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