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 #define LI_SLEEPABLE    0x00040000      /* Lock may be held while sleeping. */
 135
 136 #ifndef WITNESS_COUNT
 137 #define WITNESS_COUNT           1536
 138 #endif
 139 #define WITNESS_HASH_SIZE       251     /* Prime, gives load factor < 2 */
 140 #define WITNESS_PENDLIST        (512 + (MAXCPU * 4))
 141
 142 /* Allocate 256 KB of stack data space */
 143 #define WITNESS_LO_DATA_COUNT   2048
 144
 145 /* Prime, gives load factor of ~2 at full load */
 146 #define WITNESS_LO_HASH_SIZE    1021
 147
 148 /*
 149  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
 150  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
 151  * probably be safe for the most part, but it's still a SWAG.
 152  */
 153 #define LOCK_NCHILDREN  5
 154 #define LOCK_CHILDCOUNT 2048
 155
 156 #define MAX_W_NAME      64
 157
 158 #define FULLGRAPH_SBUF_SIZE     512
 159
 160 /*
 161  * These flags go in the witness relationship matrix and describe the
 162  * relationship between any two struct witness objects.
 163  */
 164 #define WITNESS_UNRELATED        0x00    /* No lock order relation. */
 165 #define WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
 166 #define WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
 167 #define WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
 168 #define WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
 169 #define WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
 170 #define WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
 171 #define WITNESS_RELATED_MASK                                            \
 172         (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
 173 #define WITNESS_REVERSAL         0x10    /* A lock order reversal has been
 174                                           * observed. */
 175 #define WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
 176 #define WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
 177 #define WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
 178
 179 /* Descendant to ancestor flags */
 180 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
 181
 182 /* Ancestor to descendant flags */
 183 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
 184
 185 #define WITNESS_INDEX_ASSERT(i)                                         \
 186         MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
 187
 188 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
 189
 190 /*
 191  * Lock instances.  A lock instance is the data associated with a lock while
 192  * it is held by witness.  For example, a lock instance will hold the
 193  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
 194  * are held in a per-cpu list while sleep locks are held in per-thread list.
 195  */
 196 struct lock_instance {
 197         struct lock_object      *li_lock;
 198         const char              *li_file;
 199         int                     li_line;
 200         u_int                   li_flags;
 201 };
 202
 203 /*
 204  * A simple list type used to build the list of locks held by a thread
 205  * or CPU.  We can't simply embed the list in struct lock_object since a
 206  * lock may be held by more than one thread if it is a shared lock.  Locks
 207  * are added to the head of the list, so we fill up each list entry from
 208  * "the back" logically.  To ease some of the arithmetic, we actually fill
 209  * in each list entry the normal way (children[0] then children[1], etc.) but
 210  * when we traverse the list we read children[count-1] as the first entry
 211  * down to children[0] as the final entry.
 212  */
 213 struct lock_list_entry {
 214         struct lock_list_entry  *ll_next;
 215         struct lock_instance    ll_children[LOCK_NCHILDREN];
 216         u_int                   ll_count;
 217 };
 218
 219 /*
 220  * The main witness structure. One of these per named lock type in the system
 221  * (for example, "vnode interlock").
 222  */
 223 struct witness {
 224         char                    w_name[MAX_W_NAME];
 225         uint32_t                w_index;  /* Index in the relationship matrix */
 226         struct lock_class       *w_class;
 227         STAILQ_ENTRY(witness)   w_list;         /* List of all witnesses. */
 228         STAILQ_ENTRY(witness)   w_typelist;     /* Witnesses of a type. */
 229         struct witness          *w_hash_next; /* Linked list in hash buckets. */
 230         const char              *w_file; /* File where last acquired */
 231         uint32_t                w_line; /* Line where last acquired */
 232         uint32_t                w_refcount;
 233         uint16_t                w_num_ancestors; /* direct/indirect
 234                                                   * ancestor count */
 235         uint16_t                w_num_descendants; /* direct/indirect
 236                                                     * descendant count */
 237         int16_t                 w_ddb_level;
 238         unsigned                w_displayed:1;
 239         unsigned                w_reversed:1;
 240 };
 241
 242 STAILQ_HEAD(witness_list, witness);
 243
 244 /*
 245  * The witness hash table. Keys are witness names (const char *), elements are
 246  * witness objects (struct witness *).
 247  */
 248 struct witness_hash {
 249         struct witness  *wh_array[WITNESS_HASH_SIZE];
 250         uint32_t        wh_size;
 251         uint32_t        wh_count;
 252 };
 253
 254 /*
 255  * Key type for the lock order data hash table.
 256  */
 257 struct witness_lock_order_key {
 258         uint16_t        from;
 259         uint16_t        to;
 260 };
 261
 262 struct witness_lock_order_data {
 263         struct stack                    wlod_stack;
 264         struct witness_lock_order_key   wlod_key;
 265         struct witness_lock_order_data  *wlod_next;
 266 };
 267
 268 /*
 269  * The witness lock order data hash table. Keys are witness index tuples
 270  * (struct witness_lock_order_key), elements are lock order data objects
 271  * (struct witness_lock_order_data).
 272  */
 273 struct witness_lock_order_hash {
 274         struct witness_lock_order_data  *wloh_array[WITNESS_LO_HASH_SIZE];
 275         u_int   wloh_size;
 276         u_int   wloh_count;
 277 };
 278
 279 struct witness_blessed {
 280         const char      *b_lock1;
 281         const char      *b_lock2;
 282 };
 283
 284 struct witness_pendhelp {
 285         const char              *wh_type;
 286         struct lock_object      *wh_lock;
 287 };
 288
 289 struct witness_order_list_entry {
 290         const char              *w_name;
 291         struct lock_class       *w_class;
 292 };
 293
 294 /*
 295  * Returns 0 if one of the locks is a spin lock and the other is not.
 296  * Returns 1 otherwise.
 297  */
 298 static __inline int
 299 witness_lock_type_equal(struct witness *w1, struct witness *w2)
 300 {
 301
 302         return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
 303                 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
 304 }
 305
 306 static __inline int
 307 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
 308     const struct witness_lock_order_key *b)
 309 {
 310
 311         return (a->from == b->from && a->to == b->to);
 312 }
 313
 314 static int      _isitmyx(struct witness *w1, struct witness *w2, int rmask,
 315                     const char *fname);
 316 static void     adopt(struct witness *parent, struct witness *child);
 317 static int      blessed(struct witness *, struct witness *);
 318 static void     depart(struct witness *w);
 319 static struct witness   *enroll(const char *description,
 320                             struct lock_class *lock_class);
 321 static struct lock_instance     *find_instance(struct lock_list_entry *list,
 322                                     const struct lock_object *lock);
 323 static int      isitmychild(struct witness *parent, struct witness *child);
 324 static int      isitmydescendant(struct witness *parent, struct witness *child);
 325 static void     itismychild(struct witness *parent, struct witness *child);
 326 static int      sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
 327 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
 328 static int      sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
 329 static int      sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
 330 static void     witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
 331 #ifdef DDB
 332 static void     witness_ddb_compute_levels(void);
 333 static void     witness_ddb_display(int(*)(const char *fmt, ...));
 334 static void     witness_ddb_display_descendants(int(*)(const char *fmt, ...),
 335                     struct witness *, int indent);
 336 static void     witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
 337                     struct witness_list *list);
 338 static void     witness_ddb_level_descendants(struct witness *parent, int l);
 339 static void     witness_ddb_list(struct thread *td);
 340 #endif
 341 static void     witness_debugger(int cond, const char *msg);
 342 static void     witness_free(struct witness *m);
 343 static struct witness   *witness_get(void);
 344 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
 345 static struct witness   *witness_hash_get(const char *key);
 346 static void     witness_hash_put(struct witness *w);
 347 static void     witness_init_hash_tables(void);
 348 static void     witness_increment_graph_generation(void);
 349 static void     witness_lock_list_free(struct lock_list_entry *lle);
 350 static struct lock_list_entry   *witness_lock_list_get(void);
 351 static int      witness_lock_order_add(struct witness *parent,
 352                     struct witness *child);
 353 static int      witness_lock_order_check(struct witness *parent,
 354                     struct witness *child);
 355 static struct witness_lock_order_data   *witness_lock_order_get(
 356                                             struct witness *parent,
 357                                             struct witness *child);
 358 static void     witness_list_lock(struct lock_instance *instance,
 359                     int (*prnt)(const char *fmt, ...));
 360 static int      witness_output(const char *fmt, ...) __printflike(1, 2);
 361 static int      witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
 362 static void     witness_setflag(struct lock_object *lock, int flag, int set);
 363
 364 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
 365     "Witness Locking");
 366
 367 /*
 368  * If set to 0, lock order checking is disabled.  If set to -1,
 369  * witness is completely disabled.  Otherwise witness performs full
 370  * lock order checking for all locks.  At runtime, lock order checking
 371  * may be toggled.  However, witness cannot be reenabled once it is
 372  * completely disabled.
 373  */
 374 static int witness_watch = 1;
 375 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
 376     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
 377
 378 #ifdef KDB
 379 /*
 380  * When KDB is enabled and witness_kdb is 1, it will cause the system
 381  * to drop into kdebug() when:
 382  *      - a lock hierarchy violation occurs
 383  *      - locks are held when going to sleep.
 384  */
 385 #ifdef WITNESS_KDB
 386 int     witness_kdb = 1;
 387 #else
 388 int     witness_kdb = 0;
 389 #endif
 390 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
 391 #endif /* KDB */
 392
 393 #if defined(DDB) || defined(KDB)
 394 /*
 395  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
 396  * to print a stack trace:
 397  *      - a lock hierarchy violation occurs
 398  *      - locks are held when going to sleep.
 399  */
 400 int     witness_trace = 1;
 401 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
 402 #endif /* DDB || KDB */
 403
 404 #ifdef WITNESS_SKIPSPIN
 405 int     witness_skipspin = 1;
 406 #else
 407 int     witness_skipspin = 0;
 408 #endif
 409 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
 410
 411 int badstack_sbuf_size;
 412
 413 int witness_count = WITNESS_COUNT;
 414 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
 415     &witness_count, 0, "");
 416
 417 /*
 418  * Output channel for witness messages.  By default we print to the console.
 419  */
 420 enum witness_channel {
 421         WITNESS_CONSOLE,
 422         WITNESS_LOG,
 423         WITNESS_NONE,
 424 };
 425
 426 static enum witness_channel witness_channel = WITNESS_CONSOLE;
 427 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
 428     CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
 429     "Output channel for warnings");
 430
 431 /*
 432  * Call this to print out the relations between locks.
 433  */
 434 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
 435     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
 436
 437 /*
 438  * Call this to print out the witness faulty stacks.
 439  */
 440 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
 441     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
 442
 443 static struct mtx w_mtx;
 444
 445 /* w_list */
 446 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 447 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 448
 449 /* w_typelist */
 450 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 451 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 452
 453 /* lock list */
 454 static struct lock_list_entry *w_lock_list_free = NULL;
 455 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
 456 static u_int pending_cnt;
 457
 458 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
 459 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
 460 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
 461 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
 462     "");
 463
 464 static struct witness *w_data;
 465 static uint8_t **w_rmatrix;
 466 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 467 static struct witness_hash w_hash;      /* The witness hash table. */
 468
 469 /* The lock order data hash */
 470 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
 471 static struct witness_lock_order_data *w_lofree = NULL;
 472 static struct witness_lock_order_hash w_lohash;
 473 static int w_max_used_index = 0;
 474 static unsigned int w_generation = 0;
 475 static const char w_notrunning[] = "Witness not running\n";
 476 static const char w_stillcold[] = "Witness is still cold\n";
 477 #ifdef __i386__
 478 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
 479 #endif
 480
 481 static struct witness_order_list_entry order_lists[] = {
 482         /*
 483          * sx locks
 484          */
 485         { "proctree", &lock_class_sx },
 486         { "allproc", &lock_class_sx },
 487         { "allprison", &lock_class_sx },
 488         { NULL, NULL },
 489         /*
 490          * Various mutexes
 491          */
 492         { "Giant", &lock_class_mtx_sleep },
 493         { "pipe mutex", &lock_class_mtx_sleep },
 494         { "sigio lock", &lock_class_mtx_sleep },
 495         { "process group", &lock_class_mtx_sleep },
 496 #ifdef  HWPMC_HOOKS
 497         { "pmc-sleep", &lock_class_mtx_sleep },
 498 #endif
 499         { "process lock", &lock_class_mtx_sleep },
 500         { "session", &lock_class_mtx_sleep },
 501         { "uidinfo hash", &lock_class_rw },
 502         { "time lock", &lock_class_mtx_sleep },
 503         { NULL, NULL },
 504         /*
 505          * umtx
 506          */
 507         { "umtx lock", &lock_class_mtx_sleep },
 508         { NULL, NULL },
 509         /*
 510          * Sockets
 511          */
 512         { "accept", &lock_class_mtx_sleep },
 513         { "so_snd", &lock_class_mtx_sleep },
 514         { "so_rcv", &lock_class_mtx_sleep },
 515         { "sellck", &lock_class_mtx_sleep },
 516         { NULL, NULL },
 517         /*
 518          * Routing
 519          */
 520         { "so_rcv", &lock_class_mtx_sleep },
 521         { "radix node head", &lock_class_rm },
 522         { "rtentry", &lock_class_mtx_sleep },
 523         { "ifaddr", &lock_class_mtx_sleep },
 524         { NULL, NULL },
 525         /*
 526          * IPv4 multicast:
 527          * protocol locks before interface locks, after UDP locks.
 528          */
 529         { "in_multi_sx", &lock_class_sx },
 530         { "udpinp", &lock_class_rw },
 531         { "in_multi_list_mtx", &lock_class_mtx_sleep },
 532         { "igmp_mtx", &lock_class_mtx_sleep },
 533         { "ifnet_rw", &lock_class_rw },
 534         { "if_addr_lock", &lock_class_mtx_sleep },
 535         { NULL, NULL },
 536         /*
 537          * IPv6 multicast:
 538          * protocol locks before interface locks, after UDP locks.
 539          */
 540         { "in6_multi_sx", &lock_class_sx },
 541         { "udpinp", &lock_class_rw },
 542         { "in6_multi_list_mtx", &lock_class_mtx_sleep },
 543         { "mld_mtx", &lock_class_mtx_sleep },
 544         { "ifnet_rw", &lock_class_rw },
 545         { "if_addr_lock", &lock_class_mtx_sleep },
 546         { NULL, NULL },
 547         /*
 548          * UNIX Domain Sockets
 549          */
 550         { "unp_link_rwlock", &lock_class_rw },
 551         { "unp_list_lock", &lock_class_mtx_sleep },
 552         { "unp", &lock_class_mtx_sleep },
 553         { "so_snd", &lock_class_mtx_sleep },
 554         { NULL, NULL },
 555         /*
 556          * UDP/IP
 557          */
 558         { "udp", &lock_class_mtx_sleep },
 559         { "udpinp", &lock_class_rw },
 560         { "so_snd", &lock_class_mtx_sleep },
 561         { NULL, NULL },
 562         /*
 563          * TCP/IP
 564          */
 565         { "tcp", &lock_class_mtx_sleep },
 566         { "tcpinp", &lock_class_rw },
 567         { "so_snd", &lock_class_mtx_sleep },
 568         { NULL, NULL },
 569         /*
 570          * BPF
 571          */
 572         { "bpf global lock", &lock_class_sx },
 573         { "bpf cdev lock", &lock_class_mtx_sleep },
 574         { NULL, NULL },
 575         /*
 576          * NFS server
 577          */
 578         { "nfsd_mtx", &lock_class_mtx_sleep },
 579         { "so_snd", &lock_class_mtx_sleep },
 580         { NULL, NULL },
 581
 582         /*
 583          * IEEE 802.11
 584          */
 585         { "802.11 com lock", &lock_class_mtx_sleep},
 586         { NULL, NULL },
 587         /*
 588          * Network drivers
 589          */
 590         { "network driver", &lock_class_mtx_sleep},
 591         { NULL, NULL },
 592
 593         /*
 594          * Netgraph
 595          */
 596         { "ng_node", &lock_class_mtx_sleep },
 597         { "ng_worklist", &lock_class_mtx_sleep },
 598         { NULL, NULL },
 599         /*
 600          * CDEV
 601          */
 602         { "vm map (system)", &lock_class_mtx_sleep },
 603         { "vnode interlock", &lock_class_mtx_sleep },
 604         { "cdev", &lock_class_mtx_sleep },
 605         { "devthrd", &lock_class_mtx_sleep },
 606         { NULL, NULL },
 607         /*
 608          * VM
 609          */
 610         { "vm map (user)", &lock_class_sx },
 611         { "vm object", &lock_class_rw },
 612         { "vm page", &lock_class_mtx_sleep },
 613         { "pmap pv global", &lock_class_rw },
 614         { "pmap", &lock_class_mtx_sleep },
 615         { "pmap pv list", &lock_class_rw },
 616         { "vm page free queue", &lock_class_mtx_sleep },
 617         { "vm pagequeue", &lock_class_mtx_sleep },
 618         { NULL, NULL },
 619         /*
 620          * kqueue/VFS interaction
 621          */
 622         { "kqueue", &lock_class_mtx_sleep },
 623         { "struct mount mtx", &lock_class_mtx_sleep },
 624         { "vnode interlock", &lock_class_mtx_sleep },
 625         { NULL, NULL },
 626         /*
 627          * VFS namecache
 628          */
 629         { "ncvn", &lock_class_mtx_sleep },
 630         { "ncbuc", &lock_class_rw },
 631         { "vnode interlock", &lock_class_mtx_sleep },
 632         { "ncneg", &lock_class_mtx_sleep },
 633         { NULL, NULL },
 634         /*
 635          * ZFS locking
 636          */
 637         { "dn->dn_mtx", &lock_class_sx },
 638         { "dr->dt.di.dr_mtx", &lock_class_sx },
 639         { "db->db_mtx", &lock_class_sx },
 640         { NULL, NULL },
 641         /*
 642          * TCP log locks
 643          */
 644         { "TCP ID tree", &lock_class_rw },
 645         { "tcp log id bucket", &lock_class_mtx_sleep },
 646         { "tcpinp", &lock_class_rw },
 647         { "TCP log expireq", &lock_class_mtx_sleep },
 648         { NULL, NULL },
 649         /*
 650          * spin locks
 651          */
 652 #ifdef SMP
 653         { "ap boot", &lock_class_mtx_spin },
 654 #endif
 655         { "rm.mutex_mtx", &lock_class_mtx_spin },
 656         { "sio", &lock_class_mtx_spin },
 657 #ifdef __i386__
 658         { "cy", &lock_class_mtx_spin },
 659 #endif
 660 #ifdef __sparc64__
 661         { "pcib_mtx", &lock_class_mtx_spin },
 662         { "rtc_mtx", &lock_class_mtx_spin },
 663 #endif
 664         { "scc_hwmtx", &lock_class_mtx_spin },
 665         { "uart_hwmtx", &lock_class_mtx_spin },
 666         { "fast_taskqueue", &lock_class_mtx_spin },
 667         { "intr table", &lock_class_mtx_spin },
 668         { "process slock", &lock_class_mtx_spin },
 669         { "syscons video lock", &lock_class_mtx_spin },
 670         { "sleepq chain", &lock_class_mtx_spin },
 671         { "rm_spinlock", &lock_class_mtx_spin },
 672         { "turnstile chain", &lock_class_mtx_spin },
 673         { "turnstile lock", &lock_class_mtx_spin },
 674         { "sched lock", &lock_class_mtx_spin },
 675         { "td_contested", &lock_class_mtx_spin },
 676         { "callout", &lock_class_mtx_spin },
 677         { "entropy harvest mutex", &lock_class_mtx_spin },
 678 #ifdef SMP
 679         { "smp rendezvous", &lock_class_mtx_spin },
 680 #endif
 681 #ifdef __powerpc__
 682         { "tlb0", &lock_class_mtx_spin },
 683 #endif
 684         { NULL, NULL },
 685         { "sched lock", &lock_class_mtx_spin },
 686 #ifdef  HWPMC_HOOKS
 687         { "pmc-per-proc", &lock_class_mtx_spin },
 688 #endif
 689         { NULL, NULL },
 690         /*
 691          * leaf locks
 692          */
 693         { "intrcnt", &lock_class_mtx_spin },
 694         { "icu", &lock_class_mtx_spin },
 695 #if defined(SMP) && defined(__sparc64__)
 696         { "ipi", &lock_class_mtx_spin },
 697 #endif
 698 #ifdef __i386__
 699         { "allpmaps", &lock_class_mtx_spin },
 700         { "descriptor tables", &lock_class_mtx_spin },
 701 #endif
 702         { "clk", &lock_class_mtx_spin },
 703         { "cpuset", &lock_class_mtx_spin },
 704         { "mprof lock", &lock_class_mtx_spin },
 705         { "zombie lock", &lock_class_mtx_spin },
 706         { "ALD Queue", &lock_class_mtx_spin },
 707 #if defined(__i386__) || defined(__amd64__)
 708         { "pcicfg", &lock_class_mtx_spin },
 709         { "NDIS thread lock", &lock_class_mtx_spin },
 710 #endif
 711         { "tw_osl_io_lock", &lock_class_mtx_spin },
 712         { "tw_osl_q_lock", &lock_class_mtx_spin },
 713         { "tw_cl_io_lock", &lock_class_mtx_spin },
 714         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 715         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 716 #ifdef  HWPMC_HOOKS
 717         { "pmc-leaf", &lock_class_mtx_spin },
 718 #endif
 719         { "blocked lock", &lock_class_mtx_spin },
 720         { NULL, NULL },
 721         { NULL, NULL }
 722 };
 723
 724 /*
 725  * Pairs of locks which have been blessed.  Witness does not complain about
 726  * order problems with blessed lock pairs.  Please do not add an entry to the
 727  * table without an explanatory comment.
 728  */
 729 static struct witness_blessed blessed_list[] = {
 730         /*
 731          * See the comment in ufs_dirhash.c.  Basically, a vnode lock serializes
 732          * both lock orders, so a deadlock cannot happen as a result of this
 733          * LOR.
 734          */
 735         { "dirhash",    "bufwait" },
 736
 737         /*
 738          * A UFS vnode may be locked in vget() while a buffer belonging to the
 739          * parent directory vnode is locked.
 740          */
 741         { "ufs",        "bufwait" },
 742 };
 743
 744 /*
 745  * This global is set to 0 once it becomes safe to use the witness code.
 746  */
 747 static int witness_cold = 1;
 748
 749 /*
 750  * This global is set to 1 once the static lock orders have been enrolled
 751  * so that a warning can be issued for any spin locks enrolled later.
 752  */
 753 static int witness_spin_warn = 0;
 754
 755 /* Trim useless garbage from filenames. */
 756 static const char *
 757 fixup_filename(const char *file)
 758 {
 759
 760         if (file == NULL)
 761                 return (NULL);
 762         while (strncmp(file, "../", 3) == 0)
 763                 file += 3;
 764         return (file);
 765 }
 766
 767 /*
 768  * Calculate the size of early witness structures.
 769  */
 770 int
 771 witness_startup_count(void)
 772 {
 773         int sz;
 774
 775         sz = sizeof(struct witness) * witness_count;
 776         sz += sizeof(*w_rmatrix) * (witness_count + 1);
 777         sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
 778             (witness_count + 1);
 779
 780         return (sz);
 781 }
 782
 783 /*
 784  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 785  * assume that the early boot is single-threaded at least until after this
 786  * routine is completed.
 787  */
 788 void
 789 witness_startup(void *mem)
 790 {
 791         struct lock_object *lock;
 792         struct witness_order_list_entry *order;
 793         struct witness *w, *w1;
 794         uintptr_t p;
 795         int i;
 796
 797         p = (uintptr_t)mem;
 798         w_data = (void *)p;
 799         p += sizeof(struct witness) * witness_count;
 800
 801         w_rmatrix = (void *)p;
 802         p += sizeof(*w_rmatrix) * (witness_count + 1);
 803
 804         for (i = 0; i < witness_count + 1; i++) {
 805                 w_rmatrix[i] = (void *)p;
 806                 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
 807         }
 808         badstack_sbuf_size = witness_count * 256;
 809
 810         /*
 811          * We have to release Giant before initializing its witness
 812          * structure so that WITNESS doesn't get confused.
 813          */
 814         mtx_unlock(&Giant);
 815         mtx_assert(&Giant, MA_NOTOWNED);
 816
 817         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 818         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 819             MTX_NOWITNESS | MTX_NOPROFILE);
 820         for (i = witness_count - 1; i >= 0; i--) {
 821                 w = &w_data[i];
 822                 memset(w, 0, sizeof(*w));
 823                 w_data[i].w_index = i;  /* Witness index never changes. */
 824                 witness_free(w);
 825         }
 826         KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
 827             ("%s: Invalid list of free witness objects", __func__));
 828
 829         /* Witness with index 0 is not used to aid in debugging. */
 830         STAILQ_REMOVE_HEAD(&w_free, w_list);
 831         w_free_cnt--;
 832
 833         for (i = 0; i < witness_count; i++) {
 834                 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
 835                     (witness_count + 1));
 836         }
 837
 838         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 839                 witness_lock_list_free(&w_locklistdata[i]);
 840         witness_init_hash_tables();
 841
 842         /* First add in all the specified order lists. */
 843         for (order = order_lists; order->w_name != NULL; order++) {
 844                 w = enroll(order->w_name, order->w_class);
 845                 if (w == NULL)
 846                         continue;
 847                 w->w_file = "order list";
 848                 for (order++; order->w_name != NULL; order++) {
 849                         w1 = enroll(order->w_name, order->w_class);
 850                         if (w1 == NULL)
 851                                 continue;
 852                         w1->w_file = "order list";
 853                         itismychild(w, w1);
 854                         w = w1;
 855                 }
 856         }
 857         witness_spin_warn = 1;
 858
 859         /* Iterate through all locks and add them to witness. */
 860         for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
 861                 lock = pending_locks[i].wh_lock;
 862                 KASSERT(lock->lo_flags & LO_WITNESS,
 863                     ("%s: lock %s is on pending list but not LO_WITNESS",
 864                     __func__, lock->lo_name));
 865                 lock->lo_witness = enroll(pending_locks[i].wh_type,
 866                     LOCK_CLASS(lock));
 867         }
 868
 869         /* Mark the witness code as being ready for use. */
 870         witness_cold = 0;
 871
 872         mtx_lock(&Giant);
 873 }
 874
 875 void
 876 witness_init(struct lock_object *lock, const char *type)
 877 {
 878         struct lock_class *class;
 879
 880         /* Various sanity checks. */
 881         class = LOCK_CLASS(lock);
 882         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 883             (class->lc_flags & LC_RECURSABLE) == 0)
 884                 kassert_panic("%s: lock (%s) %s can not be recursable",
 885                     __func__, class->lc_name, lock->lo_name);
 886         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 887             (class->lc_flags & LC_SLEEPABLE) == 0)
 888                 kassert_panic("%s: lock (%s) %s can not be sleepable",
 889                     __func__, class->lc_name, lock->lo_name);
 890         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 891             (class->lc_flags & LC_UPGRADABLE) == 0)
 892                 kassert_panic("%s: lock (%s) %s can not be upgradable",
 893                     __func__, class->lc_name, lock->lo_name);
 894
 895         /*
 896          * If we shouldn't watch this lock, then just clear lo_witness.
 897          * Otherwise, if witness_cold is set, then it is too early to
 898          * enroll this lock, so defer it to witness_initialize() by adding
 899          * it to the pending_locks list.  If it is not too early, then enroll
 900          * the lock now.
 901          */
 902         if (witness_watch < 1 || panicstr != NULL ||
 903             (lock->lo_flags & LO_WITNESS) == 0)
 904                 lock->lo_witness = NULL;
 905         else if (witness_cold) {
 906                 pending_locks[pending_cnt].wh_lock = lock;
 907                 pending_locks[pending_cnt++].wh_type = type;
 908                 if (pending_cnt > WITNESS_PENDLIST)
 909                         panic("%s: pending locks list is too small, "
 910                             "increase WITNESS_PENDLIST\n",
 911                             __func__);
 912         } else
 913                 lock->lo_witness = enroll(type, class);
 914 }
 915
 916 void
 917 witness_destroy(struct lock_object *lock)
 918 {
 919         struct lock_class *class;
 920         struct witness *w;
 921
 922         class = LOCK_CLASS(lock);
 923
 924         if (witness_cold)
 925                 panic("lock (%s) %s destroyed while witness_cold",
 926                     class->lc_name, lock->lo_name);
 927
 928         /* XXX: need to verify that no one holds the lock */
 929         if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
 930                 return;
 931         w = lock->lo_witness;
 932
 933         mtx_lock_spin(&w_mtx);
 934         MPASS(w->w_refcount > 0);
 935         w->w_refcount--;
 936
 937         if (w->w_refcount == 0)
 938                 depart(w);
 939         mtx_unlock_spin(&w_mtx);
 940 }
 941
 942 #ifdef DDB
 943 static void
 944 witness_ddb_compute_levels(void)
 945 {
 946         struct witness *w;
 947
 948         /*
 949          * First clear all levels.
 950          */
 951         STAILQ_FOREACH(w, &w_all, w_list)
 952                 w->w_ddb_level = -1;
 953
 954         /*
 955          * Look for locks with no parents and level all their descendants.
 956          */
 957         STAILQ_FOREACH(w, &w_all, w_list) {
 958
 959                 /* If the witness has ancestors (is not a root), skip it. */
 960                 if (w->w_num_ancestors > 0)
 961                         continue;
 962                 witness_ddb_level_descendants(w, 0);
 963         }
 964 }
 965
 966 static void
 967 witness_ddb_level_descendants(struct witness *w, int l)
 968 {
 969         int i;
 970
 971         if (w->w_ddb_level >= l)
 972                 return;
 973
 974         w->w_ddb_level = l;
 975         l++;
 976
 977         for (i = 1; i <= w_max_used_index; i++) {
 978                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 979                         witness_ddb_level_descendants(&w_data[i], l);
 980         }
 981 }
 982
 983 static void
 984 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
 985     struct witness *w, int indent)
 986 {
 987         int i;
 988
 989         for (i = 0; i < indent; i++)
 990                 prnt(" ");
 991         prnt("%s (type: %s, depth: %d, active refs: %d)",
 992              w->w_name, w->w_class->lc_name,
 993              w->w_ddb_level, w->w_refcount);
 994         if (w->w_displayed) {
 995                 prnt(" -- (already displayed)\n");
 996                 return;
 997         }
 998         w->w_displayed = 1;
 999         if (w->w_file != NULL && w->w_line != 0)
1000                 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1001                     w->w_line);
1002         else
1003                 prnt(" -- never acquired\n");
1004         indent++;
1005         WITNESS_INDEX_ASSERT(w->w_index);
1006         for (i = 1; i <= w_max_used_index; i++) {
1007                 if (db_pager_quit)
1008                         return;
1009                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1010                         witness_ddb_display_descendants(prnt, &w_data[i],
1011                             indent);
1012         }
1013 }
1014
1015 static void
1016 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1017     struct witness_list *list)
1018 {
1019         struct witness *w;
1020
1021         STAILQ_FOREACH(w, list, w_typelist) {
1022                 if (w->w_file == NULL || w->w_ddb_level > 0)
1023                         continue;
1024
1025                 /* This lock has no anscestors - display its descendants. */
1026                 witness_ddb_display_descendants(prnt, w, 0);
1027                 if (db_pager_quit)
1028                         return;
1029         }
1030 }
1031
1032 static void
1033 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1034 {
1035         struct witness *w;
1036
1037         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1038         witness_ddb_compute_levels();
1039
1040         /* Clear all the displayed flags. */
1041         STAILQ_FOREACH(w, &w_all, w_list)
1042                 w->w_displayed = 0;
1043
1044         /*
1045          * First, handle sleep locks which have been acquired at least
1046          * once.
1047          */
1048         prnt("Sleep locks:\n");
1049         witness_ddb_display_list(prnt, &w_sleep);
1050         if (db_pager_quit)
1051                 return;
1052
1053         /*
1054          * Now do spin locks which have been acquired at least once.
1055          */
1056         prnt("\nSpin locks:\n");
1057         witness_ddb_display_list(prnt, &w_spin);
1058         if (db_pager_quit)
1059                 return;
1060
1061         /*
1062          * Finally, any locks which have not been acquired yet.
1063          */
1064         prnt("\nLocks which were never acquired:\n");
1065         STAILQ_FOREACH(w, &w_all, w_list) {
1066                 if (w->w_file != NULL || w->w_refcount == 0)
1067                         continue;
1068                 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1069                     w->w_class->lc_name, w->w_ddb_level);
1070                 if (db_pager_quit)
1071                         return;
1072         }
1073 }
1074 #endif /* DDB */
1075
1076 int
1077 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1078 {
1079
1080         if (witness_watch == -1 || panicstr != NULL)
1081                 return (0);
1082
1083         /* Require locks that witness knows about. */
1084         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1085             lock2->lo_witness == NULL)
1086                 return (EINVAL);
1087
1088         mtx_assert(&w_mtx, MA_NOTOWNED);
1089         mtx_lock_spin(&w_mtx);
1090
1091         /*
1092          * If we already have either an explicit or implied lock order that
1093          * is the other way around, then return an error.
1094          */
1095         if (witness_watch &&
1096             isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1097                 mtx_unlock_spin(&w_mtx);
1098                 return (EDOOFUS);
1099         }
1100
1101         /* Try to add the new order. */
1102         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1103             lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1104         itismychild(lock1->lo_witness, lock2->lo_witness);
1105         mtx_unlock_spin(&w_mtx);
1106         return (0);
1107 }
1108
1109 void
1110 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1111     int line, struct lock_object *interlock)
1112 {
1113         struct lock_list_entry *lock_list, *lle;
1114         struct lock_instance *lock1, *lock2, *plock;
1115         struct lock_class *class, *iclass;
1116         struct witness *w, *w1;
1117         struct thread *td;
1118         int i, j;
1119
1120         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1121             panicstr != NULL)
1122                 return;
1123
1124         w = lock->lo_witness;
1125         class = LOCK_CLASS(lock);
1126         td = curthread;
1127
1128         if (class->lc_flags & LC_SLEEPLOCK) {
1129
1130                 /*
1131                  * Since spin locks include a critical section, this check
1132                  * implicitly enforces a lock order of all sleep locks before
1133                  * all spin locks.
1134                  */
1135                 if (td->td_critnest != 0 && !kdb_active)
1136                         kassert_panic("acquiring blockable sleep lock with "
1137                             "spinlock or critical section held (%s) %s @ %s:%d",
1138                             class->lc_name, lock->lo_name,
1139                             fixup_filename(file), line);
1140
1141                 /*
1142                  * If this is the first lock acquired then just return as
1143                  * no order checking is needed.
1144                  */
1145                 lock_list = td->td_sleeplocks;
1146                 if (lock_list == NULL || lock_list->ll_count == 0)
1147                         return;
1148         } else {
1149
1150                 /*
1151                  * If this is the first lock, just return as no order
1152                  * checking is needed.  Avoid problems with thread
1153                  * migration pinning the thread while checking if
1154                  * spinlocks are held.  If at least one spinlock is held
1155                  * the thread is in a safe path and it is allowed to
1156                  * unpin it.
1157                  */
1158                 sched_pin();
1159                 lock_list = PCPU_GET(spinlocks);
1160                 if (lock_list == NULL || lock_list->ll_count == 0) {
1161                         sched_unpin();
1162                         return;
1163                 }
1164                 sched_unpin();
1165         }
1166
1167         /*
1168          * Check to see if we are recursing on a lock we already own.  If
1169          * so, make sure that we don't mismatch exclusive and shared lock
1170          * acquires.
1171          */
1172         lock1 = find_instance(lock_list, lock);
1173         if (lock1 != NULL) {
1174                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1175                     (flags & LOP_EXCLUSIVE) == 0) {
1176                         witness_output("shared lock of (%s) %s @ %s:%d\n",
1177                             class->lc_name, lock->lo_name,
1178                             fixup_filename(file), line);
1179                         witness_output("while exclusively locked from %s:%d\n",
1180                             fixup_filename(lock1->li_file), lock1->li_line);
1181                         kassert_panic("excl->share");
1182                 }
1183                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1184                     (flags & LOP_EXCLUSIVE) != 0) {
1185                         witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1186                             class->lc_name, lock->lo_name,
1187                             fixup_filename(file), line);
1188                         witness_output("while share locked from %s:%d\n",
1189                             fixup_filename(lock1->li_file), lock1->li_line);
1190                         kassert_panic("share->excl");
1191                 }
1192                 return;
1193         }
1194
1195         /* Warn if the interlock is not locked exactly once. */
1196         if (interlock != NULL) {
1197                 iclass = LOCK_CLASS(interlock);
1198                 lock1 = find_instance(lock_list, interlock);
1199                 if (lock1 == NULL)
1200                         kassert_panic("interlock (%s) %s not locked @ %s:%d",
1201                             iclass->lc_name, interlock->lo_name,
1202                             fixup_filename(file), line);
1203                 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1204                         kassert_panic("interlock (%s) %s recursed @ %s:%d",
1205                             iclass->lc_name, interlock->lo_name,
1206                             fixup_filename(file), line);
1207         }
1208
1209         /*
1210          * Find the previously acquired lock, but ignore interlocks.
1211          */
1212         plock = &lock_list->ll_children[lock_list->ll_count - 1];
1213         if (interlock != NULL && plock->li_lock == interlock) {
1214                 if (lock_list->ll_count > 1)
1215                         plock =
1216                             &lock_list->ll_children[lock_list->ll_count - 2];
1217                 else {
1218                         lle = lock_list->ll_next;
1219
1220                         /*
1221                          * The interlock is the only lock we hold, so
1222                          * simply return.
1223                          */
1224                         if (lle == NULL)
1225                                 return;
1226                         plock = &lle->ll_children[lle->ll_count - 1];
1227                 }
1228         }
1229
1230         /*
1231          * Try to perform most checks without a lock.  If this succeeds we
1232          * can skip acquiring the lock and return success.  Otherwise we redo
1233          * the check with the lock held to handle races with concurrent updates.
1234          */
1235         w1 = plock->li_lock->lo_witness;
1236         if (witness_lock_order_check(w1, w))
1237                 return;
1238
1239         mtx_lock_spin(&w_mtx);
1240         if (witness_lock_order_check(w1, w)) {
1241                 mtx_unlock_spin(&w_mtx);
1242                 return;
1243         }
1244         witness_lock_order_add(w1, w);
1245
1246         /*
1247          * Check for duplicate locks of the same type.  Note that we only
1248          * have to check for this on the last lock we just acquired.  Any
1249          * other cases will be caught as lock order violations.
1250          */
1251         if (w1 == w) {
1252                 i = w->w_index;
1253                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1254                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1255                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
1256                         w->w_reversed = 1;
1257                         mtx_unlock_spin(&w_mtx);
1258                         witness_output(
1259                             "acquiring duplicate lock of same type: \"%s\"\n",
1260                             w->w_name);
1261                         witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1262                             fixup_filename(plock->li_file), plock->li_line);
1263                         witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1264                             fixup_filename(file), line);
1265                         witness_debugger(1, __func__);
1266                 } else
1267                         mtx_unlock_spin(&w_mtx);
1268                 return;
1269         }
1270         mtx_assert(&w_mtx, MA_OWNED);
1271
1272         /*
1273          * If we know that the lock we are acquiring comes after
1274          * the lock we most recently acquired in the lock order tree,
1275          * then there is no need for any further checks.
1276          */
1277         if (isitmychild(w1, w))
1278                 goto out;
1279
1280         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1281                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1282
1283                         MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1284                         lock1 = &lle->ll_children[i];
1285
1286                         /*
1287                          * Ignore the interlock.
1288                          */
1289                         if (interlock == lock1->li_lock)
1290                                 continue;
1291
1292                         /*
1293                          * If this lock doesn't undergo witness checking,
1294                          * then skip it.
1295                          */
1296                         w1 = lock1->li_lock->lo_witness;
1297                         if (w1 == NULL) {
1298                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1299                                     ("lock missing witness structure"));
1300                                 continue;
1301                         }
1302
1303                         /*
1304                          * If we are locking Giant and this is a sleepable
1305                          * lock, then skip it.
1306                          */
1307                         if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1308                             lock == &Giant.lock_object)
1309                                 continue;
1310
1311                         /*
1312                          * If we are locking a sleepable lock and this lock
1313                          * is Giant, then skip it.
1314                          */
1315                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1316                             (flags & LOP_NOSLEEP) == 0 &&
1317                             lock1->li_lock == &Giant.lock_object)
1318                                 continue;
1319
1320                         /*
1321                          * If we are locking a sleepable lock and this lock
1322                          * isn't sleepable, we want to treat it as a lock
1323                          * order violation to enfore a general lock order of
1324                          * sleepable locks before non-sleepable locks.
1325                          */
1326                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1327                             (flags & LOP_NOSLEEP) == 0 &&
1328                             (lock1->li_flags & LI_SLEEPABLE) == 0)
1329                                 goto reversal;
1330
1331                         /*
1332                          * If we are locking Giant and this is a non-sleepable
1333                          * lock, then treat it as a reversal.
1334                          */
1335                         if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1336                             lock == &Giant.lock_object)
1337                                 goto reversal;
1338
1339                         /*
1340                          * Check the lock order hierarchy for a reveresal.
1341                          */
1342                         if (!isitmydescendant(w, w1))
1343                                 continue;
1344                 reversal:
1345
1346                         /*
1347                          * We have a lock order violation, check to see if it
1348                          * is allowed or has already been yelled about.
1349                          */
1350
1351                         /* Bail if this violation is known */
1352                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1353                                 goto out;
1354
1355                         /* Record this as a violation */
1356                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1357                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1358                         w->w_reversed = w1->w_reversed = 1;
1359                         witness_increment_graph_generation();
1360
1361                         /*
1362                          * If the lock order is blessed, bail before logging
1363                          * anything.  We don't look for other lock order
1364                          * violations though, which may be a bug.
1365                          */
1366                         if (blessed(w, w1))
1367                                 goto out;
1368                         mtx_unlock_spin(&w_mtx);
1369
1370 #ifdef WITNESS_NO_VNODE
1371                         /*
1372                          * There are known LORs between VNODE locks. They are
1373                          * not an indication of a bug. VNODE locks are flagged
1374                          * as such (LO_IS_VNODE) and we don't yell if the LOR
1375                          * is between 2 VNODE locks.
1376                          */
1377                         if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1378                             (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1379                                 return;
1380 #endif
1381
1382                         /*
1383                          * Ok, yell about it.
1384                          */
1385                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1386                             (flags & LOP_NOSLEEP) == 0 &&
1387                             (lock1->li_flags & LI_SLEEPABLE) == 0)
1388                                 witness_output(
1389                 "lock order reversal: (sleepable after non-sleepable)\n");
1390                         else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1391                             && lock == &Giant.lock_object)
1392                                 witness_output(
1393                 "lock order reversal: (Giant after non-sleepable)\n");
1394                         else
1395                                 witness_output("lock order reversal:\n");
1396
1397                         /*
1398                          * Try to locate an earlier lock with
1399                          * witness w in our list.
1400                          */
1401                         do {
1402                                 lock2 = &lle->ll_children[i];
1403                                 MPASS(lock2->li_lock != NULL);
1404                                 if (lock2->li_lock->lo_witness == w)
1405                                         break;
1406                                 if (i == 0 && lle->ll_next != NULL) {
1407                                         lle = lle->ll_next;
1408                                         i = lle->ll_count - 1;
1409                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1410                                 } else
1411                                         i--;
1412                         } while (i >= 0);
1413                         if (i < 0) {
1414                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1415                                     lock1->li_lock, lock1->li_lock->lo_name,
1416                                     w1->w_name, fixup_filename(lock1->li_file),
1417                                     lock1->li_line);
1418                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1419                                     lock->lo_name, w->w_name,
1420                                     fixup_filename(file), line);
1421                         } else {
1422                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1423                                     lock2->li_lock, lock2->li_lock->lo_name,
1424                                     lock2->li_lock->lo_witness->w_name,
1425                                     fixup_filename(lock2->li_file),
1426                                     lock2->li_line);
1427                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1428                                     lock1->li_lock, lock1->li_lock->lo_name,
1429                                     w1->w_name, fixup_filename(lock1->li_file),
1430                                     lock1->li_line);
1431                                 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1432                                     lock->lo_name, w->w_name,
1433                                     fixup_filename(file), line);
1434                         }
1435                         witness_debugger(1, __func__);
1436                         return;
1437                 }
1438         }
1439
1440         /*
1441          * If requested, build a new lock order.  However, don't build a new
1442          * relationship between a sleepable lock and Giant if it is in the
1443          * wrong direction.  The correct lock order is that sleepable locks
1444          * always come before Giant.
1445          */
1446         if (flags & LOP_NEWORDER &&
1447             !(plock->li_lock == &Giant.lock_object &&
1448             (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1449             (flags & LOP_NOSLEEP) == 0)) {
1450                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1451                     w->w_name, plock->li_lock->lo_witness->w_name);
1452                 itismychild(plock->li_lock->lo_witness, w);
1453         }
1454 out:
1455         mtx_unlock_spin(&w_mtx);
1456 }
1457
1458 void
1459 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1460 {
1461         struct lock_list_entry **lock_list, *lle;
1462         struct lock_instance *instance;
1463         struct witness *w;
1464         struct thread *td;
1465
1466         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1467             panicstr != NULL)
1468                 return;
1469         w = lock->lo_witness;
1470         td = curthread;
1471
1472         /* Determine lock list for this lock. */
1473         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1474                 lock_list = &td->td_sleeplocks;
1475         else
1476                 lock_list = PCPU_PTR(spinlocks);
1477
1478         /* Check to see if we are recursing on a lock we already own. */
1479         instance = find_instance(*lock_list, lock);
1480         if (instance != NULL) {
1481                 instance->li_flags++;
1482                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1483                     td->td_proc->p_pid, lock->lo_name,
1484                     instance->li_flags & LI_RECURSEMASK);
1485                 instance->li_file = file;
1486                 instance->li_line = line;
1487                 return;
1488         }
1489
1490         /* Update per-witness last file and line acquire. */
1491         w->w_file = file;
1492         w->w_line = line;
1493
1494         /* Find the next open lock instance in the list and fill it. */
1495         lle = *lock_list;
1496         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1497                 lle = witness_lock_list_get();
1498                 if (lle == NULL)
1499                         return;
1500                 lle->ll_next = *lock_list;
1501                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1502                     td->td_proc->p_pid, lle);
1503                 *lock_list = lle;
1504         }
1505         instance = &lle->ll_children[lle->ll_count++];
1506         instance->li_lock = lock;
1507         instance->li_line = line;
1508         instance->li_file = file;
1509         instance->li_flags = 0;
1510         if ((flags & LOP_EXCLUSIVE) != 0)
1511                 instance->li_flags |= LI_EXCLUSIVE;
1512         if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1513                 instance->li_flags |= LI_SLEEPABLE;
1514         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1515             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1516 }
1517
1518 void
1519 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1520 {
1521         struct lock_instance *instance;
1522         struct lock_class *class;
1523
1524         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1525         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1526                 return;
1527         class = LOCK_CLASS(lock);
1528         if (witness_watch) {
1529                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1530                         kassert_panic(
1531                             "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1532                             class->lc_name, lock->lo_name,
1533                             fixup_filename(file), line);
1534                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1535                         kassert_panic(
1536                             "upgrade of non-sleep lock (%s) %s @ %s:%d",
1537                             class->lc_name, lock->lo_name,
1538                             fixup_filename(file), line);
1539         }
1540         instance = find_instance(curthread->td_sleeplocks, lock);
1541         if (instance == NULL) {
1542                 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1543                     class->lc_name, lock->lo_name,
1544                     fixup_filename(file), line);
1545                 return;
1546         }
1547         if (witness_watch) {
1548                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1549                         kassert_panic(
1550                             "upgrade of exclusive lock (%s) %s @ %s:%d",
1551                             class->lc_name, lock->lo_name,
1552                             fixup_filename(file), line);
1553                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1554                         kassert_panic(
1555                             "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1556                             class->lc_name, lock->lo_name,
1557                             instance->li_flags & LI_RECURSEMASK,
1558                             fixup_filename(file), line);
1559         }
1560         instance->li_flags |= LI_EXCLUSIVE;
1561 }
1562
1563 void
1564 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1565     int line)
1566 {
1567         struct lock_instance *instance;
1568         struct lock_class *class;
1569
1570         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1571         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1572                 return;
1573         class = LOCK_CLASS(lock);
1574         if (witness_watch) {
1575                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1576                         kassert_panic(
1577                             "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1578                             class->lc_name, lock->lo_name,
1579                             fixup_filename(file), line);
1580                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1581                         kassert_panic(
1582                             "downgrade of non-sleep lock (%s) %s @ %s:%d",
1583                             class->lc_name, lock->lo_name,
1584                             fixup_filename(file), line);
1585         }
1586         instance = find_instance(curthread->td_sleeplocks, lock);
1587         if (instance == NULL) {
1588                 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1589                     class->lc_name, lock->lo_name,
1590                     fixup_filename(file), line);
1591                 return;
1592         }
1593         if (witness_watch) {
1594                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1595                         kassert_panic(
1596                             "downgrade of shared lock (%s) %s @ %s:%d",
1597                             class->lc_name, lock->lo_name,
1598                             fixup_filename(file), line);
1599                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1600                         kassert_panic(
1601                             "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1602                             class->lc_name, lock->lo_name,
1603                             instance->li_flags & LI_RECURSEMASK,
1604                             fixup_filename(file), line);
1605         }
1606         instance->li_flags &= ~LI_EXCLUSIVE;
1607 }
1608
1609 void
1610 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1611 {
1612         struct lock_list_entry **lock_list, *lle;
1613         struct lock_instance *instance;
1614         struct lock_class *class;
1615         struct thread *td;
1616         register_t s;
1617         int i, j;
1618
1619         if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1620                 return;
1621         td = curthread;
1622         class = LOCK_CLASS(lock);
1623
1624         /* Find lock instance associated with this lock. */
1625         if (class->lc_flags & LC_SLEEPLOCK)
1626                 lock_list = &td->td_sleeplocks;
1627         else
1628                 lock_list = PCPU_PTR(spinlocks);
1629         lle = *lock_list;
1630         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1631                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1632                         instance = &(*lock_list)->ll_children[i];
1633                         if (instance->li_lock == lock)
1634                                 goto found;
1635                 }
1636
1637         /*
1638          * When disabling WITNESS through witness_watch we could end up in
1639          * having registered locks in the td_sleeplocks queue.
1640          * We have to make sure we flush these queues, so just search for
1641          * eventual register locks and remove them.
1642          */
1643         if (witness_watch > 0) {
1644                 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1645                     lock->lo_name, fixup_filename(file), line);
1646                 return;
1647         } else {
1648                 return;
1649         }
1650 found:
1651
1652         /* First, check for shared/exclusive mismatches. */
1653         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1654             (flags & LOP_EXCLUSIVE) == 0) {
1655                 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1656                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1657                 witness_output("while exclusively locked from %s:%d\n",
1658                     fixup_filename(instance->li_file), instance->li_line);
1659                 kassert_panic("excl->ushare");
1660         }
1661         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1662             (flags & LOP_EXCLUSIVE) != 0) {
1663                 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1664                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1665                 witness_output("while share locked from %s:%d\n",
1666                     fixup_filename(instance->li_file),
1667                     instance->li_line);
1668                 kassert_panic("share->uexcl");
1669         }
1670         /* If we are recursed, unrecurse. */
1671         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1672                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1673                     td->td_proc->p_pid, instance->li_lock->lo_name,
1674                     instance->li_flags);
1675                 instance->li_flags--;
1676                 return;
1677         }
1678         /* The lock is now being dropped, check for NORELEASE flag */
1679         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1680                 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1681                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1682                 kassert_panic("lock marked norelease");
1683         }
1684
1685         /* Otherwise, remove this item from the list. */
1686         s = intr_disable();
1687         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1688             td->td_proc->p_pid, instance->li_lock->lo_name,
1689             (*lock_list)->ll_count - 1);
1690         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1691                 (*lock_list)->ll_children[j] =
1692                     (*lock_list)->ll_children[j + 1];
1693         (*lock_list)->ll_count--;
1694         intr_restore(s);
1695
1696         /*
1697          * In order to reduce contention on w_mtx, we want to keep always an
1698          * head object into lists so that frequent allocation from the
1699          * free witness pool (and subsequent locking) is avoided.
1700          * In order to maintain the current code simple, when the head
1701          * object is totally unloaded it means also that we do not have
1702          * further objects in the list, so the list ownership needs to be
1703          * hand over to another object if the current head needs to be freed.
1704          */
1705         if ((*lock_list)->ll_count == 0) {
1706                 if (*lock_list == lle) {
1707                         if (lle->ll_next == NULL)
1708                                 return;
1709                 } else
1710                         lle = *lock_list;
1711                 *lock_list = lle->ll_next;
1712                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1713                     td->td_proc->p_pid, lle);
1714                 witness_lock_list_free(lle);
1715         }
1716 }
1717
1718 void
1719 witness_thread_exit(struct thread *td)
1720 {
1721         struct lock_list_entry *lle;
1722         int i, n;
1723
1724         lle = td->td_sleeplocks;
1725         if (lle == NULL || panicstr != NULL)
1726                 return;
1727         if (lle->ll_count != 0) {
1728                 for (n = 0; lle != NULL; lle = lle->ll_next)
1729                         for (i = lle->ll_count - 1; i >= 0; i--) {
1730                                 if (n == 0)
1731                                         witness_output(
1732                     "Thread %p exiting with the following locks held:\n", td);
1733                                 n++;
1734                                 witness_list_lock(&lle->ll_children[i],
1735                                     witness_output);
1736
1737                         }
1738                 kassert_panic(
1739                     "Thread %p cannot exit while holding sleeplocks\n", td);
1740         }
1741         witness_lock_list_free(lle);
1742 }
1743
1744 /*
1745  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1746  * exempt Giant and sleepable locks from the checks as well.  If any
1747  * non-exempt locks are held, then a supplied message is printed to the
1748  * output channel along with a list of the offending locks.  If indicated in the
1749  * flags then a failure results in a panic as well.
1750  */
1751 int
1752 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1753 {
1754         struct lock_list_entry *lock_list, *lle;
1755         struct lock_instance *lock1;
1756         struct thread *td;
1757         va_list ap;
1758         int i, n;
1759
1760         if (witness_cold || witness_watch < 1 || panicstr != NULL)
1761                 return (0);
1762         n = 0;
1763         td = curthread;
1764         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1765                 for (i = lle->ll_count - 1; i >= 0; i--) {
1766                         lock1 = &lle->ll_children[i];
1767                         if (lock1->li_lock == lock)
1768                                 continue;
1769                         if (flags & WARN_GIANTOK &&
1770                             lock1->li_lock == &Giant.lock_object)
1771                                 continue;
1772                         if (flags & WARN_SLEEPOK &&
1773                             (lock1->li_flags & LI_SLEEPABLE) != 0)
1774                                 continue;
1775                         if (n == 0) {
1776                                 va_start(ap, fmt);
1777                                 vprintf(fmt, ap);
1778                                 va_end(ap);
1779                                 printf(" with the following %slocks held:\n",
1780                                     (flags & WARN_SLEEPOK) != 0 ?
1781                                     "non-sleepable " : "");
1782                         }
1783                         n++;
1784                         witness_list_lock(lock1, printf);
1785                 }
1786
1787         /*
1788          * Pin the thread in order to avoid problems with thread migration.
1789          * Once that all verifies are passed about spinlocks ownership,
1790          * the thread is in a safe path and it can be unpinned.
1791          */
1792         sched_pin();
1793         lock_list = PCPU_GET(spinlocks);
1794         if (lock_list != NULL && lock_list->ll_count != 0) {
1795                 sched_unpin();
1796
1797                 /*
1798                  * We should only have one spinlock and as long as
1799                  * the flags cannot match for this locks class,
1800                  * check if the first spinlock is the one curthread
1801                  * should hold.
1802                  */
1803                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1804                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1805                     lock1->li_lock == lock && n == 0)
1806                         return (0);
1807
1808                 va_start(ap, fmt);
1809                 vprintf(fmt, ap);
1810                 va_end(ap);
1811                 printf(" with the following %slocks held:\n",
1812                     (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1813                 n += witness_list_locks(&lock_list, printf);
1814         } else
1815                 sched_unpin();
1816         if (flags & WARN_PANIC && n)
1817                 kassert_panic("%s", __func__);
1818         else
1819                 witness_debugger(n, __func__);
1820         return (n);
1821 }
1822
1823 const char *
1824 witness_file(struct lock_object *lock)
1825 {
1826         struct witness *w;
1827
1828         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1829                 return ("?");
1830         w = lock->lo_witness;
1831         return (w->w_file);
1832 }
1833
1834 int
1835 witness_line(struct lock_object *lock)
1836 {
1837         struct witness *w;
1838
1839         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1840                 return (0);
1841         w = lock->lo_witness;
1842         return (w->w_line);
1843 }
1844
1845 static struct witness *
1846 enroll(const char *description, struct lock_class *lock_class)
1847 {
1848         struct witness *w;
1849
1850         MPASS(description != NULL);
1851
1852         if (witness_watch == -1 || panicstr != NULL)
1853                 return (NULL);
1854         if ((lock_class->lc_flags & LC_SPINLOCK)) {
1855                 if (witness_skipspin)
1856                         return (NULL);
1857         } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1858                 kassert_panic("lock class %s is not sleep or spin",
1859                     lock_class->lc_name);
1860                 return (NULL);
1861         }
1862
1863         mtx_lock_spin(&w_mtx);
1864         w = witness_hash_get(description);
1865         if (w)
1866                 goto found;
1867         if ((w = witness_get()) == NULL)
1868                 return (NULL);
1869         MPASS(strlen(description) < MAX_W_NAME);
1870         strcpy(w->w_name, description);
1871         w->w_class = lock_class;
1872         w->w_refcount = 1;
1873         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1874         if (lock_class->lc_flags & LC_SPINLOCK) {
1875                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1876                 w_spin_cnt++;
1877         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1878                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1879                 w_sleep_cnt++;
1880         }
1881
1882         /* Insert new witness into the hash */
1883         witness_hash_put(w);
1884         witness_increment_graph_generation();
1885         mtx_unlock_spin(&w_mtx);
1886         return (w);
1887 found:
1888         w->w_refcount++;
1889         if (w->w_refcount == 1)
1890                 w->w_class = lock_class;
1891         mtx_unlock_spin(&w_mtx);
1892         if (lock_class != w->w_class)
1893                 kassert_panic(
1894                     "lock (%s) %s does not match earlier (%s) lock",
1895                     description, lock_class->lc_name,
1896                     w->w_class->lc_name);
1897         return (w);
1898 }
1899
1900 static void
1901 depart(struct witness *w)
1902 {
1903
1904         MPASS(w->w_refcount == 0);
1905         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1906                 w_sleep_cnt--;
1907         } else {
1908                 w_spin_cnt--;
1909         }
1910         /*
1911          * Set file to NULL as it may point into a loadable module.
1912          */
1913         w->w_file = NULL;
1914         w->w_line = 0;
1915         witness_increment_graph_generation();
1916 }
1917
1918
1919 static void
1920 adopt(struct witness *parent, struct witness *child)
1921 {
1922         int pi, ci, i, j;
1923
1924         if (witness_cold == 0)
1925                 mtx_assert(&w_mtx, MA_OWNED);
1926
1927         /* If the relationship is already known, there's no work to be done. */
1928         if (isitmychild(parent, child))
1929                 return;
1930
1931         /* When the structure of the graph changes, bump up the generation. */
1932         witness_increment_graph_generation();
1933
1934         /*
1935          * The hard part ... create the direct relationship, then propagate all
1936          * indirect relationships.
1937          */
1938         pi = parent->w_index;
1939         ci = child->w_index;
1940         WITNESS_INDEX_ASSERT(pi);
1941         WITNESS_INDEX_ASSERT(ci);
1942         MPASS(pi != ci);
1943         w_rmatrix[pi][ci] |= WITNESS_PARENT;
1944         w_rmatrix[ci][pi] |= WITNESS_CHILD;
1945
1946         /*
1947          * If parent was not already an ancestor of child,
1948          * then we increment the descendant and ancestor counters.
1949          */
1950         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1951                 parent->w_num_descendants++;
1952                 child->w_num_ancestors++;
1953         }
1954
1955         /*
1956          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1957          * an ancestor of 'pi' during this loop.
1958          */
1959         for (i = 1; i <= w_max_used_index; i++) {
1960                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1961                     (i != pi))
1962                         continue;
1963
1964                 /* Find each descendant of 'i' and mark it as a descendant. */
1965                 for (j = 1; j <= w_max_used_index; j++) {
1966
1967                         /*
1968                          * Skip children that are already marked as
1969                          * descendants of 'i'.
1970                          */
1971                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1972                                 continue;
1973
1974                         /*
1975                          * We are only interested in descendants of 'ci'. Note
1976                          * that 'ci' itself is counted as a descendant of 'ci'.
1977                          */
1978                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1979                             (j != ci))
1980                                 continue;
1981                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1982                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1983                         w_data[i].w_num_descendants++;
1984                         w_data[j].w_num_ancestors++;
1985
1986                         /*
1987                          * Make sure we aren't marking a node as both an
1988                          * ancestor and descendant. We should have caught
1989                          * this as a lock order reversal earlier.
1990                          */
1991                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1992                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1993                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1994                                     "both ancestor and descendant\n",
1995                                     i, j, w_rmatrix[i][j]);
1996                                 kdb_backtrace();
1997                                 printf("Witness disabled.\n");
1998                                 witness_watch = -1;
1999                         }
2000                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2001                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2002                                 printf("witness rmatrix paradox! [%d][%d]=%d "
2003                                     "both ancestor and descendant\n",
2004                                     j, i, w_rmatrix[j][i]);
2005                                 kdb_backtrace();
2006                                 printf("Witness disabled.\n");
2007                                 witness_watch = -1;
2008                         }
2009                 }
2010         }
2011 }
2012
2013 static void
2014 itismychild(struct witness *parent, struct witness *child)
2015 {
2016         int unlocked;
2017
2018         MPASS(child != NULL && parent != NULL);
2019         if (witness_cold == 0)
2020                 mtx_assert(&w_mtx, MA_OWNED);
2021
2022         if (!witness_lock_type_equal(parent, child)) {
2023                 if (witness_cold == 0) {
2024                         unlocked = 1;
2025                         mtx_unlock_spin(&w_mtx);
2026                 } else {
2027                         unlocked = 0;
2028                 }
2029                 kassert_panic(
2030                     "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2031                     "the same lock type", __func__, parent->w_name,
2032                     parent->w_class->lc_name, child->w_name,
2033                     child->w_class->lc_name);
2034                 if (unlocked)
2035                         mtx_lock_spin(&w_mtx);
2036         }
2037         adopt(parent, child);
2038 }
2039
2040 /*
2041  * Generic code for the isitmy*() functions. The rmask parameter is the
2042  * expected relationship of w1 to w2.
2043  */
2044 static int
2045 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2046 {
2047         unsigned char r1, r2;
2048         int i1, i2;
2049
2050         i1 = w1->w_index;
2051         i2 = w2->w_index;
2052         WITNESS_INDEX_ASSERT(i1);
2053         WITNESS_INDEX_ASSERT(i2);
2054         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2055         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2056
2057         /* The flags on one better be the inverse of the flags on the other */
2058         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2059             (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2060                 /* Don't squawk if we're potentially racing with an update. */
2061                 if (!mtx_owned(&w_mtx))
2062                         return (0);
2063                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2064                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
2065                     "w_rmatrix[%d][%d] == %hhx\n",
2066                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2067                     i2, i1, r2);
2068                 kdb_backtrace();
2069                 printf("Witness disabled.\n");
2070                 witness_watch = -1;
2071         }
2072         return (r1 & rmask);
2073 }
2074
2075 /*
2076  * Checks if @child is a direct child of @parent.
2077  */
2078 static int
2079 isitmychild(struct witness *parent, struct witness *child)
2080 {
2081
2082         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2083 }
2084
2085 /*
2086  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2087  */
2088 static int
2089 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2090 {
2091
2092         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2093             __func__));
2094 }
2095
2096 static int
2097 blessed(struct witness *w1, struct witness *w2)
2098 {
2099         int i;
2100         struct witness_blessed *b;
2101
2102         for (i = 0; i < nitems(blessed_list); i++) {
2103                 b = &blessed_list[i];
2104                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2105                         if (strcmp(w2->w_name, b->b_lock2) == 0)
2106                                 return (1);
2107                         continue;
2108                 }
2109                 if (strcmp(w1->w_name, b->b_lock2) == 0)
2110                         if (strcmp(w2->w_name, b->b_lock1) == 0)
2111                                 return (1);
2112         }
2113         return (0);
2114 }
2115
2116 static struct witness *
2117 witness_get(void)
2118 {
2119         struct witness *w;
2120         int index;
2121
2122         if (witness_cold == 0)
2123                 mtx_assert(&w_mtx, MA_OWNED);
2124
2125         if (witness_watch == -1) {
2126                 mtx_unlock_spin(&w_mtx);
2127                 return (NULL);
2128         }
2129         if (STAILQ_EMPTY(&w_free)) {
2130                 witness_watch = -1;
2131                 mtx_unlock_spin(&w_mtx);
2132                 printf("WITNESS: unable to allocate a new witness object\n");
2133                 return (NULL);
2134         }
2135         w = STAILQ_FIRST(&w_free);
2136         STAILQ_REMOVE_HEAD(&w_free, w_list);
2137         w_free_cnt--;
2138         index = w->w_index;
2139         MPASS(index > 0 && index == w_max_used_index+1 &&
2140             index < witness_count);
2141         bzero(w, sizeof(*w));
2142         w->w_index = index;
2143         if (index > w_max_used_index)
2144                 w_max_used_index = index;
2145         return (w);
2146 }
2147
2148 static void
2149 witness_free(struct witness *w)
2150 {
2151
2152         STAILQ_INSERT_HEAD(&w_free, w, w_list);
2153         w_free_cnt++;
2154 }
2155
2156 static struct lock_list_entry *
2157 witness_lock_list_get(void)
2158 {
2159         struct lock_list_entry *lle;
2160
2161         if (witness_watch == -1)
2162                 return (NULL);
2163         mtx_lock_spin(&w_mtx);
2164         lle = w_lock_list_free;
2165         if (lle == NULL) {
2166                 witness_watch = -1;
2167                 mtx_unlock_spin(&w_mtx);
2168                 printf("%s: witness exhausted\n", __func__);
2169                 return (NULL);
2170         }
2171         w_lock_list_free = lle->ll_next;
2172         mtx_unlock_spin(&w_mtx);
2173         bzero(lle, sizeof(*lle));
2174         return (lle);
2175 }
2176
2177 static void
2178 witness_lock_list_free(struct lock_list_entry *lle)
2179 {
2180
2181         mtx_lock_spin(&w_mtx);
2182         lle->ll_next = w_lock_list_free;
2183         w_lock_list_free = lle;
2184         mtx_unlock_spin(&w_mtx);
2185 }
2186
2187 static struct lock_instance *
2188 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2189 {
2190         struct lock_list_entry *lle;
2191         struct lock_instance *instance;
2192         int i;
2193
2194         for (lle = list; lle != NULL; lle = lle->ll_next)
2195                 for (i = lle->ll_count - 1; i >= 0; i--) {
2196                         instance = &lle->ll_children[i];
2197                         if (instance->li_lock == lock)
2198                                 return (instance);
2199                 }
2200         return (NULL);
2201 }
2202
2203 static void
2204 witness_list_lock(struct lock_instance *instance,
2205     int (*prnt)(const char *fmt, ...))
2206 {
2207         struct lock_object *lock;
2208
2209         lock = instance->li_lock;
2210         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2211             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2212         if (lock->lo_witness->w_name != lock->lo_name)
2213                 prnt(" (%s)", lock->lo_witness->w_name);
2214         prnt(" r = %d (%p) locked @ %s:%d\n",
2215             instance->li_flags & LI_RECURSEMASK, lock,
2216             fixup_filename(instance->li_file), instance->li_line);
2217 }
2218
2219 static int
2220 witness_output(const char *fmt, ...)
2221 {
2222         va_list ap;
2223         int ret;
2224
2225         va_start(ap, fmt);
2226         ret = witness_voutput(fmt, ap);
2227         va_end(ap);
2228         return (ret);
2229 }
2230
2231 static int
2232 witness_voutput(const char *fmt, va_list ap)
2233 {
2234         int ret;
2235
2236         ret = 0;
2237         switch (witness_channel) {
2238         case WITNESS_CONSOLE:
2239                 ret = vprintf(fmt, ap);
2240                 break;
2241         case WITNESS_LOG:
2242                 vlog(LOG_NOTICE, fmt, ap);
2243                 break;
2244         case WITNESS_NONE:
2245                 break;
2246         }
2247         return (ret);
2248 }
2249
2250 #ifdef DDB
2251 static int
2252 witness_thread_has_locks(struct thread *td)
2253 {
2254
2255         if (td->td_sleeplocks == NULL)
2256                 return (0);
2257         return (td->td_sleeplocks->ll_count != 0);
2258 }
2259
2260 static int
2261 witness_proc_has_locks(struct proc *p)
2262 {
2263         struct thread *td;
2264
2265         FOREACH_THREAD_IN_PROC(p, td) {
2266                 if (witness_thread_has_locks(td))
2267                         return (1);
2268         }
2269         return (0);
2270 }
2271 #endif
2272
2273 int
2274 witness_list_locks(struct lock_list_entry **lock_list,
2275     int (*prnt)(const char *fmt, ...))
2276 {
2277         struct lock_list_entry *lle;
2278         int i, nheld;
2279
2280         nheld = 0;
2281         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2282                 for (i = lle->ll_count - 1; i >= 0; i--) {
2283                         witness_list_lock(&lle->ll_children[i], prnt);
2284                         nheld++;
2285                 }
2286         return (nheld);
2287 }
2288
2289 /*
2290  * This is a bit risky at best.  We call this function when we have timed
2291  * out acquiring a spin lock, and we assume that the other CPU is stuck
2292  * with this lock held.  So, we go groveling around in the other CPU's
2293  * per-cpu data to try to find the lock instance for this spin lock to
2294  * see when it was last acquired.
2295  */
2296 void
2297 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2298     int (*prnt)(const char *fmt, ...))
2299 {
2300         struct lock_instance *instance;
2301         struct pcpu *pc;
2302
2303         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2304                 return;
2305         pc = pcpu_find(owner->td_oncpu);
2306         instance = find_instance(pc->pc_spinlocks, lock);
2307         if (instance != NULL)
2308                 witness_list_lock(instance, prnt);
2309 }
2310
2311 void
2312 witness_save(struct lock_object *lock, const char **filep, int *linep)
2313 {
2314         struct lock_list_entry *lock_list;
2315         struct lock_instance *instance;
2316         struct lock_class *class;
2317
2318         /*
2319          * This function is used independently in locking code to deal with
2320          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2321          * is gone.
2322          */
2323         if (SCHEDULER_STOPPED())
2324                 return;
2325         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2326         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2327                 return;
2328         class = LOCK_CLASS(lock);
2329         if (class->lc_flags & LC_SLEEPLOCK)
2330                 lock_list = curthread->td_sleeplocks;
2331         else {
2332                 if (witness_skipspin)
2333                         return;
2334                 lock_list = PCPU_GET(spinlocks);
2335         }
2336         instance = find_instance(lock_list, lock);
2337         if (instance == NULL) {
2338                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2339                     class->lc_name, lock->lo_name);
2340                 return;
2341         }
2342         *filep = instance->li_file;
2343         *linep = instance->li_line;
2344 }
2345
2346 void
2347 witness_restore(struct lock_object *lock, const char *file, int line)
2348 {
2349         struct lock_list_entry *lock_list;
2350         struct lock_instance *instance;
2351         struct lock_class *class;
2352
2353         /*
2354          * This function is used independently in locking code to deal with
2355          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2356          * is gone.
2357          */
2358         if (SCHEDULER_STOPPED())
2359                 return;
2360         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2361         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2362                 return;
2363         class = LOCK_CLASS(lock);
2364         if (class->lc_flags & LC_SLEEPLOCK)
2365                 lock_list = curthread->td_sleeplocks;
2366         else {
2367                 if (witness_skipspin)
2368                         return;
2369                 lock_list = PCPU_GET(spinlocks);
2370         }
2371         instance = find_instance(lock_list, lock);
2372         if (instance == NULL)
2373                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2374                     class->lc_name, lock->lo_name);
2375         lock->lo_witness->w_file = file;
2376         lock->lo_witness->w_line = line;
2377         if (instance == NULL)
2378                 return;
2379         instance->li_file = file;
2380         instance->li_line = line;
2381 }
2382
2383 void
2384 witness_assert(const struct lock_object *lock, int flags, const char *file,
2385     int line)
2386 {
2387 #ifdef INVARIANT_SUPPORT
2388         struct lock_instance *instance;
2389         struct lock_class *class;
2390
2391         if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2392                 return;
2393         class = LOCK_CLASS(lock);
2394         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2395                 instance = find_instance(curthread->td_sleeplocks, lock);
2396         else if ((class->lc_flags & LC_SPINLOCK) != 0)
2397                 instance = find_instance(PCPU_GET(spinlocks), lock);
2398         else {
2399                 kassert_panic("Lock (%s) %s is not sleep or spin!",
2400                     class->lc_name, lock->lo_name);
2401                 return;
2402         }
2403         switch (flags) {
2404         case LA_UNLOCKED:
2405                 if (instance != NULL)
2406                         kassert_panic("Lock (%s) %s locked @ %s:%d.",
2407                             class->lc_name, lock->lo_name,
2408                             fixup_filename(file), line);
2409                 break;
2410         case LA_LOCKED:
2411         case LA_LOCKED | LA_RECURSED:
2412         case LA_LOCKED | LA_NOTRECURSED:
2413         case LA_SLOCKED:
2414         case LA_SLOCKED | LA_RECURSED:
2415         case LA_SLOCKED | LA_NOTRECURSED:
2416         case LA_XLOCKED:
2417         case LA_XLOCKED | LA_RECURSED:
2418         case LA_XLOCKED | LA_NOTRECURSED:
2419                 if (instance == NULL) {
2420                         kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2421                             class->lc_name, lock->lo_name,
2422                             fixup_filename(file), line);
2423                         break;
2424                 }
2425                 if ((flags & LA_XLOCKED) != 0 &&
2426                     (instance->li_flags & LI_EXCLUSIVE) == 0)
2427                         kassert_panic(
2428                             "Lock (%s) %s not exclusively locked @ %s:%d.",
2429                             class->lc_name, lock->lo_name,
2430                             fixup_filename(file), line);
2431                 if ((flags & LA_SLOCKED) != 0 &&
2432                     (instance->li_flags & LI_EXCLUSIVE) != 0)
2433                         kassert_panic(
2434                             "Lock (%s) %s exclusively locked @ %s:%d.",
2435                             class->lc_name, lock->lo_name,
2436                             fixup_filename(file), line);
2437                 if ((flags & LA_RECURSED) != 0 &&
2438                     (instance->li_flags & LI_RECURSEMASK) == 0)
2439                         kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2440                             class->lc_name, lock->lo_name,
2441                             fixup_filename(file), line);
2442                 if ((flags & LA_NOTRECURSED) != 0 &&
2443                     (instance->li_flags & LI_RECURSEMASK) != 0)
2444                         kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2445                             class->lc_name, lock->lo_name,
2446                             fixup_filename(file), line);
2447                 break;
2448         default:
2449                 kassert_panic("Invalid lock assertion at %s:%d.",
2450                     fixup_filename(file), line);
2451
2452         }
2453 #endif  /* INVARIANT_SUPPORT */
2454 }
2455
2456 static void
2457 witness_setflag(struct lock_object *lock, int flag, int set)
2458 {
2459         struct lock_list_entry *lock_list;
2460         struct lock_instance *instance;
2461         struct lock_class *class;
2462
2463         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2464                 return;
2465         class = LOCK_CLASS(lock);
2466         if (class->lc_flags & LC_SLEEPLOCK)
2467                 lock_list = curthread->td_sleeplocks;
2468         else {
2469                 if (witness_skipspin)
2470                         return;
2471                 lock_list = PCPU_GET(spinlocks);
2472         }
2473         instance = find_instance(lock_list, lock);
2474         if (instance == NULL) {
2475                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2476                     class->lc_name, lock->lo_name);
2477                 return;
2478         }
2479
2480         if (set)
2481                 instance->li_flags |= flag;
2482         else
2483                 instance->li_flags &= ~flag;
2484 }
2485
2486 void
2487 witness_norelease(struct lock_object *lock)
2488 {
2489
2490         witness_setflag(lock, LI_NORELEASE, 1);
2491 }
2492
2493 void
2494 witness_releaseok(struct lock_object *lock)
2495 {
2496
2497         witness_setflag(lock, LI_NORELEASE, 0);
2498 }
2499
2500 #ifdef DDB
2501 static void
2502 witness_ddb_list(struct thread *td)
2503 {
2504
2505         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2506         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2507
2508         if (witness_watch < 1)
2509                 return;
2510
2511         witness_list_locks(&td->td_sleeplocks, db_printf);
2512
2513         /*
2514          * We only handle spinlocks if td == curthread.  This is somewhat broken
2515          * if td is currently executing on some other CPU and holds spin locks
2516          * as we won't display those locks.  If we had a MI way of getting
2517          * the per-cpu data for a given cpu then we could use
2518          * td->td_oncpu to get the list of spinlocks for this thread
2519          * and "fix" this.
2520          *
2521          * That still wouldn't really fix this unless we locked the scheduler
2522          * lock or stopped the other CPU to make sure it wasn't changing the
2523          * list out from under us.  It is probably best to just not try to
2524          * handle threads on other CPU's for now.
2525          */
2526         if (td == curthread && PCPU_GET(spinlocks) != NULL)
2527                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2528 }
2529
2530 DB_SHOW_COMMAND(locks, db_witness_list)
2531 {
2532         struct thread *td;
2533
2534         if (have_addr)
2535                 td = db_lookup_thread(addr, true);
2536         else
2537                 td = kdb_thread;
2538         witness_ddb_list(td);
2539 }
2540
2541 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2542 {
2543         struct thread *td;
2544         struct proc *p;
2545
2546         /*
2547          * It would be nice to list only threads and processes that actually
2548          * held sleep locks, but that information is currently not exported
2549          * by WITNESS.
2550          */
2551         FOREACH_PROC_IN_SYSTEM(p) {
2552                 if (!witness_proc_has_locks(p))
2553                         continue;
2554                 FOREACH_THREAD_IN_PROC(p, td) {
2555                         if (!witness_thread_has_locks(td))
2556                                 continue;
2557                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2558                             p->p_comm, td, td->td_tid);
2559                         witness_ddb_list(td);
2560                         if (db_pager_quit)
2561                                 return;
2562                 }
2563         }
2564 }
2565 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2566
2567 DB_SHOW_COMMAND(witness, db_witness_display)
2568 {
2569
2570         witness_ddb_display(db_printf);
2571 }
2572 #endif
2573
2574 static void
2575 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2576 {
2577         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2578         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2579         int generation, i, j;
2580
2581         tmp_data1 = NULL;
2582         tmp_data2 = NULL;
2583         tmp_w1 = NULL;
2584         tmp_w2 = NULL;
2585
2586         /* Allocate and init temporary storage space. */
2587         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2588         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2589         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2590             M_WAITOK | M_ZERO);
2591         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2592             M_WAITOK | M_ZERO);
2593         stack_zero(&tmp_data1->wlod_stack);
2594         stack_zero(&tmp_data2->wlod_stack);
2595
2596 restart:
2597         mtx_lock_spin(&w_mtx);
2598         generation = w_generation;
2599         mtx_unlock_spin(&w_mtx);
2600         sbuf_printf(sb, "Number of known direct relationships is %d\n",
2601             w_lohash.wloh_count);
2602         for (i = 1; i < w_max_used_index; i++) {
2603                 mtx_lock_spin(&w_mtx);
2604                 if (generation != w_generation) {
2605                         mtx_unlock_spin(&w_mtx);
2606
2607                         /* The graph has changed, try again. */
2608                         *oldidx = 0;
2609                         sbuf_clear(sb);
2610                         goto restart;
2611                 }
2612
2613                 w1 = &w_data[i];
2614                 if (w1->w_reversed == 0) {
2615                         mtx_unlock_spin(&w_mtx);
2616                         continue;
2617                 }
2618
2619                 /* Copy w1 locally so we can release the spin lock. */
2620                 *tmp_w1 = *w1;
2621                 mtx_unlock_spin(&w_mtx);
2622
2623                 if (tmp_w1->w_reversed == 0)
2624                         continue;
2625                 for (j = 1; j < w_max_used_index; j++) {
2626                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2627                                 continue;
2628
2629                         mtx_lock_spin(&w_mtx);
2630                         if (generation != w_generation) {
2631                                 mtx_unlock_spin(&w_mtx);
2632
2633                                 /* The graph has changed, try again. */
2634                                 *oldidx = 0;
2635                                 sbuf_clear(sb);
2636                                 goto restart;
2637                         }
2638
2639                         w2 = &w_data[j];
2640                         data1 = witness_lock_order_get(w1, w2);
2641                         data2 = witness_lock_order_get(w2, w1);
2642
2643                         /*
2644                          * Copy information locally so we can release the
2645                          * spin lock.
2646                          */
2647                         *tmp_w2 = *w2;
2648
2649                         if (data1) {
2650                                 stack_zero(&tmp_data1->wlod_stack);
2651                                 stack_copy(&data1->wlod_stack,
2652                                     &tmp_data1->wlod_stack);
2653                         }
2654                         if (data2 && data2 != data1) {
2655                                 stack_zero(&tmp_data2->wlod_stack);
2656                                 stack_copy(&data2->wlod_stack,
2657                                     &tmp_data2->wlod_stack);
2658                         }
2659                         mtx_unlock_spin(&w_mtx);
2660
2661                         if (blessed(tmp_w1, tmp_w2))
2662                                 continue;
2663
2664                         sbuf_printf(sb,
2665             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2666                             tmp_w1->w_name, tmp_w1->w_class->lc_name,
2667                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
2668                         if (data1) {
2669                                 sbuf_printf(sb,
2670                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2671                                     tmp_w1->w_name, tmp_w1->w_class->lc_name,
2672                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
2673                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2674                                 sbuf_printf(sb, "\n");
2675                         }
2676                         if (data2 && data2 != data1) {
2677                                 sbuf_printf(sb,
2678                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2679                                     tmp_w2->w_name, tmp_w2->w_class->lc_name,
2680                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
2681                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2682                                 sbuf_printf(sb, "\n");
2683                         }
2684                 }
2685         }
2686         mtx_lock_spin(&w_mtx);
2687         if (generation != w_generation) {
2688                 mtx_unlock_spin(&w_mtx);
2689
2690                 /*
2691                  * The graph changed while we were printing stack data,
2692                  * try again.
2693                  */
2694                 *oldidx = 0;
2695                 sbuf_clear(sb);
2696                 goto restart;
2697         }
2698         mtx_unlock_spin(&w_mtx);
2699
2700         /* Free temporary storage space. */
2701         free(tmp_data1, M_TEMP);
2702         free(tmp_data2, M_TEMP);
2703         free(tmp_w1, M_TEMP);
2704         free(tmp_w2, M_TEMP);
2705 }
2706
2707 static int
2708 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2709 {
2710         struct sbuf *sb;
2711         int error;
2712
2713         if (witness_watch < 1) {
2714                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2715                 return (error);
2716         }
2717         if (witness_cold) {
2718                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2719                 return (error);
2720         }
2721         error = 0;
2722         sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2723         if (sb == NULL)
2724                 return (ENOMEM);
2725
2726         sbuf_print_witness_badstacks(sb, &req->oldidx);
2727
2728         sbuf_finish(sb);
2729         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2730         sbuf_delete(sb);
2731
2732         return (error);
2733 }
2734
2735 #ifdef DDB
2736 static int
2737 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2738 {
2739
2740         return (db_printf("%.*s", len, data));
2741 }
2742
2743 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2744 {
2745         struct sbuf sb;
2746         char buffer[128];
2747         size_t dummy;
2748
2749         sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2750         sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2751         sbuf_print_witness_badstacks(&sb, &dummy);
2752         sbuf_finish(&sb);
2753 }
2754 #endif
2755
2756 static int
2757 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2758 {
2759         static const struct {
2760                 enum witness_channel channel;
2761                 const char *name;
2762         } channels[] = {
2763                 { WITNESS_CONSOLE, "console" },
2764                 { WITNESS_LOG, "log" },
2765                 { WITNESS_NONE, "none" },
2766         };
2767         char buf[16];
2768         u_int i;
2769         int error;
2770
2771         buf[0] = '\0';
2772         for (i = 0; i < nitems(channels); i++)
2773                 if (witness_channel == channels[i].channel) {
2774                         snprintf(buf, sizeof(buf), "%s", channels[i].name);
2775                         break;
2776                 }
2777
2778         error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2779         if (error != 0 || req->newptr == NULL)
2780                 return (error);
2781
2782         error = EINVAL;
2783         for (i = 0; i < nitems(channels); i++)
2784                 if (strcmp(channels[i].name, buf) == 0) {
2785                         witness_channel = channels[i].channel;
2786                         error = 0;
2787                         break;
2788                 }
2789         return (error);
2790 }
2791
2792 static int
2793 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2794 {
2795         struct witness *w;
2796         struct sbuf *sb;
2797         int error;
2798
2799 #ifdef __i386__
2800         error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2801         return (error);
2802 #endif
2803
2804         if (witness_watch < 1) {
2805                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2806                 return (error);
2807         }
2808         if (witness_cold) {
2809                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2810                 return (error);
2811         }
2812         error = 0;
2813
2814         error = sysctl_wire_old_buffer(req, 0);
2815         if (error != 0)
2816                 return (error);
2817         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2818         if (sb == NULL)
2819                 return (ENOMEM);
2820         sbuf_printf(sb, "\n");
2821
2822         mtx_lock_spin(&w_mtx);
2823         STAILQ_FOREACH(w, &w_all, w_list)
2824                 w->w_displayed = 0;
2825         STAILQ_FOREACH(w, &w_all, w_list)
2826                 witness_add_fullgraph(sb, w);
2827         mtx_unlock_spin(&w_mtx);
2828
2829         /*
2830          * Close the sbuf and return to userland.
2831          */
2832         error = sbuf_finish(sb);
2833         sbuf_delete(sb);
2834
2835         return (error);
2836 }
2837
2838 static int
2839 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2840 {
2841         int error, value;
2842
2843         value = witness_watch;
2844         error = sysctl_handle_int(oidp, &value, 0, req);
2845         if (error != 0 || req->newptr == NULL)
2846                 return (error);
2847         if (value > 1 || value < -1 ||
2848             (witness_watch == -1 && value != witness_watch))
2849                 return (EINVAL);
2850         witness_watch = value;
2851         return (0);
2852 }
2853
2854 static void
2855 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2856 {
2857         int i;
2858
2859         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2860                 return;
2861         w->w_displayed = 1;
2862
2863         WITNESS_INDEX_ASSERT(w->w_index);
2864         for (i = 1; i <= w_max_used_index; i++) {
2865                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2866                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2867                             w_data[i].w_name);
2868                         witness_add_fullgraph(sb, &w_data[i]);
2869                 }
2870         }
2871 }
2872
2873 /*
2874  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2875  * interprets the key as a string and reads until the null
2876  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2877  * hash value computed from the key.
2878  */
2879 static uint32_t
2880 witness_hash_djb2(const uint8_t *key, uint32_t size)
2881 {
2882         unsigned int hash = 5381;
2883         int i;
2884
2885         /* hash = hash * 33 + key[i] */
2886         if (size)
2887                 for (i = 0; i < size; i++)
2888                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2889         else
2890                 for (i = 0; key[i] != 0; i++)
2891                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2892
2893         return (hash);
2894 }
2895
2896
2897 /*
2898  * Initializes the two witness hash tables. Called exactly once from
2899  * witness_initialize().
2900  */
2901 static void
2902 witness_init_hash_tables(void)
2903 {
2904         int i;
2905
2906         MPASS(witness_cold);
2907
2908         /* Initialize the hash tables. */
2909         for (i = 0; i < WITNESS_HASH_SIZE; i++)
2910                 w_hash.wh_array[i] = NULL;
2911
2912         w_hash.wh_size = WITNESS_HASH_SIZE;
2913         w_hash.wh_count = 0;
2914
2915         /* Initialize the lock order data hash. */
2916         w_lofree = NULL;
2917         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2918                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2919                 w_lodata[i].wlod_next = w_lofree;
2920                 w_lofree = &w_lodata[i];
2921         }
2922         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2923         w_lohash.wloh_count = 0;
2924         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2925                 w_lohash.wloh_array[i] = NULL;
2926 }
2927
2928 static struct witness *
2929 witness_hash_get(const char *key)
2930 {
2931         struct witness *w;
2932         uint32_t hash;
2933
2934         MPASS(key != NULL);
2935         if (witness_cold == 0)
2936                 mtx_assert(&w_mtx, MA_OWNED);
2937         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2938         w = w_hash.wh_array[hash];
2939         while (w != NULL) {
2940                 if (strcmp(w->w_name, key) == 0)
2941                         goto out;
2942                 w = w->w_hash_next;
2943         }
2944
2945 out:
2946         return (w);
2947 }
2948
2949 static void
2950 witness_hash_put(struct witness *w)
2951 {
2952         uint32_t hash;
2953
2954         MPASS(w != NULL);
2955         MPASS(w->w_name != NULL);
2956         if (witness_cold == 0)
2957                 mtx_assert(&w_mtx, MA_OWNED);
2958         KASSERT(witness_hash_get(w->w_name) == NULL,
2959             ("%s: trying to add a hash entry that already exists!", __func__));
2960         KASSERT(w->w_hash_next == NULL,
2961             ("%s: w->w_hash_next != NULL", __func__));
2962
2963         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2964         w->w_hash_next = w_hash.wh_array[hash];
2965         w_hash.wh_array[hash] = w;
2966         w_hash.wh_count++;
2967 }
2968
2969
2970 static struct witness_lock_order_data *
2971 witness_lock_order_get(struct witness *parent, struct witness *child)
2972 {
2973         struct witness_lock_order_data *data = NULL;
2974         struct witness_lock_order_key key;
2975         unsigned int hash;
2976
2977         MPASS(parent != NULL && child != NULL);
2978         key.from = parent->w_index;
2979         key.to = child->w_index;
2980         WITNESS_INDEX_ASSERT(key.from);
2981         WITNESS_INDEX_ASSERT(key.to);
2982         if ((w_rmatrix[parent->w_index][child->w_index]
2983             & WITNESS_LOCK_ORDER_KNOWN) == 0)
2984                 goto out;
2985
2986         hash = witness_hash_djb2((const char*)&key,
2987             sizeof(key)) % w_lohash.wloh_size;
2988         data = w_lohash.wloh_array[hash];
2989         while (data != NULL) {
2990                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2991                         break;
2992                 data = data->wlod_next;
2993         }
2994
2995 out:
2996         return (data);
2997 }
2998
2999 /*
3000  * Verify that parent and child have a known relationship, are not the same,
3001  * and child is actually a child of parent.  This is done without w_mtx
3002  * to avoid contention in the common case.
3003  */
3004 static int
3005 witness_lock_order_check(struct witness *parent, struct witness *child)
3006 {
3007
3008         if (parent != child &&
3009             w_rmatrix[parent->w_index][child->w_index]
3010             & WITNESS_LOCK_ORDER_KNOWN &&
3011             isitmychild(parent, child))
3012                 return (1);
3013
3014         return (0);
3015 }
3016
3017 static int
3018 witness_lock_order_add(struct witness *parent, struct witness *child)
3019 {
3020         struct witness_lock_order_data *data = NULL;
3021         struct witness_lock_order_key key;
3022         unsigned int hash;
3023
3024         MPASS(parent != NULL && child != NULL);
3025         key.from = parent->w_index;
3026         key.to = child->w_index;
3027         WITNESS_INDEX_ASSERT(key.from);
3028         WITNESS_INDEX_ASSERT(key.to);
3029         if (w_rmatrix[parent->w_index][child->w_index]
3030             & WITNESS_LOCK_ORDER_KNOWN)
3031                 return (1);
3032
3033         hash = witness_hash_djb2((const char*)&key,
3034             sizeof(key)) % w_lohash.wloh_size;
3035         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3036         data = w_lofree;
3037         if (data == NULL)
3038                 return (0);
3039         w_lofree = data->wlod_next;
3040         data->wlod_next = w_lohash.wloh_array[hash];
3041         data->wlod_key = key;
3042         w_lohash.wloh_array[hash] = data;
3043         w_lohash.wloh_count++;
3044         stack_zero(&data->wlod_stack);
3045         stack_save(&data->wlod_stack);
3046         return (1);
3047 }
3048
3049 /* Call this whenever the structure of the witness graph changes. */
3050 static void
3051 witness_increment_graph_generation(void)
3052 {
3053
3054         if (witness_cold == 0)
3055                 mtx_assert(&w_mtx, MA_OWNED);
3056         w_generation++;
3057 }
3058
3059 static int
3060 witness_output_drain(void *arg __unused, const char *data, int len)
3061 {
3062
3063         witness_output("%.*s", len, data);
3064         return (len);
3065 }
3066
3067 static void
3068 witness_debugger(int cond, const char *msg)
3069 {
3070         char buf[32];
3071         struct sbuf sb;
3072         struct stack st;
3073
3074         if (!cond)
3075                 return;
3076
3077         if (witness_trace) {
3078                 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3079                 sbuf_set_drain(&sb, witness_output_drain, NULL);
3080
3081                 stack_zero(&st);
3082                 stack_save(&st);
3083                 witness_output("stack backtrace:\n");
3084                 stack_sbuf_print_ddb(&sb, &st);
3085
3086                 sbuf_finish(&sb);
3087         }
3088
3089 #ifdef KDB
3090         if (witness_kdb)
3091                 kdb_enter(KDB_WHY_WITNESS, msg);
3092 #endif
3093 }