sys/kern/subr_witness.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 2008 Isilon Systems, Inc.
   5  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
   6  * Copyright (c) 1998 Berkeley Software Design, Inc.
   7  * All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Berkeley Software Design Inc's name may not be used to endorse or
  18  *    promote products derived from this software without specific prior
  19  *    written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
  22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  24  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
  25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  *      from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
  34  *      and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
  35  */
  36
  37 /*
  38  * Implementation of the `witness' lock verifier.  Originally implemented for
  39  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
  40  * classes in FreeBSD.
  41  */
  42
  43 /*
  44  *      Main Entry: witness
  45  *      Pronunciation: 'wit-n&s
  46  *      Function: noun
  47  *      Etymology: Middle English witnesse, from Old English witnes knowledge,
  48  *          testimony, witness, from 2wit
  49  *      Date: before 12th century
  50  *      1 : attestation of a fact or event : TESTIMONY
  51  *      2 : one that gives evidence; specifically : one who testifies in
  52  *          a cause or before a judicial tribunal
  53  *      3 : one asked to be present at a transaction so as to be able to
  54  *          testify to its having taken place
  55  *      4 : one who has personal knowledge of something
  56  *      5 a : something serving as evidence or proof : SIGN
  57  *        b : public affirmation by word or example of usually
  58  *            religious faith or conviction <the heroic witness to divine
  59  *            life -- Pilot>
  60  *      6 capitalized : a member of the Jehovah's Witnesses
  61  */
  62
  63 /*
  64  * Special rules concerning Giant and lock orders:
  65  *
  66  * 1) Giant must be acquired before any other mutexes.  Stated another way,
  67  *    no other mutex may be held when Giant is acquired.
  68  *
  69  * 2) Giant must be released when blocking on a sleepable lock.
  70  *
  71  * This rule is less obvious, but is a result of Giant providing the same
  72  * semantics as spl().  Basically, when a thread sleeps, it must release
  73  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
  74  * 2).
  75  *
  76  * 3) Giant may be acquired before or after sleepable locks.
  77  *
  78  * This rule is also not quite as obvious.  Giant may be acquired after
  79  * a sleepable lock because it is a non-sleepable lock and non-sleepable
  80  * locks may always be acquired while holding a sleepable lock.  The second
  81  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
  82  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
  83  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
  84  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
  85  * execute.  Thus, acquiring Giant both before and after a sleepable lock
  86  * will not result in a lock order reversal.
  87  */
  88
  89 #include <sys/cdefs.h>
  90 __FBSDID("$FreeBSD$");
  91
  92 #include "opt_ddb.h"
  93 #include "opt_hwpmc_hooks.h"
  94 #include "opt_stack.h"
  95 #include "opt_witness.h"
  96
  97 #include <sys/param.h>
  98 #include <sys/bus.h>
  99 #include <sys/kdb.h>
 100 #include <sys/kernel.h>
 101 #include <sys/ktr.h>
 102 #include <sys/lock.h>
 103 #include <sys/malloc.h>
 104 #include <sys/mutex.h>
 105 #include <sys/priv.h>
 106 #include <sys/proc.h>
 107 #include <sys/sbuf.h>
 108 #include <sys/sched.h>
 109 #include <sys/stack.h>
 110 #include <sys/sysctl.h>
 111 #include <sys/syslog.h>
 112 #include <sys/systm.h>
 113
 114 #ifdef DDB
 115 #include <ddb/ddb.h>
 116 #endif
 117
 118 #include <machine/stdarg.h>
 119
 120 #if !defined(DDB) && !defined(STACK)
 121 #error "DDB or STACK options are required for WITNESS"
 122 #endif
 123
 124 /* Note that these traces do not work with KTR_ALQ. */
 125 #if 0
 126 #define KTR_WITNESS     KTR_SUBSYS
 127 #else
 128 #define KTR_WITNESS     0
 129 #endif
 130
 131 #define LI_RECURSEMASK  0x0000ffff      /* Recursion depth of lock instance. */
 132 #define LI_EXCLUSIVE    0x00010000      /* Exclusive lock instance. */
 133 #define LI_NORELEASE    0x00020000      /* Lock not allowed to be released. */
 134
 135 /* Define this to check for blessed mutexes */
 136 #undef BLESSING
 137
 138 #ifndef WITNESS_COUNT
 139 #define WITNESS_COUNT           1536
 140 #endif
 141 #define WITNESS_HASH_SIZE       251     /* Prime, gives load factor < 2 */
 142 #define WITNESS_PENDLIST        (512 + (MAXCPU * 4))
 143
 144 /* Allocate 256 KB of stack data space */
 145 #define WITNESS_LO_DATA_COUNT   2048
 146
 147 /* Prime, gives load factor of ~2 at full load */
 148 #define WITNESS_LO_HASH_SIZE    1021
 149
 150 /*
 151  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
 152  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
 153  * probably be safe for the most part, but it's still a SWAG.
 154  */
 155 #define LOCK_NCHILDREN  5
 156 #define LOCK_CHILDCOUNT 2048
 157
 158 #define MAX_W_NAME      64
 159
 160 #define FULLGRAPH_SBUF_SIZE     512
 161
 162 /*
 163  * These flags go in the witness relationship matrix and describe the
 164  * relationship between any two struct witness objects.
 165  */
 166 #define WITNESS_UNRELATED        0x00    /* No lock order relation. */
 167 #define WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
 168 #define WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
 169 #define WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
 170 #define WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
 171 #define WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
 172 #define WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
 173 #define WITNESS_RELATED_MASK                                            \
 174         (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
 175 #define WITNESS_REVERSAL         0x10    /* A lock order reversal has been
 176                                           * observed. */
 177 #define WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
 178 #define WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
 179 #define WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
 180
 181 /* Descendant to ancestor flags */
 182 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
 183
 184 /* Ancestor to descendant flags */
 185 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
 186
 187 #define WITNESS_INDEX_ASSERT(i)                                         \
 188         MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
 189
 190 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
 191
 192 /*
 193  * Lock instances.  A lock instance is the data associated with a lock while
 194  * it is held by witness.  For example, a lock instance will hold the
 195  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
 196  * are held in a per-cpu list while sleep locks are held in per-thread list.
 197  */
 198 struct lock_instance {
 199         struct lock_object      *li_lock;
 200         const char              *li_file;
 201         int                     li_line;
 202         u_int                   li_flags;
 203 };
 204
 205 /*
 206  * A simple list type used to build the list of locks held by a thread
 207  * or CPU.  We can't simply embed the list in struct lock_object since a
 208  * lock may be held by more than one thread if it is a shared lock.  Locks
 209  * are added to the head of the list, so we fill up each list entry from
 210  * "the back" logically.  To ease some of the arithmetic, we actually fill
 211  * in each list entry the normal way (children[0] then children[1], etc.) but
 212  * when we traverse the list we read children[count-1] as the first entry
 213  * down to children[0] as the final entry.
 214  */
 215 struct lock_list_entry {
 216         struct lock_list_entry  *ll_next;
 217         struct lock_instance    ll_children[LOCK_NCHILDREN];
 218         u_int                   ll_count;
 219 };
 220
 221 /*
 222  * The main witness structure. One of these per named lock type in the system
 223  * (for example, "vnode interlock").
 224  */
 225 struct witness {
 226         char                    w_name[MAX_W_NAME];
 227         uint32_t                w_index;  /* Index in the relationship matrix */
 228         struct lock_class       *w_class;
 229         STAILQ_ENTRY(witness)   w_list;         /* List of all witnesses. */
 230         STAILQ_ENTRY(witness)   w_typelist;     /* Witnesses of a type. */
 231         struct witness          *w_hash_next; /* Linked list in hash buckets. */
 232         const char              *w_file; /* File where last acquired */
 233         uint32_t                w_line; /* Line where last acquired */
 234         uint32_t                w_refcount;
 235         uint16_t                w_num_ancestors; /* direct/indirect
 236                                                   * ancestor count */
 237         uint16_t                w_num_descendants; /* direct/indirect
 238                                                     * descendant count */
 239         int16_t                 w_ddb_level;
 240         unsigned                w_displayed:1;
 241         unsigned                w_reversed:1;
 242 };
 243
 244 STAILQ_HEAD(witness_list, witness);
 245
 246 /*
 247  * The witness hash table. Keys are witness names (const char *), elements are
 248  * witness objects (struct witness *).
 249  */
 250 struct witness_hash {
 251         struct witness  *wh_array[WITNESS_HASH_SIZE];
 252         uint32_t        wh_size;
 253         uint32_t        wh_count;
 254 };
 255
 256 /*
 257  * Key type for the lock order data hash table.
 258  */
 259 struct witness_lock_order_key {
 260         uint16_t        from;
 261         uint16_t        to;
 262 };
 263
 264 struct witness_lock_order_data {
 265         struct stack                    wlod_stack;
 266         struct witness_lock_order_key   wlod_key;
 267         struct witness_lock_order_data  *wlod_next;
 268 };
 269
 270 /*
 271  * The witness lock order data hash table. Keys are witness index tuples
 272  * (struct witness_lock_order_key), elements are lock order data objects
 273  * (struct witness_lock_order_data).
 274  */
 275 struct witness_lock_order_hash {
 276         struct witness_lock_order_data  *wloh_array[WITNESS_LO_HASH_SIZE];
 277         u_int   wloh_size;
 278         u_int   wloh_count;
 279 };
 280
 281 #ifdef BLESSING
 282 struct witness_blessed {
 283         const char      *b_lock1;
 284         const char      *b_lock2;
 285 };
 286 #endif
 287
 288 struct witness_pendhelp {
 289         const char              *wh_type;
 290         struct lock_object      *wh_lock;
 291 };
 292
 293 struct witness_order_list_entry {
 294         const char              *w_name;
 295         struct lock_class       *w_class;
 296 };
 297
 298 /*
 299  * Returns 0 if one of the locks is a spin lock and the other is not.
 300  * Returns 1 otherwise.
 301  */
 302 static __inline int
 303 witness_lock_type_equal(struct witness *w1, struct witness *w2)
 304 {
 305
 306         return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
 307                 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
 308 }
 309
 310 static __inline int
 311 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
 312     const struct witness_lock_order_key *b)
 313 {
 314
 315         return (a->from == b->from && a->to == b->to);
 316 }
 317
 318 static int      _isitmyx(struct witness *w1, struct witness *w2, int rmask,
 319                     const char *fname);
 320 static void     adopt(struct witness *parent, struct witness *child);
 321 #ifdef BLESSING
 322 static int      blessed(struct witness *, struct witness *);
 323 #endif
 324 static void     depart(struct witness *w);
 325 static struct witness   *enroll(const char *description,
 326                             struct lock_class *lock_class);
 327 static struct lock_instance     *find_instance(struct lock_list_entry *list,
 328                                     const struct lock_object *lock);
 329 static int      isitmychild(struct witness *parent, struct witness *child);
 330 static int      isitmydescendant(struct witness *parent, struct witness *child);
 331 static void     itismychild(struct witness *parent, struct witness *child);
 332 static int      sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
 333 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
 334 static int      sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
 335 static int      sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
 336 static void     witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
 337 #ifdef DDB
 338 static void     witness_ddb_compute_levels(void);
 339 static void     witness_ddb_display(int(*)(const char *fmt, ...));
 340 static void     witness_ddb_display_descendants(int(*)(const char *fmt, ...),
 341                     struct witness *, int indent);
 342 static void     witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
 343                     struct witness_list *list);
 344 static void     witness_ddb_level_descendants(struct witness *parent, int l);
 345 static void     witness_ddb_list(struct thread *td);
 346 #endif
 347 static void     witness_debugger(int cond, const char *msg);
 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 int      witness_output(const char *fmt, ...) __printflike(1, 2);
 367 static int      witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
 368 static void     witness_setflag(struct lock_object *lock, int flag, int set);
 369
 370 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
 371     "Witness Locking");
 372
 373 /*
 374  * If set to 0, lock order checking is disabled.  If set to -1,
 375  * witness is completely disabled.  Otherwise witness performs full
 376  * lock order checking for all locks.  At runtime, lock order checking
 377  * may be toggled.  However, witness cannot be reenabled once it is
 378  * completely disabled.
 379  */
 380 static int witness_watch = 1;
 381 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
 382     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
 383
 384 #ifdef KDB
 385 /*
 386  * When KDB is enabled and witness_kdb is 1, it will cause the system
 387  * to drop into kdebug() when:
 388  *      - a lock hierarchy violation occurs
 389  *      - locks are held when going to sleep.
 390  */
 391 #ifdef WITNESS_KDB
 392 int     witness_kdb = 1;
 393 #else
 394 int     witness_kdb = 0;
 395 #endif
 396 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
 397 #endif /* KDB */
 398
 399 #if defined(DDB) || defined(KDB)
 400 /*
 401  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
 402  * to print a stack trace:
 403  *      - a lock hierarchy violation occurs
 404  *      - locks are held when going to sleep.
 405  */
 406 int     witness_trace = 1;
 407 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
 408 #endif /* DDB || KDB */
 409
 410 #ifdef WITNESS_SKIPSPIN
 411 int     witness_skipspin = 1;
 412 #else
 413 int     witness_skipspin = 0;
 414 #endif
 415 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
 416
 417 int badstack_sbuf_size;
 418
 419 int witness_count = WITNESS_COUNT;
 420 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
 421     &witness_count, 0, "");
 422
 423 /*
 424  * Output channel for witness messages.  By default we print to the console.
 425  */
 426 enum witness_channel {
 427         WITNESS_CONSOLE,
 428         WITNESS_LOG,
 429         WITNESS_NONE,
 430 };
 431
 432 static enum witness_channel witness_channel = WITNESS_CONSOLE;
 433 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
 434     CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
 435     "Output channel for warnings");
 436
 437 /*
 438  * Call this to print out the relations between locks.
 439  */
 440 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
 441     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
 442
 443 /*
 444  * Call this to print out the witness faulty stacks.
 445  */
 446 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
 447     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
 448
 449 static struct mtx w_mtx;
 450
 451 /* w_list */
 452 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 453 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 454
 455 /* w_typelist */
 456 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 457 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 458
 459 /* lock list */
 460 static struct lock_list_entry *w_lock_list_free = NULL;
 461 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
 462 static u_int pending_cnt;
 463
 464 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
 465 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
 466 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
 467 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
 468     "");
 469
 470 static struct witness *w_data;
 471 static uint8_t **w_rmatrix;
 472 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 473 static struct witness_hash w_hash;      /* The witness hash table. */
 474
 475 /* The lock order data hash */
 476 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
 477 static struct witness_lock_order_data *w_lofree = NULL;
 478 static struct witness_lock_order_hash w_lohash;
 479 static int w_max_used_index = 0;
 480 static unsigned int w_generation = 0;
 481 static const char w_notrunning[] = "Witness not running\n";
 482 static const char w_stillcold[] = "Witness is still cold\n";
 483 #ifdef __i386__
 484 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
 485 #endif
 486
 487 static struct witness_order_list_entry order_lists[] = {
 488         /*
 489          * sx locks
 490          */
 491         { "proctree", &lock_class_sx },
 492         { "allproc", &lock_class_sx },
 493         { "allprison", &lock_class_sx },
 494         { NULL, NULL },
 495         /*
 496          * Various mutexes
 497          */
 498         { "Giant", &lock_class_mtx_sleep },
 499         { "pipe mutex", &lock_class_mtx_sleep },
 500         { "sigio lock", &lock_class_mtx_sleep },
 501         { "process group", &lock_class_mtx_sleep },
 502         { "process lock", &lock_class_mtx_sleep },
 503         { "session", &lock_class_mtx_sleep },
 504         { "uidinfo hash", &lock_class_rw },
 505 #ifdef  HWPMC_HOOKS
 506         { "pmc-sleep", &lock_class_mtx_sleep },
 507 #endif
 508         { "time lock", &lock_class_mtx_sleep },
 509         { NULL, NULL },
 510         /*
 511          * umtx
 512          */
 513         { "umtx lock", &lock_class_mtx_sleep },
 514         { NULL, NULL },
 515         /*
 516          * Sockets
 517          */
 518         { "accept", &lock_class_mtx_sleep },
 519         { "so_snd", &lock_class_mtx_sleep },
 520         { "so_rcv", &lock_class_mtx_sleep },
 521         { "sellck", &lock_class_mtx_sleep },
 522         { NULL, NULL },
 523         /*
 524          * Routing
 525          */
 526         { "so_rcv", &lock_class_mtx_sleep },
 527         { "radix node head", &lock_class_rw },
 528         { "rtentry", &lock_class_mtx_sleep },
 529         { "ifaddr", &lock_class_mtx_sleep },
 530         { NULL, NULL },
 531         /*
 532          * IPv4 multicast:
 533          * protocol locks before interface locks, after UDP locks.
 534          */
 535         { "in_multi_sx", &lock_class_sx },
 536         { "udpinp", &lock_class_rw },
 537         { "in_multi_list_mtx", &lock_class_mtx_sleep },
 538         { "igmp_mtx", &lock_class_mtx_sleep },
 539         { "ifnet_rw", &lock_class_rw },
 540         { "if_addr_lock", &lock_class_rw },
 541         { NULL, NULL },
 542         /*
 543          * IPv6 multicast:
 544          * protocol locks before interface locks, after UDP locks.
 545          */
 546         { "in6_multi_sx", &lock_class_sx },
 547         { "udpinp", &lock_class_rw },
 548         { "in6_multi_list_mtx", &lock_class_mtx_sleep },
 549         { "mld_mtx", &lock_class_mtx_sleep },
 550         { "ifnet_rw", &lock_class_rw },
 551         { "if_addr_lock", &lock_class_rw },
 552         { NULL, NULL },
 553         /*
 554          * UNIX Domain Sockets
 555          */
 556         { "unp_link_rwlock", &lock_class_rw },
 557         { "unp_list_lock", &lock_class_mtx_sleep },
 558         { "unp", &lock_class_mtx_sleep },
 559         { "so_snd", &lock_class_mtx_sleep },
 560         { NULL, NULL },
 561         /*
 562          * UDP/IP
 563          */
 564         { "udp", &lock_class_rw },
 565         { "udpinp", &lock_class_rw },
 566         { "so_snd", &lock_class_mtx_sleep },
 567         { NULL, NULL },
 568         /*
 569          * TCP/IP
 570          */
 571         { "tcp", &lock_class_rw },
 572         { "tcpinp", &lock_class_rw },
 573         { "so_snd", &lock_class_mtx_sleep },
 574         { NULL, NULL },
 575         /*
 576          * BPF
 577          */
 578         { "bpf global lock", &lock_class_sx },
 579         { "bpf interface lock", &lock_class_rw },
 580         { "bpf cdev lock", &lock_class_mtx_sleep },
 581         { NULL, NULL },
 582         /*
 583          * NFS server
 584          */
 585         { "nfsd_mtx", &lock_class_mtx_sleep },
 586         { "so_snd", &lock_class_mtx_sleep },
 587         { NULL, NULL },
 588
 589         /*
 590          * IEEE 802.11
 591          */
 592         { "802.11 com lock", &lock_class_mtx_sleep},
 593         { NULL, NULL },
 594         /*
 595          * Network drivers
 596          */
 597         { "network driver", &lock_class_mtx_sleep},
 598         { NULL, NULL },
 599
 600         /*
 601          * Netgraph
 602          */
 603         { "ng_node", &lock_class_mtx_sleep },
 604         { "ng_worklist", &lock_class_mtx_sleep },
 605         { NULL, NULL },
 606         /*
 607          * CDEV
 608          */
 609         { "vm map (system)", &lock_class_mtx_sleep },
 610         { "vnode interlock", &lock_class_mtx_sleep },
 611         { "cdev", &lock_class_mtx_sleep },
 612         { NULL, NULL },
 613         /*
 614          * VM
 615          */
 616         { "vm map (user)", &lock_class_sx },
 617         { "vm object", &lock_class_rw },
 618         { "vm page", &lock_class_mtx_sleep },
 619         { "pmap pv global", &lock_class_rw },
 620         { "pmap", &lock_class_mtx_sleep },
 621         { "pmap pv list", &lock_class_rw },
 622         { "vm page free queue", &lock_class_mtx_sleep },
 623         { "vm pagequeue", &lock_class_mtx_sleep },
 624         { NULL, NULL },
 625         /*
 626          * kqueue/VFS interaction
 627          */
 628         { "kqueue", &lock_class_mtx_sleep },
 629         { "struct mount mtx", &lock_class_mtx_sleep },
 630         { "vnode interlock", &lock_class_mtx_sleep },
 631         { NULL, NULL },
 632         /*
 633          * VFS namecache
 634          */
 635         { "ncvn", &lock_class_mtx_sleep },
 636         { "ncbuc", &lock_class_rw },
 637         { "vnode interlock", &lock_class_mtx_sleep },
 638         { "ncneg", &lock_class_mtx_sleep },
 639         { NULL, NULL },
 640         /*
 641          * ZFS locking
 642          */
 643         { "dn->dn_mtx", &lock_class_sx },
 644         { "dr->dt.di.dr_mtx", &lock_class_sx },
 645         { "db->db_mtx", &lock_class_sx },
 646         { NULL, NULL },
 647         /*
 648          * TCP log locks
 649          */
 650         { "TCP ID tree", &lock_class_rw },
 651         { "tcp log id bucket", &lock_class_mtx_sleep },
 652         { "tcpinp", &lock_class_rw },
 653         { "TCP log expireq", &lock_class_mtx_sleep },
 654         { NULL, NULL },
 655         /*
 656          * spin locks
 657          */
 658 #ifdef SMP
 659         { "ap boot", &lock_class_mtx_spin },
 660 #endif
 661         { "rm.mutex_mtx", &lock_class_mtx_spin },
 662         { "sio", &lock_class_mtx_spin },
 663 #ifdef __i386__
 664         { "cy", &lock_class_mtx_spin },
 665 #endif
 666 #ifdef __sparc64__
 667         { "pcib_mtx", &lock_class_mtx_spin },
 668         { "rtc_mtx", &lock_class_mtx_spin },
 669 #endif
 670         { "scc_hwmtx", &lock_class_mtx_spin },
 671         { "uart_hwmtx", &lock_class_mtx_spin },
 672         { "fast_taskqueue", &lock_class_mtx_spin },
 673         { "intr table", &lock_class_mtx_spin },
 674 #ifdef  HWPMC_HOOKS
 675         { "pmc-per-proc", &lock_class_mtx_spin },
 676 #endif
 677         { "process slock", &lock_class_mtx_spin },
 678         { "syscons video lock", &lock_class_mtx_spin },
 679         { "sleepq chain", &lock_class_mtx_spin },
 680         { "rm_spinlock", &lock_class_mtx_spin },
 681         { "turnstile chain", &lock_class_mtx_spin },
 682         { "turnstile lock", &lock_class_mtx_spin },
 683         { "sched lock", &lock_class_mtx_spin },
 684         { "td_contested", &lock_class_mtx_spin },
 685         { "callout", &lock_class_mtx_spin },
 686         { "entropy harvest mutex", &lock_class_mtx_spin },
 687 #ifdef SMP
 688         { "smp rendezvous", &lock_class_mtx_spin },
 689 #endif
 690 #ifdef __powerpc__
 691         { "tlb0", &lock_class_mtx_spin },
 692 #endif
 693         /*
 694          * leaf locks
 695          */
 696         { "intrcnt", &lock_class_mtx_spin },
 697         { "icu", &lock_class_mtx_spin },
 698 #if defined(SMP) && defined(__sparc64__)
 699         { "ipi", &lock_class_mtx_spin },
 700 #endif
 701 #ifdef __i386__
 702         { "allpmaps", &lock_class_mtx_spin },
 703         { "descriptor tables", &lock_class_mtx_spin },
 704 #endif
 705         { "clk", &lock_class_mtx_spin },
 706         { "cpuset", &lock_class_mtx_spin },
 707         { "mprof lock", &lock_class_mtx_spin },
 708         { "zombie lock", &lock_class_mtx_spin },
 709         { "ALD Queue", &lock_class_mtx_spin },
 710 #if defined(__i386__) || defined(__amd64__)
 711         { "pcicfg", &lock_class_mtx_spin },
 712         { "NDIS thread lock", &lock_class_mtx_spin },
 713 #endif
 714         { "tw_osl_io_lock", &lock_class_mtx_spin },
 715         { "tw_osl_q_lock", &lock_class_mtx_spin },
 716         { "tw_cl_io_lock", &lock_class_mtx_spin },
 717         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 718         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 719 #ifdef  HWPMC_HOOKS
 720         { "pmc-leaf", &lock_class_mtx_spin },
 721 #endif
 722         { "blocked lock", &lock_class_mtx_spin },
 723         { NULL, NULL },
 724         { NULL, NULL }
 725 };
 726
 727 #ifdef BLESSING
 728 /*
 729  * Pairs of locks which have been blessed
 730  * Don't complain about order problems with blessed locks
 731  */
 732 static struct witness_blessed blessed_list[] = {
 733 };
 734 #endif
 735
 736 /*
 737  * This global is set to 0 once it becomes safe to use the witness code.
 738  */
 739 static int witness_cold = 1;
 740
 741 /*
 742  * This global is set to 1 once the static lock orders have been enrolled
 743  * so that a warning can be issued for any spin locks enrolled later.
 744  */
 745 static int witness_spin_warn = 0;
 746
 747 /* Trim useless garbage from filenames. */
 748 static const char *
 749 fixup_filename(const char *file)
 750 {
 751
 752         if (file == NULL)
 753                 return (NULL);
 754         while (strncmp(file, "../", 3) == 0)
 755                 file += 3;
 756         return (file);
 757 }
 758
 759 /*
 760  * Calculate the size of early witness structures.
 761  */
 762 int
 763 witness_startup_count(void)
 764 {
 765         int sz;
 766
 767         sz = sizeof(struct witness) * witness_count;
 768         sz += sizeof(*w_rmatrix) * (witness_count + 1);
 769         sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
 770             (witness_count + 1);
 771
 772         return (sz);
 773 }
 774
 775 /*
 776  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 777  * assume that the early boot is single-threaded at least until after this
 778  * routine is completed.
 779  */
 780 void
 781 witness_startup(void *mem)
 782 {
 783         struct lock_object *lock;
 784         struct witness_order_list_entry *order;
 785         struct witness *w, *w1;
 786         uintptr_t p;
 787         int i;
 788
 789         p = (uintptr_t)mem;
 790         w_data = (void *)p;
 791         p += sizeof(struct witness) * witness_count;
 792
 793         w_rmatrix = (void *)p;
 794         p += sizeof(*w_rmatrix) * (witness_count + 1);
 795
 796         for (i = 0; i < witness_count + 1; i++) {
 797                 w_rmatrix[i] = (void *)p;
 798                 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
 799         }
 800         badstack_sbuf_size = witness_count * 256;
 801
 802         /*
 803          * We have to release Giant before initializing its witness
 804          * structure so that WITNESS doesn't get confused.
 805          */
 806         mtx_unlock(&Giant);
 807         mtx_assert(&Giant, MA_NOTOWNED);
 808
 809         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 810         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 811             MTX_NOWITNESS | MTX_NOPROFILE);
 812         for (i = witness_count - 1; i >= 0; i--) {
 813                 w = &w_data[i];
 814                 memset(w, 0, sizeof(*w));
 815                 w_data[i].w_index = i;  /* Witness index never changes. */
 816                 witness_free(w);
 817         }
 818         KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
 819             ("%s: Invalid list of free witness objects", __func__));
 820
 821         /* Witness with index 0 is not used to aid in debugging. */
 822         STAILQ_REMOVE_HEAD(&w_free, w_list);
 823         w_free_cnt--;
 824
 825         for (i = 0; i < witness_count; i++) {
 826                 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
 827                     (witness_count + 1));
 828         }
 829
 830         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 831                 witness_lock_list_free(&w_locklistdata[i]);
 832         witness_init_hash_tables();
 833
 834         /* First add in all the specified order lists. */
 835         for (order = order_lists; order->w_name != NULL; order++) {
 836                 w = enroll(order->w_name, order->w_class);
 837                 if (w == NULL)
 838                         continue;
 839                 w->w_file = "order list";
 840                 for (order++; order->w_name != NULL; order++) {
 841                         w1 = enroll(order->w_name, order->w_class);
 842                         if (w1 == NULL)
 843                                 continue;
 844                         w1->w_file = "order list";
 845                         itismychild(w, w1);
 846                         w = w1;
 847                 }
 848         }
 849         witness_spin_warn = 1;
 850
 851         /* Iterate through all locks and add them to witness. */
 852         for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
 853                 lock = pending_locks[i].wh_lock;
 854                 KASSERT(lock->lo_flags & LO_WITNESS,
 855                     ("%s: lock %s is on pending list but not LO_WITNESS",
 856                     __func__, lock->lo_name));
 857                 lock->lo_witness = enroll(pending_locks[i].wh_type,
 858                     LOCK_CLASS(lock));
 859         }
 860
 861         /* Mark the witness code as being ready for use. */
 862         witness_cold = 0;
 863
 864         mtx_lock(&Giant);
 865 }
 866
 867 void
 868 witness_init(struct lock_object *lock, const char *type)
 869 {
 870         struct lock_class *class;
 871
 872         /* Various sanity checks. */
 873         class = LOCK_CLASS(lock);
 874         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 875             (class->lc_flags & LC_RECURSABLE) == 0)
 876                 kassert_panic("%s: lock (%s) %s can not be recursable",
 877                     __func__, class->lc_name, lock->lo_name);
 878         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 879             (class->lc_flags & LC_SLEEPABLE) == 0)
 880                 kassert_panic("%s: lock (%s) %s can not be sleepable",
 881                     __func__, class->lc_name, lock->lo_name);
 882         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 883             (class->lc_flags & LC_UPGRADABLE) == 0)
 884                 kassert_panic("%s: lock (%s) %s can not be upgradable",
 885                     __func__, class->lc_name, lock->lo_name);
 886
 887         /*
 888          * If we shouldn't watch this lock, then just clear lo_witness.
 889          * Otherwise, if witness_cold is set, then it is too early to
 890          * enroll this lock, so defer it to witness_initialize() by adding
 891          * it to the pending_locks list.  If it is not too early, then enroll
 892          * the lock now.
 893          */
 894         if (witness_watch < 1 || panicstr != NULL ||
 895             (lock->lo_flags & LO_WITNESS) == 0)
 896                 lock->lo_witness = NULL;
 897         else if (witness_cold) {
 898                 pending_locks[pending_cnt].wh_lock = lock;
 899                 pending_locks[pending_cnt++].wh_type = type;
 900                 if (pending_cnt > WITNESS_PENDLIST)
 901                         panic("%s: pending locks list is too small, "
 902                             "increase WITNESS_PENDLIST\n",
 903                             __func__);
 904         } else
 905                 lock->lo_witness = enroll(type, class);
 906 }
 907
 908 void
 909 witness_destroy(struct lock_object *lock)
 910 {
 911         struct lock_class *class;
 912         struct witness *w;
 913
 914         class = LOCK_CLASS(lock);
 915
 916         if (witness_cold)
 917                 panic("lock (%s) %s destroyed while witness_cold",
 918                     class->lc_name, lock->lo_name);
 919
 920         /* XXX: need to verify that no one holds the lock */
 921         if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
 922                 return;
 923         w = lock->lo_witness;
 924
 925         mtx_lock_spin(&w_mtx);
 926         MPASS(w->w_refcount > 0);
 927         w->w_refcount--;
 928
 929         if (w->w_refcount == 0)
 930                 depart(w);
 931         mtx_unlock_spin(&w_mtx);
 932 }
 933
 934 #ifdef DDB
 935 static void
 936 witness_ddb_compute_levels(void)
 937 {
 938         struct witness *w;
 939
 940         /*
 941          * First clear all levels.
 942          */
 943         STAILQ_FOREACH(w, &w_all, w_list)
 944                 w->w_ddb_level = -1;
 945
 946         /*
 947          * Look for locks with no parents and level all their descendants.
 948          */
 949         STAILQ_FOREACH(w, &w_all, w_list) {
 950
 951                 /* If the witness has ancestors (is not a root), skip it. */
 952                 if (w->w_num_ancestors > 0)
 953                         continue;
 954                 witness_ddb_level_descendants(w, 0);
 955         }
 956 }
 957
 958 static void
 959 witness_ddb_level_descendants(struct witness *w, int l)
 960 {
 961         int i;
 962
 963         if (w->w_ddb_level >= l)
 964                 return;
 965
 966         w->w_ddb_level = l;
 967         l++;
 968
 969         for (i = 1; i <= w_max_used_index; i++) {
 970                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 971                         witness_ddb_level_descendants(&w_data[i], l);
 972         }
 973 }
 974
 975 static void
 976 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
 977     struct witness *w, int indent)
 978 {
 979         int i;
 980
 981         for (i = 0; i < indent; i++)
 982                 prnt(" ");
 983         prnt("%s (type: %s, depth: %d, active refs: %d)",
 984              w->w_name, w->w_class->lc_name,
 985              w->w_ddb_level, w->w_refcount);
 986         if (w->w_displayed) {
 987                 prnt(" -- (already displayed)\n");
 988                 return;
 989         }
 990         w->w_displayed = 1;
 991         if (w->w_file != NULL && w->w_line != 0)
 992                 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
 993                     w->w_line);
 994         else
 995                 prnt(" -- never acquired\n");
 996         indent++;
 997         WITNESS_INDEX_ASSERT(w->w_index);
 998         for (i = 1; i <= w_max_used_index; i++) {
 999                 if (db_pager_quit)
1000                         return;
1001                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1002                         witness_ddb_display_descendants(prnt, &w_data[i],
1003                             indent);
1004         }
1005 }
1006
1007 static void
1008 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1009     struct witness_list *list)
1010 {
1011         struct witness *w;
1012
1013         STAILQ_FOREACH(w, list, w_typelist) {
1014                 if (w->w_file == NULL || w->w_ddb_level > 0)
1015                         continue;
1016
1017                 /* This lock has no anscestors - display its descendants. */
1018                 witness_ddb_display_descendants(prnt, w, 0);
1019                 if (db_pager_quit)
1020                         return;
1021         }
1022 }
1023
1024 static void
1025 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1026 {
1027         struct witness *w;
1028
1029         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1030         witness_ddb_compute_levels();
1031
1032         /* Clear all the displayed flags. */
1033         STAILQ_FOREACH(w, &w_all, w_list)
1034                 w->w_displayed = 0;
1035
1036         /*
1037          * First, handle sleep locks which have been acquired at least
1038          * once.
1039          */
1040         prnt("Sleep locks:\n");
1041         witness_ddb_display_list(prnt, &w_sleep);
1042         if (db_pager_quit)
1043                 return;
1044
1045         /*
1046          * Now do spin locks which have been acquired at least once.
1047          */
1048         prnt("\nSpin locks:\n");
1049         witness_ddb_display_list(prnt, &w_spin);
1050         if (db_pager_quit)
1051                 return;
1052
1053         /*
1054          * Finally, any locks which have not been acquired yet.
1055          */
1056         prnt("\nLocks which were never acquired:\n");
1057         STAILQ_FOREACH(w, &w_all, w_list) {
1058                 if (w->w_file != NULL || w->w_refcount == 0)
1059                         continue;
1060                 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1061                     w->w_class->lc_name, w->w_ddb_level);
1062                 if (db_pager_quit)
1063                         return;
1064         }
1065 }
1066 #endif /* DDB */
1067
1068 int
1069 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1070 {
1071
1072         if (witness_watch == -1 || panicstr != NULL)
1073                 return (0);
1074
1075         /* Require locks that witness knows about. */
1076         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1077             lock2->lo_witness == NULL)
1078                 return (EINVAL);
1079
1080         mtx_assert(&w_mtx, MA_NOTOWNED);
1081         mtx_lock_spin(&w_mtx);
1082
1083         /*
1084          * If we already have either an explicit or implied lock order that
1085          * is the other way around, then return an error.
1086          */
1087         if (witness_watch &&
1088             isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1089                 mtx_unlock_spin(&w_mtx);
1090                 return (EDOOFUS);
1091         }
1092
1093         /* Try to add the new order. */
1094         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1095             lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1096         itismychild(lock1->lo_witness, lock2->lo_witness);
1097         mtx_unlock_spin(&w_mtx);
1098         return (0);
1099 }
1100
1101 void
1102 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1103     int line, struct lock_object *interlock)
1104 {
1105         struct lock_list_entry *lock_list, *lle;
1106         struct lock_instance *lock1, *lock2, *plock;
1107         struct lock_class *class, *iclass;
1108         struct witness *w, *w1;
1109         struct thread *td;
1110         int i, j;
1111
1112         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1113             panicstr != NULL)
1114                 return;
1115
1116         w = lock->lo_witness;
1117         class = LOCK_CLASS(lock);
1118         td = curthread;
1119
1120         if (class->lc_flags & LC_SLEEPLOCK) {
1121
1122                 /*
1123                  * Since spin locks include a critical section, this check
1124                  * implicitly enforces a lock order of all sleep locks before
1125                  * all spin locks.
1126                  */
1127                 if (td->td_critnest != 0 && !kdb_active)
1128                         kassert_panic("acquiring blockable sleep lock with "
1129                             "spinlock or critical section held (%s) %s @ %s:%d",
1130                             class->lc_name, lock->lo_name,
1131                             fixup_filename(file), line);
1132
1133                 /*
1134                  * If this is the first lock acquired then just return as
1135                  * no order checking is needed.
1136                  */
1137                 lock_list = td->td_sleeplocks;
1138                 if (lock_list == NULL || lock_list->ll_count == 0)
1139                         return;
1140         } else {
1141
1142                 /*
1143                  * If this is the first lock, just return as no order
1144                  * checking is needed.  Avoid problems with thread
1145                  * migration pinning the thread while checking if
1146                  * spinlocks are held.  If at least one spinlock is held
1147                  * the thread is in a safe path and it is allowed to
1148                  * unpin it.
1149                  */
1150                 sched_pin();
1151                 lock_list = PCPU_GET(spinlocks);
1152                 if (lock_list == NULL || lock_list->ll_count == 0) {
1153                         sched_unpin();
1154                         return;
1155                 }
1156                 sched_unpin();
1157         }
1158
1159         /*
1160          * Check to see if we are recursing on a lock we already own.  If
1161          * so, make sure that we don't mismatch exclusive and shared lock
1162          * acquires.
1163          */
1164         lock1 = find_instance(lock_list, lock);
1165         if (lock1 != NULL) {
1166                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1167                     (flags & LOP_EXCLUSIVE) == 0) {
1168                         witness_output("shared lock of (%s) %s @ %s:%d\n",
1169                             class->lc_name, lock->lo_name,
1170                             fixup_filename(file), line);
1171                         witness_output("while exclusively locked from %s:%d\n",
1172                             fixup_filename(lock1->li_file), lock1->li_line);
1173                         kassert_panic("excl->share");
1174                 }
1175                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1176                     (flags & LOP_EXCLUSIVE) != 0) {
1177                         witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1178                             class->lc_name, lock->lo_name,
1179                             fixup_filename(file), line);
1180                         witness_output("while share locked from %s:%d\n",
1181                             fixup_filename(lock1->li_file), lock1->li_line);
1182                         kassert_panic("share->excl");
1183                 }
1184                 return;
1185         }
1186
1187         /* Warn if the interlock is not locked exactly once. */
1188         if (interlock != NULL) {
1189                 iclass = LOCK_CLASS(interlock);
1190                 lock1 = find_instance(lock_list, interlock);
1191                 if (lock1 == NULL)
1192                         kassert_panic("interlock (%s) %s not locked @ %s:%d",
1193                             iclass->lc_name, interlock->lo_name,
1194                             fixup_filename(file), line);
1195                 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1196                         kassert_panic("interlock (%s) %s recursed @ %s:%d",
1197                             iclass->lc_name, interlock->lo_name,
1198                             fixup_filename(file), line);
1199         }
1200
1201         /*
1202          * Find the previously acquired lock, but ignore interlocks.
1203          */
1204         plock = &lock_list->ll_children[lock_list->ll_count - 1];
1205         if (interlock != NULL && plock->li_lock == interlock) {
1206                 if (lock_list->ll_count > 1)
1207                         plock =
1208                             &lock_list->ll_children[lock_list->ll_count - 2];
1209                 else {
1210                         lle = lock_list->ll_next;
1211
1212                         /*
1213                          * The interlock is the only lock we hold, so
1214                          * simply return.
1215                          */
1216                         if (lle == NULL)
1217                                 return;
1218                         plock = &lle->ll_children[lle->ll_count - 1];
1219                 }
1220         }
1221
1222         /*
1223          * Try to perform most checks without a lock.  If this succeeds we
1224          * can skip acquiring the lock and return success.  Otherwise we redo
1225          * the check with the lock held to handle races with concurrent updates.
1226          */
1227         w1 = plock->li_lock->lo_witness;
1228         if (witness_lock_order_check(w1, w))
1229                 return;
1230
1231         mtx_lock_spin(&w_mtx);
1232         if (witness_lock_order_check(w1, w)) {
1233                 mtx_unlock_spin(&w_mtx);
1234                 return;
1235         }
1236         witness_lock_order_add(w1, w);
1237
1238         /*
1239          * Check for duplicate locks of the same type.  Note that we only
1240          * have to check for this on the last lock we just acquired.  Any
1241          * other cases will be caught as lock order violations.
1242          */
1243         if (w1 == w) {
1244                 i = w->w_index;
1245                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1246                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1247                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
1248                         w->w_reversed = 1;
1249                         mtx_unlock_spin(&w_mtx);
1250                         witness_output(
1251                             "acquiring duplicate lock of same type: \"%s\"\n",
1252                             w->w_name);
1253                         witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1254                             fixup_filename(plock->li_file), plock->li_line);
1255                         witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1256                             fixup_filename(file), line);
1257                         witness_debugger(1, __func__);
1258                 } else
1259                         mtx_unlock_spin(&w_mtx);
1260                 return;
1261         }
1262         mtx_assert(&w_mtx, MA_OWNED);
1263
1264         /*
1265          * If we know that the lock we are acquiring comes after
1266          * the lock we most recently acquired in the lock order tree,
1267          * then there is no need for any further checks.
1268          */
1269         if (isitmychild(w1, w))
1270                 goto out;
1271
1272         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1273                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1274
1275                         MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1276                         lock1 = &lle->ll_children[i];
1277
1278                         /*
1279                          * Ignore the interlock.
1280                          */
1281                         if (interlock == lock1->li_lock)
1282                                 continue;
1283
1284                         /*
1285                          * If this lock doesn't undergo witness checking,
1286                          * then skip it.
1287                          */
1288                         w1 = lock1->li_lock->lo_witness;
1289                         if (w1 == NULL) {
1290                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1291                                     ("lock missing witness structure"));
1292                                 continue;
1293                         }
1294
1295                         /*
1296                          * If we are locking Giant and this is a sleepable
1297                          * lock, then skip it.
1298                          */
1299                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1300                             lock == &Giant.lock_object)
1301                                 continue;
1302
1303                         /*
1304                          * If we are locking a sleepable lock and this lock
1305                          * is Giant, then skip it.
1306                          */
1307                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1308                             lock1->li_lock == &Giant.lock_object)
1309                                 continue;
1310
1311                         /*
1312                          * If we are locking a sleepable lock and this lock
1313                          * isn't sleepable, we want to treat it as a lock
1314                          * order violation to enfore a general lock order of
1315                          * sleepable locks before non-sleepable locks.
1316                          */
1317                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1318                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1319                                 goto reversal;
1320
1321                         /*
1322                          * If we are locking Giant and this is a non-sleepable
1323                          * lock, then treat it as a reversal.
1324                          */
1325                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1326                             lock == &Giant.lock_object)
1327                                 goto reversal;
1328
1329                         /*
1330                          * Check the lock order hierarchy for a reveresal.
1331                          */
1332                         if (!isitmydescendant(w, w1))
1333                                 continue;
1334                 reversal:
1335
1336                         /*
1337                          * We have a lock order violation, check to see if it
1338                          * is allowed or has already been yelled about.
1339                          */
1340 #ifdef BLESSING
1341
1342                         /*
1343                          * If the lock order is blessed, just bail.  We don't
1344                          * look for other lock order violations though, which
1345                          * may be a bug.
1346                          */
1347                         if (blessed(w, w1))
1348                                 goto out;
1349 #endif
1350
1351                         /* Bail if this violation is known */
1352                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1353                                 goto out;
1354
1355                         /* Record this as a violation */
1356                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1357                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1358                         w->w_reversed = w1->w_reversed = 1;
1359                         witness_increment_graph_generation();
1360                         mtx_unlock_spin(&w_mtx);
1361
1362 #ifdef WITNESS_NO_VNODE
1363                         /*
1364                          * There are known LORs between VNODE locks. They are
1365                          * not an indication of a bug. VNODE locks are flagged
1366                          * as such (LO_IS_VNODE) and we don't yell if the LOR
1367                          * is between 2 VNODE locks.
1368                          */
1369                         if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1370                             (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1371                                 return;
1372 #endif
1373
1374                         /*
1375                          * Ok, yell about it.
1376                          */
1377                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1378                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1379                                 witness_output(
1380                 "lock order reversal: (sleepable after non-sleepable)\n");
1381                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1382                             && lock == &Giant.lock_object)
1383                                 witness_output(
1384                 "lock order reversal: (Giant after non-sleepable)\n");
1385                         else
1386                                 witness_output("lock order reversal:\n");
1387
1388                         /*
1389                          * Try to locate an earlier lock with
1390                          * witness w in our list.
1391                          */
1392                         do {
1393                                 lock2 = &lle->ll_children[i];
1394                                 MPASS(lock2->li_lock != NULL);
1395                                 if (lock2->li_lock->lo_witness == w)
1396                                         break;
1397                                 if (i == 0 && lle->ll_next != NULL) {
1398                                         lle = lle->ll_next;
1399                                         i = lle->ll_count - 1;
1400                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1401                                 } else
1402                                         i--;
1403                         } while (i >= 0);
1404                         if (i < 0) {
1405                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1406                                     lock1->li_lock, lock1->li_lock->lo_name,
1407                                     w1->w_name, fixup_filename(lock1->li_file),
1408                                     lock1->li_line);
1409                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1410                                     lock->lo_name, w->w_name,
1411                                     fixup_filename(file), line);
1412                         } else {
1413                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1414                                     lock2->li_lock, lock2->li_lock->lo_name,
1415                                     lock2->li_lock->lo_witness->w_name,
1416                                     fixup_filename(lock2->li_file),
1417                                     lock2->li_line);
1418                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1419                                     lock1->li_lock, lock1->li_lock->lo_name,
1420                                     w1->w_name, fixup_filename(lock1->li_file),
1421                                     lock1->li_line);
1422                                 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1423                                     lock->lo_name, w->w_name,
1424                                     fixup_filename(file), line);
1425                         }
1426                         witness_debugger(1, __func__);
1427                         return;
1428                 }
1429         }
1430
1431         /*
1432          * If requested, build a new lock order.  However, don't build a new
1433          * relationship between a sleepable lock and Giant if it is in the
1434          * wrong direction.  The correct lock order is that sleepable locks
1435          * always come before Giant.
1436          */
1437         if (flags & LOP_NEWORDER &&
1438             !(plock->li_lock == &Giant.lock_object &&
1439             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1440                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1441                     w->w_name, plock->li_lock->lo_witness->w_name);
1442                 itismychild(plock->li_lock->lo_witness, w);
1443         }
1444 out:
1445         mtx_unlock_spin(&w_mtx);
1446 }
1447
1448 void
1449 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1450 {
1451         struct lock_list_entry **lock_list, *lle;
1452         struct lock_instance *instance;
1453         struct witness *w;
1454         struct thread *td;
1455
1456         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1457             panicstr != NULL)
1458                 return;
1459         w = lock->lo_witness;
1460         td = curthread;
1461
1462         /* Determine lock list for this lock. */
1463         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1464                 lock_list = &td->td_sleeplocks;
1465         else
1466                 lock_list = PCPU_PTR(spinlocks);
1467
1468         /* Check to see if we are recursing on a lock we already own. */
1469         instance = find_instance(*lock_list, lock);
1470         if (instance != NULL) {
1471                 instance->li_flags++;
1472                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1473                     td->td_proc->p_pid, lock->lo_name,
1474                     instance->li_flags & LI_RECURSEMASK);
1475                 instance->li_file = file;
1476                 instance->li_line = line;
1477                 return;
1478         }
1479
1480         /* Update per-witness last file and line acquire. */
1481         w->w_file = file;
1482         w->w_line = line;
1483
1484         /* Find the next open lock instance in the list and fill it. */
1485         lle = *lock_list;
1486         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1487                 lle = witness_lock_list_get();
1488                 if (lle == NULL)
1489                         return;
1490                 lle->ll_next = *lock_list;
1491                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1492                     td->td_proc->p_pid, lle);
1493                 *lock_list = lle;
1494         }
1495         instance = &lle->ll_children[lle->ll_count++];
1496         instance->li_lock = lock;
1497         instance->li_line = line;
1498         instance->li_file = file;
1499         if ((flags & LOP_EXCLUSIVE) != 0)
1500                 instance->li_flags = LI_EXCLUSIVE;
1501         else
1502                 instance->li_flags = 0;
1503         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1504             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1505 }
1506
1507 void
1508 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1509 {
1510         struct lock_instance *instance;
1511         struct lock_class *class;
1512
1513         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1514         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1515                 return;
1516         class = LOCK_CLASS(lock);
1517         if (witness_watch) {
1518                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1519                         kassert_panic(
1520                             "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1521                             class->lc_name, lock->lo_name,
1522                             fixup_filename(file), line);
1523                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1524                         kassert_panic(
1525                             "upgrade of non-sleep lock (%s) %s @ %s:%d",
1526                             class->lc_name, lock->lo_name,
1527                             fixup_filename(file), line);
1528         }
1529         instance = find_instance(curthread->td_sleeplocks, lock);
1530         if (instance == NULL) {
1531                 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1532                     class->lc_name, lock->lo_name,
1533                     fixup_filename(file), line);
1534                 return;
1535         }
1536         if (witness_watch) {
1537                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1538                         kassert_panic(
1539                             "upgrade of exclusive lock (%s) %s @ %s:%d",
1540                             class->lc_name, lock->lo_name,
1541                             fixup_filename(file), line);
1542                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1543                         kassert_panic(
1544                             "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1545                             class->lc_name, lock->lo_name,
1546                             instance->li_flags & LI_RECURSEMASK,
1547                             fixup_filename(file), line);
1548         }
1549         instance->li_flags |= LI_EXCLUSIVE;
1550 }
1551
1552 void
1553 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1554     int line)
1555 {
1556         struct lock_instance *instance;
1557         struct lock_class *class;
1558
1559         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1560         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1561                 return;
1562         class = LOCK_CLASS(lock);
1563         if (witness_watch) {
1564                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1565                         kassert_panic(
1566                             "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1567                             class->lc_name, lock->lo_name,
1568                             fixup_filename(file), line);
1569                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1570                         kassert_panic(
1571                             "downgrade of non-sleep lock (%s) %s @ %s:%d",
1572                             class->lc_name, lock->lo_name,
1573                             fixup_filename(file), line);
1574         }
1575         instance = find_instance(curthread->td_sleeplocks, lock);
1576         if (instance == NULL) {
1577                 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1578                     class->lc_name, lock->lo_name,
1579                     fixup_filename(file), line);
1580                 return;
1581         }
1582         if (witness_watch) {
1583                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1584                         kassert_panic(
1585                             "downgrade of shared lock (%s) %s @ %s:%d",
1586                             class->lc_name, lock->lo_name,
1587                             fixup_filename(file), line);
1588                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1589                         kassert_panic(
1590                             "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1591                             class->lc_name, lock->lo_name,
1592                             instance->li_flags & LI_RECURSEMASK,
1593                             fixup_filename(file), line);
1594         }
1595         instance->li_flags &= ~LI_EXCLUSIVE;
1596 }
1597
1598 void
1599 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1600 {
1601         struct lock_list_entry **lock_list, *lle;
1602         struct lock_instance *instance;
1603         struct lock_class *class;
1604         struct thread *td;
1605         register_t s;
1606         int i, j;
1607
1608         if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1609                 return;
1610         td = curthread;
1611         class = LOCK_CLASS(lock);
1612
1613         /* Find lock instance associated with this lock. */
1614         if (class->lc_flags & LC_SLEEPLOCK)
1615                 lock_list = &td->td_sleeplocks;
1616         else
1617                 lock_list = PCPU_PTR(spinlocks);
1618         lle = *lock_list;
1619         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1620                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1621                         instance = &(*lock_list)->ll_children[i];
1622                         if (instance->li_lock == lock)
1623                                 goto found;
1624                 }
1625
1626         /*
1627          * When disabling WITNESS through witness_watch we could end up in
1628          * having registered locks in the td_sleeplocks queue.
1629          * We have to make sure we flush these queues, so just search for
1630          * eventual register locks and remove them.
1631          */
1632         if (witness_watch > 0) {
1633                 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1634                     lock->lo_name, fixup_filename(file), line);
1635                 return;
1636         } else {
1637                 return;
1638         }
1639 found:
1640
1641         /* First, check for shared/exclusive mismatches. */
1642         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1643             (flags & LOP_EXCLUSIVE) == 0) {
1644                 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1645                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1646                 witness_output("while exclusively locked from %s:%d\n",
1647                     fixup_filename(instance->li_file), instance->li_line);
1648                 kassert_panic("excl->ushare");
1649         }
1650         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1651             (flags & LOP_EXCLUSIVE) != 0) {
1652                 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1653                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1654                 witness_output("while share locked from %s:%d\n",
1655                     fixup_filename(instance->li_file),
1656                     instance->li_line);
1657                 kassert_panic("share->uexcl");
1658         }
1659         /* If we are recursed, unrecurse. */
1660         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1661                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1662                     td->td_proc->p_pid, instance->li_lock->lo_name,
1663                     instance->li_flags);
1664                 instance->li_flags--;
1665                 return;
1666         }
1667         /* The lock is now being dropped, check for NORELEASE flag */
1668         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1669                 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1670                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1671                 kassert_panic("lock marked norelease");
1672         }
1673
1674         /* Otherwise, remove this item from the list. */
1675         s = intr_disable();
1676         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1677             td->td_proc->p_pid, instance->li_lock->lo_name,
1678             (*lock_list)->ll_count - 1);
1679         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1680                 (*lock_list)->ll_children[j] =
1681                     (*lock_list)->ll_children[j + 1];
1682         (*lock_list)->ll_count--;
1683         intr_restore(s);
1684
1685         /*
1686          * In order to reduce contention on w_mtx, we want to keep always an
1687          * head object into lists so that frequent allocation from the
1688          * free witness pool (and subsequent locking) is avoided.
1689          * In order to maintain the current code simple, when the head
1690          * object is totally unloaded it means also that we do not have
1691          * further objects in the list, so the list ownership needs to be
1692          * hand over to another object if the current head needs to be freed.
1693          */
1694         if ((*lock_list)->ll_count == 0) {
1695                 if (*lock_list == lle) {
1696                         if (lle->ll_next == NULL)
1697                                 return;
1698                 } else
1699                         lle = *lock_list;
1700                 *lock_list = lle->ll_next;
1701                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1702                     td->td_proc->p_pid, lle);
1703                 witness_lock_list_free(lle);
1704         }
1705 }
1706
1707 void
1708 witness_thread_exit(struct thread *td)
1709 {
1710         struct lock_list_entry *lle;
1711         int i, n;
1712
1713         lle = td->td_sleeplocks;
1714         if (lle == NULL || panicstr != NULL)
1715                 return;
1716         if (lle->ll_count != 0) {
1717                 for (n = 0; lle != NULL; lle = lle->ll_next)
1718                         for (i = lle->ll_count - 1; i >= 0; i--) {
1719                                 if (n == 0)
1720                                         witness_output(
1721                     "Thread %p exiting with the following locks held:\n", td);
1722                                 n++;
1723                                 witness_list_lock(&lle->ll_children[i],
1724                                     witness_output);
1725
1726                         }
1727                 kassert_panic(
1728                     "Thread %p cannot exit while holding sleeplocks\n", td);
1729         }
1730         witness_lock_list_free(lle);
1731 }
1732
1733 /*
1734  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1735  * exempt Giant and sleepable locks from the checks as well.  If any
1736  * non-exempt locks are held, then a supplied message is printed to the
1737  * output channel along with a list of the offending locks.  If indicated in the
1738  * flags then a failure results in a panic as well.
1739  */
1740 int
1741 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1742 {
1743         struct lock_list_entry *lock_list, *lle;
1744         struct lock_instance *lock1;
1745         struct thread *td;
1746         va_list ap;
1747         int i, n;
1748
1749         if (witness_cold || witness_watch < 1 || panicstr != NULL)
1750                 return (0);
1751         n = 0;
1752         td = curthread;
1753         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1754                 for (i = lle->ll_count - 1; i >= 0; i--) {
1755                         lock1 = &lle->ll_children[i];
1756                         if (lock1->li_lock == lock)
1757                                 continue;
1758                         if (flags & WARN_GIANTOK &&
1759                             lock1->li_lock == &Giant.lock_object)
1760                                 continue;
1761                         if (flags & WARN_SLEEPOK &&
1762                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1763                                 continue;
1764                         if (n == 0) {
1765                                 va_start(ap, fmt);
1766                                 vprintf(fmt, ap);
1767                                 va_end(ap);
1768                                 printf(" with the following %slocks held:\n",
1769                                     (flags & WARN_SLEEPOK) != 0 ?
1770                                     "non-sleepable " : "");
1771                         }
1772                         n++;
1773                         witness_list_lock(lock1, printf);
1774                 }
1775
1776         /*
1777          * Pin the thread in order to avoid problems with thread migration.
1778          * Once that all verifies are passed about spinlocks ownership,
1779          * the thread is in a safe path and it can be unpinned.
1780          */
1781         sched_pin();
1782         lock_list = PCPU_GET(spinlocks);
1783         if (lock_list != NULL && lock_list->ll_count != 0) {
1784                 sched_unpin();
1785
1786                 /*
1787                  * We should only have one spinlock and as long as
1788                  * the flags cannot match for this locks class,
1789                  * check if the first spinlock is the one curthread
1790                  * should hold.
1791                  */
1792                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1793                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1794                     lock1->li_lock == lock && n == 0)
1795                         return (0);
1796
1797                 va_start(ap, fmt);
1798                 vprintf(fmt, ap);
1799                 va_end(ap);
1800                 printf(" with the following %slocks held:\n",
1801                     (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1802                 n += witness_list_locks(&lock_list, printf);
1803         } else
1804                 sched_unpin();
1805         if (flags & WARN_PANIC && n)
1806                 kassert_panic("%s", __func__);
1807         else
1808                 witness_debugger(n, __func__);
1809         return (n);
1810 }
1811
1812 const char *
1813 witness_file(struct lock_object *lock)
1814 {
1815         struct witness *w;
1816
1817         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1818                 return ("?");
1819         w = lock->lo_witness;
1820         return (w->w_file);
1821 }
1822
1823 int
1824 witness_line(struct lock_object *lock)
1825 {
1826         struct witness *w;
1827
1828         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1829                 return (0);
1830         w = lock->lo_witness;
1831         return (w->w_line);
1832 }
1833
1834 static struct witness *
1835 enroll(const char *description, struct lock_class *lock_class)
1836 {
1837         struct witness *w;
1838
1839         MPASS(description != NULL);
1840
1841         if (witness_watch == -1 || panicstr != NULL)
1842                 return (NULL);
1843         if ((lock_class->lc_flags & LC_SPINLOCK)) {
1844                 if (witness_skipspin)
1845                         return (NULL);
1846         } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1847                 kassert_panic("lock class %s is not sleep or spin",
1848                     lock_class->lc_name);
1849                 return (NULL);
1850         }
1851
1852         mtx_lock_spin(&w_mtx);
1853         w = witness_hash_get(description);
1854         if (w)
1855                 goto found;
1856         if ((w = witness_get()) == NULL)
1857                 return (NULL);
1858         MPASS(strlen(description) < MAX_W_NAME);
1859         strcpy(w->w_name, description);
1860         w->w_class = lock_class;
1861         w->w_refcount = 1;
1862         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1863         if (lock_class->lc_flags & LC_SPINLOCK) {
1864                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1865                 w_spin_cnt++;
1866         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1867                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1868                 w_sleep_cnt++;
1869         }
1870
1871         /* Insert new witness into the hash */
1872         witness_hash_put(w);
1873         witness_increment_graph_generation();
1874         mtx_unlock_spin(&w_mtx);
1875         return (w);
1876 found:
1877         w->w_refcount++;
1878         if (w->w_refcount == 1)
1879                 w->w_class = lock_class;
1880         mtx_unlock_spin(&w_mtx);
1881         if (lock_class != w->w_class)
1882                 kassert_panic(
1883                     "lock (%s) %s does not match earlier (%s) lock",
1884                     description, lock_class->lc_name,
1885                     w->w_class->lc_name);
1886         return (w);
1887 }
1888
1889 static void
1890 depart(struct witness *w)
1891 {
1892
1893         MPASS(w->w_refcount == 0);
1894         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1895                 w_sleep_cnt--;
1896         } else {
1897                 w_spin_cnt--;
1898         }
1899         /*
1900          * Set file to NULL as it may point into a loadable module.
1901          */
1902         w->w_file = NULL;
1903         w->w_line = 0;
1904         witness_increment_graph_generation();
1905 }
1906
1907
1908 static void
1909 adopt(struct witness *parent, struct witness *child)
1910 {
1911         int pi, ci, i, j;
1912
1913         if (witness_cold == 0)
1914                 mtx_assert(&w_mtx, MA_OWNED);
1915
1916         /* If the relationship is already known, there's no work to be done. */
1917         if (isitmychild(parent, child))
1918                 return;
1919
1920         /* When the structure of the graph changes, bump up the generation. */
1921         witness_increment_graph_generation();
1922
1923         /*
1924          * The hard part ... create the direct relationship, then propagate all
1925          * indirect relationships.
1926          */
1927         pi = parent->w_index;
1928         ci = child->w_index;
1929         WITNESS_INDEX_ASSERT(pi);
1930         WITNESS_INDEX_ASSERT(ci);
1931         MPASS(pi != ci);
1932         w_rmatrix[pi][ci] |= WITNESS_PARENT;
1933         w_rmatrix[ci][pi] |= WITNESS_CHILD;
1934
1935         /*
1936          * If parent was not already an ancestor of child,
1937          * then we increment the descendant and ancestor counters.
1938          */
1939         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1940                 parent->w_num_descendants++;
1941                 child->w_num_ancestors++;
1942         }
1943
1944         /*
1945          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1946          * an ancestor of 'pi' during this loop.
1947          */
1948         for (i = 1; i <= w_max_used_index; i++) {
1949                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1950                     (i != pi))
1951                         continue;
1952
1953                 /* Find each descendant of 'i' and mark it as a descendant. */
1954                 for (j = 1; j <= w_max_used_index; j++) {
1955
1956                         /*
1957                          * Skip children that are already marked as
1958                          * descendants of 'i'.
1959                          */
1960                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1961                                 continue;
1962
1963                         /*
1964                          * We are only interested in descendants of 'ci'. Note
1965                          * that 'ci' itself is counted as a descendant of 'ci'.
1966                          */
1967                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1968                             (j != ci))
1969                                 continue;
1970                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1971                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1972                         w_data[i].w_num_descendants++;
1973                         w_data[j].w_num_ancestors++;
1974
1975                         /*
1976                          * Make sure we aren't marking a node as both an
1977                          * ancestor and descendant. We should have caught
1978                          * this as a lock order reversal earlier.
1979                          */
1980                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1981                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1982                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1983                                     "both ancestor and descendant\n",
1984                                     i, j, w_rmatrix[i][j]);
1985                                 kdb_backtrace();
1986                                 printf("Witness disabled.\n");
1987                                 witness_watch = -1;
1988                         }
1989                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1990                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1991                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1992                                     "both ancestor and descendant\n",
1993                                     j, i, w_rmatrix[j][i]);
1994                                 kdb_backtrace();
1995                                 printf("Witness disabled.\n");
1996                                 witness_watch = -1;
1997                         }
1998                 }
1999         }
2000 }
2001
2002 static void
2003 itismychild(struct witness *parent, struct witness *child)
2004 {
2005         int unlocked;
2006
2007         MPASS(child != NULL && parent != NULL);
2008         if (witness_cold == 0)
2009                 mtx_assert(&w_mtx, MA_OWNED);
2010
2011         if (!witness_lock_type_equal(parent, child)) {
2012                 if (witness_cold == 0) {
2013                         unlocked = 1;
2014                         mtx_unlock_spin(&w_mtx);
2015                 } else {
2016                         unlocked = 0;
2017                 }
2018                 kassert_panic(
2019                     "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2020                     "the same lock type", __func__, parent->w_name,
2021                     parent->w_class->lc_name, child->w_name,
2022                     child->w_class->lc_name);
2023                 if (unlocked)
2024                         mtx_lock_spin(&w_mtx);
2025         }
2026         adopt(parent, child);
2027 }
2028
2029 /*
2030  * Generic code for the isitmy*() functions. The rmask parameter is the
2031  * expected relationship of w1 to w2.
2032  */
2033 static int
2034 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2035 {
2036         unsigned char r1, r2;
2037         int i1, i2;
2038
2039         i1 = w1->w_index;
2040         i2 = w2->w_index;
2041         WITNESS_INDEX_ASSERT(i1);
2042         WITNESS_INDEX_ASSERT(i2);
2043         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2044         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2045
2046         /* The flags on one better be the inverse of the flags on the other */
2047         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2048             (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2049                 /* Don't squawk if we're potentially racing with an update. */
2050                 if (!mtx_owned(&w_mtx))
2051                         return (0);
2052                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2053                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
2054                     "w_rmatrix[%d][%d] == %hhx\n",
2055                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2056                     i2, i1, r2);
2057                 kdb_backtrace();
2058                 printf("Witness disabled.\n");
2059                 witness_watch = -1;
2060         }
2061         return (r1 & rmask);
2062 }
2063
2064 /*
2065  * Checks if @child is a direct child of @parent.
2066  */
2067 static int
2068 isitmychild(struct witness *parent, struct witness *child)
2069 {
2070
2071         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2072 }
2073
2074 /*
2075  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2076  */
2077 static int
2078 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2079 {
2080
2081         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2082             __func__));
2083 }
2084
2085 #ifdef BLESSING
2086 static int
2087 blessed(struct witness *w1, struct witness *w2)
2088 {
2089         int i;
2090         struct witness_blessed *b;
2091
2092         for (i = 0; i < nitems(blessed_list); i++) {
2093                 b = &blessed_list[i];
2094                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2095                         if (strcmp(w2->w_name, b->b_lock2) == 0)
2096                                 return (1);
2097                         continue;
2098                 }
2099                 if (strcmp(w1->w_name, b->b_lock2) == 0)
2100                         if (strcmp(w2->w_name, b->b_lock1) == 0)
2101                                 return (1);
2102         }
2103         return (0);
2104 }
2105 #endif
2106
2107 static struct witness *
2108 witness_get(void)
2109 {
2110         struct witness *w;
2111         int index;
2112
2113         if (witness_cold == 0)
2114                 mtx_assert(&w_mtx, MA_OWNED);
2115
2116         if (witness_watch == -1) {
2117                 mtx_unlock_spin(&w_mtx);
2118                 return (NULL);
2119         }
2120         if (STAILQ_EMPTY(&w_free)) {
2121                 witness_watch = -1;
2122                 mtx_unlock_spin(&w_mtx);
2123                 printf("WITNESS: unable to allocate a new witness object\n");
2124                 return (NULL);
2125         }
2126         w = STAILQ_FIRST(&w_free);
2127         STAILQ_REMOVE_HEAD(&w_free, w_list);
2128         w_free_cnt--;
2129         index = w->w_index;
2130         MPASS(index > 0 && index == w_max_used_index+1 &&
2131             index < witness_count);
2132         bzero(w, sizeof(*w));
2133         w->w_index = index;
2134         if (index > w_max_used_index)
2135                 w_max_used_index = index;
2136         return (w);
2137 }
2138
2139 static void
2140 witness_free(struct witness *w)
2141 {
2142
2143         STAILQ_INSERT_HEAD(&w_free, w, w_list);
2144         w_free_cnt++;
2145 }
2146
2147 static struct lock_list_entry *
2148 witness_lock_list_get(void)
2149 {
2150         struct lock_list_entry *lle;
2151
2152         if (witness_watch == -1)
2153                 return (NULL);
2154         mtx_lock_spin(&w_mtx);
2155         lle = w_lock_list_free;
2156         if (lle == NULL) {
2157                 witness_watch = -1;
2158                 mtx_unlock_spin(&w_mtx);
2159                 printf("%s: witness exhausted\n", __func__);
2160                 return (NULL);
2161         }
2162         w_lock_list_free = lle->ll_next;
2163         mtx_unlock_spin(&w_mtx);
2164         bzero(lle, sizeof(*lle));
2165         return (lle);
2166 }
2167
2168 static void
2169 witness_lock_list_free(struct lock_list_entry *lle)
2170 {
2171
2172         mtx_lock_spin(&w_mtx);
2173         lle->ll_next = w_lock_list_free;
2174         w_lock_list_free = lle;
2175         mtx_unlock_spin(&w_mtx);
2176 }
2177
2178 static struct lock_instance *
2179 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2180 {
2181         struct lock_list_entry *lle;
2182         struct lock_instance *instance;
2183         int i;
2184
2185         for (lle = list; lle != NULL; lle = lle->ll_next)
2186                 for (i = lle->ll_count - 1; i >= 0; i--) {
2187                         instance = &lle->ll_children[i];
2188                         if (instance->li_lock == lock)
2189                                 return (instance);
2190                 }
2191         return (NULL);
2192 }
2193
2194 static void
2195 witness_list_lock(struct lock_instance *instance,
2196     int (*prnt)(const char *fmt, ...))
2197 {
2198         struct lock_object *lock;
2199
2200         lock = instance->li_lock;
2201         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2202             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2203         if (lock->lo_witness->w_name != lock->lo_name)
2204                 prnt(" (%s)", lock->lo_witness->w_name);
2205         prnt(" r = %d (%p) locked @ %s:%d\n",
2206             instance->li_flags & LI_RECURSEMASK, lock,
2207             fixup_filename(instance->li_file), instance->li_line);
2208 }
2209
2210 static int
2211 witness_output(const char *fmt, ...)
2212 {
2213         va_list ap;
2214         int ret;
2215
2216         va_start(ap, fmt);
2217         ret = witness_voutput(fmt, ap);
2218         va_end(ap);
2219         return (ret);
2220 }
2221
2222 static int
2223 witness_voutput(const char *fmt, va_list ap)
2224 {
2225         int ret;
2226
2227         ret = 0;
2228         switch (witness_channel) {
2229         case WITNESS_CONSOLE:
2230                 ret = vprintf(fmt, ap);
2231                 break;
2232         case WITNESS_LOG:
2233                 vlog(LOG_NOTICE, fmt, ap);
2234                 break;
2235         case WITNESS_NONE:
2236                 break;
2237         }
2238         return (ret);
2239 }
2240
2241 #ifdef DDB
2242 static int
2243 witness_thread_has_locks(struct thread *td)
2244 {
2245
2246         if (td->td_sleeplocks == NULL)
2247                 return (0);
2248         return (td->td_sleeplocks->ll_count != 0);
2249 }
2250
2251 static int
2252 witness_proc_has_locks(struct proc *p)
2253 {
2254         struct thread *td;
2255
2256         FOREACH_THREAD_IN_PROC(p, td) {
2257                 if (witness_thread_has_locks(td))
2258                         return (1);
2259         }
2260         return (0);
2261 }
2262 #endif
2263
2264 int
2265 witness_list_locks(struct lock_list_entry **lock_list,
2266     int (*prnt)(const char *fmt, ...))
2267 {
2268         struct lock_list_entry *lle;
2269         int i, nheld;
2270
2271         nheld = 0;
2272         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2273                 for (i = lle->ll_count - 1; i >= 0; i--) {
2274                         witness_list_lock(&lle->ll_children[i], prnt);
2275                         nheld++;
2276                 }
2277         return (nheld);
2278 }
2279
2280 /*
2281  * This is a bit risky at best.  We call this function when we have timed
2282  * out acquiring a spin lock, and we assume that the other CPU is stuck
2283  * with this lock held.  So, we go groveling around in the other CPU's
2284  * per-cpu data to try to find the lock instance for this spin lock to
2285  * see when it was last acquired.
2286  */
2287 void
2288 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2289     int (*prnt)(const char *fmt, ...))
2290 {
2291         struct lock_instance *instance;
2292         struct pcpu *pc;
2293
2294         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2295                 return;
2296         pc = pcpu_find(owner->td_oncpu);
2297         instance = find_instance(pc->pc_spinlocks, lock);
2298         if (instance != NULL)
2299                 witness_list_lock(instance, prnt);
2300 }
2301
2302 void
2303 witness_save(struct lock_object *lock, const char **filep, int *linep)
2304 {
2305         struct lock_list_entry *lock_list;
2306         struct lock_instance *instance;
2307         struct lock_class *class;
2308
2309         /*
2310          * This function is used independently in locking code to deal with
2311          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2312          * is gone.
2313          */
2314         if (SCHEDULER_STOPPED())
2315                 return;
2316         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2317         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2318                 return;
2319         class = LOCK_CLASS(lock);
2320         if (class->lc_flags & LC_SLEEPLOCK)
2321                 lock_list = curthread->td_sleeplocks;
2322         else {
2323                 if (witness_skipspin)
2324                         return;
2325                 lock_list = PCPU_GET(spinlocks);
2326         }
2327         instance = find_instance(lock_list, lock);
2328         if (instance == NULL) {
2329                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2330                     class->lc_name, lock->lo_name);
2331                 return;
2332         }
2333         *filep = instance->li_file;
2334         *linep = instance->li_line;
2335 }
2336
2337 void
2338 witness_restore(struct lock_object *lock, const char *file, int line)
2339 {
2340         struct lock_list_entry *lock_list;
2341         struct lock_instance *instance;
2342         struct lock_class *class;
2343
2344         /*
2345          * This function is used independently in locking code to deal with
2346          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2347          * is gone.
2348          */
2349         if (SCHEDULER_STOPPED())
2350                 return;
2351         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
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         lock->lo_witness->w_file = file;
2367         lock->lo_witness->w_line = line;
2368         if (instance == NULL)
2369                 return;
2370         instance->li_file = file;
2371         instance->li_line = line;
2372 }
2373
2374 void
2375 witness_assert(const struct lock_object *lock, int flags, const char *file,
2376     int line)
2377 {
2378 #ifdef INVARIANT_SUPPORT
2379         struct lock_instance *instance;
2380         struct lock_class *class;
2381
2382         if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2383                 return;
2384         class = LOCK_CLASS(lock);
2385         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2386                 instance = find_instance(curthread->td_sleeplocks, lock);
2387         else if ((class->lc_flags & LC_SPINLOCK) != 0)
2388                 instance = find_instance(PCPU_GET(spinlocks), lock);
2389         else {
2390                 kassert_panic("Lock (%s) %s is not sleep or spin!",
2391                     class->lc_name, lock->lo_name);
2392                 return;
2393         }
2394         switch (flags) {
2395         case LA_UNLOCKED:
2396                 if (instance != NULL)
2397                         kassert_panic("Lock (%s) %s locked @ %s:%d.",
2398                             class->lc_name, lock->lo_name,
2399                             fixup_filename(file), line);
2400                 break;
2401         case LA_LOCKED:
2402         case LA_LOCKED | LA_RECURSED:
2403         case LA_LOCKED | LA_NOTRECURSED:
2404         case LA_SLOCKED:
2405         case LA_SLOCKED | LA_RECURSED:
2406         case LA_SLOCKED | LA_NOTRECURSED:
2407         case LA_XLOCKED:
2408         case LA_XLOCKED | LA_RECURSED:
2409         case LA_XLOCKED | LA_NOTRECURSED:
2410                 if (instance == NULL) {
2411                         kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2412                             class->lc_name, lock->lo_name,
2413                             fixup_filename(file), line);
2414                         break;
2415                 }
2416                 if ((flags & LA_XLOCKED) != 0 &&
2417                     (instance->li_flags & LI_EXCLUSIVE) == 0)
2418                         kassert_panic(
2419                             "Lock (%s) %s not exclusively locked @ %s:%d.",
2420                             class->lc_name, lock->lo_name,
2421                             fixup_filename(file), line);
2422                 if ((flags & LA_SLOCKED) != 0 &&
2423                     (instance->li_flags & LI_EXCLUSIVE) != 0)
2424                         kassert_panic(
2425                             "Lock (%s) %s exclusively locked @ %s:%d.",
2426                             class->lc_name, lock->lo_name,
2427                             fixup_filename(file), line);
2428                 if ((flags & LA_RECURSED) != 0 &&
2429                     (instance->li_flags & LI_RECURSEMASK) == 0)
2430                         kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2431                             class->lc_name, lock->lo_name,
2432                             fixup_filename(file), line);
2433                 if ((flags & LA_NOTRECURSED) != 0 &&
2434                     (instance->li_flags & LI_RECURSEMASK) != 0)
2435                         kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2436                             class->lc_name, lock->lo_name,
2437                             fixup_filename(file), line);
2438                 break;
2439         default:
2440                 kassert_panic("Invalid lock assertion at %s:%d.",
2441                     fixup_filename(file), line);
2442
2443         }
2444 #endif  /* INVARIANT_SUPPORT */
2445 }
2446
2447 static void
2448 witness_setflag(struct lock_object *lock, int flag, int set)
2449 {
2450         struct lock_list_entry *lock_list;
2451         struct lock_instance *instance;
2452         struct lock_class *class;
2453
2454         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2455                 return;
2456         class = LOCK_CLASS(lock);
2457         if (class->lc_flags & LC_SLEEPLOCK)
2458                 lock_list = curthread->td_sleeplocks;
2459         else {
2460                 if (witness_skipspin)
2461                         return;
2462                 lock_list = PCPU_GET(spinlocks);
2463         }
2464         instance = find_instance(lock_list, lock);
2465         if (instance == NULL) {
2466                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2467                     class->lc_name, lock->lo_name);
2468                 return;
2469         }
2470
2471         if (set)
2472                 instance->li_flags |= flag;
2473         else
2474                 instance->li_flags &= ~flag;
2475 }
2476
2477 void
2478 witness_norelease(struct lock_object *lock)
2479 {
2480
2481         witness_setflag(lock, LI_NORELEASE, 1);
2482 }
2483
2484 void
2485 witness_releaseok(struct lock_object *lock)
2486 {
2487
2488         witness_setflag(lock, LI_NORELEASE, 0);
2489 }
2490
2491 #ifdef DDB
2492 static void
2493 witness_ddb_list(struct thread *td)
2494 {
2495
2496         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2497         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2498
2499         if (witness_watch < 1)
2500                 return;
2501
2502         witness_list_locks(&td->td_sleeplocks, db_printf);
2503
2504         /*
2505          * We only handle spinlocks if td == curthread.  This is somewhat broken
2506          * if td is currently executing on some other CPU and holds spin locks
2507          * as we won't display those locks.  If we had a MI way of getting
2508          * the per-cpu data for a given cpu then we could use
2509          * td->td_oncpu to get the list of spinlocks for this thread
2510          * and "fix" this.
2511          *
2512          * That still wouldn't really fix this unless we locked the scheduler
2513          * lock or stopped the other CPU to make sure it wasn't changing the
2514          * list out from under us.  It is probably best to just not try to
2515          * handle threads on other CPU's for now.
2516          */
2517         if (td == curthread && PCPU_GET(spinlocks) != NULL)
2518                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2519 }
2520
2521 DB_SHOW_COMMAND(locks, db_witness_list)
2522 {
2523         struct thread *td;
2524
2525         if (have_addr)
2526                 td = db_lookup_thread(addr, true);
2527         else
2528                 td = kdb_thread;
2529         witness_ddb_list(td);
2530 }
2531
2532 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2533 {
2534         struct thread *td;
2535         struct proc *p;
2536
2537         /*
2538          * It would be nice to list only threads and processes that actually
2539          * held sleep locks, but that information is currently not exported
2540          * by WITNESS.
2541          */
2542         FOREACH_PROC_IN_SYSTEM(p) {
2543                 if (!witness_proc_has_locks(p))
2544                         continue;
2545                 FOREACH_THREAD_IN_PROC(p, td) {
2546                         if (!witness_thread_has_locks(td))
2547                                 continue;
2548                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2549                             p->p_comm, td, td->td_tid);
2550                         witness_ddb_list(td);
2551                         if (db_pager_quit)
2552                                 return;
2553                 }
2554         }
2555 }
2556 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2557
2558 DB_SHOW_COMMAND(witness, db_witness_display)
2559 {
2560
2561         witness_ddb_display(db_printf);
2562 }
2563 #endif
2564
2565 static void
2566 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2567 {
2568         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2569         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2570         int generation, i, j;
2571
2572         tmp_data1 = NULL;
2573         tmp_data2 = NULL;
2574         tmp_w1 = NULL;
2575         tmp_w2 = NULL;
2576
2577         /* Allocate and init temporary storage space. */
2578         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2579         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2580         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2581             M_WAITOK | M_ZERO);
2582         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2583             M_WAITOK | M_ZERO);
2584         stack_zero(&tmp_data1->wlod_stack);
2585         stack_zero(&tmp_data2->wlod_stack);
2586
2587 restart:
2588         mtx_lock_spin(&w_mtx);
2589         generation = w_generation;
2590         mtx_unlock_spin(&w_mtx);
2591         sbuf_printf(sb, "Number of known direct relationships is %d\n",
2592             w_lohash.wloh_count);
2593         for (i = 1; i < w_max_used_index; i++) {
2594                 mtx_lock_spin(&w_mtx);
2595                 if (generation != w_generation) {
2596                         mtx_unlock_spin(&w_mtx);
2597
2598                         /* The graph has changed, try again. */
2599                         *oldidx = 0;
2600                         sbuf_clear(sb);
2601                         goto restart;
2602                 }
2603
2604                 w1 = &w_data[i];
2605                 if (w1->w_reversed == 0) {
2606                         mtx_unlock_spin(&w_mtx);
2607                         continue;
2608                 }
2609
2610                 /* Copy w1 locally so we can release the spin lock. */
2611                 *tmp_w1 = *w1;
2612                 mtx_unlock_spin(&w_mtx);
2613
2614                 if (tmp_w1->w_reversed == 0)
2615                         continue;
2616                 for (j = 1; j < w_max_used_index; j++) {
2617                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2618                                 continue;
2619
2620                         mtx_lock_spin(&w_mtx);
2621                         if (generation != w_generation) {
2622                                 mtx_unlock_spin(&w_mtx);
2623
2624                                 /* The graph has changed, try again. */
2625                                 *oldidx = 0;
2626                                 sbuf_clear(sb);
2627                                 goto restart;
2628                         }
2629
2630                         w2 = &w_data[j];
2631                         data1 = witness_lock_order_get(w1, w2);
2632                         data2 = witness_lock_order_get(w2, w1);
2633
2634                         /*
2635                          * Copy information locally so we can release the
2636                          * spin lock.
2637                          */
2638                         *tmp_w2 = *w2;
2639
2640                         if (data1) {
2641                                 stack_zero(&tmp_data1->wlod_stack);
2642                                 stack_copy(&data1->wlod_stack,
2643                                     &tmp_data1->wlod_stack);
2644                         }
2645                         if (data2 && data2 != data1) {
2646                                 stack_zero(&tmp_data2->wlod_stack);
2647                                 stack_copy(&data2->wlod_stack,
2648                                     &tmp_data2->wlod_stack);
2649                         }
2650                         mtx_unlock_spin(&w_mtx);
2651
2652                         sbuf_printf(sb,
2653             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2654                             tmp_w1->w_name, tmp_w1->w_class->lc_name,
2655                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
2656                         if (data1) {
2657                                 sbuf_printf(sb,
2658                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2659                                     tmp_w1->w_name, tmp_w1->w_class->lc_name,
2660                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
2661                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2662                                 sbuf_printf(sb, "\n");
2663                         }
2664                         if (data2 && data2 != data1) {
2665                                 sbuf_printf(sb,
2666                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2667                                     tmp_w2->w_name, tmp_w2->w_class->lc_name,
2668                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
2669                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2670                                 sbuf_printf(sb, "\n");
2671                         }
2672                 }
2673         }
2674         mtx_lock_spin(&w_mtx);
2675         if (generation != w_generation) {
2676                 mtx_unlock_spin(&w_mtx);
2677
2678                 /*
2679                  * The graph changed while we were printing stack data,
2680                  * try again.
2681                  */
2682                 *oldidx = 0;
2683                 sbuf_clear(sb);
2684                 goto restart;
2685         }
2686         mtx_unlock_spin(&w_mtx);
2687
2688         /* Free temporary storage space. */
2689         free(tmp_data1, M_TEMP);
2690         free(tmp_data2, M_TEMP);
2691         free(tmp_w1, M_TEMP);
2692         free(tmp_w2, M_TEMP);
2693 }
2694
2695 static int
2696 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2697 {
2698         struct sbuf *sb;
2699         int error;
2700
2701         if (witness_watch < 1) {
2702                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2703                 return (error);
2704         }
2705         if (witness_cold) {
2706                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2707                 return (error);
2708         }
2709         error = 0;
2710         sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2711         if (sb == NULL)
2712                 return (ENOMEM);
2713
2714         sbuf_print_witness_badstacks(sb, &req->oldidx);
2715
2716         sbuf_finish(sb);
2717         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2718         sbuf_delete(sb);
2719
2720         return (error);
2721 }
2722
2723 #ifdef DDB
2724 static int
2725 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2726 {
2727
2728         return (db_printf("%.*s", len, data));
2729 }
2730
2731 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2732 {
2733         struct sbuf sb;
2734         char buffer[128];
2735         size_t dummy;
2736
2737         sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2738         sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2739         sbuf_print_witness_badstacks(&sb, &dummy);
2740         sbuf_finish(&sb);
2741 }
2742 #endif
2743
2744 static int
2745 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2746 {
2747         static const struct {
2748                 enum witness_channel channel;
2749                 const char *name;
2750         } channels[] = {
2751                 { WITNESS_CONSOLE, "console" },
2752                 { WITNESS_LOG, "log" },
2753                 { WITNESS_NONE, "none" },
2754         };
2755         char buf[16];
2756         u_int i;
2757         int error;
2758
2759         buf[0] = '\0';
2760         for (i = 0; i < nitems(channels); i++)
2761                 if (witness_channel == channels[i].channel) {
2762                         snprintf(buf, sizeof(buf), "%s", channels[i].name);
2763                         break;
2764                 }
2765
2766         error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2767         if (error != 0 || req->newptr == NULL)
2768                 return (error);
2769
2770         error = EINVAL;
2771         for (i = 0; i < nitems(channels); i++)
2772                 if (strcmp(channels[i].name, buf) == 0) {
2773                         witness_channel = channels[i].channel;
2774                         error = 0;
2775                         break;
2776                 }
2777         return (error);
2778 }
2779
2780 static int
2781 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2782 {
2783         struct witness *w;
2784         struct sbuf *sb;
2785         int error;
2786
2787 #ifdef __i386__
2788         error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2789         return (error);
2790 #endif
2791
2792         if (witness_watch < 1) {
2793                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2794                 return (error);
2795         }
2796         if (witness_cold) {
2797                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2798                 return (error);
2799         }
2800         error = 0;
2801
2802         error = sysctl_wire_old_buffer(req, 0);
2803         if (error != 0)
2804                 return (error);
2805         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2806         if (sb == NULL)
2807                 return (ENOMEM);
2808         sbuf_printf(sb, "\n");
2809
2810         mtx_lock_spin(&w_mtx);
2811         STAILQ_FOREACH(w, &w_all, w_list)
2812                 w->w_displayed = 0;
2813         STAILQ_FOREACH(w, &w_all, w_list)
2814                 witness_add_fullgraph(sb, w);
2815         mtx_unlock_spin(&w_mtx);
2816
2817         /*
2818          * Close the sbuf and return to userland.
2819          */
2820         error = sbuf_finish(sb);
2821         sbuf_delete(sb);
2822
2823         return (error);
2824 }
2825
2826 static int
2827 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2828 {
2829         int error, value;
2830
2831         value = witness_watch;
2832         error = sysctl_handle_int(oidp, &value, 0, req);
2833         if (error != 0 || req->newptr == NULL)
2834                 return (error);
2835         if (value > 1 || value < -1 ||
2836             (witness_watch == -1 && value != witness_watch))
2837                 return (EINVAL);
2838         witness_watch = value;
2839         return (0);
2840 }
2841
2842 static void
2843 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2844 {
2845         int i;
2846
2847         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2848                 return;
2849         w->w_displayed = 1;
2850
2851         WITNESS_INDEX_ASSERT(w->w_index);
2852         for (i = 1; i <= w_max_used_index; i++) {
2853                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2854                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2855                             w_data[i].w_name);
2856                         witness_add_fullgraph(sb, &w_data[i]);
2857                 }
2858         }
2859 }
2860
2861 /*
2862  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2863  * interprets the key as a string and reads until the null
2864  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2865  * hash value computed from the key.
2866  */
2867 static uint32_t
2868 witness_hash_djb2(const uint8_t *key, uint32_t size)
2869 {
2870         unsigned int hash = 5381;
2871         int i;
2872
2873         /* hash = hash * 33 + key[i] */
2874         if (size)
2875                 for (i = 0; i < size; i++)
2876                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2877         else
2878                 for (i = 0; key[i] != 0; i++)
2879                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2880
2881         return (hash);
2882 }
2883
2884
2885 /*
2886  * Initializes the two witness hash tables. Called exactly once from
2887  * witness_initialize().
2888  */
2889 static void
2890 witness_init_hash_tables(void)
2891 {
2892         int i;
2893
2894         MPASS(witness_cold);
2895
2896         /* Initialize the hash tables. */
2897         for (i = 0; i < WITNESS_HASH_SIZE; i++)
2898                 w_hash.wh_array[i] = NULL;
2899
2900         w_hash.wh_size = WITNESS_HASH_SIZE;
2901         w_hash.wh_count = 0;
2902
2903         /* Initialize the lock order data hash. */
2904         w_lofree = NULL;
2905         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2906                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2907                 w_lodata[i].wlod_next = w_lofree;
2908                 w_lofree = &w_lodata[i];
2909         }
2910         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2911         w_lohash.wloh_count = 0;
2912         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2913                 w_lohash.wloh_array[i] = NULL;
2914 }
2915
2916 static struct witness *
2917 witness_hash_get(const char *key)
2918 {
2919         struct witness *w;
2920         uint32_t hash;
2921
2922         MPASS(key != NULL);
2923         if (witness_cold == 0)
2924                 mtx_assert(&w_mtx, MA_OWNED);
2925         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2926         w = w_hash.wh_array[hash];
2927         while (w != NULL) {
2928                 if (strcmp(w->w_name, key) == 0)
2929                         goto out;
2930                 w = w->w_hash_next;
2931         }
2932
2933 out:
2934         return (w);
2935 }
2936
2937 static void
2938 witness_hash_put(struct witness *w)
2939 {
2940         uint32_t hash;
2941
2942         MPASS(w != NULL);
2943         MPASS(w->w_name != NULL);
2944         if (witness_cold == 0)
2945                 mtx_assert(&w_mtx, MA_OWNED);
2946         KASSERT(witness_hash_get(w->w_name) == NULL,
2947             ("%s: trying to add a hash entry that already exists!", __func__));
2948         KASSERT(w->w_hash_next == NULL,
2949             ("%s: w->w_hash_next != NULL", __func__));
2950
2951         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2952         w->w_hash_next = w_hash.wh_array[hash];
2953         w_hash.wh_array[hash] = w;
2954         w_hash.wh_count++;
2955 }
2956
2957
2958 static struct witness_lock_order_data *
2959 witness_lock_order_get(struct witness *parent, struct witness *child)
2960 {
2961         struct witness_lock_order_data *data = NULL;
2962         struct witness_lock_order_key key;
2963         unsigned int hash;
2964
2965         MPASS(parent != NULL && child != NULL);
2966         key.from = parent->w_index;
2967         key.to = child->w_index;
2968         WITNESS_INDEX_ASSERT(key.from);
2969         WITNESS_INDEX_ASSERT(key.to);
2970         if ((w_rmatrix[parent->w_index][child->w_index]
2971             & WITNESS_LOCK_ORDER_KNOWN) == 0)
2972                 goto out;
2973
2974         hash = witness_hash_djb2((const char*)&key,
2975             sizeof(key)) % w_lohash.wloh_size;
2976         data = w_lohash.wloh_array[hash];
2977         while (data != NULL) {
2978                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2979                         break;
2980                 data = data->wlod_next;
2981         }
2982
2983 out:
2984         return (data);
2985 }
2986
2987 /*
2988  * Verify that parent and child have a known relationship, are not the same,
2989  * and child is actually a child of parent.  This is done without w_mtx
2990  * to avoid contention in the common case.
2991  */
2992 static int
2993 witness_lock_order_check(struct witness *parent, struct witness *child)
2994 {
2995
2996         if (parent != child &&
2997             w_rmatrix[parent->w_index][child->w_index]
2998             & WITNESS_LOCK_ORDER_KNOWN &&
2999             isitmychild(parent, child))
3000                 return (1);
3001
3002         return (0);
3003 }
3004
3005 static int
3006 witness_lock_order_add(struct witness *parent, struct witness *child)
3007 {
3008         struct witness_lock_order_data *data = NULL;
3009         struct witness_lock_order_key key;
3010         unsigned int hash;
3011
3012         MPASS(parent != NULL && child != NULL);
3013         key.from = parent->w_index;
3014         key.to = child->w_index;
3015         WITNESS_INDEX_ASSERT(key.from);
3016         WITNESS_INDEX_ASSERT(key.to);
3017         if (w_rmatrix[parent->w_index][child->w_index]
3018             & WITNESS_LOCK_ORDER_KNOWN)
3019                 return (1);
3020
3021         hash = witness_hash_djb2((const char*)&key,
3022             sizeof(key)) % w_lohash.wloh_size;
3023         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3024         data = w_lofree;
3025         if (data == NULL)
3026                 return (0);
3027         w_lofree = data->wlod_next;
3028         data->wlod_next = w_lohash.wloh_array[hash];
3029         data->wlod_key = key;
3030         w_lohash.wloh_array[hash] = data;
3031         w_lohash.wloh_count++;
3032         stack_zero(&data->wlod_stack);
3033         stack_save(&data->wlod_stack);
3034         return (1);
3035 }
3036
3037 /* Call this whenever the structure of the witness graph changes. */
3038 static void
3039 witness_increment_graph_generation(void)
3040 {
3041
3042         if (witness_cold == 0)
3043                 mtx_assert(&w_mtx, MA_OWNED);
3044         w_generation++;
3045 }
3046
3047 static int
3048 witness_output_drain(void *arg __unused, const char *data, int len)
3049 {
3050
3051         witness_output("%.*s", len, data);
3052         return (len);
3053 }
3054
3055 static void
3056 witness_debugger(int cond, const char *msg)
3057 {
3058         char buf[32];
3059         struct sbuf sb;
3060         struct stack st;
3061
3062         if (!cond)
3063                 return;
3064
3065         if (witness_trace) {
3066                 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3067                 sbuf_set_drain(&sb, witness_output_drain, NULL);
3068
3069                 stack_zero(&st);
3070                 stack_save(&st);
3071                 witness_output("stack backtrace:\n");
3072                 stack_sbuf_print_ddb(&sb, &st);
3073
3074                 sbuf_finish(&sb);
3075         }
3076
3077 #ifdef KDB
3078         if (witness_kdb)
3079                 kdb_enter(KDB_WHY_WITNESS, msg);
3080 #endif
3081 }