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 #ifdef  HWPMC_HOOKS
 503         { "pmc-sleep", &lock_class_mtx_sleep },
 504 #endif
 505         { "process lock", &lock_class_mtx_sleep },
 506         { "session", &lock_class_mtx_sleep },
 507         { "uidinfo hash", &lock_class_rw },
 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_rm },
 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_mtx_sleep },
 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_mtx_sleep },
 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_mtx_sleep },
 565         { "udpinp", &lock_class_rw },
 566         { "so_snd", &lock_class_mtx_sleep },
 567         { NULL, NULL },
 568         /*
 569          * TCP/IP
 570          */
 571         { "tcp", &lock_class_mtx_sleep },
 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         { "process slock", &lock_class_mtx_spin },
 675         { "syscons video lock", &lock_class_mtx_spin },
 676         { "sleepq chain", &lock_class_mtx_spin },
 677         { "rm_spinlock", &lock_class_mtx_spin },
 678         { "turnstile chain", &lock_class_mtx_spin },
 679         { "turnstile lock", &lock_class_mtx_spin },
 680         { "sched lock", &lock_class_mtx_spin },
 681         { "td_contested", &lock_class_mtx_spin },
 682         { "callout", &lock_class_mtx_spin },
 683         { "entropy harvest mutex", &lock_class_mtx_spin },
 684 #ifdef SMP
 685         { "smp rendezvous", &lock_class_mtx_spin },
 686 #endif
 687 #ifdef __powerpc__
 688         { "tlb0", &lock_class_mtx_spin },
 689 #endif
 690         { NULL, NULL },
 691         { "sched lock", &lock_class_mtx_spin },
 692 #ifdef  HWPMC_HOOKS
 693         { "pmc-per-proc", &lock_class_mtx_spin },
 694 #endif
 695         { NULL, NULL },
 696         /*
 697          * leaf locks
 698          */
 699         { "intrcnt", &lock_class_mtx_spin },
 700         { "icu", &lock_class_mtx_spin },
 701 #if defined(SMP) && defined(__sparc64__)
 702         { "ipi", &lock_class_mtx_spin },
 703 #endif
 704 #ifdef __i386__
 705         { "allpmaps", &lock_class_mtx_spin },
 706         { "descriptor tables", &lock_class_mtx_spin },
 707 #endif
 708         { "clk", &lock_class_mtx_spin },
 709         { "cpuset", &lock_class_mtx_spin },
 710         { "mprof lock", &lock_class_mtx_spin },
 711         { "zombie lock", &lock_class_mtx_spin },
 712         { "ALD Queue", &lock_class_mtx_spin },
 713 #if defined(__i386__) || defined(__amd64__)
 714         { "pcicfg", &lock_class_mtx_spin },
 715         { "NDIS thread lock", &lock_class_mtx_spin },
 716 #endif
 717         { "tw_osl_io_lock", &lock_class_mtx_spin },
 718         { "tw_osl_q_lock", &lock_class_mtx_spin },
 719         { "tw_cl_io_lock", &lock_class_mtx_spin },
 720         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 721         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 722 #ifdef  HWPMC_HOOKS
 723         { "pmc-leaf", &lock_class_mtx_spin },
 724 #endif
 725         { "blocked lock", &lock_class_mtx_spin },
 726         { NULL, NULL },
 727         { NULL, NULL }
 728 };
 729
 730 #ifdef BLESSING
 731 /*
 732  * Pairs of locks which have been blessed
 733  * Don't complain about order problems with blessed locks
 734  */
 735 static struct witness_blessed blessed_list[] = {
 736 };
 737 #endif
 738
 739 /*
 740  * This global is set to 0 once it becomes safe to use the witness code.
 741  */
 742 static int witness_cold = 1;
 743
 744 /*
 745  * This global is set to 1 once the static lock orders have been enrolled
 746  * so that a warning can be issued for any spin locks enrolled later.
 747  */
 748 static int witness_spin_warn = 0;
 749
 750 /* Trim useless garbage from filenames. */
 751 static const char *
 752 fixup_filename(const char *file)
 753 {
 754
 755         if (file == NULL)
 756                 return (NULL);
 757         while (strncmp(file, "../", 3) == 0)
 758                 file += 3;
 759         return (file);
 760 }
 761
 762 /*
 763  * Calculate the size of early witness structures.
 764  */
 765 int
 766 witness_startup_count(void)
 767 {
 768         int sz;
 769
 770         sz = sizeof(struct witness) * witness_count;
 771         sz += sizeof(*w_rmatrix) * (witness_count + 1);
 772         sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
 773             (witness_count + 1);
 774
 775         return (sz);
 776 }
 777
 778 /*
 779  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 780  * assume that the early boot is single-threaded at least until after this
 781  * routine is completed.
 782  */
 783 void
 784 witness_startup(void *mem)
 785 {
 786         struct lock_object *lock;
 787         struct witness_order_list_entry *order;
 788         struct witness *w, *w1;
 789         uintptr_t p;
 790         int i;
 791
 792         p = (uintptr_t)mem;
 793         w_data = (void *)p;
 794         p += sizeof(struct witness) * witness_count;
 795
 796         w_rmatrix = (void *)p;
 797         p += sizeof(*w_rmatrix) * (witness_count + 1);
 798
 799         for (i = 0; i < witness_count + 1; i++) {
 800                 w_rmatrix[i] = (void *)p;
 801                 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
 802         }
 803         badstack_sbuf_size = witness_count * 256;
 804
 805         /*
 806          * We have to release Giant before initializing its witness
 807          * structure so that WITNESS doesn't get confused.
 808          */
 809         mtx_unlock(&Giant);
 810         mtx_assert(&Giant, MA_NOTOWNED);
 811
 812         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 813         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 814             MTX_NOWITNESS | MTX_NOPROFILE);
 815         for (i = witness_count - 1; i >= 0; i--) {
 816                 w = &w_data[i];
 817                 memset(w, 0, sizeof(*w));
 818                 w_data[i].w_index = i;  /* Witness index never changes. */
 819                 witness_free(w);
 820         }
 821         KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
 822             ("%s: Invalid list of free witness objects", __func__));
 823
 824         /* Witness with index 0 is not used to aid in debugging. */
 825         STAILQ_REMOVE_HEAD(&w_free, w_list);
 826         w_free_cnt--;
 827
 828         for (i = 0; i < witness_count; i++) {
 829                 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
 830                     (witness_count + 1));
 831         }
 832
 833         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 834                 witness_lock_list_free(&w_locklistdata[i]);
 835         witness_init_hash_tables();
 836
 837         /* First add in all the specified order lists. */
 838         for (order = order_lists; order->w_name != NULL; order++) {
 839                 w = enroll(order->w_name, order->w_class);
 840                 if (w == NULL)
 841                         continue;
 842                 w->w_file = "order list";
 843                 for (order++; order->w_name != NULL; order++) {
 844                         w1 = enroll(order->w_name, order->w_class);
 845                         if (w1 == NULL)
 846                                 continue;
 847                         w1->w_file = "order list";
 848                         itismychild(w, w1);
 849                         w = w1;
 850                 }
 851         }
 852         witness_spin_warn = 1;
 853
 854         /* Iterate through all locks and add them to witness. */
 855         for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
 856                 lock = pending_locks[i].wh_lock;
 857                 KASSERT(lock->lo_flags & LO_WITNESS,
 858                     ("%s: lock %s is on pending list but not LO_WITNESS",
 859                     __func__, lock->lo_name));
 860                 lock->lo_witness = enroll(pending_locks[i].wh_type,
 861                     LOCK_CLASS(lock));
 862         }
 863
 864         /* Mark the witness code as being ready for use. */
 865         witness_cold = 0;
 866
 867         mtx_lock(&Giant);
 868 }
 869
 870 void
 871 witness_init(struct lock_object *lock, const char *type)
 872 {
 873         struct lock_class *class;
 874
 875         /* Various sanity checks. */
 876         class = LOCK_CLASS(lock);
 877         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 878             (class->lc_flags & LC_RECURSABLE) == 0)
 879                 kassert_panic("%s: lock (%s) %s can not be recursable",
 880                     __func__, class->lc_name, lock->lo_name);
 881         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 882             (class->lc_flags & LC_SLEEPABLE) == 0)
 883                 kassert_panic("%s: lock (%s) %s can not be sleepable",
 884                     __func__, class->lc_name, lock->lo_name);
 885         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 886             (class->lc_flags & LC_UPGRADABLE) == 0)
 887                 kassert_panic("%s: lock (%s) %s can not be upgradable",
 888                     __func__, class->lc_name, lock->lo_name);
 889
 890         /*
 891          * If we shouldn't watch this lock, then just clear lo_witness.
 892          * Otherwise, if witness_cold is set, then it is too early to
 893          * enroll this lock, so defer it to witness_initialize() by adding
 894          * it to the pending_locks list.  If it is not too early, then enroll
 895          * the lock now.
 896          */
 897         if (witness_watch < 1 || panicstr != NULL ||
 898             (lock->lo_flags & LO_WITNESS) == 0)
 899                 lock->lo_witness = NULL;
 900         else if (witness_cold) {
 901                 pending_locks[pending_cnt].wh_lock = lock;
 902                 pending_locks[pending_cnt++].wh_type = type;
 903                 if (pending_cnt > WITNESS_PENDLIST)
 904                         panic("%s: pending locks list is too small, "
 905                             "increase WITNESS_PENDLIST\n",
 906                             __func__);
 907         } else
 908                 lock->lo_witness = enroll(type, class);
 909 }
 910
 911 void
 912 witness_destroy(struct lock_object *lock)
 913 {
 914         struct lock_class *class;
 915         struct witness *w;
 916
 917         class = LOCK_CLASS(lock);
 918
 919         if (witness_cold)
 920                 panic("lock (%s) %s destroyed while witness_cold",
 921                     class->lc_name, lock->lo_name);
 922
 923         /* XXX: need to verify that no one holds the lock */
 924         if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
 925                 return;
 926         w = lock->lo_witness;
 927
 928         mtx_lock_spin(&w_mtx);
 929         MPASS(w->w_refcount > 0);
 930         w->w_refcount--;
 931
 932         if (w->w_refcount == 0)
 933                 depart(w);
 934         mtx_unlock_spin(&w_mtx);
 935 }
 936
 937 #ifdef DDB
 938 static void
 939 witness_ddb_compute_levels(void)
 940 {
 941         struct witness *w;
 942
 943         /*
 944          * First clear all levels.
 945          */
 946         STAILQ_FOREACH(w, &w_all, w_list)
 947                 w->w_ddb_level = -1;
 948
 949         /*
 950          * Look for locks with no parents and level all their descendants.
 951          */
 952         STAILQ_FOREACH(w, &w_all, w_list) {
 953
 954                 /* If the witness has ancestors (is not a root), skip it. */
 955                 if (w->w_num_ancestors > 0)
 956                         continue;
 957                 witness_ddb_level_descendants(w, 0);
 958         }
 959 }
 960
 961 static void
 962 witness_ddb_level_descendants(struct witness *w, int l)
 963 {
 964         int i;
 965
 966         if (w->w_ddb_level >= l)
 967                 return;
 968
 969         w->w_ddb_level = l;
 970         l++;
 971
 972         for (i = 1; i <= w_max_used_index; i++) {
 973                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 974                         witness_ddb_level_descendants(&w_data[i], l);
 975         }
 976 }
 977
 978 static void
 979 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
 980     struct witness *w, int indent)
 981 {
 982         int i;
 983
 984         for (i = 0; i < indent; i++)
 985                 prnt(" ");
 986         prnt("%s (type: %s, depth: %d, active refs: %d)",
 987              w->w_name, w->w_class->lc_name,
 988              w->w_ddb_level, w->w_refcount);
 989         if (w->w_displayed) {
 990                 prnt(" -- (already displayed)\n");
 991                 return;
 992         }
 993         w->w_displayed = 1;
 994         if (w->w_file != NULL && w->w_line != 0)
 995                 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
 996                     w->w_line);
 997         else
 998                 prnt(" -- never acquired\n");
 999         indent++;
1000         WITNESS_INDEX_ASSERT(w->w_index);
1001         for (i = 1; i <= w_max_used_index; i++) {
1002                 if (db_pager_quit)
1003                         return;
1004                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1005                         witness_ddb_display_descendants(prnt, &w_data[i],
1006                             indent);
1007         }
1008 }
1009
1010 static void
1011 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1012     struct witness_list *list)
1013 {
1014         struct witness *w;
1015
1016         STAILQ_FOREACH(w, list, w_typelist) {
1017                 if (w->w_file == NULL || w->w_ddb_level > 0)
1018                         continue;
1019
1020                 /* This lock has no anscestors - display its descendants. */
1021                 witness_ddb_display_descendants(prnt, w, 0);
1022                 if (db_pager_quit)
1023                         return;
1024         }
1025 }
1026
1027 static void
1028 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1029 {
1030         struct witness *w;
1031
1032         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1033         witness_ddb_compute_levels();
1034
1035         /* Clear all the displayed flags. */
1036         STAILQ_FOREACH(w, &w_all, w_list)
1037                 w->w_displayed = 0;
1038
1039         /*
1040          * First, handle sleep locks which have been acquired at least
1041          * once.
1042          */
1043         prnt("Sleep locks:\n");
1044         witness_ddb_display_list(prnt, &w_sleep);
1045         if (db_pager_quit)
1046                 return;
1047
1048         /*
1049          * Now do spin locks which have been acquired at least once.
1050          */
1051         prnt("\nSpin locks:\n");
1052         witness_ddb_display_list(prnt, &w_spin);
1053         if (db_pager_quit)
1054                 return;
1055
1056         /*
1057          * Finally, any locks which have not been acquired yet.
1058          */
1059         prnt("\nLocks which were never acquired:\n");
1060         STAILQ_FOREACH(w, &w_all, w_list) {
1061                 if (w->w_file != NULL || w->w_refcount == 0)
1062                         continue;
1063                 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1064                     w->w_class->lc_name, w->w_ddb_level);
1065                 if (db_pager_quit)
1066                         return;
1067         }
1068 }
1069 #endif /* DDB */
1070
1071 int
1072 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1073 {
1074
1075         if (witness_watch == -1 || panicstr != NULL)
1076                 return (0);
1077
1078         /* Require locks that witness knows about. */
1079         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1080             lock2->lo_witness == NULL)
1081                 return (EINVAL);
1082
1083         mtx_assert(&w_mtx, MA_NOTOWNED);
1084         mtx_lock_spin(&w_mtx);
1085
1086         /*
1087          * If we already have either an explicit or implied lock order that
1088          * is the other way around, then return an error.
1089          */
1090         if (witness_watch &&
1091             isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1092                 mtx_unlock_spin(&w_mtx);
1093                 return (EDOOFUS);
1094         }
1095
1096         /* Try to add the new order. */
1097         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1098             lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1099         itismychild(lock1->lo_witness, lock2->lo_witness);
1100         mtx_unlock_spin(&w_mtx);
1101         return (0);
1102 }
1103
1104 void
1105 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1106     int line, struct lock_object *interlock)
1107 {
1108         struct lock_list_entry *lock_list, *lle;
1109         struct lock_instance *lock1, *lock2, *plock;
1110         struct lock_class *class, *iclass;
1111         struct witness *w, *w1;
1112         struct thread *td;
1113         int i, j;
1114
1115         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1116             panicstr != NULL)
1117                 return;
1118
1119         w = lock->lo_witness;
1120         class = LOCK_CLASS(lock);
1121         td = curthread;
1122
1123         if (class->lc_flags & LC_SLEEPLOCK) {
1124
1125                 /*
1126                  * Since spin locks include a critical section, this check
1127                  * implicitly enforces a lock order of all sleep locks before
1128                  * all spin locks.
1129                  */
1130                 if (td->td_critnest != 0 && !kdb_active)
1131                         kassert_panic("acquiring blockable sleep lock with "
1132                             "spinlock or critical section held (%s) %s @ %s:%d",
1133                             class->lc_name, lock->lo_name,
1134                             fixup_filename(file), line);
1135
1136                 /*
1137                  * If this is the first lock acquired then just return as
1138                  * no order checking is needed.
1139                  */
1140                 lock_list = td->td_sleeplocks;
1141                 if (lock_list == NULL || lock_list->ll_count == 0)
1142                         return;
1143         } else {
1144
1145                 /*
1146                  * If this is the first lock, just return as no order
1147                  * checking is needed.  Avoid problems with thread
1148                  * migration pinning the thread while checking if
1149                  * spinlocks are held.  If at least one spinlock is held
1150                  * the thread is in a safe path and it is allowed to
1151                  * unpin it.
1152                  */
1153                 sched_pin();
1154                 lock_list = PCPU_GET(spinlocks);
1155                 if (lock_list == NULL || lock_list->ll_count == 0) {
1156                         sched_unpin();
1157                         return;
1158                 }
1159                 sched_unpin();
1160         }
1161
1162         /*
1163          * Check to see if we are recursing on a lock we already own.  If
1164          * so, make sure that we don't mismatch exclusive and shared lock
1165          * acquires.
1166          */
1167         lock1 = find_instance(lock_list, lock);
1168         if (lock1 != NULL) {
1169                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1170                     (flags & LOP_EXCLUSIVE) == 0) {
1171                         witness_output("shared lock of (%s) %s @ %s:%d\n",
1172                             class->lc_name, lock->lo_name,
1173                             fixup_filename(file), line);
1174                         witness_output("while exclusively locked from %s:%d\n",
1175                             fixup_filename(lock1->li_file), lock1->li_line);
1176                         kassert_panic("excl->share");
1177                 }
1178                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1179                     (flags & LOP_EXCLUSIVE) != 0) {
1180                         witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1181                             class->lc_name, lock->lo_name,
1182                             fixup_filename(file), line);
1183                         witness_output("while share locked from %s:%d\n",
1184                             fixup_filename(lock1->li_file), lock1->li_line);
1185                         kassert_panic("share->excl");
1186                 }
1187                 return;
1188         }
1189
1190         /* Warn if the interlock is not locked exactly once. */
1191         if (interlock != NULL) {
1192                 iclass = LOCK_CLASS(interlock);
1193                 lock1 = find_instance(lock_list, interlock);
1194                 if (lock1 == NULL)
1195                         kassert_panic("interlock (%s) %s not locked @ %s:%d",
1196                             iclass->lc_name, interlock->lo_name,
1197                             fixup_filename(file), line);
1198                 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1199                         kassert_panic("interlock (%s) %s recursed @ %s:%d",
1200                             iclass->lc_name, interlock->lo_name,
1201                             fixup_filename(file), line);
1202         }
1203
1204         /*
1205          * Find the previously acquired lock, but ignore interlocks.
1206          */
1207         plock = &lock_list->ll_children[lock_list->ll_count - 1];
1208         if (interlock != NULL && plock->li_lock == interlock) {
1209                 if (lock_list->ll_count > 1)
1210                         plock =
1211                             &lock_list->ll_children[lock_list->ll_count - 2];
1212                 else {
1213                         lle = lock_list->ll_next;
1214
1215                         /*
1216                          * The interlock is the only lock we hold, so
1217                          * simply return.
1218                          */
1219                         if (lle == NULL)
1220                                 return;
1221                         plock = &lle->ll_children[lle->ll_count - 1];
1222                 }
1223         }
1224
1225         /*
1226          * Try to perform most checks without a lock.  If this succeeds we
1227          * can skip acquiring the lock and return success.  Otherwise we redo
1228          * the check with the lock held to handle races with concurrent updates.
1229          */
1230         w1 = plock->li_lock->lo_witness;
1231         if (witness_lock_order_check(w1, w))
1232                 return;
1233
1234         mtx_lock_spin(&w_mtx);
1235         if (witness_lock_order_check(w1, w)) {
1236                 mtx_unlock_spin(&w_mtx);
1237                 return;
1238         }
1239         witness_lock_order_add(w1, w);
1240
1241         /*
1242          * Check for duplicate locks of the same type.  Note that we only
1243          * have to check for this on the last lock we just acquired.  Any
1244          * other cases will be caught as lock order violations.
1245          */
1246         if (w1 == w) {
1247                 i = w->w_index;
1248                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1249                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1250                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
1251                         w->w_reversed = 1;
1252                         mtx_unlock_spin(&w_mtx);
1253                         witness_output(
1254                             "acquiring duplicate lock of same type: \"%s\"\n",
1255                             w->w_name);
1256                         witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1257                             fixup_filename(plock->li_file), plock->li_line);
1258                         witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1259                             fixup_filename(file), line);
1260                         witness_debugger(1, __func__);
1261                 } else
1262                         mtx_unlock_spin(&w_mtx);
1263                 return;
1264         }
1265         mtx_assert(&w_mtx, MA_OWNED);
1266
1267         /*
1268          * If we know that the lock we are acquiring comes after
1269          * the lock we most recently acquired in the lock order tree,
1270          * then there is no need for any further checks.
1271          */
1272         if (isitmychild(w1, w))
1273                 goto out;
1274
1275         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1276                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1277
1278                         MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1279                         lock1 = &lle->ll_children[i];
1280
1281                         /*
1282                          * Ignore the interlock.
1283                          */
1284                         if (interlock == lock1->li_lock)
1285                                 continue;
1286
1287                         /*
1288                          * If this lock doesn't undergo witness checking,
1289                          * then skip it.
1290                          */
1291                         w1 = lock1->li_lock->lo_witness;
1292                         if (w1 == NULL) {
1293                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1294                                     ("lock missing witness structure"));
1295                                 continue;
1296                         }
1297
1298                         /*
1299                          * If we are locking Giant and this is a sleepable
1300                          * lock, then skip it.
1301                          */
1302                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1303                             lock == &Giant.lock_object)
1304                                 continue;
1305
1306                         /*
1307                          * If we are locking a sleepable lock and this lock
1308                          * is Giant, then skip it.
1309                          */
1310                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1311                             lock1->li_lock == &Giant.lock_object)
1312                                 continue;
1313
1314                         /*
1315                          * If we are locking a sleepable lock and this lock
1316                          * isn't sleepable, we want to treat it as a lock
1317                          * order violation to enfore a general lock order of
1318                          * sleepable locks before non-sleepable locks.
1319                          */
1320                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1321                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1322                                 goto reversal;
1323
1324                         /*
1325                          * If we are locking Giant and this is a non-sleepable
1326                          * lock, then treat it as a reversal.
1327                          */
1328                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1329                             lock == &Giant.lock_object)
1330                                 goto reversal;
1331
1332                         /*
1333                          * Check the lock order hierarchy for a reveresal.
1334                          */
1335                         if (!isitmydescendant(w, w1))
1336                                 continue;
1337                 reversal:
1338
1339                         /*
1340                          * We have a lock order violation, check to see if it
1341                          * is allowed or has already been yelled about.
1342                          */
1343 #ifdef BLESSING
1344
1345                         /*
1346                          * If the lock order is blessed, just bail.  We don't
1347                          * look for other lock order violations though, which
1348                          * may be a bug.
1349                          */
1350                         if (blessed(w, w1))
1351                                 goto out;
1352 #endif
1353
1354                         /* Bail if this violation is known */
1355                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1356                                 goto out;
1357
1358                         /* Record this as a violation */
1359                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1360                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1361                         w->w_reversed = w1->w_reversed = 1;
1362                         witness_increment_graph_generation();
1363                         mtx_unlock_spin(&w_mtx);
1364
1365 #ifdef WITNESS_NO_VNODE
1366                         /*
1367                          * There are known LORs between VNODE locks. They are
1368                          * not an indication of a bug. VNODE locks are flagged
1369                          * as such (LO_IS_VNODE) and we don't yell if the LOR
1370                          * is between 2 VNODE locks.
1371                          */
1372                         if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1373                             (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1374                                 return;
1375 #endif
1376
1377                         /*
1378                          * Ok, yell about it.
1379                          */
1380                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1381                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1382                                 witness_output(
1383                 "lock order reversal: (sleepable after non-sleepable)\n");
1384                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1385                             && lock == &Giant.lock_object)
1386                                 witness_output(
1387                 "lock order reversal: (Giant after non-sleepable)\n");
1388                         else
1389                                 witness_output("lock order reversal:\n");
1390
1391                         /*
1392                          * Try to locate an earlier lock with
1393                          * witness w in our list.
1394                          */
1395                         do {
1396                                 lock2 = &lle->ll_children[i];
1397                                 MPASS(lock2->li_lock != NULL);
1398                                 if (lock2->li_lock->lo_witness == w)
1399                                         break;
1400                                 if (i == 0 && lle->ll_next != NULL) {
1401                                         lle = lle->ll_next;
1402                                         i = lle->ll_count - 1;
1403                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1404                                 } else
1405                                         i--;
1406                         } while (i >= 0);
1407                         if (i < 0) {
1408                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1409                                     lock1->li_lock, lock1->li_lock->lo_name,
1410                                     w1->w_name, fixup_filename(lock1->li_file),
1411                                     lock1->li_line);
1412                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1413                                     lock->lo_name, w->w_name,
1414                                     fixup_filename(file), line);
1415                         } else {
1416                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1417                                     lock2->li_lock, lock2->li_lock->lo_name,
1418                                     lock2->li_lock->lo_witness->w_name,
1419                                     fixup_filename(lock2->li_file),
1420                                     lock2->li_line);
1421                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1422                                     lock1->li_lock, lock1->li_lock->lo_name,
1423                                     w1->w_name, fixup_filename(lock1->li_file),
1424                                     lock1->li_line);
1425                                 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1426                                     lock->lo_name, w->w_name,
1427                                     fixup_filename(file), line);
1428                         }
1429                         witness_debugger(1, __func__);
1430                         return;
1431                 }
1432         }
1433
1434         /*
1435          * If requested, build a new lock order.  However, don't build a new
1436          * relationship between a sleepable lock and Giant if it is in the
1437          * wrong direction.  The correct lock order is that sleepable locks
1438          * always come before Giant.
1439          */
1440         if (flags & LOP_NEWORDER &&
1441             !(plock->li_lock == &Giant.lock_object &&
1442             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1443                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1444                     w->w_name, plock->li_lock->lo_witness->w_name);
1445                 itismychild(plock->li_lock->lo_witness, w);
1446         }
1447 out:
1448         mtx_unlock_spin(&w_mtx);
1449 }
1450
1451 void
1452 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1453 {
1454         struct lock_list_entry **lock_list, *lle;
1455         struct lock_instance *instance;
1456         struct witness *w;
1457         struct thread *td;
1458
1459         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1460             panicstr != NULL)
1461                 return;
1462         w = lock->lo_witness;
1463         td = curthread;
1464
1465         /* Determine lock list for this lock. */
1466         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1467                 lock_list = &td->td_sleeplocks;
1468         else
1469                 lock_list = PCPU_PTR(spinlocks);
1470
1471         /* Check to see if we are recursing on a lock we already own. */
1472         instance = find_instance(*lock_list, lock);
1473         if (instance != NULL) {
1474                 instance->li_flags++;
1475                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1476                     td->td_proc->p_pid, lock->lo_name,
1477                     instance->li_flags & LI_RECURSEMASK);
1478                 instance->li_file = file;
1479                 instance->li_line = line;
1480                 return;
1481         }
1482
1483         /* Update per-witness last file and line acquire. */
1484         w->w_file = file;
1485         w->w_line = line;
1486
1487         /* Find the next open lock instance in the list and fill it. */
1488         lle = *lock_list;
1489         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1490                 lle = witness_lock_list_get();
1491                 if (lle == NULL)
1492                         return;
1493                 lle->ll_next = *lock_list;
1494                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1495                     td->td_proc->p_pid, lle);
1496                 *lock_list = lle;
1497         }
1498         instance = &lle->ll_children[lle->ll_count++];
1499         instance->li_lock = lock;
1500         instance->li_line = line;
1501         instance->li_file = file;
1502         if ((flags & LOP_EXCLUSIVE) != 0)
1503                 instance->li_flags = LI_EXCLUSIVE;
1504         else
1505                 instance->li_flags = 0;
1506         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1507             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1508 }
1509
1510 void
1511 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1512 {
1513         struct lock_instance *instance;
1514         struct lock_class *class;
1515
1516         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1517         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1518                 return;
1519         class = LOCK_CLASS(lock);
1520         if (witness_watch) {
1521                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1522                         kassert_panic(
1523                             "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1524                             class->lc_name, lock->lo_name,
1525                             fixup_filename(file), line);
1526                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1527                         kassert_panic(
1528                             "upgrade of non-sleep lock (%s) %s @ %s:%d",
1529                             class->lc_name, lock->lo_name,
1530                             fixup_filename(file), line);
1531         }
1532         instance = find_instance(curthread->td_sleeplocks, lock);
1533         if (instance == NULL) {
1534                 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1535                     class->lc_name, lock->lo_name,
1536                     fixup_filename(file), line);
1537                 return;
1538         }
1539         if (witness_watch) {
1540                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1541                         kassert_panic(
1542                             "upgrade of exclusive lock (%s) %s @ %s:%d",
1543                             class->lc_name, lock->lo_name,
1544                             fixup_filename(file), line);
1545                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1546                         kassert_panic(
1547                             "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1548                             class->lc_name, lock->lo_name,
1549                             instance->li_flags & LI_RECURSEMASK,
1550                             fixup_filename(file), line);
1551         }
1552         instance->li_flags |= LI_EXCLUSIVE;
1553 }
1554
1555 void
1556 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1557     int line)
1558 {
1559         struct lock_instance *instance;
1560         struct lock_class *class;
1561
1562         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1563         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1564                 return;
1565         class = LOCK_CLASS(lock);
1566         if (witness_watch) {
1567                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1568                         kassert_panic(
1569                             "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1570                             class->lc_name, lock->lo_name,
1571                             fixup_filename(file), line);
1572                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1573                         kassert_panic(
1574                             "downgrade of non-sleep lock (%s) %s @ %s:%d",
1575                             class->lc_name, lock->lo_name,
1576                             fixup_filename(file), line);
1577         }
1578         instance = find_instance(curthread->td_sleeplocks, lock);
1579         if (instance == NULL) {
1580                 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1581                     class->lc_name, lock->lo_name,
1582                     fixup_filename(file), line);
1583                 return;
1584         }
1585         if (witness_watch) {
1586                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1587                         kassert_panic(
1588                             "downgrade of shared lock (%s) %s @ %s:%d",
1589                             class->lc_name, lock->lo_name,
1590                             fixup_filename(file), line);
1591                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1592                         kassert_panic(
1593                             "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1594                             class->lc_name, lock->lo_name,
1595                             instance->li_flags & LI_RECURSEMASK,
1596                             fixup_filename(file), line);
1597         }
1598         instance->li_flags &= ~LI_EXCLUSIVE;
1599 }
1600
1601 void
1602 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1603 {
1604         struct lock_list_entry **lock_list, *lle;
1605         struct lock_instance *instance;
1606         struct lock_class *class;
1607         struct thread *td;
1608         register_t s;
1609         int i, j;
1610
1611         if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1612                 return;
1613         td = curthread;
1614         class = LOCK_CLASS(lock);
1615
1616         /* Find lock instance associated with this lock. */
1617         if (class->lc_flags & LC_SLEEPLOCK)
1618                 lock_list = &td->td_sleeplocks;
1619         else
1620                 lock_list = PCPU_PTR(spinlocks);
1621         lle = *lock_list;
1622         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1623                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1624                         instance = &(*lock_list)->ll_children[i];
1625                         if (instance->li_lock == lock)
1626                                 goto found;
1627                 }
1628
1629         /*
1630          * When disabling WITNESS through witness_watch we could end up in
1631          * having registered locks in the td_sleeplocks queue.
1632          * We have to make sure we flush these queues, so just search for
1633          * eventual register locks and remove them.
1634          */
1635         if (witness_watch > 0) {
1636                 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1637                     lock->lo_name, fixup_filename(file), line);
1638                 return;
1639         } else {
1640                 return;
1641         }
1642 found:
1643
1644         /* First, check for shared/exclusive mismatches. */
1645         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1646             (flags & LOP_EXCLUSIVE) == 0) {
1647                 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1648                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1649                 witness_output("while exclusively locked from %s:%d\n",
1650                     fixup_filename(instance->li_file), instance->li_line);
1651                 kassert_panic("excl->ushare");
1652         }
1653         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1654             (flags & LOP_EXCLUSIVE) != 0) {
1655                 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1656                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1657                 witness_output("while share locked from %s:%d\n",
1658                     fixup_filename(instance->li_file),
1659                     instance->li_line);
1660                 kassert_panic("share->uexcl");
1661         }
1662         /* If we are recursed, unrecurse. */
1663         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1664                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1665                     td->td_proc->p_pid, instance->li_lock->lo_name,
1666                     instance->li_flags);
1667                 instance->li_flags--;
1668                 return;
1669         }
1670         /* The lock is now being dropped, check for NORELEASE flag */
1671         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1672                 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1673                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1674                 kassert_panic("lock marked norelease");
1675         }
1676
1677         /* Otherwise, remove this item from the list. */
1678         s = intr_disable();
1679         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1680             td->td_proc->p_pid, instance->li_lock->lo_name,
1681             (*lock_list)->ll_count - 1);
1682         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1683                 (*lock_list)->ll_children[j] =
1684                     (*lock_list)->ll_children[j + 1];
1685         (*lock_list)->ll_count--;
1686         intr_restore(s);
1687
1688         /*
1689          * In order to reduce contention on w_mtx, we want to keep always an
1690          * head object into lists so that frequent allocation from the
1691          * free witness pool (and subsequent locking) is avoided.
1692          * In order to maintain the current code simple, when the head
1693          * object is totally unloaded it means also that we do not have
1694          * further objects in the list, so the list ownership needs to be
1695          * hand over to another object if the current head needs to be freed.
1696          */
1697         if ((*lock_list)->ll_count == 0) {
1698                 if (*lock_list == lle) {
1699                         if (lle->ll_next == NULL)
1700                                 return;
1701                 } else
1702                         lle = *lock_list;
1703                 *lock_list = lle->ll_next;
1704                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1705                     td->td_proc->p_pid, lle);
1706                 witness_lock_list_free(lle);
1707         }
1708 }
1709
1710 void
1711 witness_thread_exit(struct thread *td)
1712 {
1713         struct lock_list_entry *lle;
1714         int i, n;
1715
1716         lle = td->td_sleeplocks;
1717         if (lle == NULL || panicstr != NULL)
1718                 return;
1719         if (lle->ll_count != 0) {
1720                 for (n = 0; lle != NULL; lle = lle->ll_next)
1721                         for (i = lle->ll_count - 1; i >= 0; i--) {
1722                                 if (n == 0)
1723                                         witness_output(
1724                     "Thread %p exiting with the following locks held:\n", td);
1725                                 n++;
1726                                 witness_list_lock(&lle->ll_children[i],
1727                                     witness_output);
1728
1729                         }
1730                 kassert_panic(
1731                     "Thread %p cannot exit while holding sleeplocks\n", td);
1732         }
1733         witness_lock_list_free(lle);
1734 }
1735
1736 /*
1737  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1738  * exempt Giant and sleepable locks from the checks as well.  If any
1739  * non-exempt locks are held, then a supplied message is printed to the
1740  * output channel along with a list of the offending locks.  If indicated in the
1741  * flags then a failure results in a panic as well.
1742  */
1743 int
1744 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1745 {
1746         struct lock_list_entry *lock_list, *lle;
1747         struct lock_instance *lock1;
1748         struct thread *td;
1749         va_list ap;
1750         int i, n;
1751
1752         if (witness_cold || witness_watch < 1 || panicstr != NULL)
1753                 return (0);
1754         n = 0;
1755         td = curthread;
1756         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1757                 for (i = lle->ll_count - 1; i >= 0; i--) {
1758                         lock1 = &lle->ll_children[i];
1759                         if (lock1->li_lock == lock)
1760                                 continue;
1761                         if (flags & WARN_GIANTOK &&
1762                             lock1->li_lock == &Giant.lock_object)
1763                                 continue;
1764                         if (flags & WARN_SLEEPOK &&
1765                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1766                                 continue;
1767                         if (n == 0) {
1768                                 va_start(ap, fmt);
1769                                 vprintf(fmt, ap);
1770                                 va_end(ap);
1771                                 printf(" with the following %slocks held:\n",
1772                                     (flags & WARN_SLEEPOK) != 0 ?
1773                                     "non-sleepable " : "");
1774                         }
1775                         n++;
1776                         witness_list_lock(lock1, printf);
1777                 }
1778
1779         /*
1780          * Pin the thread in order to avoid problems with thread migration.
1781          * Once that all verifies are passed about spinlocks ownership,
1782          * the thread is in a safe path and it can be unpinned.
1783          */
1784         sched_pin();
1785         lock_list = PCPU_GET(spinlocks);
1786         if (lock_list != NULL && lock_list->ll_count != 0) {
1787                 sched_unpin();
1788
1789                 /*
1790                  * We should only have one spinlock and as long as
1791                  * the flags cannot match for this locks class,
1792                  * check if the first spinlock is the one curthread
1793                  * should hold.
1794                  */
1795                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1796                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1797                     lock1->li_lock == lock && n == 0)
1798                         return (0);
1799
1800                 va_start(ap, fmt);
1801                 vprintf(fmt, ap);
1802                 va_end(ap);
1803                 printf(" with the following %slocks held:\n",
1804                     (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1805                 n += witness_list_locks(&lock_list, printf);
1806         } else
1807                 sched_unpin();
1808         if (flags & WARN_PANIC && n)
1809                 kassert_panic("%s", __func__);
1810         else
1811                 witness_debugger(n, __func__);
1812         return (n);
1813 }
1814
1815 const char *
1816 witness_file(struct lock_object *lock)
1817 {
1818         struct witness *w;
1819
1820         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1821                 return ("?");
1822         w = lock->lo_witness;
1823         return (w->w_file);
1824 }
1825
1826 int
1827 witness_line(struct lock_object *lock)
1828 {
1829         struct witness *w;
1830
1831         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1832                 return (0);
1833         w = lock->lo_witness;
1834         return (w->w_line);
1835 }
1836
1837 static struct witness *
1838 enroll(const char *description, struct lock_class *lock_class)
1839 {
1840         struct witness *w;
1841
1842         MPASS(description != NULL);
1843
1844         if (witness_watch == -1 || panicstr != NULL)
1845                 return (NULL);
1846         if ((lock_class->lc_flags & LC_SPINLOCK)) {
1847                 if (witness_skipspin)
1848                         return (NULL);
1849         } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1850                 kassert_panic("lock class %s is not sleep or spin",
1851                     lock_class->lc_name);
1852                 return (NULL);
1853         }
1854
1855         mtx_lock_spin(&w_mtx);
1856         w = witness_hash_get(description);
1857         if (w)
1858                 goto found;
1859         if ((w = witness_get()) == NULL)
1860                 return (NULL);
1861         MPASS(strlen(description) < MAX_W_NAME);
1862         strcpy(w->w_name, description);
1863         w->w_class = lock_class;
1864         w->w_refcount = 1;
1865         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1866         if (lock_class->lc_flags & LC_SPINLOCK) {
1867                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1868                 w_spin_cnt++;
1869         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1870                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1871                 w_sleep_cnt++;
1872         }
1873
1874         /* Insert new witness into the hash */
1875         witness_hash_put(w);
1876         witness_increment_graph_generation();
1877         mtx_unlock_spin(&w_mtx);
1878         return (w);
1879 found:
1880         w->w_refcount++;
1881         if (w->w_refcount == 1)
1882                 w->w_class = lock_class;
1883         mtx_unlock_spin(&w_mtx);
1884         if (lock_class != w->w_class)
1885                 kassert_panic(
1886                     "lock (%s) %s does not match earlier (%s) lock",
1887                     description, lock_class->lc_name,
1888                     w->w_class->lc_name);
1889         return (w);
1890 }
1891
1892 static void
1893 depart(struct witness *w)
1894 {
1895
1896         MPASS(w->w_refcount == 0);
1897         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1898                 w_sleep_cnt--;
1899         } else {
1900                 w_spin_cnt--;
1901         }
1902         /*
1903          * Set file to NULL as it may point into a loadable module.
1904          */
1905         w->w_file = NULL;
1906         w->w_line = 0;
1907         witness_increment_graph_generation();
1908 }
1909
1910
1911 static void
1912 adopt(struct witness *parent, struct witness *child)
1913 {
1914         int pi, ci, i, j;
1915
1916         if (witness_cold == 0)
1917                 mtx_assert(&w_mtx, MA_OWNED);
1918
1919         /* If the relationship is already known, there's no work to be done. */
1920         if (isitmychild(parent, child))
1921                 return;
1922
1923         /* When the structure of the graph changes, bump up the generation. */
1924         witness_increment_graph_generation();
1925
1926         /*
1927          * The hard part ... create the direct relationship, then propagate all
1928          * indirect relationships.
1929          */
1930         pi = parent->w_index;
1931         ci = child->w_index;
1932         WITNESS_INDEX_ASSERT(pi);
1933         WITNESS_INDEX_ASSERT(ci);
1934         MPASS(pi != ci);
1935         w_rmatrix[pi][ci] |= WITNESS_PARENT;
1936         w_rmatrix[ci][pi] |= WITNESS_CHILD;
1937
1938         /*
1939          * If parent was not already an ancestor of child,
1940          * then we increment the descendant and ancestor counters.
1941          */
1942         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1943                 parent->w_num_descendants++;
1944                 child->w_num_ancestors++;
1945         }
1946
1947         /*
1948          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1949          * an ancestor of 'pi' during this loop.
1950          */
1951         for (i = 1; i <= w_max_used_index; i++) {
1952                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1953                     (i != pi))
1954                         continue;
1955
1956                 /* Find each descendant of 'i' and mark it as a descendant. */
1957                 for (j = 1; j <= w_max_used_index; j++) {
1958
1959                         /*
1960                          * Skip children that are already marked as
1961                          * descendants of 'i'.
1962                          */
1963                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1964                                 continue;
1965
1966                         /*
1967                          * We are only interested in descendants of 'ci'. Note
1968                          * that 'ci' itself is counted as a descendant of 'ci'.
1969                          */
1970                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1971                             (j != ci))
1972                                 continue;
1973                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1974                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1975                         w_data[i].w_num_descendants++;
1976                         w_data[j].w_num_ancestors++;
1977
1978                         /*
1979                          * Make sure we aren't marking a node as both an
1980                          * ancestor and descendant. We should have caught
1981                          * this as a lock order reversal earlier.
1982                          */
1983                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1984                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1985                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1986                                     "both ancestor and descendant\n",
1987                                     i, j, w_rmatrix[i][j]);
1988                                 kdb_backtrace();
1989                                 printf("Witness disabled.\n");
1990                                 witness_watch = -1;
1991                         }
1992                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1993                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1994                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1995                                     "both ancestor and descendant\n",
1996                                     j, i, w_rmatrix[j][i]);
1997                                 kdb_backtrace();
1998                                 printf("Witness disabled.\n");
1999                                 witness_watch = -1;
2000                         }
2001                 }
2002         }
2003 }
2004
2005 static void
2006 itismychild(struct witness *parent, struct witness *child)
2007 {
2008         int unlocked;
2009
2010         MPASS(child != NULL && parent != NULL);
2011         if (witness_cold == 0)
2012                 mtx_assert(&w_mtx, MA_OWNED);
2013
2014         if (!witness_lock_type_equal(parent, child)) {
2015                 if (witness_cold == 0) {
2016                         unlocked = 1;
2017                         mtx_unlock_spin(&w_mtx);
2018                 } else {
2019                         unlocked = 0;
2020                 }
2021                 kassert_panic(
2022                     "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2023                     "the same lock type", __func__, parent->w_name,
2024                     parent->w_class->lc_name, child->w_name,
2025                     child->w_class->lc_name);
2026                 if (unlocked)
2027                         mtx_lock_spin(&w_mtx);
2028         }
2029         adopt(parent, child);
2030 }
2031
2032 /*
2033  * Generic code for the isitmy*() functions. The rmask parameter is the
2034  * expected relationship of w1 to w2.
2035  */
2036 static int
2037 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2038 {
2039         unsigned char r1, r2;
2040         int i1, i2;
2041
2042         i1 = w1->w_index;
2043         i2 = w2->w_index;
2044         WITNESS_INDEX_ASSERT(i1);
2045         WITNESS_INDEX_ASSERT(i2);
2046         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2047         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2048
2049         /* The flags on one better be the inverse of the flags on the other */
2050         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2051             (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2052                 /* Don't squawk if we're potentially racing with an update. */
2053                 if (!mtx_owned(&w_mtx))
2054                         return (0);
2055                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2056                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
2057                     "w_rmatrix[%d][%d] == %hhx\n",
2058                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2059                     i2, i1, r2);
2060                 kdb_backtrace();
2061                 printf("Witness disabled.\n");
2062                 witness_watch = -1;
2063         }
2064         return (r1 & rmask);
2065 }
2066
2067 /*
2068  * Checks if @child is a direct child of @parent.
2069  */
2070 static int
2071 isitmychild(struct witness *parent, struct witness *child)
2072 {
2073
2074         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2075 }
2076
2077 /*
2078  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2079  */
2080 static int
2081 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2082 {
2083
2084         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2085             __func__));
2086 }
2087
2088 #ifdef BLESSING
2089 static int
2090 blessed(struct witness *w1, struct witness *w2)
2091 {
2092         int i;
2093         struct witness_blessed *b;
2094
2095         for (i = 0; i < nitems(blessed_list); i++) {
2096                 b = &blessed_list[i];
2097                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2098                         if (strcmp(w2->w_name, b->b_lock2) == 0)
2099                                 return (1);
2100                         continue;
2101                 }
2102                 if (strcmp(w1->w_name, b->b_lock2) == 0)
2103                         if (strcmp(w2->w_name, b->b_lock1) == 0)
2104                                 return (1);
2105         }
2106         return (0);
2107 }
2108 #endif
2109
2110 static struct witness *
2111 witness_get(void)
2112 {
2113         struct witness *w;
2114         int index;
2115
2116         if (witness_cold == 0)
2117                 mtx_assert(&w_mtx, MA_OWNED);
2118
2119         if (witness_watch == -1) {
2120                 mtx_unlock_spin(&w_mtx);
2121                 return (NULL);
2122         }
2123         if (STAILQ_EMPTY(&w_free)) {
2124                 witness_watch = -1;
2125                 mtx_unlock_spin(&w_mtx);
2126                 printf("WITNESS: unable to allocate a new witness object\n");
2127                 return (NULL);
2128         }
2129         w = STAILQ_FIRST(&w_free);
2130         STAILQ_REMOVE_HEAD(&w_free, w_list);
2131         w_free_cnt--;
2132         index = w->w_index;
2133         MPASS(index > 0 && index == w_max_used_index+1 &&
2134             index < witness_count);
2135         bzero(w, sizeof(*w));
2136         w->w_index = index;
2137         if (index > w_max_used_index)
2138                 w_max_used_index = index;
2139         return (w);
2140 }
2141
2142 static void
2143 witness_free(struct witness *w)
2144 {
2145
2146         STAILQ_INSERT_HEAD(&w_free, w, w_list);
2147         w_free_cnt++;
2148 }
2149
2150 static struct lock_list_entry *
2151 witness_lock_list_get(void)
2152 {
2153         struct lock_list_entry *lle;
2154
2155         if (witness_watch == -1)
2156                 return (NULL);
2157         mtx_lock_spin(&w_mtx);
2158         lle = w_lock_list_free;
2159         if (lle == NULL) {
2160                 witness_watch = -1;
2161                 mtx_unlock_spin(&w_mtx);
2162                 printf("%s: witness exhausted\n", __func__);
2163                 return (NULL);
2164         }
2165         w_lock_list_free = lle->ll_next;
2166         mtx_unlock_spin(&w_mtx);
2167         bzero(lle, sizeof(*lle));
2168         return (lle);
2169 }
2170
2171 static void
2172 witness_lock_list_free(struct lock_list_entry *lle)
2173 {
2174
2175         mtx_lock_spin(&w_mtx);
2176         lle->ll_next = w_lock_list_free;
2177         w_lock_list_free = lle;
2178         mtx_unlock_spin(&w_mtx);
2179 }
2180
2181 static struct lock_instance *
2182 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2183 {
2184         struct lock_list_entry *lle;
2185         struct lock_instance *instance;
2186         int i;
2187
2188         for (lle = list; lle != NULL; lle = lle->ll_next)
2189                 for (i = lle->ll_count - 1; i >= 0; i--) {
2190                         instance = &lle->ll_children[i];
2191                         if (instance->li_lock == lock)
2192                                 return (instance);
2193                 }
2194         return (NULL);
2195 }
2196
2197 static void
2198 witness_list_lock(struct lock_instance *instance,
2199     int (*prnt)(const char *fmt, ...))
2200 {
2201         struct lock_object *lock;
2202
2203         lock = instance->li_lock;
2204         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2205             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2206         if (lock->lo_witness->w_name != lock->lo_name)
2207                 prnt(" (%s)", lock->lo_witness->w_name);
2208         prnt(" r = %d (%p) locked @ %s:%d\n",
2209             instance->li_flags & LI_RECURSEMASK, lock,
2210             fixup_filename(instance->li_file), instance->li_line);
2211 }
2212
2213 static int
2214 witness_output(const char *fmt, ...)
2215 {
2216         va_list ap;
2217         int ret;
2218
2219         va_start(ap, fmt);
2220         ret = witness_voutput(fmt, ap);
2221         va_end(ap);
2222         return (ret);
2223 }
2224
2225 static int
2226 witness_voutput(const char *fmt, va_list ap)
2227 {
2228         int ret;
2229
2230         ret = 0;
2231         switch (witness_channel) {
2232         case WITNESS_CONSOLE:
2233                 ret = vprintf(fmt, ap);
2234                 break;
2235         case WITNESS_LOG:
2236                 vlog(LOG_NOTICE, fmt, ap);
2237                 break;
2238         case WITNESS_NONE:
2239                 break;
2240         }
2241         return (ret);
2242 }
2243
2244 #ifdef DDB
2245 static int
2246 witness_thread_has_locks(struct thread *td)
2247 {
2248
2249         if (td->td_sleeplocks == NULL)
2250                 return (0);
2251         return (td->td_sleeplocks->ll_count != 0);
2252 }
2253
2254 static int
2255 witness_proc_has_locks(struct proc *p)
2256 {
2257         struct thread *td;
2258
2259         FOREACH_THREAD_IN_PROC(p, td) {
2260                 if (witness_thread_has_locks(td))
2261                         return (1);
2262         }
2263         return (0);
2264 }
2265 #endif
2266
2267 int
2268 witness_list_locks(struct lock_list_entry **lock_list,
2269     int (*prnt)(const char *fmt, ...))
2270 {
2271         struct lock_list_entry *lle;
2272         int i, nheld;
2273
2274         nheld = 0;
2275         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2276                 for (i = lle->ll_count - 1; i >= 0; i--) {
2277                         witness_list_lock(&lle->ll_children[i], prnt);
2278                         nheld++;
2279                 }
2280         return (nheld);
2281 }
2282
2283 /*
2284  * This is a bit risky at best.  We call this function when we have timed
2285  * out acquiring a spin lock, and we assume that the other CPU is stuck
2286  * with this lock held.  So, we go groveling around in the other CPU's
2287  * per-cpu data to try to find the lock instance for this spin lock to
2288  * see when it was last acquired.
2289  */
2290 void
2291 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2292     int (*prnt)(const char *fmt, ...))
2293 {
2294         struct lock_instance *instance;
2295         struct pcpu *pc;
2296
2297         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2298                 return;
2299         pc = pcpu_find(owner->td_oncpu);
2300         instance = find_instance(pc->pc_spinlocks, lock);
2301         if (instance != NULL)
2302                 witness_list_lock(instance, prnt);
2303 }
2304
2305 void
2306 witness_save(struct lock_object *lock, const char **filep, int *linep)
2307 {
2308         struct lock_list_entry *lock_list;
2309         struct lock_instance *instance;
2310         struct lock_class *class;
2311
2312         /*
2313          * This function is used independently in locking code to deal with
2314          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2315          * is gone.
2316          */
2317         if (SCHEDULER_STOPPED())
2318                 return;
2319         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2320         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2321                 return;
2322         class = LOCK_CLASS(lock);
2323         if (class->lc_flags & LC_SLEEPLOCK)
2324                 lock_list = curthread->td_sleeplocks;
2325         else {
2326                 if (witness_skipspin)
2327                         return;
2328                 lock_list = PCPU_GET(spinlocks);
2329         }
2330         instance = find_instance(lock_list, lock);
2331         if (instance == NULL) {
2332                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2333                     class->lc_name, lock->lo_name);
2334                 return;
2335         }
2336         *filep = instance->li_file;
2337         *linep = instance->li_line;
2338 }
2339
2340 void
2341 witness_restore(struct lock_object *lock, const char *file, int line)
2342 {
2343         struct lock_list_entry *lock_list;
2344         struct lock_instance *instance;
2345         struct lock_class *class;
2346
2347         /*
2348          * This function is used independently in locking code to deal with
2349          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2350          * is gone.
2351          */
2352         if (SCHEDULER_STOPPED())
2353                 return;
2354         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2355         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2356                 return;
2357         class = LOCK_CLASS(lock);
2358         if (class->lc_flags & LC_SLEEPLOCK)
2359                 lock_list = curthread->td_sleeplocks;
2360         else {
2361                 if (witness_skipspin)
2362                         return;
2363                 lock_list = PCPU_GET(spinlocks);
2364         }
2365         instance = find_instance(lock_list, lock);
2366         if (instance == NULL)
2367                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2368                     class->lc_name, lock->lo_name);
2369         lock->lo_witness->w_file = file;
2370         lock->lo_witness->w_line = line;
2371         if (instance == NULL)
2372                 return;
2373         instance->li_file = file;
2374         instance->li_line = line;
2375 }
2376
2377 void
2378 witness_assert(const struct lock_object *lock, int flags, const char *file,
2379     int line)
2380 {
2381 #ifdef INVARIANT_SUPPORT
2382         struct lock_instance *instance;
2383         struct lock_class *class;
2384
2385         if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2386                 return;
2387         class = LOCK_CLASS(lock);
2388         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2389                 instance = find_instance(curthread->td_sleeplocks, lock);
2390         else if ((class->lc_flags & LC_SPINLOCK) != 0)
2391                 instance = find_instance(PCPU_GET(spinlocks), lock);
2392         else {
2393                 kassert_panic("Lock (%s) %s is not sleep or spin!",
2394                     class->lc_name, lock->lo_name);
2395                 return;
2396         }
2397         switch (flags) {
2398         case LA_UNLOCKED:
2399                 if (instance != NULL)
2400                         kassert_panic("Lock (%s) %s locked @ %s:%d.",
2401                             class->lc_name, lock->lo_name,
2402                             fixup_filename(file), line);
2403                 break;
2404         case LA_LOCKED:
2405         case LA_LOCKED | LA_RECURSED:
2406         case LA_LOCKED | LA_NOTRECURSED:
2407         case LA_SLOCKED:
2408         case LA_SLOCKED | LA_RECURSED:
2409         case LA_SLOCKED | LA_NOTRECURSED:
2410         case LA_XLOCKED:
2411         case LA_XLOCKED | LA_RECURSED:
2412         case LA_XLOCKED | LA_NOTRECURSED:
2413                 if (instance == NULL) {
2414                         kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2415                             class->lc_name, lock->lo_name,
2416                             fixup_filename(file), line);
2417                         break;
2418                 }
2419                 if ((flags & LA_XLOCKED) != 0 &&
2420                     (instance->li_flags & LI_EXCLUSIVE) == 0)
2421                         kassert_panic(
2422                             "Lock (%s) %s not exclusively locked @ %s:%d.",
2423                             class->lc_name, lock->lo_name,
2424                             fixup_filename(file), line);
2425                 if ((flags & LA_SLOCKED) != 0 &&
2426                     (instance->li_flags & LI_EXCLUSIVE) != 0)
2427                         kassert_panic(
2428                             "Lock (%s) %s exclusively locked @ %s:%d.",
2429                             class->lc_name, lock->lo_name,
2430                             fixup_filename(file), line);
2431                 if ((flags & LA_RECURSED) != 0 &&
2432                     (instance->li_flags & LI_RECURSEMASK) == 0)
2433                         kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2434                             class->lc_name, lock->lo_name,
2435                             fixup_filename(file), line);
2436                 if ((flags & LA_NOTRECURSED) != 0 &&
2437                     (instance->li_flags & LI_RECURSEMASK) != 0)
2438                         kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2439                             class->lc_name, lock->lo_name,
2440                             fixup_filename(file), line);
2441                 break;
2442         default:
2443                 kassert_panic("Invalid lock assertion at %s:%d.",
2444                     fixup_filename(file), line);
2445
2446         }
2447 #endif  /* INVARIANT_SUPPORT */
2448 }
2449
2450 static void
2451 witness_setflag(struct lock_object *lock, int flag, int set)
2452 {
2453         struct lock_list_entry *lock_list;
2454         struct lock_instance *instance;
2455         struct lock_class *class;
2456
2457         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2458                 return;
2459         class = LOCK_CLASS(lock);
2460         if (class->lc_flags & LC_SLEEPLOCK)
2461                 lock_list = curthread->td_sleeplocks;
2462         else {
2463                 if (witness_skipspin)
2464                         return;
2465                 lock_list = PCPU_GET(spinlocks);
2466         }
2467         instance = find_instance(lock_list, lock);
2468         if (instance == NULL) {
2469                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2470                     class->lc_name, lock->lo_name);
2471                 return;
2472         }
2473
2474         if (set)
2475                 instance->li_flags |= flag;
2476         else
2477                 instance->li_flags &= ~flag;
2478 }
2479
2480 void
2481 witness_norelease(struct lock_object *lock)
2482 {
2483
2484         witness_setflag(lock, LI_NORELEASE, 1);
2485 }
2486
2487 void
2488 witness_releaseok(struct lock_object *lock)
2489 {
2490
2491         witness_setflag(lock, LI_NORELEASE, 0);
2492 }
2493
2494 #ifdef DDB
2495 static void
2496 witness_ddb_list(struct thread *td)
2497 {
2498
2499         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2500         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2501
2502         if (witness_watch < 1)
2503                 return;
2504
2505         witness_list_locks(&td->td_sleeplocks, db_printf);
2506
2507         /*
2508          * We only handle spinlocks if td == curthread.  This is somewhat broken
2509          * if td is currently executing on some other CPU and holds spin locks
2510          * as we won't display those locks.  If we had a MI way of getting
2511          * the per-cpu data for a given cpu then we could use
2512          * td->td_oncpu to get the list of spinlocks for this thread
2513          * and "fix" this.
2514          *
2515          * That still wouldn't really fix this unless we locked the scheduler
2516          * lock or stopped the other CPU to make sure it wasn't changing the
2517          * list out from under us.  It is probably best to just not try to
2518          * handle threads on other CPU's for now.
2519          */
2520         if (td == curthread && PCPU_GET(spinlocks) != NULL)
2521                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2522 }
2523
2524 DB_SHOW_COMMAND(locks, db_witness_list)
2525 {
2526         struct thread *td;
2527
2528         if (have_addr)
2529                 td = db_lookup_thread(addr, true);
2530         else
2531                 td = kdb_thread;
2532         witness_ddb_list(td);
2533 }
2534
2535 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2536 {
2537         struct thread *td;
2538         struct proc *p;
2539
2540         /*
2541          * It would be nice to list only threads and processes that actually
2542          * held sleep locks, but that information is currently not exported
2543          * by WITNESS.
2544          */
2545         FOREACH_PROC_IN_SYSTEM(p) {
2546                 if (!witness_proc_has_locks(p))
2547                         continue;
2548                 FOREACH_THREAD_IN_PROC(p, td) {
2549                         if (!witness_thread_has_locks(td))
2550                                 continue;
2551                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2552                             p->p_comm, td, td->td_tid);
2553                         witness_ddb_list(td);
2554                         if (db_pager_quit)
2555                                 return;
2556                 }
2557         }
2558 }
2559 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2560
2561 DB_SHOW_COMMAND(witness, db_witness_display)
2562 {
2563
2564         witness_ddb_display(db_printf);
2565 }
2566 #endif
2567
2568 static void
2569 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2570 {
2571         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2572         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2573         int generation, i, j;
2574
2575         tmp_data1 = NULL;
2576         tmp_data2 = NULL;
2577         tmp_w1 = NULL;
2578         tmp_w2 = NULL;
2579
2580         /* Allocate and init temporary storage space. */
2581         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2582         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2583         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2584             M_WAITOK | M_ZERO);
2585         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2586             M_WAITOK | M_ZERO);
2587         stack_zero(&tmp_data1->wlod_stack);
2588         stack_zero(&tmp_data2->wlod_stack);
2589
2590 restart:
2591         mtx_lock_spin(&w_mtx);
2592         generation = w_generation;
2593         mtx_unlock_spin(&w_mtx);
2594         sbuf_printf(sb, "Number of known direct relationships is %d\n",
2595             w_lohash.wloh_count);
2596         for (i = 1; i < w_max_used_index; i++) {
2597                 mtx_lock_spin(&w_mtx);
2598                 if (generation != w_generation) {
2599                         mtx_unlock_spin(&w_mtx);
2600
2601                         /* The graph has changed, try again. */
2602                         *oldidx = 0;
2603                         sbuf_clear(sb);
2604                         goto restart;
2605                 }
2606
2607                 w1 = &w_data[i];
2608                 if (w1->w_reversed == 0) {
2609                         mtx_unlock_spin(&w_mtx);
2610                         continue;
2611                 }
2612
2613                 /* Copy w1 locally so we can release the spin lock. */
2614                 *tmp_w1 = *w1;
2615                 mtx_unlock_spin(&w_mtx);
2616
2617                 if (tmp_w1->w_reversed == 0)
2618                         continue;
2619                 for (j = 1; j < w_max_used_index; j++) {
2620                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2621                                 continue;
2622
2623                         mtx_lock_spin(&w_mtx);
2624                         if (generation != w_generation) {
2625                                 mtx_unlock_spin(&w_mtx);
2626
2627                                 /* The graph has changed, try again. */
2628                                 *oldidx = 0;
2629                                 sbuf_clear(sb);
2630                                 goto restart;
2631                         }
2632
2633                         w2 = &w_data[j];
2634                         data1 = witness_lock_order_get(w1, w2);
2635                         data2 = witness_lock_order_get(w2, w1);
2636
2637                         /*
2638                          * Copy information locally so we can release the
2639                          * spin lock.
2640                          */
2641                         *tmp_w2 = *w2;
2642
2643                         if (data1) {
2644                                 stack_zero(&tmp_data1->wlod_stack);
2645                                 stack_copy(&data1->wlod_stack,
2646                                     &tmp_data1->wlod_stack);
2647                         }
2648                         if (data2 && data2 != data1) {
2649                                 stack_zero(&tmp_data2->wlod_stack);
2650                                 stack_copy(&data2->wlod_stack,
2651                                     &tmp_data2->wlod_stack);
2652                         }
2653                         mtx_unlock_spin(&w_mtx);
2654
2655                         sbuf_printf(sb,
2656             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2657                             tmp_w1->w_name, tmp_w1->w_class->lc_name,
2658                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
2659                         if (data1) {
2660                                 sbuf_printf(sb,
2661                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2662                                     tmp_w1->w_name, tmp_w1->w_class->lc_name,
2663                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
2664                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2665                                 sbuf_printf(sb, "\n");
2666                         }
2667                         if (data2 && data2 != data1) {
2668                                 sbuf_printf(sb,
2669                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2670                                     tmp_w2->w_name, tmp_w2->w_class->lc_name,
2671                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
2672                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2673                                 sbuf_printf(sb, "\n");
2674                         }
2675                 }
2676         }
2677         mtx_lock_spin(&w_mtx);
2678         if (generation != w_generation) {
2679                 mtx_unlock_spin(&w_mtx);
2680
2681                 /*
2682                  * The graph changed while we were printing stack data,
2683                  * try again.
2684                  */
2685                 *oldidx = 0;
2686                 sbuf_clear(sb);
2687                 goto restart;
2688         }
2689         mtx_unlock_spin(&w_mtx);
2690
2691         /* Free temporary storage space. */
2692         free(tmp_data1, M_TEMP);
2693         free(tmp_data2, M_TEMP);
2694         free(tmp_w1, M_TEMP);
2695         free(tmp_w2, M_TEMP);
2696 }
2697
2698 static int
2699 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2700 {
2701         struct sbuf *sb;
2702         int error;
2703
2704         if (witness_watch < 1) {
2705                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2706                 return (error);
2707         }
2708         if (witness_cold) {
2709                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2710                 return (error);
2711         }
2712         error = 0;
2713         sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2714         if (sb == NULL)
2715                 return (ENOMEM);
2716
2717         sbuf_print_witness_badstacks(sb, &req->oldidx);
2718
2719         sbuf_finish(sb);
2720         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2721         sbuf_delete(sb);
2722
2723         return (error);
2724 }
2725
2726 #ifdef DDB
2727 static int
2728 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2729 {
2730
2731         return (db_printf("%.*s", len, data));
2732 }
2733
2734 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2735 {
2736         struct sbuf sb;
2737         char buffer[128];
2738         size_t dummy;
2739
2740         sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2741         sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2742         sbuf_print_witness_badstacks(&sb, &dummy);
2743         sbuf_finish(&sb);
2744 }
2745 #endif
2746
2747 static int
2748 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2749 {
2750         static const struct {
2751                 enum witness_channel channel;
2752                 const char *name;
2753         } channels[] = {
2754                 { WITNESS_CONSOLE, "console" },
2755                 { WITNESS_LOG, "log" },
2756                 { WITNESS_NONE, "none" },
2757         };
2758         char buf[16];
2759         u_int i;
2760         int error;
2761
2762         buf[0] = '\0';
2763         for (i = 0; i < nitems(channels); i++)
2764                 if (witness_channel == channels[i].channel) {
2765                         snprintf(buf, sizeof(buf), "%s", channels[i].name);
2766                         break;
2767                 }
2768
2769         error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2770         if (error != 0 || req->newptr == NULL)
2771                 return (error);
2772
2773         error = EINVAL;
2774         for (i = 0; i < nitems(channels); i++)
2775                 if (strcmp(channels[i].name, buf) == 0) {
2776                         witness_channel = channels[i].channel;
2777                         error = 0;
2778                         break;
2779                 }
2780         return (error);
2781 }
2782
2783 static int
2784 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2785 {
2786         struct witness *w;
2787         struct sbuf *sb;
2788         int error;
2789
2790 #ifdef __i386__
2791         error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2792         return (error);
2793 #endif
2794
2795         if (witness_watch < 1) {
2796                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2797                 return (error);
2798         }
2799         if (witness_cold) {
2800                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2801                 return (error);
2802         }
2803         error = 0;
2804
2805         error = sysctl_wire_old_buffer(req, 0);
2806         if (error != 0)
2807                 return (error);
2808         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2809         if (sb == NULL)
2810                 return (ENOMEM);
2811         sbuf_printf(sb, "\n");
2812
2813         mtx_lock_spin(&w_mtx);
2814         STAILQ_FOREACH(w, &w_all, w_list)
2815                 w->w_displayed = 0;
2816         STAILQ_FOREACH(w, &w_all, w_list)
2817                 witness_add_fullgraph(sb, w);
2818         mtx_unlock_spin(&w_mtx);
2819
2820         /*
2821          * Close the sbuf and return to userland.
2822          */
2823         error = sbuf_finish(sb);
2824         sbuf_delete(sb);
2825
2826         return (error);
2827 }
2828
2829 static int
2830 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2831 {
2832         int error, value;
2833
2834         value = witness_watch;
2835         error = sysctl_handle_int(oidp, &value, 0, req);
2836         if (error != 0 || req->newptr == NULL)
2837                 return (error);
2838         if (value > 1 || value < -1 ||
2839             (witness_watch == -1 && value != witness_watch))
2840                 return (EINVAL);
2841         witness_watch = value;
2842         return (0);
2843 }
2844
2845 static void
2846 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2847 {
2848         int i;
2849
2850         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2851                 return;
2852         w->w_displayed = 1;
2853
2854         WITNESS_INDEX_ASSERT(w->w_index);
2855         for (i = 1; i <= w_max_used_index; i++) {
2856                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2857                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2858                             w_data[i].w_name);
2859                         witness_add_fullgraph(sb, &w_data[i]);
2860                 }
2861         }
2862 }
2863
2864 /*
2865  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2866  * interprets the key as a string and reads until the null
2867  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2868  * hash value computed from the key.
2869  */
2870 static uint32_t
2871 witness_hash_djb2(const uint8_t *key, uint32_t size)
2872 {
2873         unsigned int hash = 5381;
2874         int i;
2875
2876         /* hash = hash * 33 + key[i] */
2877         if (size)
2878                 for (i = 0; i < size; i++)
2879                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2880         else
2881                 for (i = 0; key[i] != 0; i++)
2882                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2883
2884         return (hash);
2885 }
2886
2887
2888 /*
2889  * Initializes the two witness hash tables. Called exactly once from
2890  * witness_initialize().
2891  */
2892 static void
2893 witness_init_hash_tables(void)
2894 {
2895         int i;
2896
2897         MPASS(witness_cold);
2898
2899         /* Initialize the hash tables. */
2900         for (i = 0; i < WITNESS_HASH_SIZE; i++)
2901                 w_hash.wh_array[i] = NULL;
2902
2903         w_hash.wh_size = WITNESS_HASH_SIZE;
2904         w_hash.wh_count = 0;
2905
2906         /* Initialize the lock order data hash. */
2907         w_lofree = NULL;
2908         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2909                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2910                 w_lodata[i].wlod_next = w_lofree;
2911                 w_lofree = &w_lodata[i];
2912         }
2913         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2914         w_lohash.wloh_count = 0;
2915         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2916                 w_lohash.wloh_array[i] = NULL;
2917 }
2918
2919 static struct witness *
2920 witness_hash_get(const char *key)
2921 {
2922         struct witness *w;
2923         uint32_t hash;
2924
2925         MPASS(key != NULL);
2926         if (witness_cold == 0)
2927                 mtx_assert(&w_mtx, MA_OWNED);
2928         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2929         w = w_hash.wh_array[hash];
2930         while (w != NULL) {
2931                 if (strcmp(w->w_name, key) == 0)
2932                         goto out;
2933                 w = w->w_hash_next;
2934         }
2935
2936 out:
2937         return (w);
2938 }
2939
2940 static void
2941 witness_hash_put(struct witness *w)
2942 {
2943         uint32_t hash;
2944
2945         MPASS(w != NULL);
2946         MPASS(w->w_name != NULL);
2947         if (witness_cold == 0)
2948                 mtx_assert(&w_mtx, MA_OWNED);
2949         KASSERT(witness_hash_get(w->w_name) == NULL,
2950             ("%s: trying to add a hash entry that already exists!", __func__));
2951         KASSERT(w->w_hash_next == NULL,
2952             ("%s: w->w_hash_next != NULL", __func__));
2953
2954         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2955         w->w_hash_next = w_hash.wh_array[hash];
2956         w_hash.wh_array[hash] = w;
2957         w_hash.wh_count++;
2958 }
2959
2960
2961 static struct witness_lock_order_data *
2962 witness_lock_order_get(struct witness *parent, struct witness *child)
2963 {
2964         struct witness_lock_order_data *data = NULL;
2965         struct witness_lock_order_key key;
2966         unsigned int hash;
2967
2968         MPASS(parent != NULL && child != NULL);
2969         key.from = parent->w_index;
2970         key.to = child->w_index;
2971         WITNESS_INDEX_ASSERT(key.from);
2972         WITNESS_INDEX_ASSERT(key.to);
2973         if ((w_rmatrix[parent->w_index][child->w_index]
2974             & WITNESS_LOCK_ORDER_KNOWN) == 0)
2975                 goto out;
2976
2977         hash = witness_hash_djb2((const char*)&key,
2978             sizeof(key)) % w_lohash.wloh_size;
2979         data = w_lohash.wloh_array[hash];
2980         while (data != NULL) {
2981                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2982                         break;
2983                 data = data->wlod_next;
2984         }
2985
2986 out:
2987         return (data);
2988 }
2989
2990 /*
2991  * Verify that parent and child have a known relationship, are not the same,
2992  * and child is actually a child of parent.  This is done without w_mtx
2993  * to avoid contention in the common case.
2994  */
2995 static int
2996 witness_lock_order_check(struct witness *parent, struct witness *child)
2997 {
2998
2999         if (parent != child &&
3000             w_rmatrix[parent->w_index][child->w_index]
3001             & WITNESS_LOCK_ORDER_KNOWN &&
3002             isitmychild(parent, child))
3003                 return (1);
3004
3005         return (0);
3006 }
3007
3008 static int
3009 witness_lock_order_add(struct witness *parent, struct witness *child)
3010 {
3011         struct witness_lock_order_data *data = NULL;
3012         struct witness_lock_order_key key;
3013         unsigned int hash;
3014
3015         MPASS(parent != NULL && child != NULL);
3016         key.from = parent->w_index;
3017         key.to = child->w_index;
3018         WITNESS_INDEX_ASSERT(key.from);
3019         WITNESS_INDEX_ASSERT(key.to);
3020         if (w_rmatrix[parent->w_index][child->w_index]
3021             & WITNESS_LOCK_ORDER_KNOWN)
3022                 return (1);
3023
3024         hash = witness_hash_djb2((const char*)&key,
3025             sizeof(key)) % w_lohash.wloh_size;
3026         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3027         data = w_lofree;
3028         if (data == NULL)
3029                 return (0);
3030         w_lofree = data->wlod_next;
3031         data->wlod_next = w_lohash.wloh_array[hash];
3032         data->wlod_key = key;
3033         w_lohash.wloh_array[hash] = data;
3034         w_lohash.wloh_count++;
3035         stack_zero(&data->wlod_stack);
3036         stack_save(&data->wlod_stack);
3037         return (1);
3038 }
3039
3040 /* Call this whenever the structure of the witness graph changes. */
3041 static void
3042 witness_increment_graph_generation(void)
3043 {
3044
3045         if (witness_cold == 0)
3046                 mtx_assert(&w_mtx, MA_OWNED);
3047         w_generation++;
3048 }
3049
3050 static int
3051 witness_output_drain(void *arg __unused, const char *data, int len)
3052 {
3053
3054         witness_output("%.*s", len, data);
3055         return (len);
3056 }
3057
3058 static void
3059 witness_debugger(int cond, const char *msg)
3060 {
3061         char buf[32];
3062         struct sbuf sb;
3063         struct stack st;
3064
3065         if (!cond)
3066                 return;
3067
3068         if (witness_trace) {
3069                 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3070                 sbuf_set_drain(&sb, witness_output_drain, NULL);
3071
3072                 stack_zero(&st);
3073                 stack_save(&st);
3074                 witness_output("stack backtrace:\n");
3075                 stack_sbuf_print_ddb(&sb, &st);
3076
3077                 sbuf_finish(&sb);
3078         }
3079
3080 #ifdef KDB
3081         if (witness_kdb)
3082                 kdb_enter(KDB_WHY_WITNESS, msg);
3083 #endif
3084 }