1136                 if (lock_list->ll_count > 1)
1137                         plock =
1138                             &lock_list->ll_children[lock_list->ll_count - 2];
1139                 else {
1140                         lle = lock_list->ll_next;
1141
1142                         /*
1143                          * The interlock is the only lock we hold, so
1144                          * simply return.
1145                          */
1146                         if (lle == NULL)
1147                                 return;
1148                         plock = &lle->ll_children[lle->ll_count - 1];
1149                 }
1150         }
1151
1152         /*
1153          * Try to perform most checks without a lock.  If this succeeds we
1154          * can skip acquiring the lock and return success.
1155          */
1156         w1 = plock->li_lock->lo_witness;
1157         if (witness_lock_order_check(w1, w))
1158                 return;
1159
1160         /*
1161          * Check for duplicate locks of the same type.  Note that we only
1162          * have to check for this on the last lock we just acquired.  Any
1163          * other cases will be caught as lock order violations.
1164          */
1165         mtx_lock_spin(&w_mtx);
1166         witness_lock_order_add(w1, w);
1167         if (w1 == w) {
1168                 i = w->w_index;
1169                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1170                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1171                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
1172                         w->w_reversed = 1;
1173                         mtx_unlock_spin(&w_mtx);
1174                         printf(
1175                             "acquiring duplicate lock of same type: \"%s\"\n",
1176                             w->w_name);
1177                         printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1178                             fixup_filename(plock->li_file), plock->li_line);
1179                         printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1180                             fixup_filename(file), line);
1181                         witness_debugger(1);
1182                 } else
1183                         mtx_unlock_spin(&w_mtx);
1184                 return;
1185         }
1186         mtx_assert(&w_mtx, MA_OWNED);
1187
1188         /*
1189          * If we know that the lock we are acquiring comes after
1190          * the lock we most recently acquired in the lock order tree,
1191          * then there is no need for any further checks.
1192          */
1193         if (isitmychild(w1, w))
1194                 goto out;
1195
1196         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1197                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1198
1199                         MPASS(j < WITNESS_COUNT);
1200                         lock1 = &lle->ll_children[i];
1201
1202                         /*
1203                          * Ignore the interlock.
1204                          */
1205                         if (interlock == lock1->li_lock)
1206                                 continue;
1207
1208                         /*
1209                          * If this lock doesn't undergo witness checking,
1210                          * then skip it.
1211                          */
1212                         w1 = lock1->li_lock->lo_witness;
1213                         if (w1 == NULL) {
1214                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1215                                     ("lock missing witness structure"));
1216                                 continue;
1217                         }
1218
1219                         /*
1220                          * If we are locking Giant and this is a sleepable
1221                          * lock, then skip it.
1222                          */
1223                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1224                             lock == &Giant.lock_object)
1225                                 continue;
1226
1227                         /*
1228                          * If we are locking a sleepable lock and this lock
1229                          * is Giant, then skip it.
1230                          */
1231                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1232                             lock1->li_lock == &Giant.lock_object)
1233                                 continue;
1234
1235                         /*
1236                          * If we are locking a sleepable lock and this lock
1237                          * isn't sleepable, we want to treat it as a lock
1238                          * order violation to enfore a general lock order of
1239                          * sleepable locks before non-sleepable locks.
1240                          */
1241                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1242                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1243                                 goto reversal;
1244
1245                         /*
1246                          * If we are locking Giant and this is a non-sleepable
1247                          * lock, then treat it as a reversal.
1248                          */
1249                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1250                             lock == &Giant.lock_object)
1251                                 goto reversal;
1252
1253                         /*
1254                          * Check the lock order hierarchy for a reveresal.
1255                          */
1256                         if (!isitmydescendant(w, w1))
1257                                 continue;
1258                 reversal:
1259
1260                         /*
1261                          * We have a lock order violation, check to see if it
1262                          * is allowed or has already been yelled about.
1263                          */
1264 #ifdef BLESSING
1265
1266                         /*
1267                          * If the lock order is blessed, just bail.  We don't
1268                          * look for other lock order violations though, which
1269                          * may be a bug.
1270                          */
1271                         if (blessed(w, w1))
1272                                 goto out;
1273 #endif
1274
1275                         /* Bail if this violation is known */
1276                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1277                                 goto out;
1278
1279                         /* Record this as a violation */
1280                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1281                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1282                         w->w_reversed = w1->w_reversed = 1;
1283                         witness_increment_graph_generation();
1284                         mtx_unlock_spin(&w_mtx);
1285
1286 #ifdef WITNESS_NO_VNODE
1287                         /*
1288                          * There are known LORs between VNODE locks. They are
1289                          * not an indication of a bug. VNODE locks are flagged
1290                          * as such (LO_IS_VNODE) and we don't yell if the LOR
1291                          * is between 2 VNODE locks.
1292                          */
1293                         if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1294                             (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1295                                 return;
1296 #endif
1297
1298                         /*
1299                          * Ok, yell about it.
1300                          */
1301                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1302                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1303                                 printf(
1304                 "lock order reversal: (sleepable after non-sleepable)\n");
1305                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1306                             && lock == &Giant.lock_object)
1307                                 printf(
1308                 "lock order reversal: (Giant after non-sleepable)\n");
1309                         else
1310                                 printf("lock order reversal:\n");
1311
1312                         /*
1313                          * Try to locate an earlier lock with
1314                          * witness w in our list.
1315                          */
1316                         do {
1317                                 lock2 = &lle->ll_children[i];
1318                                 MPASS(lock2->li_lock != NULL);
1319                                 if (lock2->li_lock->lo_witness == w)
1320                                         break;
1321                                 if (i == 0 && lle->ll_next != NULL) {
1322                                         lle = lle->ll_next;
1323                                         i = lle->ll_count - 1;
1324                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1325                                 } else
1326                                         i--;
1327                         } while (i >= 0);
1328                         if (i < 0) {
1329                                 printf(" 1st %p %s (%s) @ %s:%d\n",
1330                                     lock1->li_lock, lock1->li_lock->lo_name,
1331                                     w1->w_name, fixup_filename(lock1->li_file),
1332                                     lock1->li_line);
1333                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1334                                     lock->lo_name, w->w_name,
1335                                     fixup_filename(file), line);
1336                         } else {
1337                                 printf(" 1st %p %s (%s) @ %s:%d\n",
1338                                     lock2->li_lock, lock2->li_lock->lo_name,
1339                                     lock2->li_lock->lo_witness->w_name,
1340                                     fixup_filename(lock2->li_file),
1341                                     lock2->li_line);
1342                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
1343                                     lock1->li_lock, lock1->li_lock->lo_name,
1344                                     w1->w_name, fixup_filename(lock1->li_file),
1345                                     lock1->li_line);
1346                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1347                                     lock->lo_name, w->w_name,
1348                                     fixup_filename(file), line);
1349                         }
1350                         witness_debugger(1);
1351                         return;
1352                 }
1353         }
1354
1355         /*
1356          * If requested, build a new lock order.  However, don't build a new
1357          * relationship between a sleepable lock and Giant if it is in the
1358          * wrong direction.  The correct lock order is that sleepable locks
1359          * always come before Giant.
1360          */
1361         if (flags & LOP_NEWORDER &&
1362             !(plock->li_lock == &Giant.lock_object &&
1363             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1364                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1365                     w->w_name, plock->li_lock->lo_witness->w_name);
1366                 itismychild(plock->li_lock->lo_witness, w);
1367         }
1368 out:
1369         mtx_unlock_spin(&w_mtx);
1370 }
1371
1372 void
1373 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1374 {
1375         struct lock_list_entry **lock_list, *lle;
1376         struct lock_instance *instance;
1377         struct witness *w;
1378         struct thread *td;
1379
1380         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1381             panicstr != NULL)
1382                 return;
1383         w = lock->lo_witness;
1384         td = curthread;
1385
1386         /* Determine lock list for this lock. */
1387         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1388                 lock_list = &td->td_sleeplocks;
1389         else
1390                 lock_list = PCPU_PTR(spinlocks);
1391
1392         /* Check to see if we are recursing on a lock we already own. */
1393         instance = find_instance(*lock_list, lock);
1394         if (instance != NULL) {
1395                 instance->li_flags++;
1396                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1397                     td->td_proc->p_pid, lock->lo_name,
1398                     instance->li_flags & LI_RECURSEMASK);
1399                 instance->li_file = file;
1400                 instance->li_line = line;
1401                 return;
1402         }
1403
1404         /* Update per-witness last file and line acquire. */
1405         w->w_file = file;
1406         w->w_line = line;
1407
1408         /* Find the next open lock instance in the list and fill it. */
1409         lle = *lock_list;
1410         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1411                 lle = witness_lock_list_get();
1412                 if (lle == NULL)
1413                         return;
1414                 lle->ll_next = *lock_list;
1415                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1416                     td->td_proc->p_pid, lle);
1417                 *lock_list = lle;
1418         }
1419         instance = &lle->ll_children[lle->ll_count++];
1420         instance->li_lock = lock;
1421         instance->li_line = line;
1422         instance->li_file = file;
1423         if ((flags & LOP_EXCLUSIVE) != 0)
1424                 instance->li_flags = LI_EXCLUSIVE;
1425         else
1426                 instance->li_flags = 0;
1427         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1428             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1429 }
1430
1431 void
1432 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1433 {
1434         struct lock_instance *instance;
1435         struct lock_class *class;
1436
1437         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1438         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1439                 return;
1440         class = LOCK_CLASS(lock);
1441         if (witness_watch) {
1442                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1443                         kassert_panic(
1444                             "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1445                             class->lc_name, lock->lo_name,
1446                             fixup_filename(file), line);
1447                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1448                         kassert_panic(
1449                             "upgrade of non-sleep lock (%s) %s @ %s:%d",
1450                             class->lc_name, lock->lo_name,
1451                             fixup_filename(file), line);
1452         }
1453         instance = find_instance(curthread->td_sleeplocks, lock);
1454         if (instance == NULL) {
1455                 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1456                     class->lc_name, lock->lo_name,
1457                     fixup_filename(file), line);
1458                 return;
1459         }
1460         if (witness_watch) {
1461                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1462                         kassert_panic(
1463                             "upgrade of exclusive lock (%s) %s @ %s:%d",
1464                             class->lc_name, lock->lo_name,
1465                             fixup_filename(file), line);
1466                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1467                         kassert_panic(
1468                             "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1469                             class->lc_name, lock->lo_name,
1470                             instance->li_flags & LI_RECURSEMASK,
1471                             fixup_filename(file), line);
1472         }
1473         instance->li_flags |= LI_EXCLUSIVE;
1474 }
1475
1476 void
1477 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1478     int line)
1479 {
1480         struct lock_instance *instance;
1481         struct lock_class *class;
1482
1483         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1484         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1485                 return;
1486         class = LOCK_CLASS(lock);
1487         if (witness_watch) {
1488                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1489                         kassert_panic(
1490                             "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1491                             class->lc_name, lock->lo_name,
1492                             fixup_filename(file), line);
1493                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1494                         kassert_panic(
1495                             "downgrade of non-sleep lock (%s) %s @ %s:%d",
1496                             class->lc_name, lock->lo_name,
1497                             fixup_filename(file), line);
1498         }
1499         instance = find_instance(curthread->td_sleeplocks, lock);
1500         if (instance == NULL) {
1501                 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1502                     class->lc_name, lock->lo_name,
1503                     fixup_filename(file), line);
1504                 return;
1505         }
1506         if (witness_watch) {
1507                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1508                         kassert_panic(
1509                             "downgrade of shared lock (%s) %s @ %s:%d",
1510                             class->lc_name, lock->lo_name,
1511                             fixup_filename(file), line);
1512                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1513                         kassert_panic(
1514                             "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1515                             class->lc_name, lock->lo_name,
1516                             instance->li_flags & LI_RECURSEMASK,
1517                             fixup_filename(file), line);
1518         }
1519         instance->li_flags &= ~LI_EXCLUSIVE;
1520 }
1521
1522 void
1523 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1524 {
1525         struct lock_list_entry **lock_list, *lle;
1526         struct lock_instance *instance;
1527         struct lock_class *class;
1528         struct thread *td;
1529         register_t s;
1530         int i, j;
1531
1532         if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1533                 return;
1534         td = curthread;
1535         class = LOCK_CLASS(lock);
1536
1537         /* Find lock instance associated with this lock. */
1538         if (class->lc_flags & LC_SLEEPLOCK)
1539                 lock_list = &td->td_sleeplocks;
1540         else
1541                 lock_list = PCPU_PTR(spinlocks);
1542         lle = *lock_list;
1543         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1544                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1545                         instance = &(*lock_list)->ll_children[i];
1546                         if (instance->li_lock == lock)
1547                                 goto found;
1548                 }
1549
1550         /*
1551          * When disabling WITNESS through witness_watch we could end up in
1552          * having registered locks in the td_sleeplocks queue.
1553          * We have to make sure we flush these queues, so just search for
1554          * eventual register locks and remove them.
1555          */
1556         if (witness_watch > 0) {
1557                 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1558                     lock->lo_name, fixup_filename(file), line);
1559                 return;
1560         } else {
1561                 return;
1562         }
1563 found:
1564
1565         /* First, check for shared/exclusive mismatches. */
1566         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1567             (flags & LOP_EXCLUSIVE) == 0) {
1568                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1569                     lock->lo_name, fixup_filename(file), line);
1570                 printf("while exclusively locked from %s:%d\n",
1571                     fixup_filename(instance->li_file), instance->li_line);
1572                 kassert_panic("excl->ushare");
1573         }
1574         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1575             (flags & LOP_EXCLUSIVE) != 0) {
1576                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1577                     lock->lo_name, fixup_filename(file), line);
1578                 printf("while share locked from %s:%d\n",
1579                     fixup_filename(instance->li_file),
1580                     instance->li_line);
1581                 kassert_panic("share->uexcl");
1582         }
1583         /* If we are recursed, unrecurse. */
1584         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1585                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1586                     td->td_proc->p_pid, instance->li_lock->lo_name,
1587                     instance->li_flags);
1588                 instance->li_flags--;
1589                 return;
1590         }
1591         /* The lock is now being dropped, check for NORELEASE flag */
1592         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1593                 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1594                     lock->lo_name, fixup_filename(file), line);
1595                 kassert_panic("lock marked norelease");
1596         }
1597
1598         /* Otherwise, remove this item from the list. */
1599         s = intr_disable();
1600         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1601             td->td_proc->p_pid, instance->li_lock->lo_name,
1602             (*lock_list)->ll_count - 1);
1603         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1604                 (*lock_list)->ll_children[j] =
1605                     (*lock_list)->ll_children[j + 1];
1606         (*lock_list)->ll_count--;
1607         intr_restore(s);
1608
1609         /*
1610          * In order to reduce contention on w_mtx, we want to keep always an
1611          * head object into lists so that frequent allocation from the
1612          * free witness pool (and subsequent locking) is avoided.
1613          * In order to maintain the current code simple, when the head
1614          * object is totally unloaded it means also that we do not have
1615          * further objects in the list, so the list ownership needs to be
1616          * hand over to another object if the current head needs to be freed.
1617          */
1618         if ((*lock_list)->ll_count == 0) {
1619                 if (*lock_list == lle) {
1620                         if (lle->ll_next == NULL)
1621                                 return;
1622                 } else
1623                         lle = *lock_list;
1624                 *lock_list = lle->ll_next;
1625                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1626                     td->td_proc->p_pid, lle);
1627                 witness_lock_list_free(lle);
1628         }
1629 }
1630
1631 void
1632 witness_thread_exit(struct thread *td)
1633 {
1634         struct lock_list_entry *lle;
1635         int i, n;
1636
1637         lle = td->td_sleeplocks;
1638         if (lle == NULL || panicstr != NULL)
1639                 return;
1640         if (lle->ll_count != 0) {
1641                 for (n = 0; lle != NULL; lle = lle->ll_next)
1642                         for (i = lle->ll_count - 1; i >= 0; i--) {
1643                                 if (n == 0)
1644                 printf("Thread %p exiting with the following locks held:\n",
1645                                             td);
1646                                 n++;
1647                                 witness_list_lock(&lle->ll_children[i], printf);
1648
1649                         }
1650                 kassert_panic(
1651                     "Thread %p cannot exit while holding sleeplocks\n", td);
1652         }
1653         witness_lock_list_free(lle);
1654 }
1655
1656 /*
1657  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1658  * exempt Giant and sleepable locks from the checks as well.  If any
1659  * non-exempt locks are held, then a supplied message is printed to the
1660  * console along with a list of the offending locks.  If indicated in the
1661  * flags then a failure results in a panic as well.
1662  */
1663 int
1664 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1665 {
1666         struct lock_list_entry *lock_list, *lle;
1667         struct lock_instance *lock1;
1668         struct thread *td;
1669         va_list ap;
1670         int i, n;
1671
1672         if (witness_cold || witness_watch < 1 || panicstr != NULL)
1673                 return (0);
1674         n = 0;
1675         td = curthread;
1676         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1677                 for (i = lle->ll_count - 1; i >= 0; i--) {
1678                         lock1 = &lle->ll_children[i];
1679                         if (lock1->li_lock == lock)
1680                                 continue;
1681                         if (flags & WARN_GIANTOK &&
1682                             lock1->li_lock == &Giant.lock_object)
1683                                 continue;
1684                         if (flags & WARN_SLEEPOK &&
1685                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1686                                 continue;
1687                         if (n == 0) {
1688                                 va_start(ap, fmt);
1689                                 vprintf(fmt, ap);
1690                                 va_end(ap);
1691                                 printf(" with the following");
1692                                 if (flags & WARN_SLEEPOK)
1693                                         printf(" non-sleepable");
1694                                 printf(" locks held:\n");
1695                         }
1696                         n++;
1697                         witness_list_lock(lock1, printf);
1698                 }
1699
1700         /*
1701          * Pin the thread in order to avoid problems with thread migration.
1702          * Once that all verifies are passed about spinlocks ownership,
1703          * the thread is in a safe path and it can be unpinned.
1704          */
1705         sched_pin();
1706         lock_list = PCPU_GET(spinlocks);
1707         if (lock_list != NULL && lock_list->ll_count != 0) {
1708                 sched_unpin();
1709
1710                 /*
1711                  * We should only have one spinlock and as long as
1712                  * the flags cannot match for this locks class,
1713                  * check if the first spinlock is the one curthread
1714                  * should hold.
1715                  */
1716                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1717                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1718                     lock1->li_lock == lock && n == 0)
1719                         return (0);
1720
1721                 va_start(ap, fmt);
1722                 vprintf(fmt, ap);
1723                 va_end(ap);
1724                 printf(" with the following");
1725                 if (flags & WARN_SLEEPOK)
1726                         printf(" non-sleepable");
1727                 printf(" locks held:\n");
1728                 n += witness_list_locks(&lock_list, printf);
1729         } else
1730                 sched_unpin();
1731         if (flags & WARN_PANIC && n)
1732                 kassert_panic("%s", __func__);
1733         else
1734                 witness_debugger(n);
1735         return (n);
1736 }
1737
1738 const char *
1739 witness_file(struct lock_object *lock)
1740 {
1741         struct witness *w;
1742
1743         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1744                 return ("?");
1745         w = lock->lo_witness;
1746         return (w->w_file);
1747 }
1748
1749 int
1750 witness_line(struct lock_object *lock)
1751 {
1752         struct witness *w;
1753
1754         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1755                 return (0);
1756         w = lock->lo_witness;
1757         return (w->w_line);
1758 }
1759
1760 static struct witness *
1761 enroll(const char *description, struct lock_class *lock_class)
1762 {
1763         struct witness *w;
1764         struct witness_list *typelist;
1765
1766         MPASS(description != NULL);
1767
1768         if (witness_watch == -1 || panicstr != NULL)
1769                 return (NULL);
1770         if ((lock_class->lc_flags & LC_SPINLOCK)) {
1771                 if (witness_skipspin)
1772                         return (NULL);
1773                 else
1774                         typelist = &w_spin;
1775         } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1776                 typelist = &w_sleep;
1777         } else {
1778                 kassert_panic("lock class %s is not sleep or spin",
1779                     lock_class->lc_name);
1780                 return (NULL);
1781         }
1782
1783         mtx_lock_spin(&w_mtx);
1784         w = witness_hash_get(description);
1785         if (w)
1786                 goto found;
1787         if ((w = witness_get()) == NULL)
1788                 return (NULL);
1789         MPASS(strlen(description) < MAX_W_NAME);
1790         strcpy(w->w_name, description);
1791         w->w_class = lock_class;
1792         w->w_refcount = 1;
1793         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1794         if (lock_class->lc_flags & LC_SPINLOCK) {
1795                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1796                 w_spin_cnt++;
1797         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1798                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1799                 w_sleep_cnt++;
1800         }
1801
1802         /* Insert new witness into the hash */
1803         witness_hash_put(w);
1804         witness_increment_graph_generation();
1805         mtx_unlock_spin(&w_mtx);
1806         return (w);
1807 found:
1808         w->w_refcount++;
1809         mtx_unlock_spin(&w_mtx);
1810         if (lock_class != w->w_class)
1811                 kassert_panic(
1812                         "lock (%s) %s does not match earlier (%s) lock",
1813                         description, lock_class->lc_name,
1814                         w->w_class->lc_name);
1815         return (w);
1816 }
1817
1818 static void
1819 depart(struct witness *w)
1820 {
1821         struct witness_list *list;
1822
1823         MPASS(w->w_refcount == 0);
1824         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1825                 list = &w_sleep;
1826                 w_sleep_cnt--;
1827         } else {
1828                 list = &w_spin;
1829                 w_spin_cnt--;
1830         }
1831         /*
1832          * Set file to NULL as it may point into a loadable module.
1833          */
1834         w->w_file = NULL;
1835         w->w_line = 0;
1836         witness_increment_graph_generation();
1837 }
1838
1839
1840 static void
1841 adopt(struct witness *parent, struct witness *child)
1842 {
1843         int pi, ci, i, j;
1844
1845         if (witness_cold == 0)
1846                 mtx_assert(&w_mtx, MA_OWNED);
1847
1848         /* If the relationship is already known, there's no work to be done. */
1849         if (isitmychild(parent, child))
1850                 return;
1851
1852         /* When the structure of the graph changes, bump up the generation. */
1853         witness_increment_graph_generation();
1854
1855         /*
1856          * The hard part ... create the direct relationship, then propagate all
1857          * indirect relationships.
1858          */
1859         pi = parent->w_index;
1860         ci = child->w_index;
1861         WITNESS_INDEX_ASSERT(pi);
1862         WITNESS_INDEX_ASSERT(ci);
1863         MPASS(pi != ci);
1864         w_rmatrix[pi][ci] |= WITNESS_PARENT;
1865         w_rmatrix[ci][pi] |= WITNESS_CHILD;
1866
1867         /*
1868          * If parent was not already an ancestor of child,
1869          * then we increment the descendant and ancestor counters.
1870          */
1871         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1872                 parent->w_num_descendants++;
1873                 child->w_num_ancestors++;
1874         }
1875
1876         /*
1877          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1878          * an ancestor of 'pi' during this loop.
1879          */
1880         for (i = 1; i <= w_max_used_index; i++) {
1881                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1882                     (i != pi))
1883                         continue;
1884
1885                 /* Find each descendant of 'i' and mark it as a descendant. */
1886                 for (j = 1; j <= w_max_used_index; j++) {
1887
1888                         /*
1889                          * Skip children that are already marked as
1890                          * descendants of 'i'.
1891                          */
1892                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1893                                 continue;
1894
1895                         /*
1896                          * We are only interested in descendants of 'ci'. Note
1897                          * that 'ci' itself is counted as a descendant of 'ci'.
1898                          */
1899                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1900                             (j != ci))
1901                                 continue;
1902                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1903                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1904                         w_data[i].w_num_descendants++;
1905                         w_data[j].w_num_ancestors++;
1906
1907                         /*
1908                          * Make sure we aren't marking a node as both an
1909                          * ancestor and descendant. We should have caught
1910                          * this as a lock order reversal earlier.
1911                          */
1912                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1913                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1914                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1915                                     "both ancestor and descendant\n",
1916                                     i, j, w_rmatrix[i][j]);
1917                                 kdb_backtrace();
1918                                 printf("Witness disabled.\n");
1919                                 witness_watch = -1;
1920                         }
1921                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1922                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1923                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1924                                     "both ancestor and descendant\n",
1925                                     j, i, w_rmatrix[j][i]);
1926                                 kdb_backtrace();
1927                                 printf("Witness disabled.\n");
1928                                 witness_watch = -1;
1929                         }
1930                 }
1931         }
1932 }
1933
1934 static void
1935 itismychild(struct witness *parent, struct witness *child)
1936 {
1937         int unlocked;
1938
1939         MPASS(child != NULL && parent != NULL);
1940         if (witness_cold == 0)
1941                 mtx_assert(&w_mtx, MA_OWNED);
1942
1943         if (!witness_lock_type_equal(parent, child)) {
1944                 if (witness_cold == 0) {
1945                         unlocked = 1;
1946                         mtx_unlock_spin(&w_mtx);
1947                 } else {
1948                         unlocked = 0;
1949                 }
1950                 kassert_panic(
1951                     "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1952                     "the same lock type", __func__, parent->w_name,
1953                     parent->w_class->lc_name, child->w_name,
1954                     child->w_class->lc_name);
1955                 if (unlocked)
1956                         mtx_lock_spin(&w_mtx);
1957         }
1958         adopt(parent, child);
1959 }
1960
1961 /*
1962  * Generic code for the isitmy*() functions. The rmask parameter is the
1963  * expected relationship of w1 to w2.
1964  */
1965 static int
1966 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1967 {
1968         unsigned char r1, r2;
1969         int i1, i2;
1970
1971         i1 = w1->w_index;
1972         i2 = w2->w_index;
1973         WITNESS_INDEX_ASSERT(i1);
1974         WITNESS_INDEX_ASSERT(i2);
1975         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1976         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1977
1978         /* The flags on one better be the inverse of the flags on the other */
1979         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1980                 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1981                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
1982                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
1983                     "w_rmatrix[%d][%d] == %hhx\n",
1984                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1985                     i2, i1, r2);
1986                 kdb_backtrace();
1987                 printf("Witness disabled.\n");
1988                 witness_watch = -1;
1989         }
1990         return (r1 & rmask);
1991 }
1992
1993 /*
1994  * Checks if @child is a direct child of @parent.
1995  */
1996 static int
1997 isitmychild(struct witness *parent, struct witness *child)
1998 {
1999
2000         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2001 }
2002
2003 /*
2004  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2005  */
2006 static int
2007 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2008 {
2009
2010         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2011             __func__));
2012 }
2013
2014 #ifdef BLESSING
2015 static int
2016 blessed(struct witness *w1, struct witness *w2)
2017 {
2018         int i;
2019         struct witness_blessed *b;
2020
2021         for (i = 0; i < blessed_count; i++) {
2022                 b = &blessed_list[i];
2023                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2024                         if (strcmp(w2->w_name, b->b_lock2) == 0)
2025                                 return (1);
2026                         continue;
2027                 }
2028                 if (strcmp(w1->w_name, b->b_lock2) == 0)
2029                         if (strcmp(w2->w_name, b->b_lock1) == 0)
2030                                 return (1);
2031         }
2032         return (0);
2033 }
2034 #endif
2035
2036 static struct witness *
2037 witness_get(void)
2038 {
2039         struct witness *w;
2040         int index;
2041
2042         if (witness_cold == 0)
2043                 mtx_assert(&w_mtx, MA_OWNED);
2044
2045         if (witness_watch == -1) {
2046                 mtx_unlock_spin(&w_mtx);
2047                 return (NULL);
2048         }
2049         if (STAILQ_EMPTY(&w_free)) {
2050                 witness_watch = -1;
2051                 mtx_unlock_spin(&w_mtx);
2052                 printf("WITNESS: unable to allocate a new witness object\n");
2053                 return (NULL);
2054         }
2055         w = STAILQ_FIRST(&w_free);
2056         STAILQ_REMOVE_HEAD(&w_free, w_list);
2057         w_free_cnt--;
2058         index = w->w_index;
2059         MPASS(index > 0 && index == w_max_used_index+1 &&
2060             index < WITNESS_COUNT);
2061         bzero(w, sizeof(*w));
2062         w->w_index = index;
2063         if (index > w_max_used_index)
2064                 w_max_used_index = index;
2065         return (w);
2066 }
2067
2068 static void
2069 witness_free(struct witness *w)
2070 {
2071
2072         STAILQ_INSERT_HEAD(&w_free, w, w_list);
2073         w_free_cnt++;
2074 }
2075
2076 static struct lock_list_entry *
2077 witness_lock_list_get(void)
2078 {
2079         struct lock_list_entry *lle;
2080
2081         if (witness_watch == -1)
2082                 return (NULL);
2083         mtx_lock_spin(&w_mtx);
2084         lle = w_lock_list_free;
2085         if (lle == NULL) {
2086                 witness_watch = -1;
2087                 mtx_unlock_spin(&w_mtx);
2088                 printf("%s: witness exhausted\n", __func__);
2089                 return (NULL);
2090         }
2091         w_lock_list_free = lle->ll_next;
2092         mtx_unlock_spin(&w_mtx);
2093         bzero(lle, sizeof(*lle));
2094         return (lle);
2095 }
2096
2097 static void
2098 witness_lock_list_free(struct lock_list_entry *lle)
2099 {
2100
2101         mtx_lock_spin(&w_mtx);
2102         lle->ll_next = w_lock_list_free;
2103         w_lock_list_free = lle;
2104         mtx_unlock_spin(&w_mtx);
2105 }
2106
2107 static struct lock_instance *
2108 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2109 {
2110         struct lock_list_entry *lle;
2111         struct lock_instance *instance;
2112         int i;
2113
2114         for (lle = list; lle != NULL; lle = lle->ll_next)
2115                 for (i = lle->ll_count - 1; i >= 0; i--) {
2116                         instance = &lle->ll_children[i];
2117                         if (instance->li_lock == lock)
2118                                 return (instance);
2119                 }
2120         return (NULL);
2121 }
2122
2123 static void
2124 witness_list_lock(struct lock_instance *instance,
2125     int (*prnt)(const char *fmt, ...))
2126 {
2127         struct lock_object *lock;
2128
2129         lock = instance->li_lock;
2130         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2131             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2132         if (lock->lo_witness->w_name != lock->lo_name)
2133                 prnt(" (%s)", lock->lo_witness->w_name);
2134         prnt(" r = %d (%p) locked @ %s:%d\n",
2135             instance->li_flags & LI_RECURSEMASK, lock,
2136             fixup_filename(instance->li_file), instance->li_line);
2137 }
2138
2139 #ifdef DDB
2140 static int
2141 witness_thread_has_locks(struct thread *td)
2142 {
2143
2144         if (td->td_sleeplocks == NULL)
2145                 return (0);
2146         return (td->td_sleeplocks->ll_count != 0);
2147 }
2148
2149 static int
2150 witness_proc_has_locks(struct proc *p)
2151 {
2152         struct thread *td;
2153
2154         FOREACH_THREAD_IN_PROC(p, td) {
2155                 if (witness_thread_has_locks(td))
2156                         return (1);
2157         }
2158         return (0);
2159 }
2160 #endif
2161
2162 int
2163 witness_list_locks(struct lock_list_entry **lock_list,
2164     int (*prnt)(const char *fmt, ...))
2165 {
2166         struct lock_list_entry *lle;
2167         int i, nheld;
2168
2169         nheld = 0;
2170         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2171                 for (i = lle->ll_count - 1; i >= 0; i--) {
2172                         witness_list_lock(&lle->ll_children[i], prnt);
2173                         nheld++;
2174                 }
2175         return (nheld);
2176 }
2177
2178 /*
2179  * This is a bit risky at best.  We call this function when we have timed
2180  * out acquiring a spin lock, and we assume that the other CPU is stuck
2181  * with this lock held.  So, we go groveling around in the other CPU's
2182  * per-cpu data to try to find the lock instance for this spin lock to
2183  * see when it was last acquired.
2184  */
2185 void
2186 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2187     int (*prnt)(const char *fmt, ...))
2188 {
2189         struct lock_instance *instance;
2190         struct pcpu *pc;
2191
2192         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2193                 return;
2194         pc = pcpu_find(owner->td_oncpu);
2195         instance = find_instance(pc->pc_spinlocks, lock);
2196         if (instance != NULL)
2197                 witness_list_lock(instance, prnt);
2198 }
2199
2200 void
2201 witness_save(struct lock_object *lock, const char **filep, int *linep)
2202 {
2203         struct lock_list_entry *lock_list;
2204         struct lock_instance *instance;
2205         struct lock_class *class;
2206
2207         /*
2208          * This function is used independently in locking code to deal with
2209          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2210          * is gone.
2211          */
2212         if (SCHEDULER_STOPPED())
2213                 return;
2214         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2215         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2216                 return;
2217         class = LOCK_CLASS(lock);
2218         if (class->lc_flags & LC_SLEEPLOCK)
2219                 lock_list = curthread->td_sleeplocks;
2220         else {
2221                 if (witness_skipspin)
2222                         return;
2223                 lock_list = PCPU_GET(spinlocks);
2224         }
2225         instance = find_instance(lock_list, lock);
2226         if (instance == NULL) {
2227                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2228                     class->lc_name, lock->lo_name);
2229                 return;
2230         }
2231         *filep = instance->li_file;
2232         *linep = instance->li_line;
2233 }
2234
2235 void
2236 witness_restore(struct lock_object *lock, const char *file, int line)
2237 {
2238         struct lock_list_entry *lock_list;
2239         struct lock_instance *instance;
2240         struct lock_class *class;
2241
2242         /*
2243          * This function is used independently in locking code to deal with
2244          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2245          * is gone.
2246          */
2247         if (SCHEDULER_STOPPED())
2248                 return;
2249         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2250         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2251                 return;
2252         class = LOCK_CLASS(lock);
2253         if (class->lc_flags & LC_SLEEPLOCK)
2254                 lock_list = curthread->td_sleeplocks;
2255         else {
2256                 if (witness_skipspin)
2257                         return;
2258                 lock_list = PCPU_GET(spinlocks);
2259         }
2260         instance = find_instance(lock_list, lock);
2261         if (instance == NULL)
2262                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2263                     class->lc_name, lock->lo_name);
2264         lock->lo_witness->w_file = file;
2265         lock->lo_witness->w_line = line;
2266         if (instance == NULL)
2267                 return;
2268         instance->li_file = file;
2269         instance->li_line = line;
2270 }
2271
2272 void
2273 witness_assert(const struct lock_object *lock, int flags, const char *file,
2274     int line)
2275 {
2276 #ifdef INVARIANT_SUPPORT
2277         struct lock_instance *instance;
2278         struct lock_class *class;
2279
2280         if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2281                 return;
2282         class = LOCK_CLASS(lock);
2283         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2284                 instance = find_instance(curthread->td_sleeplocks, lock);
2285         else if ((class->lc_flags & LC_SPINLOCK) != 0)
2286                 instance = find_instance(PCPU_GET(spinlocks), lock);
2287         else {
2288                 kassert_panic("Lock (%s) %s is not sleep or spin!",
2289                     class->lc_name, lock->lo_name);
2290                 return;
2291         }
2292         switch (flags) {
2293         case LA_UNLOCKED:
2294                 if (instance != NULL)
2295                         kassert_panic("Lock (%s) %s locked @ %s:%d.",
2296                             class->lc_name, lock->lo_name,
2297                             fixup_filename(file), line);
2298                 break;
2299         case LA_LOCKED:
2300         case LA_LOCKED | LA_RECURSED:
2301         case LA_LOCKED | LA_NOTRECURSED:
2302         case LA_SLOCKED:
2303         case LA_SLOCKED | LA_RECURSED:
2304         case LA_SLOCKED | LA_NOTRECURSED:
2305         case LA_XLOCKED:
2306         case LA_XLOCKED | LA_RECURSED:
2307         case LA_XLOCKED | LA_NOTRECURSED:
2308                 if (instance == NULL) {
2309                         kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2310                             class->lc_name, lock->lo_name,
2311                             fixup_filename(file), line);
2312                         break;
2313                 }
2314                 if ((flags & LA_XLOCKED) != 0 &&
2315                     (instance->li_flags & LI_EXCLUSIVE) == 0)
2316                         kassert_panic(
2317                             "Lock (%s) %s not exclusively locked @ %s:%d.",
2318                             class->lc_name, lock->lo_name,
2319                             fixup_filename(file), line);
2320                 if ((flags & LA_SLOCKED) != 0 &&
2321                     (instance->li_flags & LI_EXCLUSIVE) != 0)
2322                         kassert_panic(
2323                             "Lock (%s) %s exclusively locked @ %s:%d.",
2324                             class->lc_name, lock->lo_name,
2325                             fixup_filename(file), line);
2326                 if ((flags & LA_RECURSED) != 0 &&
2327                     (instance->li_flags & LI_RECURSEMASK) == 0)
2328                         kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2329                             class->lc_name, lock->lo_name,
2330                             fixup_filename(file), line);
2331                 if ((flags & LA_NOTRECURSED) != 0 &&
2332                     (instance->li_flags & LI_RECURSEMASK) != 0)
2333                         kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2334                             class->lc_name, lock->lo_name,
2335                             fixup_filename(file), line);
2336                 break;
2337         default:
2338                 kassert_panic("Invalid lock assertion at %s:%d.",
2339                     fixup_filename(file), line);
2340
2341         }
2342 #endif  /* INVARIANT_SUPPORT */
2343 }
2344
2345 static void
2346 witness_setflag(struct lock_object *lock, int flag, int set)
2347 {
2348         struct lock_list_entry *lock_list;
2349         struct lock_instance *instance;
2350         struct lock_class *class;
2351
2352         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2353                 return;
2354         class = LOCK_CLASS(lock);
2355         if (class->lc_flags & LC_SLEEPLOCK)
2356                 lock_list = curthread->td_sleeplocks;
2357         else {
2358                 if (witness_skipspin)
2359                         return;
2360                 lock_list = PCPU_GET(spinlocks);
2361         }
2362         instance = find_instance(lock_list, lock);
2363         if (instance == NULL) {
2364                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2365                     class->lc_name, lock->lo_name);
2366                 return;
2367         }
2368
2369         if (set)
2370                 instance->li_flags |= flag;
2371         else
2372                 instance->li_flags &= ~flag;
2373 }
2374
2375 void
2376 witness_norelease(struct lock_object *lock)
2377 {
2378
2379         witness_setflag(lock, LI_NORELEASE, 1);
2380 }
2381
2382 void
2383 witness_releaseok(struct lock_object *lock)
2384 {
2385
2386         witness_setflag(lock, LI_NORELEASE, 0);
2387 }
2388
2389 #ifdef DDB
2390 static void
2391 witness_ddb_list(struct thread *td)
2392 {
2393
2394         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2395         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2396
2397         if (witness_watch < 1)
2398                 return;
2399
2400         witness_list_locks(&td->td_sleeplocks, db_printf);
2401
2402         /*
2403          * We only handle spinlocks if td == curthread.  This is somewhat broken
2404          * if td is currently executing on some other CPU and holds spin locks
2405          * as we won't display those locks.  If we had a MI way of getting
2406          * the per-cpu data for a given cpu then we could use
2407          * td->td_oncpu to get the list of spinlocks for this thread
2408          * and "fix" this.
2409          *
2410          * That still wouldn't really fix this unless we locked the scheduler
2411          * lock or stopped the other CPU to make sure it wasn't changing the
2412          * list out from under us.  It is probably best to just not try to
2413          * handle threads on other CPU's for now.
2414          */
2415         if (td == curthread && PCPU_GET(spinlocks) != NULL)
2416                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2417 }
2418
2419 DB_SHOW_COMMAND(locks, db_witness_list)
2420 {
2421         struct thread *td;
2422
2423         if (have_addr)
2424                 td = db_lookup_thread(addr, TRUE);
2425         else
2426                 td = kdb_thread;
2427         witness_ddb_list(td);
2428 }
2429
2430 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2431 {
2432         struct thread *td;
2433         struct proc *p;
2434
2435         /*
2436          * It would be nice to list only threads and processes that actually
2437          * held sleep locks, but that information is currently not exported
2438          * by WITNESS.
2439          */
2440         FOREACH_PROC_IN_SYSTEM(p) {
2441                 if (!witness_proc_has_locks(p))
2442                         continue;
2443                 FOREACH_THREAD_IN_PROC(p, td) {
2444                         if (!witness_thread_has_locks(td))
2445                                 continue;
2446                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2447                             p->p_comm, td, td->td_tid);
2448                         witness_ddb_list(td);
2449                         if (db_pager_quit)
2450                                 return;
2451                 }
2452         }
2453 }
2454 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2455
2456 DB_SHOW_COMMAND(witness, db_witness_display)
2457 {
2458
2459         witness_ddb_display(db_printf);
2460 }
2461 #endif
2462
2463 static int
2464 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2465 {
2466         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2467         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2468         struct sbuf *sb;
2469         u_int w_rmatrix1, w_rmatrix2;
2470         int error, generation, i, j;
2471
2472         tmp_data1 = NULL;
2473         tmp_data2 = NULL;
2474         tmp_w1 = NULL;
2475         tmp_w2 = NULL;
2476         if (witness_watch < 1) {
2477                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2478                 return (error);
2479         }
2480         if (witness_cold) {
2481                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2482                 return (error);
2483         }
2484         error = 0;
2485         sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2486         if (sb == NULL)
2487                 return (ENOMEM);
2488
2489         /* Allocate and init temporary storage space. */
2490         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2491         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2492         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2493             M_WAITOK | M_ZERO);
2494         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2495             M_WAITOK | M_ZERO);
2496         stack_zero(&tmp_data1->wlod_stack);
2497         stack_zero(&tmp_data2->wlod_stack);
2498
2499 restart:
2500         mtx_lock_spin(&w_mtx);
2501         generation = w_generation;
2502         mtx_unlock_spin(&w_mtx);
2503         sbuf_printf(sb, "Number of known direct relationships is %d\n",
2504             w_lohash.wloh_count);
2505         for (i = 1; i < w_max_used_index; i++) {
2506                 mtx_lock_spin(&w_mtx);
2507                 if (generation != w_generation) {
2508                         mtx_unlock_spin(&w_mtx);
2509
2510                         /* The graph has changed, try again. */
2511                         req->oldidx = 0;
2512                         sbuf_clear(sb);
2513                         goto restart;
2514                 }
2515
2516                 w1 = &w_data[i];
2517                 if (w1->w_reversed == 0) {
2518                         mtx_unlock_spin(&w_mtx);
2519                         continue;
2520                 }
2521
2522                 /* Copy w1 locally so we can release the spin lock. */
2523                 *tmp_w1 = *w1;
2524                 mtx_unlock_spin(&w_mtx);
2525
2526                 if (tmp_w1->w_reversed == 0)
2527                         continue;
2528                 for (j = 1; j < w_max_used_index; j++) {
2529                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2530                                 continue;
2531
2532                         mtx_lock_spin(&w_mtx);
2533                         if (generation != w_generation) {
2534                                 mtx_unlock_spin(&w_mtx);
2535
2536                                 /* The graph has changed, try again. */
2537                                 req->oldidx = 0;
2538                                 sbuf_clear(sb);
2539                                 goto restart;
2540                         }
2541
2542                         w2 = &w_data[j];
2543                         data1 = witness_lock_order_get(w1, w2);
2544                         data2 = witness_lock_order_get(w2, w1);
2545
2546                         /*
2547                          * Copy information locally so we can release the
2548                          * spin lock.
2549                          */
2550                         *tmp_w2 = *w2;
2551                         w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2552                         w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2553
2554                         if (data1) {
2555                                 stack_zero(&tmp_data1->wlod_stack);
2556                                 stack_copy(&data1->wlod_stack,
2557                                     &tmp_data1->wlod_stack);
2558                         }
2559                         if (data2 && data2 != data1) {
2560                                 stack_zero(&tmp_data2->wlod_stack);
2561                                 stack_copy(&data2->wlod_stack,
2562                                     &tmp_data2->wlod_stack);
2563                         }
2564                         mtx_unlock_spin(&w_mtx);
2565
2566                         sbuf_printf(sb,
2567             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2568                             tmp_w1->w_name, tmp_w1->w_class->lc_name,
2569                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
2570 #if 0
2571                         sbuf_printf(sb,
2572                         "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2573                             tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2574                             tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2575 #endif
2576                         if (data1) {
2577                                 sbuf_printf(sb,
2578                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2579                                     tmp_w1->w_name, tmp_w1->w_class->lc_name,
2580                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
2581                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2582                                 sbuf_printf(sb, "\n");
2583                         }
2584                         if (data2 && data2 != data1) {
2585                                 sbuf_printf(sb,
2586                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2587                                     tmp_w2->w_name, tmp_w2->w_class->lc_name,
2588                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
2589                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2590                                 sbuf_printf(sb, "\n");
2591                         }
2592                 }
2593         }
2594         mtx_lock_spin(&w_mtx);
2595         if (generation != w_generation) {
2596                 mtx_unlock_spin(&w_mtx);
2597
2598                 /*
2599                  * The graph changed while we were printing stack data,
2600                  * try again.
2601                  */
2602                 req->oldidx = 0;
2603                 sbuf_clear(sb);
2604                 goto restart;
2605         }
2606         mtx_unlock_spin(&w_mtx);
2607
2608         /* Free temporary storage space. */
2609         free(tmp_data1, M_TEMP);
2610         free(tmp_data2, M_TEMP);
2611         free(tmp_w1, M_TEMP);
2612         free(tmp_w2, M_TEMP);
2613
2614         sbuf_finish(sb);
2615         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2616         sbuf_delete(sb);
2617
2618         return (error);
2619 }
2620
2621 static int
2622 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2623 {
2624         struct witness *w;
2625         struct sbuf *sb;
2626         int error;
2627
2628         if (witness_watch < 1) {
2629                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2630                 return (error);
2631         }
2632         if (witness_cold) {
2633                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2634                 return (error);
2635         }
2636         error = 0;
2637
2638         error = sysctl_wire_old_buffer(req, 0);
2639         if (error != 0)
2640                 return (error);
2641         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2642         if (sb == NULL)
2643                 return (ENOMEM);
2644         sbuf_printf(sb, "\n");
2645
2646         mtx_lock_spin(&w_mtx);
2647         STAILQ_FOREACH(w, &w_all, w_list)
2648                 w->w_displayed = 0;
2649         STAILQ_FOREACH(w, &w_all, w_list)
2650                 witness_add_fullgraph(sb, w);
2651         mtx_unlock_spin(&w_mtx);
2652
2653         /*
2654          * Close the sbuf and return to userland.
2655          */
2656         error = sbuf_finish(sb);
2657         sbuf_delete(sb);
2658
2659         return (error);
2660 }
2661
2662 static int
2663 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2664 {
2665         int error, value;
2666
2667         value = witness_watch;
2668         error = sysctl_handle_int(oidp, &value, 0, req);
2669         if (error != 0 || req->newptr == NULL)
2670                 return (error);
2671         if (value > 1 || value < -1 ||
2672             (witness_watch == -1 && value != witness_watch))
2673                 return (EINVAL);
2674         witness_watch = value;
2675         return (0);
2676 }
2677
2678 static void
2679 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2680 {
2681         int i;
2682
2683         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2684                 return;
2685         w->w_displayed = 1;
2686
2687         WITNESS_INDEX_ASSERT(w->w_index);
2688         for (i = 1; i <= w_max_used_index; i++) {
2689                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2690                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2691                             w_data[i].w_name);
2692                         witness_add_fullgraph(sb, &w_data[i]);
2693                 }
2694         }
2695 }
2696
2697 /*
2698  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2699  * interprets the key as a string and reads until the null
2700  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2701  * hash value computed from the key.
2702  */
2703 static uint32_t
2704 witness_hash_djb2(const uint8_t *key, uint32_t size)
2705 {
2706         unsigned int hash = 5381;
2707         int i;
2708
2709         /* hash = hash * 33 + key[i] */
2710         if (size)
2711                 for (i = 0; i < size; i++)
2712                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2713         else
2714                 for (i = 0; key[i] != 0; i++)
2715                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2716
2717         return (hash);
2718 }
2719
2720
2721 /*
2722  * Initializes the two witness hash tables. Called exactly once from
2723  * witness_initialize().
2724  */
2725 static void
2726 witness_init_hash_tables(void)
2727 {
2728         int i;
2729
2730         MPASS(witness_cold);
2731
2732         /* Initialize the hash tables. */
2733         for (i = 0; i < WITNESS_HASH_SIZE; i++)
2734                 w_hash.wh_array[i] = NULL;
2735
2736         w_hash.wh_size = WITNESS_HASH_SIZE;
2737         w_hash.wh_count = 0;
2738
2739         /* Initialize the lock order data hash. */
2740         w_lofree = NULL;
2741         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2742                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2743                 w_lodata[i].wlod_next = w_lofree;
2744                 w_lofree = &w_lodata[i];
2745         }
2746         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2747         w_lohash.wloh_count = 0;
2748         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2749                 w_lohash.wloh_array[i] = NULL;
2750 }
2751
2752 static struct witness *
2753 witness_hash_get(const char *key)
2754 {
2755         struct witness *w;
2756         uint32_t hash;
2757
2758         MPASS(key != NULL);
2759         if (witness_cold == 0)
2760                 mtx_assert(&w_mtx, MA_OWNED);
2761         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2762         w = w_hash.wh_array[hash];
2763         while (w != NULL) {
2764                 if (strcmp(w->w_name, key) == 0)
2765                         goto out;
2766                 w = w->w_hash_next;
2767         }
2768
2769 out:
2770         return (w);
2771 }
2772
2773 static void
2774 witness_hash_put(struct witness *w)
2775 {
2776         uint32_t hash;
2777
2778         MPASS(w != NULL);
2779         MPASS(w->w_name != NULL);
2780         if (witness_cold == 0)
2781                 mtx_assert(&w_mtx, MA_OWNED);
2782         KASSERT(witness_hash_get(w->w_name) == NULL,
2783             ("%s: trying to add a hash entry that already exists!", __func__));
2784         KASSERT(w->w_hash_next == NULL,
2785             ("%s: w->w_hash_next != NULL", __func__));
2786
2787         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2788         w->w_hash_next = w_hash.wh_array[hash];
2789         w_hash.wh_array[hash] = w;
2790         w_hash.wh_count++;
2791 }
2792
2793
2794 static struct witness_lock_order_data *
2795 witness_lock_order_get(struct witness *parent, struct witness *child)
2796 {
2797         struct witness_lock_order_data *data = NULL;
2798         struct witness_lock_order_key key;
2799         unsigned int hash;
2800
2801         MPASS(parent != NULL && child != NULL);
2802         key.from = parent->w_index;
2803         key.to = child->w_index;
2804         WITNESS_INDEX_ASSERT(key.from);
2805         WITNESS_INDEX_ASSERT(key.to);
2806         if ((w_rmatrix[parent->w_index][child->w_index]
2807             & WITNESS_LOCK_ORDER_KNOWN) == 0)
2808                 goto out;
2809
2810         hash = witness_hash_djb2((const char*)&key,
2811             sizeof(key)) % w_lohash.wloh_size;
2812         data = w_lohash.wloh_array[hash];
2813         while (data != NULL) {
2814                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2815                         break;
2816                 data = data->wlod_next;
2817         }
2818
2819 out:
2820         return (data);
2821 }
2822
2823 /*
2824  * Verify that parent and child have a known relationship, are not the same,
2825  * and child is actually a child of parent.  This is done without w_mtx
2826  * to avoid contention in the common case.
2827  */
2828 static int
2829 witness_lock_order_check(struct witness *parent, struct witness *child)
2830 {
2831
2832         if (parent != child &&
2833             w_rmatrix[parent->w_index][child->w_index]
2834             & WITNESS_LOCK_ORDER_KNOWN &&
2835             isitmychild(parent, child))
2836                 return (1);
2837
2838         return (0);
2839 }
2840
2841 static int
2842 witness_lock_order_add(struct witness *parent, struct witness *child)
2843 {
2844         struct witness_lock_order_data *data = NULL;
2845         struct witness_lock_order_key key;
2846         unsigned int hash;
2847
2848         MPASS(parent != NULL && child != NULL);
2849         key.from = parent->w_index;
2850         key.to = child->w_index;
2851         WITNESS_INDEX_ASSERT(key.from);
2852         WITNESS_INDEX_ASSERT(key.to);
2853         if (w_rmatrix[parent->w_index][child->w_index]
2854             & WITNESS_LOCK_ORDER_KNOWN)
2855                 return (1);
2856
2857         hash = witness_hash_djb2((const char*)&key,
2858             sizeof(key)) % w_lohash.wloh_size;
2859         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2860         data = w_lofree;
2861         if (data == NULL)
2862                 return (0);
2863         w_lofree = data->wlod_next;
2864         data->wlod_next = w_lohash.wloh_array[hash];
2865         data->wlod_key = key;
2866         w_lohash.wloh_array[hash] = data;
2867         w_lohash.wloh_count++;
2868         stack_zero(&data->wlod_stack);
2869         stack_save(&data->wlod_stack);
2870         return (1);
2871 }
2872
2873 /* Call this whenver the structure of the witness graph changes. */
2874 static void
2875 witness_increment_graph_generation(void)
2876 {
2877
2878         if (witness_cold == 0)
2879                 mtx_assert(&w_mtx, MA_OWNED);
2880         w_generation++;
2881 }
2882
2883 #ifdef KDB
2884 static void
2885 _witness_debugger(int cond, const char *msg)
2886 {
2887
2888         if (witness_trace && cond)
2889                 kdb_backtrace();
2890         if (witness_kdb && cond)
2891                 kdb_enter(KDB_WHY_WITNESS, msg);
2892 }
2893 #endif