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 FEATURE(witness, "kernel has witness(9) support");
 365
 366 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
 367     "Witness Locking");
 368
 369 /*
 370  * If set to 0, lock order checking is disabled.  If set to -1,
 371  * witness is completely disabled.  Otherwise witness performs full
 372  * lock order checking for all locks.  At runtime, lock order checking
 373  * may be toggled.  However, witness cannot be reenabled once it is
 374  * completely disabled.
 375  */
 376 static int witness_watch = 1;
 377 SYSCTL_PROC(_debug_witness, OID_AUTO, watch,
 378     CTLFLAG_RWTUN | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0,
 379     sysctl_debug_witness_watch, "I",
 380     "witness is watching lock operations");
 381
 382 #ifdef KDB
 383 /*
 384  * When KDB is enabled and witness_kdb is 1, it will cause the system
 385  * to drop into kdebug() when:
 386  *      - a lock hierarchy violation occurs
 387  *      - locks are held when going to sleep.
 388  */
 389 #ifdef WITNESS_KDB
 390 int     witness_kdb = 1;
 391 #else
 392 int     witness_kdb = 0;
 393 #endif
 394 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
 395 #endif /* KDB */
 396
 397 #if defined(DDB) || defined(KDB)
 398 /*
 399  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
 400  * to print a stack trace:
 401  *      - a lock hierarchy violation occurs
 402  *      - locks are held when going to sleep.
 403  */
 404 int     witness_trace = 1;
 405 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
 406 #endif /* DDB || KDB */
 407
 408 #ifdef WITNESS_SKIPSPIN
 409 int     witness_skipspin = 1;
 410 #else
 411 int     witness_skipspin = 0;
 412 #endif
 413 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
 414
 415 int badstack_sbuf_size;
 416
 417 int witness_count = WITNESS_COUNT;
 418 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
 419     &witness_count, 0, "");
 420
 421 /*
 422  * Output channel for witness messages.  By default we print to the console.
 423  */
 424 enum witness_channel {
 425         WITNESS_CONSOLE,
 426         WITNESS_LOG,
 427         WITNESS_NONE,
 428 };
 429
 430 static enum witness_channel witness_channel = WITNESS_CONSOLE;
 431 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel,
 432     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, NULL, 0,
 433     sysctl_debug_witness_channel, "A",
 434     "Output channel for warnings");
 435
 436 /*
 437  * Call this to print out the relations between locks.
 438  */
 439 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph,
 440     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
 441     sysctl_debug_witness_fullgraph, "A",
 442     "Show locks relation graphs");
 443
 444 /*
 445  * Call this to print out the witness faulty stacks.
 446  */
 447 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks,
 448     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
 449     sysctl_debug_witness_badstacks, "A",
 450     "Show bad witness stacks");
 451
 452 static struct mtx w_mtx;
 453
 454 /* w_list */
 455 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 456 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 457
 458 /* w_typelist */
 459 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 460 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 461
 462 /* lock list */
 463 static struct lock_list_entry *w_lock_list_free = NULL;
 464 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
 465 static u_int pending_cnt;
 466
 467 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
 468 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
 469 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
 470 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
 471     "");
 472
 473 static struct witness *w_data;
 474 static uint8_t **w_rmatrix;
 475 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 476 static struct witness_hash w_hash;      /* The witness hash table. */
 477
 478 /* The lock order data hash */
 479 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
 480 static struct witness_lock_order_data *w_lofree = NULL;
 481 static struct witness_lock_order_hash w_lohash;
 482 static int w_max_used_index = 0;
 483 static unsigned int w_generation = 0;
 484 static const char w_notrunning[] = "Witness not running\n";
 485 static const char w_stillcold[] = "Witness is still cold\n";
 486 #ifdef __i386__
 487 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
 488 #endif
 489
 490 static struct witness_order_list_entry order_lists[] = {
 491         /*
 492          * sx locks
 493          */
 494         { "proctree", &lock_class_sx },
 495         { "allproc", &lock_class_sx },
 496         { "allprison", &lock_class_sx },
 497         { NULL, NULL },
 498         /*
 499          * Various mutexes
 500          */
 501         { "Giant", &lock_class_mtx_sleep },
 502         { "pipe mutex", &lock_class_mtx_sleep },
 503         { "sigio lock", &lock_class_mtx_sleep },
 504         { "process group", &lock_class_mtx_sleep },
 505 #ifdef  HWPMC_HOOKS
 506         { "pmc-sleep", &lock_class_mtx_sleep },
 507 #endif
 508         { "process lock", &lock_class_mtx_sleep },
 509         { "session", &lock_class_mtx_sleep },
 510         { "uidinfo hash", &lock_class_rw },
 511         { "time lock", &lock_class_mtx_sleep },
 512         { NULL, NULL },
 513         /*
 514          * umtx
 515          */
 516         { "umtx lock", &lock_class_mtx_sleep },
 517         { NULL, NULL },
 518         /*
 519          * Sockets
 520          */
 521         { "accept", &lock_class_mtx_sleep },
 522         { "so_snd", &lock_class_mtx_sleep },
 523         { "so_rcv", &lock_class_mtx_sleep },
 524         { "sellck", &lock_class_mtx_sleep },
 525         { NULL, NULL },
 526         /*
 527          * Routing
 528          */
 529         { "so_rcv", &lock_class_mtx_sleep },
 530         { "radix node head", &lock_class_rm },
 531         { "rtentry", &lock_class_mtx_sleep },
 532         { "ifaddr", &lock_class_mtx_sleep },
 533         { NULL, NULL },
 534         /*
 535          * IPv4 multicast:
 536          * protocol locks before interface locks, after UDP locks.
 537          */
 538         { "in_multi_sx", &lock_class_sx },
 539         { "udpinp", &lock_class_rw },
 540         { "in_multi_list_mtx", &lock_class_mtx_sleep },
 541         { "igmp_mtx", &lock_class_mtx_sleep },
 542         { "ifnet_rw", &lock_class_rw },
 543         { "if_addr_lock", &lock_class_mtx_sleep },
 544         { NULL, NULL },
 545         /*
 546          * IPv6 multicast:
 547          * protocol locks before interface locks, after UDP locks.
 548          */
 549         { "in6_multi_sx", &lock_class_sx },
 550         { "udpinp", &lock_class_rw },
 551         { "in6_multi_list_mtx", &lock_class_mtx_sleep },
 552         { "mld_mtx", &lock_class_mtx_sleep },
 553         { "ifnet_rw", &lock_class_rw },
 554         { "if_addr_lock", &lock_class_mtx_sleep },
 555         { NULL, NULL },
 556         /*
 557          * UNIX Domain Sockets
 558          */
 559         { "unp_link_rwlock", &lock_class_rw },
 560         { "unp_list_lock", &lock_class_mtx_sleep },
 561         { "unp", &lock_class_mtx_sleep },
 562         { "so_snd", &lock_class_mtx_sleep },
 563         { NULL, NULL },
 564         /*
 565          * UDP/IP
 566          */
 567         { "udp", &lock_class_mtx_sleep },
 568         { "udpinp", &lock_class_rw },
 569         { "so_snd", &lock_class_mtx_sleep },
 570         { NULL, NULL },
 571         /*
 572          * TCP/IP
 573          */
 574         { "tcp", &lock_class_mtx_sleep },
 575         { "tcpinp", &lock_class_rw },
 576         { "so_snd", &lock_class_mtx_sleep },
 577         { NULL, NULL },
 578         /*
 579          * BPF
 580          */
 581         { "bpf global lock", &lock_class_sx },
 582         { "bpf cdev lock", &lock_class_mtx_sleep },
 583         { NULL, NULL },
 584         /*
 585          * NFS server
 586          */
 587         { "nfsd_mtx", &lock_class_mtx_sleep },
 588         { "so_snd", &lock_class_mtx_sleep },
 589         { NULL, NULL },
 590
 591         /*
 592          * IEEE 802.11
 593          */
 594         { "802.11 com lock", &lock_class_mtx_sleep},
 595         { NULL, NULL },
 596         /*
 597          * Network drivers
 598          */
 599         { "network driver", &lock_class_mtx_sleep},
 600         { NULL, NULL },
 601
 602         /*
 603          * Netgraph
 604          */
 605         { "ng_node", &lock_class_mtx_sleep },
 606         { "ng_worklist", &lock_class_mtx_sleep },
 607         { NULL, NULL },
 608         /*
 609          * CDEV
 610          */
 611         { "vm map (system)", &lock_class_mtx_sleep },
 612         { "vnode interlock", &lock_class_mtx_sleep },
 613         { "cdev", &lock_class_mtx_sleep },
 614         { "devthrd", &lock_class_mtx_sleep },
 615         { NULL, NULL },
 616         /*
 617          * VM
 618          */
 619         { "vm map (user)", &lock_class_sx },
 620         { "vm object", &lock_class_rw },
 621         { "vm page", &lock_class_mtx_sleep },
 622         { "pmap pv global", &lock_class_rw },
 623         { "pmap", &lock_class_mtx_sleep },
 624         { "pmap pv list", &lock_class_rw },
 625         { "vm page free queue", &lock_class_mtx_sleep },
 626         { "vm pagequeue", &lock_class_mtx_sleep },
 627         { NULL, NULL },
 628         /*
 629          * kqueue/VFS interaction
 630          */
 631         { "kqueue", &lock_class_mtx_sleep },
 632         { "struct mount mtx", &lock_class_mtx_sleep },
 633         { "vnode interlock", &lock_class_mtx_sleep },
 634         { NULL, NULL },
 635         /*
 636          * VFS namecache
 637          */
 638         { "ncvn", &lock_class_mtx_sleep },
 639         { "ncbuc", &lock_class_rw },
 640         { "vnode interlock", &lock_class_mtx_sleep },
 641         { "ncneg", &lock_class_mtx_sleep },
 642         { NULL, NULL },
 643         /*
 644          * ZFS locking
 645          */
 646         { "dn->dn_mtx", &lock_class_sx },
 647         { "dr->dt.di.dr_mtx", &lock_class_sx },
 648         { "db->db_mtx", &lock_class_sx },
 649         { NULL, NULL },
 650         /*
 651          * TCP log locks
 652          */
 653         { "TCP ID tree", &lock_class_rw },
 654         { "tcp log id bucket", &lock_class_mtx_sleep },
 655         { "tcpinp", &lock_class_rw },
 656         { "TCP log expireq", &lock_class_mtx_sleep },
 657         { NULL, NULL },
 658         /*
 659          * spin locks
 660          */
 661 #ifdef SMP
 662         { "ap boot", &lock_class_mtx_spin },
 663 #endif
 664         { "rm.mutex_mtx", &lock_class_mtx_spin },
 665         { "sio", &lock_class_mtx_spin },
 666 #ifdef __i386__
 667         { "cy", &lock_class_mtx_spin },
 668 #endif
 669         { "scc_hwmtx", &lock_class_mtx_spin },
 670         { "uart_hwmtx", &lock_class_mtx_spin },
 671         { "fast_taskqueue", &lock_class_mtx_spin },
 672         { "intr table", &lock_class_mtx_spin },
 673         { "process slock", &lock_class_mtx_spin },
 674         { "syscons video lock", &lock_class_mtx_spin },
 675         { "sleepq chain", &lock_class_mtx_spin },
 676         { "rm_spinlock", &lock_class_mtx_spin },
 677         { "turnstile chain", &lock_class_mtx_spin },
 678         { "turnstile lock", &lock_class_mtx_spin },
 679         { "sched lock", &lock_class_mtx_spin },
 680         { "td_contested", &lock_class_mtx_spin },
 681         { "callout", &lock_class_mtx_spin },
 682         { "entropy harvest mutex", &lock_class_mtx_spin },
 683 #ifdef SMP
 684         { "smp rendezvous", &lock_class_mtx_spin },
 685 #endif
 686 #ifdef __powerpc__
 687         { "tlb0", &lock_class_mtx_spin },
 688 #endif
 689         { NULL, NULL },
 690         { "sched lock", &lock_class_mtx_spin },
 691 #ifdef  HWPMC_HOOKS
 692         { "pmc-per-proc", &lock_class_mtx_spin },
 693 #endif
 694         { NULL, NULL },
 695         /*
 696          * leaf locks
 697          */
 698         { "intrcnt", &lock_class_mtx_spin },
 699         { "icu", &lock_class_mtx_spin },
 700 #ifdef __i386__
 701         { "allpmaps", &lock_class_mtx_spin },
 702         { "descriptor tables", &lock_class_mtx_spin },
 703 #endif
 704         { "clk", &lock_class_mtx_spin },
 705         { "cpuset", &lock_class_mtx_spin },
 706         { "mprof lock", &lock_class_mtx_spin },
 707         { "zombie lock", &lock_class_mtx_spin },
 708         { "ALD Queue", &lock_class_mtx_spin },
 709 #if defined(__i386__) || defined(__amd64__)
 710         { "pcicfg", &lock_class_mtx_spin },
 711         { "NDIS thread lock", &lock_class_mtx_spin },
 712 #endif
 713         { "tw_osl_io_lock", &lock_class_mtx_spin },
 714         { "tw_osl_q_lock", &lock_class_mtx_spin },
 715         { "tw_cl_io_lock", &lock_class_mtx_spin },
 716         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 717         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 718 #ifdef  HWPMC_HOOKS
 719         { "pmc-leaf", &lock_class_mtx_spin },
 720 #endif
 721         { "blocked lock", &lock_class_mtx_spin },
 722         { NULL, NULL },
 723         { NULL, NULL }
 724 };
 725
 726 /*
 727  * Pairs of locks which have been blessed.  Witness does not complain about
 728  * order problems with blessed lock pairs.  Please do not add an entry to the
 729  * table without an explanatory comment.
 730  */
 731 static struct witness_blessed blessed_list[] = {
 732         /*
 733          * See the comment in ufs_dirhash.c.  Basically, a vnode lock serializes
 734          * both lock orders, so a deadlock cannot happen as a result of this
 735          * LOR.
 736          */
 737         { "dirhash",    "bufwait" },
 738
 739         /*
 740          * A UFS vnode may be locked in vget() while a buffer belonging to the
 741          * parent directory vnode is locked.
 742          */
 743         { "ufs",        "bufwait" },
 744 };
 745
 746 /*
 747  * This global is set to 0 once it becomes safe to use the witness code.
 748  */
 749 static int witness_cold = 1;
 750
 751 /*
 752  * This global is set to 1 once the static lock orders have been enrolled
 753  * so that a warning can be issued for any spin locks enrolled later.
 754  */
 755 static int witness_spin_warn = 0;
 756
 757 /* Trim useless garbage from filenames. */
 758 static const char *
 759 fixup_filename(const char *file)
 760 {
 761
 762         if (file == NULL)
 763                 return (NULL);
 764         while (strncmp(file, "../", 3) == 0)
 765                 file += 3;
 766         return (file);
 767 }
 768
 769 /*
 770  * Calculate the size of early witness structures.
 771  */
 772 int
 773 witness_startup_count(void)
 774 {
 775         int sz;
 776
 777         sz = sizeof(struct witness) * witness_count;
 778         sz += sizeof(*w_rmatrix) * (witness_count + 1);
 779         sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
 780             (witness_count + 1);
 781
 782         return (sz);
 783 }
 784
 785 /*
 786  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 787  * assume that the early boot is single-threaded at least until after this
 788  * routine is completed.
 789  */
 790 void
 791 witness_startup(void *mem)
 792 {
 793         struct lock_object *lock;
 794         struct witness_order_list_entry *order;
 795         struct witness *w, *w1;
 796         uintptr_t p;
 797         int i;
 798
 799         p = (uintptr_t)mem;
 800         w_data = (void *)p;
 801         p += sizeof(struct witness) * witness_count;
 802
 803         w_rmatrix = (void *)p;
 804         p += sizeof(*w_rmatrix) * (witness_count + 1);
 805
 806         for (i = 0; i < witness_count + 1; i++) {
 807                 w_rmatrix[i] = (void *)p;
 808                 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
 809         }
 810         badstack_sbuf_size = witness_count * 256;
 811
 812         /*
 813          * We have to release Giant before initializing its witness
 814          * structure so that WITNESS doesn't get confused.
 815          */
 816         mtx_unlock(&Giant);
 817         mtx_assert(&Giant, MA_NOTOWNED);
 818
 819         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 820         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 821             MTX_NOWITNESS | MTX_NOPROFILE);
 822         for (i = witness_count - 1; i >= 0; i--) {
 823                 w = &w_data[i];
 824                 memset(w, 0, sizeof(*w));
 825                 w_data[i].w_index = i;  /* Witness index never changes. */
 826                 witness_free(w);
 827         }
 828         KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
 829             ("%s: Invalid list of free witness objects", __func__));
 830
 831         /* Witness with index 0 is not used to aid in debugging. */
 832         STAILQ_REMOVE_HEAD(&w_free, w_list);
 833         w_free_cnt--;
 834
 835         for (i = 0; i < witness_count; i++) {
 836                 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
 837                     (witness_count + 1));
 838         }
 839
 840         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 841                 witness_lock_list_free(&w_locklistdata[i]);
 842         witness_init_hash_tables();
 843
 844         /* First add in all the specified order lists. */
 845         for (order = order_lists; order->w_name != NULL; order++) {
 846                 w = enroll(order->w_name, order->w_class);
 847                 if (w == NULL)
 848                         continue;
 849                 w->w_file = "order list";
 850                 for (order++; order->w_name != NULL; order++) {
 851                         w1 = enroll(order->w_name, order->w_class);
 852                         if (w1 == NULL)
 853                                 continue;
 854                         w1->w_file = "order list";
 855                         itismychild(w, w1);
 856                         w = w1;
 857                 }
 858         }
 859         witness_spin_warn = 1;
 860
 861         /* Iterate through all locks and add them to witness. */
 862         for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
 863                 lock = pending_locks[i].wh_lock;
 864                 KASSERT(lock->lo_flags & LO_WITNESS,
 865                     ("%s: lock %s is on pending list but not LO_WITNESS",
 866                     __func__, lock->lo_name));
 867                 lock->lo_witness = enroll(pending_locks[i].wh_type,
 868                     LOCK_CLASS(lock));
 869         }
 870
 871         /* Mark the witness code as being ready for use. */
 872         witness_cold = 0;
 873
 874         mtx_lock(&Giant);
 875 }
 876
 877 void
 878 witness_init(struct lock_object *lock, const char *type)
 879 {
 880         struct lock_class *class;
 881
 882         /* Various sanity checks. */
 883         class = LOCK_CLASS(lock);
 884         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 885             (class->lc_flags & LC_RECURSABLE) == 0)
 886                 kassert_panic("%s: lock (%s) %s can not be recursable",
 887                     __func__, class->lc_name, lock->lo_name);
 888         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 889             (class->lc_flags & LC_SLEEPABLE) == 0)
 890                 kassert_panic("%s: lock (%s) %s can not be sleepable",
 891                     __func__, class->lc_name, lock->lo_name);
 892         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 893             (class->lc_flags & LC_UPGRADABLE) == 0)
 894                 kassert_panic("%s: lock (%s) %s can not be upgradable",
 895                     __func__, class->lc_name, lock->lo_name);
 896
 897         /*
 898          * If we shouldn't watch this lock, then just clear lo_witness.
 899          * Otherwise, if witness_cold is set, then it is too early to
 900          * enroll this lock, so defer it to witness_initialize() by adding
 901          * it to the pending_locks list.  If it is not too early, then enroll
 902          * the lock now.
 903          */
 904         if (witness_watch < 1 || KERNEL_PANICKED() ||
 905             (lock->lo_flags & LO_WITNESS) == 0)
 906                 lock->lo_witness = NULL;
 907         else if (witness_cold) {
 908                 pending_locks[pending_cnt].wh_lock = lock;
 909                 pending_locks[pending_cnt++].wh_type = type;
 910                 if (pending_cnt > WITNESS_PENDLIST)
 911                         panic("%s: pending locks list is too small, "
 912                             "increase WITNESS_PENDLIST\n",
 913                             __func__);
 914         } else
 915                 lock->lo_witness = enroll(type, class);
 916 }
 917
 918 void
 919 witness_destroy(struct lock_object *lock)
 920 {
 921         struct lock_class *class;
 922         struct witness *w;
 923
 924         class = LOCK_CLASS(lock);
 925
 926         if (witness_cold)
 927                 panic("lock (%s) %s destroyed while witness_cold",
 928                     class->lc_name, lock->lo_name);
 929
 930         /* XXX: need to verify that no one holds the lock */
 931         if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
 932                 return;
 933         w = lock->lo_witness;
 934
 935         mtx_lock_spin(&w_mtx);
 936         MPASS(w->w_refcount > 0);
 937         w->w_refcount--;
 938
 939         if (w->w_refcount == 0)
 940                 depart(w);
 941         mtx_unlock_spin(&w_mtx);
 942 }
 943
 944 #ifdef DDB
 945 static void
 946 witness_ddb_compute_levels(void)
 947 {
 948         struct witness *w;
 949
 950         /*
 951          * First clear all levels.
 952          */
 953         STAILQ_FOREACH(w, &w_all, w_list)
 954                 w->w_ddb_level = -1;
 955
 956         /*
 957          * Look for locks with no parents and level all their descendants.
 958          */
 959         STAILQ_FOREACH(w, &w_all, w_list) {
 960
 961                 /* If the witness has ancestors (is not a root), skip it. */
 962                 if (w->w_num_ancestors > 0)
 963                         continue;
 964                 witness_ddb_level_descendants(w, 0);
 965         }
 966 }
 967
 968 static void
 969 witness_ddb_level_descendants(struct witness *w, int l)
 970 {
 971         int i;
 972
 973         if (w->w_ddb_level >= l)
 974                 return;
 975
 976         w->w_ddb_level = l;
 977         l++;
 978
 979         for (i = 1; i <= w_max_used_index; i++) {
 980                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 981                         witness_ddb_level_descendants(&w_data[i], l);
 982         }
 983 }
 984
 985 static void
 986 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
 987     struct witness *w, int indent)
 988 {
 989         int i;
 990
 991         for (i = 0; i < indent; i++)
 992                 prnt(" ");
 993         prnt("%s (type: %s, depth: %d, active refs: %d)",
 994              w->w_name, w->w_class->lc_name,
 995              w->w_ddb_level, w->w_refcount);
 996         if (w->w_displayed) {
 997                 prnt(" -- (already displayed)\n");
 998                 return;
 999         }
1000         w->w_displayed = 1;
1001         if (w->w_file != NULL && w->w_line != 0)
1002                 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1003                     w->w_line);
1004         else
1005                 prnt(" -- never acquired\n");
1006         indent++;
1007         WITNESS_INDEX_ASSERT(w->w_index);
1008         for (i = 1; i <= w_max_used_index; i++) {
1009                 if (db_pager_quit)
1010                         return;
1011                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1012                         witness_ddb_display_descendants(prnt, &w_data[i],
1013                             indent);
1014         }
1015 }
1016
1017 static void
1018 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1019     struct witness_list *list)
1020 {
1021         struct witness *w;
1022
1023         STAILQ_FOREACH(w, list, w_typelist) {
1024                 if (w->w_file == NULL || w->w_ddb_level > 0)
1025                         continue;
1026
1027                 /* This lock has no anscestors - display its descendants. */
1028                 witness_ddb_display_descendants(prnt, w, 0);
1029                 if (db_pager_quit)
1030                         return;
1031         }
1032 }
1033
1034 static void
1035 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1036 {
1037         struct witness *w;
1038
1039         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1040         witness_ddb_compute_levels();
1041
1042         /* Clear all the displayed flags. */
1043         STAILQ_FOREACH(w, &w_all, w_list)
1044                 w->w_displayed = 0;
1045
1046         /*
1047          * First, handle sleep locks which have been acquired at least
1048          * once.
1049          */
1050         prnt("Sleep locks:\n");
1051         witness_ddb_display_list(prnt, &w_sleep);
1052         if (db_pager_quit)
1053                 return;
1054
1055         /*
1056          * Now do spin locks which have been acquired at least once.
1057          */
1058         prnt("\nSpin locks:\n");
1059         witness_ddb_display_list(prnt, &w_spin);
1060         if (db_pager_quit)
1061                 return;
1062
1063         /*
1064          * Finally, any locks which have not been acquired yet.
1065          */
1066         prnt("\nLocks which were never acquired:\n");
1067         STAILQ_FOREACH(w, &w_all, w_list) {
1068                 if (w->w_file != NULL || w->w_refcount == 0)
1069                         continue;
1070                 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1071                     w->w_class->lc_name, w->w_ddb_level);
1072                 if (db_pager_quit)
1073                         return;
1074         }
1075 }
1076 #endif /* DDB */
1077
1078 int
1079 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1080 {
1081
1082         if (witness_watch == -1 || KERNEL_PANICKED())
1083                 return (0);
1084
1085         /* Require locks that witness knows about. */
1086         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1087             lock2->lo_witness == NULL)
1088                 return (EINVAL);
1089
1090         mtx_assert(&w_mtx, MA_NOTOWNED);
1091         mtx_lock_spin(&w_mtx);
1092
1093         /*
1094          * If we already have either an explicit or implied lock order that
1095          * is the other way around, then return an error.
1096          */
1097         if (witness_watch &&
1098             isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1099                 mtx_unlock_spin(&w_mtx);
1100                 return (EDOOFUS);
1101         }
1102
1103         /* Try to add the new order. */
1104         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1105             lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1106         itismychild(lock1->lo_witness, lock2->lo_witness);
1107         mtx_unlock_spin(&w_mtx);
1108         return (0);
1109 }
1110
1111 void
1112 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1113     int line, struct lock_object *interlock)
1114 {
1115         struct lock_list_entry *lock_list, *lle;
1116         struct lock_instance *lock1, *lock2, *plock;
1117         struct lock_class *class, *iclass;
1118         struct witness *w, *w1;
1119         struct thread *td;
1120         int i, j;
1121
1122         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1123             KERNEL_PANICKED())
1124                 return;
1125
1126         w = lock->lo_witness;
1127         class = LOCK_CLASS(lock);
1128         td = curthread;
1129
1130         if (class->lc_flags & LC_SLEEPLOCK) {
1131
1132                 /*
1133                  * Since spin locks include a critical section, this check
1134                  * implicitly enforces a lock order of all sleep locks before
1135                  * all spin locks.
1136                  */
1137                 if (td->td_critnest != 0 && !kdb_active)
1138                         kassert_panic("acquiring blockable sleep lock with "
1139                             "spinlock or critical section held (%s) %s @ %s:%d",
1140                             class->lc_name, lock->lo_name,
1141                             fixup_filename(file), line);
1142
1143                 /*
1144                  * If this is the first lock acquired then just return as
1145                  * no order checking is needed.
1146                  */
1147                 lock_list = td->td_sleeplocks;
1148                 if (lock_list == NULL || lock_list->ll_count == 0)
1149                         return;
1150         } else {
1151
1152                 /*
1153                  * If this is the first lock, just return as no order
1154                  * checking is needed.  Avoid problems with thread
1155                  * migration pinning the thread while checking if
1156                  * spinlocks are held.  If at least one spinlock is held
1157                  * the thread is in a safe path and it is allowed to
1158                  * unpin it.
1159                  */
1160                 sched_pin();
1161                 lock_list = PCPU_GET(spinlocks);
1162                 if (lock_list == NULL || lock_list->ll_count == 0) {
1163                         sched_unpin();
1164                         return;
1165                 }
1166                 sched_unpin();
1167         }
1168
1169         /*
1170          * Check to see if we are recursing on a lock we already own.  If
1171          * so, make sure that we don't mismatch exclusive and shared lock
1172          * acquires.
1173          */
1174         lock1 = find_instance(lock_list, lock);
1175         if (lock1 != NULL) {
1176                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1177                     (flags & LOP_EXCLUSIVE) == 0) {
1178                         witness_output("shared lock of (%s) %s @ %s:%d\n",
1179                             class->lc_name, lock->lo_name,
1180                             fixup_filename(file), line);
1181                         witness_output("while exclusively locked from %s:%d\n",
1182                             fixup_filename(lock1->li_file), lock1->li_line);
1183                         kassert_panic("excl->share");
1184                 }
1185                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1186                     (flags & LOP_EXCLUSIVE) != 0) {
1187                         witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1188                             class->lc_name, lock->lo_name,
1189                             fixup_filename(file), line);
1190                         witness_output("while share locked from %s:%d\n",
1191                             fixup_filename(lock1->li_file), lock1->li_line);
1192                         kassert_panic("share->excl");
1193                 }
1194                 return;
1195         }
1196
1197         /* Warn if the interlock is not locked exactly once. */
1198         if (interlock != NULL) {
1199                 iclass = LOCK_CLASS(interlock);
1200                 lock1 = find_instance(lock_list, interlock);
1201                 if (lock1 == NULL)
1202                         kassert_panic("interlock (%s) %s not locked @ %s:%d",
1203                             iclass->lc_name, interlock->lo_name,
1204                             fixup_filename(file), line);
1205                 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1206                         kassert_panic("interlock (%s) %s recursed @ %s:%d",
1207                             iclass->lc_name, interlock->lo_name,
1208                             fixup_filename(file), line);
1209         }
1210
1211         /*
1212          * Find the previously acquired lock, but ignore interlocks.
1213          */
1214         plock = &lock_list->ll_children[lock_list->ll_count - 1];
1215         if (interlock != NULL && plock->li_lock == interlock) {
1216                 if (lock_list->ll_count > 1)
1217                         plock =
1218                             &lock_list->ll_children[lock_list->ll_count - 2];
1219                 else {
1220                         lle = lock_list->ll_next;
1221
1222                         /*
1223                          * The interlock is the only lock we hold, so
1224                          * simply return.
1225                          */
1226                         if (lle == NULL)
1227                                 return;
1228                         plock = &lle->ll_children[lle->ll_count - 1];
1229                 }
1230         }
1231
1232         /*
1233          * Try to perform most checks without a lock.  If this succeeds we
1234          * can skip acquiring the lock and return success.  Otherwise we redo
1235          * the check with the lock held to handle races with concurrent updates.
1236          */
1237         w1 = plock->li_lock->lo_witness;
1238         if (witness_lock_order_check(w1, w))
1239                 return;
1240
1241         mtx_lock_spin(&w_mtx);
1242         if (witness_lock_order_check(w1, w)) {
1243                 mtx_unlock_spin(&w_mtx);
1244                 return;
1245         }
1246         witness_lock_order_add(w1, w);
1247
1248         /*
1249          * Check for duplicate locks of the same type.  Note that we only
1250          * have to check for this on the last lock we just acquired.  Any
1251          * other cases will be caught as lock order violations.
1252          */
1253         if (w1 == w) {
1254                 i = w->w_index;
1255                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1256                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1257                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
1258                         w->w_reversed = 1;
1259                         mtx_unlock_spin(&w_mtx);
1260                         witness_output(
1261                             "acquiring duplicate lock of same type: \"%s\"\n",
1262                             w->w_name);
1263                         witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1264                             fixup_filename(plock->li_file), plock->li_line);
1265                         witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1266                             fixup_filename(file), line);
1267                         witness_debugger(1, __func__);
1268                 } else
1269                         mtx_unlock_spin(&w_mtx);
1270                 return;
1271         }
1272         mtx_assert(&w_mtx, MA_OWNED);
1273
1274         /*
1275          * If we know that the lock we are acquiring comes after
1276          * the lock we most recently acquired in the lock order tree,
1277          * then there is no need for any further checks.
1278          */
1279         if (isitmychild(w1, w))
1280                 goto out;
1281
1282         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1283                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1284
1285                         MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1286                         lock1 = &lle->ll_children[i];
1287
1288                         /*
1289                          * Ignore the interlock.
1290                          */
1291                         if (interlock == lock1->li_lock)
1292                                 continue;
1293
1294                         /*
1295                          * If this lock doesn't undergo witness checking,
1296                          * then skip it.
1297                          */
1298                         w1 = lock1->li_lock->lo_witness;
1299                         if (w1 == NULL) {
1300                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1301                                     ("lock missing witness structure"));
1302                                 continue;
1303                         }
1304
1305                         /*
1306                          * If we are locking Giant and this is a sleepable
1307                          * lock, then skip it.
1308                          */
1309                         if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1310                             lock == &Giant.lock_object)
1311                                 continue;
1312
1313                         /*
1314                          * If we are locking a sleepable lock and this lock
1315                          * is Giant, then skip it.
1316                          */
1317                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1318                             (flags & LOP_NOSLEEP) == 0 &&
1319                             lock1->li_lock == &Giant.lock_object)
1320                                 continue;
1321
1322                         /*
1323                          * If we are locking a sleepable lock and this lock
1324                          * isn't sleepable, we want to treat it as a lock
1325                          * order violation to enfore a general lock order of
1326                          * sleepable locks before non-sleepable locks.
1327                          */
1328                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1329                             (flags & LOP_NOSLEEP) == 0 &&
1330                             (lock1->li_flags & LI_SLEEPABLE) == 0)
1331                                 goto reversal;
1332
1333                         /*
1334                          * If we are locking Giant and this is a non-sleepable
1335                          * lock, then treat it as a reversal.
1336                          */
1337                         if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1338                             lock == &Giant.lock_object)
1339                                 goto reversal;
1340
1341                         /*
1342                          * Check the lock order hierarchy for a reveresal.
1343                          */
1344                         if (!isitmydescendant(w, w1))
1345                                 continue;
1346                 reversal:
1347
1348                         /*
1349                          * We have a lock order violation, check to see if it
1350                          * is allowed or has already been yelled about.
1351                          */
1352
1353                         /* Bail if this violation is known */
1354                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1355                                 goto out;
1356
1357                         /* Record this as a violation */
1358                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1359                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1360                         w->w_reversed = w1->w_reversed = 1;
1361                         witness_increment_graph_generation();
1362
1363                         /*
1364                          * If the lock order is blessed, bail before logging
1365                          * anything.  We don't look for other lock order
1366                          * violations though, which may be a bug.
1367                          */
1368                         if (blessed(w, w1))
1369                                 goto out;
1370                         mtx_unlock_spin(&w_mtx);
1371
1372 #ifdef WITNESS_NO_VNODE
1373                         /*
1374                          * There are known LORs between VNODE locks. They are
1375                          * not an indication of a bug. VNODE locks are flagged
1376                          * as such (LO_IS_VNODE) and we don't yell if the LOR
1377                          * is between 2 VNODE locks.
1378                          */
1379                         if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1380                             (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1381                                 return;
1382 #endif
1383
1384                         /*
1385                          * Ok, yell about it.
1386                          */
1387                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1388                             (flags & LOP_NOSLEEP) == 0 &&
1389                             (lock1->li_flags & LI_SLEEPABLE) == 0)
1390                                 witness_output(
1391                 "lock order reversal: (sleepable after non-sleepable)\n");
1392                         else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1393                             && lock == &Giant.lock_object)
1394                                 witness_output(
1395                 "lock order reversal: (Giant after non-sleepable)\n");
1396                         else
1397                                 witness_output("lock order reversal:\n");
1398
1399                         /*
1400                          * Try to locate an earlier lock with
1401                          * witness w in our list.
1402                          */
1403                         do {
1404                                 lock2 = &lle->ll_children[i];
1405                                 MPASS(lock2->li_lock != NULL);
1406                                 if (lock2->li_lock->lo_witness == w)
1407                                         break;
1408                                 if (i == 0 && lle->ll_next != NULL) {
1409                                         lle = lle->ll_next;
1410                                         i = lle->ll_count - 1;
1411                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1412                                 } else
1413                                         i--;
1414                         } while (i >= 0);
1415                         if (i < 0) {
1416                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1417                                     lock1->li_lock, lock1->li_lock->lo_name,
1418                                     w1->w_name, fixup_filename(lock1->li_file),
1419                                     lock1->li_line);
1420                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1421                                     lock->lo_name, w->w_name,
1422                                     fixup_filename(file), line);
1423                         } else {
1424                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1425                                     lock2->li_lock, lock2->li_lock->lo_name,
1426                                     lock2->li_lock->lo_witness->w_name,
1427                                     fixup_filename(lock2->li_file),
1428                                     lock2->li_line);
1429                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1430                                     lock1->li_lock, lock1->li_lock->lo_name,
1431                                     w1->w_name, fixup_filename(lock1->li_file),
1432                                     lock1->li_line);
1433                                 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1434                                     lock->lo_name, w->w_name,
1435                                     fixup_filename(file), line);
1436                         }
1437                         witness_debugger(1, __func__);
1438                         return;
1439                 }
1440         }
1441
1442         /*
1443          * If requested, build a new lock order.  However, don't build a new
1444          * relationship between a sleepable lock and Giant if it is in the
1445          * wrong direction.  The correct lock order is that sleepable locks
1446          * always come before Giant.
1447          */
1448         if (flags & LOP_NEWORDER &&
1449             !(plock->li_lock == &Giant.lock_object &&
1450             (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1451             (flags & LOP_NOSLEEP) == 0)) {
1452                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1453                     w->w_name, plock->li_lock->lo_witness->w_name);
1454                 itismychild(plock->li_lock->lo_witness, w);
1455         }
1456 out:
1457         mtx_unlock_spin(&w_mtx);
1458 }
1459
1460 void
1461 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1462 {
1463         struct lock_list_entry **lock_list, *lle;
1464         struct lock_instance *instance;
1465         struct witness *w;
1466         struct thread *td;
1467
1468         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1469             KERNEL_PANICKED())
1470                 return;
1471         w = lock->lo_witness;
1472         td = curthread;
1473
1474         /* Determine lock list for this lock. */
1475         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1476                 lock_list = &td->td_sleeplocks;
1477         else
1478                 lock_list = PCPU_PTR(spinlocks);
1479
1480         /* Check to see if we are recursing on a lock we already own. */
1481         instance = find_instance(*lock_list, lock);
1482         if (instance != NULL) {
1483                 instance->li_flags++;
1484                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1485                     td->td_proc->p_pid, lock->lo_name,
1486                     instance->li_flags & LI_RECURSEMASK);
1487                 instance->li_file = file;
1488                 instance->li_line = line;
1489                 return;
1490         }
1491
1492         /* Update per-witness last file and line acquire. */
1493         w->w_file = file;
1494         w->w_line = line;
1495
1496         /* Find the next open lock instance in the list and fill it. */
1497         lle = *lock_list;
1498         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1499                 lle = witness_lock_list_get();
1500                 if (lle == NULL)
1501                         return;
1502                 lle->ll_next = *lock_list;
1503                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1504                     td->td_proc->p_pid, lle);
1505                 *lock_list = lle;
1506         }
1507         instance = &lle->ll_children[lle->ll_count++];
1508         instance->li_lock = lock;
1509         instance->li_line = line;
1510         instance->li_file = file;
1511         instance->li_flags = 0;
1512         if ((flags & LOP_EXCLUSIVE) != 0)
1513                 instance->li_flags |= LI_EXCLUSIVE;
1514         if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1515                 instance->li_flags |= LI_SLEEPABLE;
1516         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1517             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1518 }
1519
1520 void
1521 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1522 {
1523         struct lock_instance *instance;
1524         struct lock_class *class;
1525
1526         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1527         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1528                 return;
1529         class = LOCK_CLASS(lock);
1530         if (witness_watch) {
1531                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1532                         kassert_panic(
1533                             "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1534                             class->lc_name, lock->lo_name,
1535                             fixup_filename(file), line);
1536                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1537                         kassert_panic(
1538                             "upgrade of non-sleep lock (%s) %s @ %s:%d",
1539                             class->lc_name, lock->lo_name,
1540                             fixup_filename(file), line);
1541         }
1542         instance = find_instance(curthread->td_sleeplocks, lock);
1543         if (instance == NULL) {
1544                 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1545                     class->lc_name, lock->lo_name,
1546                     fixup_filename(file), line);
1547                 return;
1548         }
1549         if (witness_watch) {
1550                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1551                         kassert_panic(
1552                             "upgrade of exclusive lock (%s) %s @ %s:%d",
1553                             class->lc_name, lock->lo_name,
1554                             fixup_filename(file), line);
1555                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1556                         kassert_panic(
1557                             "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1558                             class->lc_name, lock->lo_name,
1559                             instance->li_flags & LI_RECURSEMASK,
1560                             fixup_filename(file), line);
1561         }
1562         instance->li_flags |= LI_EXCLUSIVE;
1563 }
1564
1565 void
1566 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1567     int line)
1568 {
1569         struct lock_instance *instance;
1570         struct lock_class *class;
1571
1572         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1573         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1574                 return;
1575         class = LOCK_CLASS(lock);
1576         if (witness_watch) {
1577                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1578                         kassert_panic(
1579                             "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1580                             class->lc_name, lock->lo_name,
1581                             fixup_filename(file), line);
1582                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1583                         kassert_panic(
1584                             "downgrade of non-sleep lock (%s) %s @ %s:%d",
1585                             class->lc_name, lock->lo_name,
1586                             fixup_filename(file), line);
1587         }
1588         instance = find_instance(curthread->td_sleeplocks, lock);
1589         if (instance == NULL) {
1590                 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1591                     class->lc_name, lock->lo_name,
1592                     fixup_filename(file), line);
1593                 return;
1594         }
1595         if (witness_watch) {
1596                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1597                         kassert_panic(
1598                             "downgrade of shared lock (%s) %s @ %s:%d",
1599                             class->lc_name, lock->lo_name,
1600                             fixup_filename(file), line);
1601                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1602                         kassert_panic(
1603                             "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1604                             class->lc_name, lock->lo_name,
1605                             instance->li_flags & LI_RECURSEMASK,
1606                             fixup_filename(file), line);
1607         }
1608         instance->li_flags &= ~LI_EXCLUSIVE;
1609 }
1610
1611 void
1612 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1613 {
1614         struct lock_list_entry **lock_list, *lle;
1615         struct lock_instance *instance;
1616         struct lock_class *class;
1617         struct thread *td;
1618         register_t s;
1619         int i, j;
1620
1621         if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1622                 return;
1623         td = curthread;
1624         class = LOCK_CLASS(lock);
1625
1626         /* Find lock instance associated with this lock. */
1627         if (class->lc_flags & LC_SLEEPLOCK)
1628                 lock_list = &td->td_sleeplocks;
1629         else
1630                 lock_list = PCPU_PTR(spinlocks);
1631         lle = *lock_list;
1632         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1633                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1634                         instance = &(*lock_list)->ll_children[i];
1635                         if (instance->li_lock == lock)
1636                                 goto found;
1637                 }
1638
1639         /*
1640          * When disabling WITNESS through witness_watch we could end up in
1641          * having registered locks in the td_sleeplocks queue.
1642          * We have to make sure we flush these queues, so just search for
1643          * eventual register locks and remove them.
1644          */
1645         if (witness_watch > 0) {
1646                 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1647                     lock->lo_name, fixup_filename(file), line);
1648                 return;
1649         } else {
1650                 return;
1651         }
1652 found:
1653
1654         /* First, check for shared/exclusive mismatches. */
1655         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1656             (flags & LOP_EXCLUSIVE) == 0) {
1657                 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1658                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1659                 witness_output("while exclusively locked from %s:%d\n",
1660                     fixup_filename(instance->li_file), instance->li_line);
1661                 kassert_panic("excl->ushare");
1662         }
1663         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1664             (flags & LOP_EXCLUSIVE) != 0) {
1665                 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1666                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1667                 witness_output("while share locked from %s:%d\n",
1668                     fixup_filename(instance->li_file),
1669                     instance->li_line);
1670                 kassert_panic("share->uexcl");
1671         }
1672         /* If we are recursed, unrecurse. */
1673         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1674                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1675                     td->td_proc->p_pid, instance->li_lock->lo_name,
1676                     instance->li_flags);
1677                 instance->li_flags--;
1678                 return;
1679         }
1680         /* The lock is now being dropped, check for NORELEASE flag */
1681         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1682                 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1683                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1684                 kassert_panic("lock marked norelease");
1685         }
1686
1687         /* Otherwise, remove this item from the list. */
1688         s = intr_disable();
1689         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1690             td->td_proc->p_pid, instance->li_lock->lo_name,
1691             (*lock_list)->ll_count - 1);
1692         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1693                 (*lock_list)->ll_children[j] =
1694                     (*lock_list)->ll_children[j + 1];
1695         (*lock_list)->ll_count--;
1696         intr_restore(s);
1697
1698         /*
1699          * In order to reduce contention on w_mtx, we want to keep always an
1700          * head object into lists so that frequent allocation from the
1701          * free witness pool (and subsequent locking) is avoided.
1702          * In order to maintain the current code simple, when the head
1703          * object is totally unloaded it means also that we do not have
1704          * further objects in the list, so the list ownership needs to be
1705          * hand over to another object if the current head needs to be freed.
1706          */
1707         if ((*lock_list)->ll_count == 0) {
1708                 if (*lock_list == lle) {
1709                         if (lle->ll_next == NULL)
1710                                 return;
1711                 } else
1712                         lle = *lock_list;
1713                 *lock_list = lle->ll_next;
1714                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1715                     td->td_proc->p_pid, lle);
1716                 witness_lock_list_free(lle);
1717         }
1718 }
1719
1720 void
1721 witness_thread_exit(struct thread *td)
1722 {
1723         struct lock_list_entry *lle;
1724         int i, n;
1725
1726         lle = td->td_sleeplocks;
1727         if (lle == NULL || KERNEL_PANICKED())
1728                 return;
1729         if (lle->ll_count != 0) {
1730                 for (n = 0; lle != NULL; lle = lle->ll_next)
1731                         for (i = lle->ll_count - 1; i >= 0; i--) {
1732                                 if (n == 0)
1733                                         witness_output(
1734                     "Thread %p exiting with the following locks held:\n", td);
1735                                 n++;
1736                                 witness_list_lock(&lle->ll_children[i],
1737                                     witness_output);
1738
1739                         }
1740                 kassert_panic(
1741                     "Thread %p cannot exit while holding sleeplocks\n", td);
1742         }
1743         witness_lock_list_free(lle);
1744 }
1745
1746 /*
1747  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1748  * exempt Giant and sleepable locks from the checks as well.  If any
1749  * non-exempt locks are held, then a supplied message is printed to the
1750  * output channel along with a list of the offending locks.  If indicated in the
1751  * flags then a failure results in a panic as well.
1752  */
1753 int
1754 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1755 {
1756         struct lock_list_entry *lock_list, *lle;
1757         struct lock_instance *lock1;
1758         struct thread *td;
1759         va_list ap;
1760         int i, n;
1761
1762         if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1763                 return (0);
1764         n = 0;
1765         td = curthread;
1766         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1767                 for (i = lle->ll_count - 1; i >= 0; i--) {
1768                         lock1 = &lle->ll_children[i];
1769                         if (lock1->li_lock == lock)
1770                                 continue;
1771                         if (flags & WARN_GIANTOK &&
1772                             lock1->li_lock == &Giant.lock_object)
1773                                 continue;
1774                         if (flags & WARN_SLEEPOK &&
1775                             (lock1->li_flags & LI_SLEEPABLE) != 0)
1776                                 continue;
1777                         if (n == 0) {
1778                                 va_start(ap, fmt);
1779                                 vprintf(fmt, ap);
1780                                 va_end(ap);
1781                                 printf(" with the following %slocks held:\n",
1782                                     (flags & WARN_SLEEPOK) != 0 ?
1783                                     "non-sleepable " : "");
1784                         }
1785                         n++;
1786                         witness_list_lock(lock1, printf);
1787                 }
1788
1789         /*
1790          * Pin the thread in order to avoid problems with thread migration.
1791          * Once that all verifies are passed about spinlocks ownership,
1792          * the thread is in a safe path and it can be unpinned.
1793          */
1794         sched_pin();
1795         lock_list = PCPU_GET(spinlocks);
1796         if (lock_list != NULL && lock_list->ll_count != 0) {
1797                 sched_unpin();
1798
1799                 /*
1800                  * We should only have one spinlock and as long as
1801                  * the flags cannot match for this locks class,
1802                  * check if the first spinlock is the one curthread
1803                  * should hold.
1804                  */
1805                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1806                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1807                     lock1->li_lock == lock && n == 0)
1808                         return (0);
1809
1810                 va_start(ap, fmt);
1811                 vprintf(fmt, ap);
1812                 va_end(ap);
1813                 printf(" with the following %slocks held:\n",
1814                     (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1815                 n += witness_list_locks(&lock_list, printf);
1816         } else
1817                 sched_unpin();
1818         if (flags & WARN_PANIC && n)
1819                 kassert_panic("%s", __func__);
1820         else
1821                 witness_debugger(n, __func__);
1822         return (n);
1823 }
1824
1825 const char *
1826 witness_file(struct lock_object *lock)
1827 {
1828         struct witness *w;
1829
1830         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1831                 return ("?");
1832         w = lock->lo_witness;
1833         return (w->w_file);
1834 }
1835
1836 int
1837 witness_line(struct lock_object *lock)
1838 {
1839         struct witness *w;
1840
1841         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1842                 return (0);
1843         w = lock->lo_witness;
1844         return (w->w_line);
1845 }
1846
1847 static struct witness *
1848 enroll(const char *description, struct lock_class *lock_class)
1849 {
1850         struct witness *w;
1851
1852         MPASS(description != NULL);
1853
1854         if (witness_watch == -1 || KERNEL_PANICKED())
1855                 return (NULL);
1856         if ((lock_class->lc_flags & LC_SPINLOCK)) {
1857                 if (witness_skipspin)
1858                         return (NULL);
1859         } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1860                 kassert_panic("lock class %s is not sleep or spin",
1861                     lock_class->lc_name);
1862                 return (NULL);
1863         }
1864
1865         mtx_lock_spin(&w_mtx);
1866         w = witness_hash_get(description);
1867         if (w)
1868                 goto found;
1869         if ((w = witness_get()) == NULL)
1870                 return (NULL);
1871         MPASS(strlen(description) < MAX_W_NAME);
1872         strcpy(w->w_name, description);
1873         w->w_class = lock_class;
1874         w->w_refcount = 1;
1875         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1876         if (lock_class->lc_flags & LC_SPINLOCK) {
1877                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1878                 w_spin_cnt++;
1879         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1880                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1881                 w_sleep_cnt++;
1882         }
1883
1884         /* Insert new witness into the hash */
1885         witness_hash_put(w);
1886         witness_increment_graph_generation();
1887         mtx_unlock_spin(&w_mtx);
1888         return (w);
1889 found:
1890         w->w_refcount++;
1891         if (w->w_refcount == 1)
1892                 w->w_class = lock_class;
1893         mtx_unlock_spin(&w_mtx);
1894         if (lock_class != w->w_class)
1895                 kassert_panic(
1896                     "lock (%s) %s does not match earlier (%s) lock",
1897                     description, lock_class->lc_name,
1898                     w->w_class->lc_name);
1899         return (w);
1900 }
1901
1902 static void
1903 depart(struct witness *w)
1904 {
1905
1906         MPASS(w->w_refcount == 0);
1907         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1908                 w_sleep_cnt--;
1909         } else {
1910                 w_spin_cnt--;
1911         }
1912         /*
1913          * Set file to NULL as it may point into a loadable module.
1914          */
1915         w->w_file = NULL;
1916         w->w_line = 0;
1917         witness_increment_graph_generation();
1918 }
1919
1920 static void
1921 adopt(struct witness *parent, struct witness *child)
1922 {
1923         int pi, ci, i, j;
1924
1925         if (witness_cold == 0)
1926                 mtx_assert(&w_mtx, MA_OWNED);
1927
1928         /* If the relationship is already known, there's no work to be done. */
1929         if (isitmychild(parent, child))
1930                 return;
1931
1932         /* When the structure of the graph changes, bump up the generation. */
1933         witness_increment_graph_generation();
1934
1935         /*
1936          * The hard part ... create the direct relationship, then propagate all
1937          * indirect relationships.
1938          */
1939         pi = parent->w_index;
1940         ci = child->w_index;
1941         WITNESS_INDEX_ASSERT(pi);
1942         WITNESS_INDEX_ASSERT(ci);
1943         MPASS(pi != ci);
1944         w_rmatrix[pi][ci] |= WITNESS_PARENT;
1945         w_rmatrix[ci][pi] |= WITNESS_CHILD;
1946
1947         /*
1948          * If parent was not already an ancestor of child,
1949          * then we increment the descendant and ancestor counters.
1950          */
1951         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1952                 parent->w_num_descendants++;
1953                 child->w_num_ancestors++;
1954         }
1955
1956         /*
1957          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1958          * an ancestor of 'pi' during this loop.
1959          */
1960         for (i = 1; i <= w_max_used_index; i++) {
1961                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1962                     (i != pi))
1963                         continue;
1964
1965                 /* Find each descendant of 'i' and mark it as a descendant. */
1966                 for (j = 1; j <= w_max_used_index; j++) {
1967
1968                         /*
1969                          * Skip children that are already marked as
1970                          * descendants of 'i'.
1971                          */
1972                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1973                                 continue;
1974
1975                         /*
1976                          * We are only interested in descendants of 'ci'. Note
1977                          * that 'ci' itself is counted as a descendant of 'ci'.
1978                          */
1979                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1980                             (j != ci))
1981                                 continue;
1982                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1983                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1984                         w_data[i].w_num_descendants++;
1985                         w_data[j].w_num_ancestors++;
1986
1987                         /*
1988                          * Make sure we aren't marking a node as both an
1989                          * ancestor and descendant. We should have caught
1990                          * this as a lock order reversal earlier.
1991                          */
1992                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1993                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1994                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1995                                     "both ancestor and descendant\n",
1996                                     i, j, w_rmatrix[i][j]);
1997                                 kdb_backtrace();
1998                                 printf("Witness disabled.\n");
1999                                 witness_watch = -1;
2000                         }
2001                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2002                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2003                                 printf("witness rmatrix paradox! [%d][%d]=%d "
2004                                     "both ancestor and descendant\n",
2005                                     j, i, w_rmatrix[j][i]);
2006                                 kdb_backtrace();
2007                                 printf("Witness disabled.\n");
2008                                 witness_watch = -1;
2009                         }
2010                 }
2011         }
2012 }
2013
2014 static void
2015 itismychild(struct witness *parent, struct witness *child)
2016 {
2017         int unlocked;
2018
2019         MPASS(child != NULL && parent != NULL);
2020         if (witness_cold == 0)
2021                 mtx_assert(&w_mtx, MA_OWNED);
2022
2023         if (!witness_lock_type_equal(parent, child)) {
2024                 if (witness_cold == 0) {
2025                         unlocked = 1;
2026                         mtx_unlock_spin(&w_mtx);
2027                 } else {
2028                         unlocked = 0;
2029                 }
2030                 kassert_panic(
2031                     "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2032                     "the same lock type", __func__, parent->w_name,
2033                     parent->w_class->lc_name, child->w_name,
2034                     child->w_class->lc_name);
2035                 if (unlocked)
2036                         mtx_lock_spin(&w_mtx);
2037         }
2038         adopt(parent, child);
2039 }
2040
2041 /*
2042  * Generic code for the isitmy*() functions. The rmask parameter is the
2043  * expected relationship of w1 to w2.
2044  */
2045 static int
2046 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2047 {
2048         unsigned char r1, r2;
2049         int i1, i2;
2050
2051         i1 = w1->w_index;
2052         i2 = w2->w_index;
2053         WITNESS_INDEX_ASSERT(i1);
2054         WITNESS_INDEX_ASSERT(i2);
2055         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2056         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2057
2058         /* The flags on one better be the inverse of the flags on the other */
2059         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2060             (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2061                 /* Don't squawk if we're potentially racing with an update. */
2062                 if (!mtx_owned(&w_mtx))
2063                         return (0);
2064                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2065                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
2066                     "w_rmatrix[%d][%d] == %hhx\n",
2067                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2068                     i2, i1, r2);
2069                 kdb_backtrace();
2070                 printf("Witness disabled.\n");
2071                 witness_watch = -1;
2072         }
2073         return (r1 & rmask);
2074 }
2075
2076 /*
2077  * Checks if @child is a direct child of @parent.
2078  */
2079 static int
2080 isitmychild(struct witness *parent, struct witness *child)
2081 {
2082
2083         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2084 }
2085
2086 /*
2087  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2088  */
2089 static int
2090 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2091 {
2092
2093         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2094             __func__));
2095 }
2096
2097 static int
2098 blessed(struct witness *w1, struct witness *w2)
2099 {
2100         int i;
2101         struct witness_blessed *b;
2102
2103         for (i = 0; i < nitems(blessed_list); i++) {
2104                 b = &blessed_list[i];
2105                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2106                         if (strcmp(w2->w_name, b->b_lock2) == 0)
2107                                 return (1);
2108                         continue;
2109                 }
2110                 if (strcmp(w1->w_name, b->b_lock2) == 0)
2111                         if (strcmp(w2->w_name, b->b_lock1) == 0)
2112                                 return (1);
2113         }
2114         return (0);
2115 }
2116
2117 static struct witness *
2118 witness_get(void)
2119 {
2120         struct witness *w;
2121         int index;
2122
2123         if (witness_cold == 0)
2124                 mtx_assert(&w_mtx, MA_OWNED);
2125
2126         if (witness_watch == -1) {
2127                 mtx_unlock_spin(&w_mtx);
2128                 return (NULL);
2129         }
2130         if (STAILQ_EMPTY(&w_free)) {
2131                 witness_watch = -1;
2132                 mtx_unlock_spin(&w_mtx);
2133                 printf("WITNESS: unable to allocate a new witness object\n");
2134                 return (NULL);
2135         }
2136         w = STAILQ_FIRST(&w_free);
2137         STAILQ_REMOVE_HEAD(&w_free, w_list);
2138         w_free_cnt--;
2139         index = w->w_index;
2140         MPASS(index > 0 && index == w_max_used_index+1 &&
2141             index < witness_count);
2142         bzero(w, sizeof(*w));
2143         w->w_index = index;
2144         if (index > w_max_used_index)
2145                 w_max_used_index = index;
2146         return (w);
2147 }
2148
2149 static void
2150 witness_free(struct witness *w)
2151 {
2152
2153         STAILQ_INSERT_HEAD(&w_free, w, w_list);
2154         w_free_cnt++;
2155 }
2156
2157 static struct lock_list_entry *
2158 witness_lock_list_get(void)
2159 {
2160         struct lock_list_entry *lle;
2161
2162         if (witness_watch == -1)
2163                 return (NULL);
2164         mtx_lock_spin(&w_mtx);
2165         lle = w_lock_list_free;
2166         if (lle == NULL) {
2167                 witness_watch = -1;
2168                 mtx_unlock_spin(&w_mtx);
2169                 printf("%s: witness exhausted\n", __func__);
2170                 return (NULL);
2171         }
2172         w_lock_list_free = lle->ll_next;
2173         mtx_unlock_spin(&w_mtx);
2174         bzero(lle, sizeof(*lle));
2175         return (lle);
2176 }
2177
2178 static void
2179 witness_lock_list_free(struct lock_list_entry *lle)
2180 {
2181
2182         mtx_lock_spin(&w_mtx);
2183         lle->ll_next = w_lock_list_free;
2184         w_lock_list_free = lle;
2185         mtx_unlock_spin(&w_mtx);
2186 }
2187
2188 static struct lock_instance *
2189 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2190 {
2191         struct lock_list_entry *lle;
2192         struct lock_instance *instance;
2193         int i;
2194
2195         for (lle = list; lle != NULL; lle = lle->ll_next)
2196                 for (i = lle->ll_count - 1; i >= 0; i--) {
2197                         instance = &lle->ll_children[i];
2198                         if (instance->li_lock == lock)
2199                                 return (instance);
2200                 }
2201         return (NULL);
2202 }
2203
2204 static void
2205 witness_list_lock(struct lock_instance *instance,
2206     int (*prnt)(const char *fmt, ...))
2207 {
2208         struct lock_object *lock;
2209
2210         lock = instance->li_lock;
2211         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2212             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2213         if (lock->lo_witness->w_name != lock->lo_name)
2214                 prnt(" (%s)", lock->lo_witness->w_name);
2215         prnt(" r = %d (%p) locked @ %s:%d\n",
2216             instance->li_flags & LI_RECURSEMASK, lock,
2217             fixup_filename(instance->li_file), instance->li_line);
2218 }
2219
2220 static int
2221 witness_output(const char *fmt, ...)
2222 {
2223         va_list ap;
2224         int ret;
2225
2226         va_start(ap, fmt);
2227         ret = witness_voutput(fmt, ap);
2228         va_end(ap);
2229         return (ret);
2230 }
2231
2232 static int
2233 witness_voutput(const char *fmt, va_list ap)
2234 {
2235         int ret;
2236
2237         ret = 0;
2238         switch (witness_channel) {
2239         case WITNESS_CONSOLE:
2240                 ret = vprintf(fmt, ap);
2241                 break;
2242         case WITNESS_LOG:
2243                 vlog(LOG_NOTICE, fmt, ap);
2244                 break;
2245         case WITNESS_NONE:
2246                 break;
2247         }
2248         return (ret);
2249 }
2250
2251 #ifdef DDB
2252 static int
2253 witness_thread_has_locks(struct thread *td)
2254 {
2255
2256         if (td->td_sleeplocks == NULL)
2257                 return (0);
2258         return (td->td_sleeplocks->ll_count != 0);
2259 }
2260
2261 static int
2262 witness_proc_has_locks(struct proc *p)
2263 {
2264         struct thread *td;
2265
2266         FOREACH_THREAD_IN_PROC(p, td) {
2267                 if (witness_thread_has_locks(td))
2268                         return (1);
2269         }
2270         return (0);
2271 }
2272 #endif
2273
2274 int
2275 witness_list_locks(struct lock_list_entry **lock_list,
2276     int (*prnt)(const char *fmt, ...))
2277 {
2278         struct lock_list_entry *lle;
2279         int i, nheld;
2280
2281         nheld = 0;
2282         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2283                 for (i = lle->ll_count - 1; i >= 0; i--) {
2284                         witness_list_lock(&lle->ll_children[i], prnt);
2285                         nheld++;
2286                 }
2287         return (nheld);
2288 }
2289
2290 /*
2291  * This is a bit risky at best.  We call this function when we have timed
2292  * out acquiring a spin lock, and we assume that the other CPU is stuck
2293  * with this lock held.  So, we go groveling around in the other CPU's
2294  * per-cpu data to try to find the lock instance for this spin lock to
2295  * see when it was last acquired.
2296  */
2297 void
2298 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2299     int (*prnt)(const char *fmt, ...))
2300 {
2301         struct lock_instance *instance;
2302         struct pcpu *pc;
2303
2304         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2305                 return;
2306         pc = pcpu_find(owner->td_oncpu);
2307         instance = find_instance(pc->pc_spinlocks, lock);
2308         if (instance != NULL)
2309                 witness_list_lock(instance, prnt);
2310 }
2311
2312 void
2313 witness_save(struct lock_object *lock, const char **filep, int *linep)
2314 {
2315         struct lock_list_entry *lock_list;
2316         struct lock_instance *instance;
2317         struct lock_class *class;
2318
2319         /*
2320          * This function is used independently in locking code to deal with
2321          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2322          * is gone.
2323          */
2324         if (SCHEDULER_STOPPED())
2325                 return;
2326         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2327         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2328                 return;
2329         class = LOCK_CLASS(lock);
2330         if (class->lc_flags & LC_SLEEPLOCK)
2331                 lock_list = curthread->td_sleeplocks;
2332         else {
2333                 if (witness_skipspin)
2334                         return;
2335                 lock_list = PCPU_GET(spinlocks);
2336         }
2337         instance = find_instance(lock_list, lock);
2338         if (instance == NULL) {
2339                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2340                     class->lc_name, lock->lo_name);
2341                 return;
2342         }
2343         *filep = instance->li_file;
2344         *linep = instance->li_line;
2345 }
2346
2347 void
2348 witness_restore(struct lock_object *lock, const char *file, int line)
2349 {
2350         struct lock_list_entry *lock_list;
2351         struct lock_instance *instance;
2352         struct lock_class *class;
2353
2354         /*
2355          * This function is used independently in locking code to deal with
2356          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2357          * is gone.
2358          */
2359         if (SCHEDULER_STOPPED())
2360                 return;
2361         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2362         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2363                 return;
2364         class = LOCK_CLASS(lock);
2365         if (class->lc_flags & LC_SLEEPLOCK)
2366                 lock_list = curthread->td_sleeplocks;
2367         else {
2368                 if (witness_skipspin)
2369                         return;
2370                 lock_list = PCPU_GET(spinlocks);
2371         }
2372         instance = find_instance(lock_list, lock);
2373         if (instance == NULL)
2374                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2375                     class->lc_name, lock->lo_name);
2376         lock->lo_witness->w_file = file;
2377         lock->lo_witness->w_line = line;
2378         if (instance == NULL)
2379                 return;
2380         instance->li_file = file;
2381         instance->li_line = line;
2382 }
2383
2384 void
2385 witness_assert(const struct lock_object *lock, int flags, const char *file,
2386     int line)
2387 {
2388 #ifdef INVARIANT_SUPPORT
2389         struct lock_instance *instance;
2390         struct lock_class *class;
2391
2392         if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2393                 return;
2394         class = LOCK_CLASS(lock);
2395         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2396                 instance = find_instance(curthread->td_sleeplocks, lock);
2397         else if ((class->lc_flags & LC_SPINLOCK) != 0)
2398                 instance = find_instance(PCPU_GET(spinlocks), lock);
2399         else {
2400                 kassert_panic("Lock (%s) %s is not sleep or spin!",
2401                     class->lc_name, lock->lo_name);
2402                 return;
2403         }
2404         switch (flags) {
2405         case LA_UNLOCKED:
2406                 if (instance != NULL)
2407                         kassert_panic("Lock (%s) %s locked @ %s:%d.",
2408                             class->lc_name, lock->lo_name,
2409                             fixup_filename(file), line);
2410                 break;
2411         case LA_LOCKED:
2412         case LA_LOCKED | LA_RECURSED:
2413         case LA_LOCKED | LA_NOTRECURSED:
2414         case LA_SLOCKED:
2415         case LA_SLOCKED | LA_RECURSED:
2416         case LA_SLOCKED | LA_NOTRECURSED:
2417         case LA_XLOCKED:
2418         case LA_XLOCKED | LA_RECURSED:
2419         case LA_XLOCKED | LA_NOTRECURSED:
2420                 if (instance == NULL) {
2421                         kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2422                             class->lc_name, lock->lo_name,
2423                             fixup_filename(file), line);
2424                         break;
2425                 }
2426                 if ((flags & LA_XLOCKED) != 0 &&
2427                     (instance->li_flags & LI_EXCLUSIVE) == 0)
2428                         kassert_panic(
2429                             "Lock (%s) %s not exclusively locked @ %s:%d.",
2430                             class->lc_name, lock->lo_name,
2431                             fixup_filename(file), line);
2432                 if ((flags & LA_SLOCKED) != 0 &&
2433                     (instance->li_flags & LI_EXCLUSIVE) != 0)
2434                         kassert_panic(
2435                             "Lock (%s) %s exclusively locked @ %s:%d.",
2436                             class->lc_name, lock->lo_name,
2437                             fixup_filename(file), line);
2438                 if ((flags & LA_RECURSED) != 0 &&
2439                     (instance->li_flags & LI_RECURSEMASK) == 0)
2440                         kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2441                             class->lc_name, lock->lo_name,
2442                             fixup_filename(file), line);
2443                 if ((flags & LA_NOTRECURSED) != 0 &&
2444                     (instance->li_flags & LI_RECURSEMASK) != 0)
2445                         kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2446                             class->lc_name, lock->lo_name,
2447                             fixup_filename(file), line);
2448                 break;
2449         default:
2450                 kassert_panic("Invalid lock assertion at %s:%d.",
2451                     fixup_filename(file), line);
2452
2453         }
2454 #endif  /* INVARIANT_SUPPORT */
2455 }
2456
2457 static void
2458 witness_setflag(struct lock_object *lock, int flag, int set)
2459 {
2460         struct lock_list_entry *lock_list;
2461         struct lock_instance *instance;
2462         struct lock_class *class;
2463
2464         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2465                 return;
2466         class = LOCK_CLASS(lock);
2467         if (class->lc_flags & LC_SLEEPLOCK)
2468                 lock_list = curthread->td_sleeplocks;
2469         else {
2470                 if (witness_skipspin)
2471                         return;
2472                 lock_list = PCPU_GET(spinlocks);
2473         }
2474         instance = find_instance(lock_list, lock);
2475         if (instance == NULL) {
2476                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2477                     class->lc_name, lock->lo_name);
2478                 return;
2479         }
2480
2481         if (set)
2482                 instance->li_flags |= flag;
2483         else
2484                 instance->li_flags &= ~flag;
2485 }
2486
2487 void
2488 witness_norelease(struct lock_object *lock)
2489 {
2490
2491         witness_setflag(lock, LI_NORELEASE, 1);
2492 }
2493
2494 void
2495 witness_releaseok(struct lock_object *lock)
2496 {
2497
2498         witness_setflag(lock, LI_NORELEASE, 0);
2499 }
2500
2501 #ifdef DDB
2502 static void
2503 witness_ddb_list(struct thread *td)
2504 {
2505
2506         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2507         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2508
2509         if (witness_watch < 1)
2510                 return;
2511
2512         witness_list_locks(&td->td_sleeplocks, db_printf);
2513
2514         /*
2515          * We only handle spinlocks if td == curthread.  This is somewhat broken
2516          * if td is currently executing on some other CPU and holds spin locks
2517          * as we won't display those locks.  If we had a MI way of getting
2518          * the per-cpu data for a given cpu then we could use
2519          * td->td_oncpu to get the list of spinlocks for this thread
2520          * and "fix" this.
2521          *
2522          * That still wouldn't really fix this unless we locked the scheduler
2523          * lock or stopped the other CPU to make sure it wasn't changing the
2524          * list out from under us.  It is probably best to just not try to
2525          * handle threads on other CPU's for now.
2526          */
2527         if (td == curthread && PCPU_GET(spinlocks) != NULL)
2528                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2529 }
2530
2531 DB_SHOW_COMMAND(locks, db_witness_list)
2532 {
2533         struct thread *td;
2534
2535         if (have_addr)
2536                 td = db_lookup_thread(addr, true);
2537         else
2538                 td = kdb_thread;
2539         witness_ddb_list(td);
2540 }
2541
2542 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2543 {
2544         struct thread *td;
2545         struct proc *p;
2546
2547         /*
2548          * It would be nice to list only threads and processes that actually
2549          * held sleep locks, but that information is currently not exported
2550          * by WITNESS.
2551          */
2552         FOREACH_PROC_IN_SYSTEM(p) {
2553                 if (!witness_proc_has_locks(p))
2554                         continue;
2555                 FOREACH_THREAD_IN_PROC(p, td) {
2556                         if (!witness_thread_has_locks(td))
2557                                 continue;
2558                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2559                             p->p_comm, td, td->td_tid);
2560                         witness_ddb_list(td);
2561                         if (db_pager_quit)
2562                                 return;
2563                 }
2564         }
2565 }
2566 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2567
2568 DB_SHOW_COMMAND(witness, db_witness_display)
2569 {
2570
2571         witness_ddb_display(db_printf);
2572 }
2573 #endif
2574
2575 static void
2576 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2577 {
2578         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2579         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2580         int generation, i, j;
2581
2582         tmp_data1 = NULL;
2583         tmp_data2 = NULL;
2584         tmp_w1 = NULL;
2585         tmp_w2 = NULL;
2586
2587         /* Allocate and init temporary storage space. */
2588         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2589         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2590         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2591             M_WAITOK | M_ZERO);
2592         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2593             M_WAITOK | M_ZERO);
2594         stack_zero(&tmp_data1->wlod_stack);
2595         stack_zero(&tmp_data2->wlod_stack);
2596
2597 restart:
2598         mtx_lock_spin(&w_mtx);
2599         generation = w_generation;
2600         mtx_unlock_spin(&w_mtx);
2601         sbuf_printf(sb, "Number of known direct relationships is %d\n",
2602             w_lohash.wloh_count);
2603         for (i = 1; i < w_max_used_index; i++) {
2604                 mtx_lock_spin(&w_mtx);
2605                 if (generation != w_generation) {
2606                         mtx_unlock_spin(&w_mtx);
2607
2608                         /* The graph has changed, try again. */
2609                         *oldidx = 0;
2610                         sbuf_clear(sb);
2611                         goto restart;
2612                 }
2613
2614                 w1 = &w_data[i];
2615                 if (w1->w_reversed == 0) {
2616                         mtx_unlock_spin(&w_mtx);
2617                         continue;
2618                 }
2619
2620                 /* Copy w1 locally so we can release the spin lock. */
2621                 *tmp_w1 = *w1;
2622                 mtx_unlock_spin(&w_mtx);
2623
2624                 if (tmp_w1->w_reversed == 0)
2625                         continue;
2626                 for (j = 1; j < w_max_used_index; j++) {
2627                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2628                                 continue;
2629
2630                         mtx_lock_spin(&w_mtx);
2631                         if (generation != w_generation) {
2632                                 mtx_unlock_spin(&w_mtx);
2633
2634                                 /* The graph has changed, try again. */
2635                                 *oldidx = 0;
2636                                 sbuf_clear(sb);
2637                                 goto restart;
2638                         }
2639
2640                         w2 = &w_data[j];
2641                         data1 = witness_lock_order_get(w1, w2);
2642                         data2 = witness_lock_order_get(w2, w1);
2643
2644                         /*
2645                          * Copy information locally so we can release the
2646                          * spin lock.
2647                          */
2648                         *tmp_w2 = *w2;
2649
2650                         if (data1) {
2651                                 stack_zero(&tmp_data1->wlod_stack);
2652                                 stack_copy(&data1->wlod_stack,
2653                                     &tmp_data1->wlod_stack);
2654                         }
2655                         if (data2 && data2 != data1) {
2656                                 stack_zero(&tmp_data2->wlod_stack);
2657                                 stack_copy(&data2->wlod_stack,
2658                                     &tmp_data2->wlod_stack);
2659                         }
2660                         mtx_unlock_spin(&w_mtx);
2661
2662                         if (blessed(tmp_w1, tmp_w2))
2663                                 continue;
2664
2665                         sbuf_printf(sb,
2666             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2667                             tmp_w1->w_name, tmp_w1->w_class->lc_name,
2668                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
2669                         if (data1) {
2670                                 sbuf_printf(sb,
2671                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2672                                     tmp_w1->w_name, tmp_w1->w_class->lc_name,
2673                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
2674                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2675                                 sbuf_printf(sb, "\n");
2676                         }
2677                         if (data2 && data2 != data1) {
2678                                 sbuf_printf(sb,
2679                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2680                                     tmp_w2->w_name, tmp_w2->w_class->lc_name,
2681                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
2682                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2683                                 sbuf_printf(sb, "\n");
2684                         }
2685                 }
2686         }
2687         mtx_lock_spin(&w_mtx);
2688         if (generation != w_generation) {
2689                 mtx_unlock_spin(&w_mtx);
2690
2691                 /*
2692                  * The graph changed while we were printing stack data,
2693                  * try again.
2694                  */
2695                 *oldidx = 0;
2696                 sbuf_clear(sb);
2697                 goto restart;
2698         }
2699         mtx_unlock_spin(&w_mtx);
2700
2701         /* Free temporary storage space. */
2702         free(tmp_data1, M_TEMP);
2703         free(tmp_data2, M_TEMP);
2704         free(tmp_w1, M_TEMP);
2705         free(tmp_w2, M_TEMP);
2706 }
2707
2708 static int
2709 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2710 {
2711         struct sbuf *sb;
2712         int error;
2713
2714         if (witness_watch < 1) {
2715                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2716                 return (error);
2717         }
2718         if (witness_cold) {
2719                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2720                 return (error);
2721         }
2722         error = 0;
2723         sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2724         if (sb == NULL)
2725                 return (ENOMEM);
2726
2727         sbuf_print_witness_badstacks(sb, &req->oldidx);
2728
2729         sbuf_finish(sb);
2730         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2731         sbuf_delete(sb);
2732
2733         return (error);
2734 }
2735
2736 #ifdef DDB
2737 static int
2738 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2739 {
2740
2741         return (db_printf("%.*s", len, data));
2742 }
2743
2744 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2745 {
2746         struct sbuf sb;
2747         char buffer[128];
2748         size_t dummy;
2749
2750         sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2751         sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2752         sbuf_print_witness_badstacks(&sb, &dummy);
2753         sbuf_finish(&sb);
2754 }
2755 #endif
2756
2757 static int
2758 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2759 {
2760         static const struct {
2761                 enum witness_channel channel;
2762                 const char *name;
2763         } channels[] = {
2764                 { WITNESS_CONSOLE, "console" },
2765                 { WITNESS_LOG, "log" },
2766                 { WITNESS_NONE, "none" },
2767         };
2768         char buf[16];
2769         u_int i;
2770         int error;
2771
2772         buf[0] = '\0';
2773         for (i = 0; i < nitems(channels); i++)
2774                 if (witness_channel == channels[i].channel) {
2775                         snprintf(buf, sizeof(buf), "%s", channels[i].name);
2776                         break;
2777                 }
2778
2779         error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2780         if (error != 0 || req->newptr == NULL)
2781                 return (error);
2782
2783         error = EINVAL;
2784         for (i = 0; i < nitems(channels); i++)
2785                 if (strcmp(channels[i].name, buf) == 0) {
2786                         witness_channel = channels[i].channel;
2787                         error = 0;
2788                         break;
2789                 }
2790         return (error);
2791 }
2792
2793 static int
2794 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2795 {
2796         struct witness *w;
2797         struct sbuf *sb;
2798         int error;
2799
2800 #ifdef __i386__
2801         error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2802         return (error);
2803 #endif
2804
2805         if (witness_watch < 1) {
2806                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2807                 return (error);
2808         }
2809         if (witness_cold) {
2810                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2811                 return (error);
2812         }
2813         error = 0;
2814
2815         error = sysctl_wire_old_buffer(req, 0);
2816         if (error != 0)
2817                 return (error);
2818         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2819         if (sb == NULL)
2820                 return (ENOMEM);
2821         sbuf_printf(sb, "\n");
2822
2823         mtx_lock_spin(&w_mtx);
2824         STAILQ_FOREACH(w, &w_all, w_list)
2825                 w->w_displayed = 0;
2826         STAILQ_FOREACH(w, &w_all, w_list)
2827                 witness_add_fullgraph(sb, w);
2828         mtx_unlock_spin(&w_mtx);
2829
2830         /*
2831          * Close the sbuf and return to userland.
2832          */
2833         error = sbuf_finish(sb);
2834         sbuf_delete(sb);
2835
2836         return (error);
2837 }
2838
2839 static int
2840 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2841 {
2842         int error, value;
2843
2844         value = witness_watch;
2845         error = sysctl_handle_int(oidp, &value, 0, req);
2846         if (error != 0 || req->newptr == NULL)
2847                 return (error);
2848         if (value > 1 || value < -1 ||
2849             (witness_watch == -1 && value != witness_watch))
2850                 return (EINVAL);
2851         witness_watch = value;
2852         return (0);
2853 }
2854
2855 static void
2856 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2857 {
2858         int i;
2859
2860         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2861                 return;
2862         w->w_displayed = 1;
2863
2864         WITNESS_INDEX_ASSERT(w->w_index);
2865         for (i = 1; i <= w_max_used_index; i++) {
2866                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2867                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2868                             w_data[i].w_name);
2869                         witness_add_fullgraph(sb, &w_data[i]);
2870                 }
2871         }
2872 }
2873
2874 /*
2875  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2876  * interprets the key as a string and reads until the null
2877  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2878  * hash value computed from the key.
2879  */
2880 static uint32_t
2881 witness_hash_djb2(const uint8_t *key, uint32_t size)
2882 {
2883         unsigned int hash = 5381;
2884         int i;
2885
2886         /* hash = hash * 33 + key[i] */
2887         if (size)
2888                 for (i = 0; i < size; i++)
2889                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2890         else
2891                 for (i = 0; key[i] != 0; i++)
2892                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2893
2894         return (hash);
2895 }
2896
2897 /*
2898  * Initializes the two witness hash tables. Called exactly once from
2899  * witness_initialize().
2900  */
2901 static void
2902 witness_init_hash_tables(void)
2903 {
2904         int i;
2905
2906         MPASS(witness_cold);
2907
2908         /* Initialize the hash tables. */
2909         for (i = 0; i < WITNESS_HASH_SIZE; i++)
2910                 w_hash.wh_array[i] = NULL;
2911
2912         w_hash.wh_size = WITNESS_HASH_SIZE;
2913         w_hash.wh_count = 0;
2914
2915         /* Initialize the lock order data hash. */
2916         w_lofree = NULL;
2917         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2918                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2919                 w_lodata[i].wlod_next = w_lofree;
2920                 w_lofree = &w_lodata[i];
2921         }
2922         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2923         w_lohash.wloh_count = 0;
2924         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2925                 w_lohash.wloh_array[i] = NULL;
2926 }
2927
2928 static struct witness *
2929 witness_hash_get(const char *key)
2930 {
2931         struct witness *w;
2932         uint32_t hash;
2933
2934         MPASS(key != NULL);
2935         if (witness_cold == 0)
2936                 mtx_assert(&w_mtx, MA_OWNED);
2937         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2938         w = w_hash.wh_array[hash];
2939         while (w != NULL) {
2940                 if (strcmp(w->w_name, key) == 0)
2941                         goto out;
2942                 w = w->w_hash_next;
2943         }
2944
2945 out:
2946         return (w);
2947 }
2948
2949 static void
2950 witness_hash_put(struct witness *w)
2951 {
2952         uint32_t hash;
2953
2954         MPASS(w != NULL);
2955         MPASS(w->w_name != NULL);
2956         if (witness_cold == 0)
2957                 mtx_assert(&w_mtx, MA_OWNED);
2958         KASSERT(witness_hash_get(w->w_name) == NULL,
2959             ("%s: trying to add a hash entry that already exists!", __func__));
2960         KASSERT(w->w_hash_next == NULL,
2961             ("%s: w->w_hash_next != NULL", __func__));
2962
2963         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2964         w->w_hash_next = w_hash.wh_array[hash];
2965         w_hash.wh_array[hash] = w;
2966         w_hash.wh_count++;
2967 }
2968
2969 static struct witness_lock_order_data *
2970 witness_lock_order_get(struct witness *parent, struct witness *child)
2971 {
2972         struct witness_lock_order_data *data = NULL;
2973         struct witness_lock_order_key key;
2974         unsigned int hash;
2975
2976         MPASS(parent != NULL && child != NULL);
2977         key.from = parent->w_index;
2978         key.to = child->w_index;
2979         WITNESS_INDEX_ASSERT(key.from);
2980         WITNESS_INDEX_ASSERT(key.to);
2981         if ((w_rmatrix[parent->w_index][child->w_index]
2982             & WITNESS_LOCK_ORDER_KNOWN) == 0)
2983                 goto out;
2984
2985         hash = witness_hash_djb2((const char*)&key,
2986             sizeof(key)) % w_lohash.wloh_size;
2987         data = w_lohash.wloh_array[hash];
2988         while (data != NULL) {
2989                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2990                         break;
2991                 data = data->wlod_next;
2992         }
2993
2994 out:
2995         return (data);
2996 }
2997
2998 /*
2999  * Verify that parent and child have a known relationship, are not the same,
3000  * and child is actually a child of parent.  This is done without w_mtx
3001  * to avoid contention in the common case.
3002  */
3003 static int
3004 witness_lock_order_check(struct witness *parent, struct witness *child)
3005 {
3006
3007         if (parent != child &&
3008             w_rmatrix[parent->w_index][child->w_index]
3009             & WITNESS_LOCK_ORDER_KNOWN &&
3010             isitmychild(parent, child))
3011                 return (1);
3012
3013         return (0);
3014 }
3015
3016 static int
3017 witness_lock_order_add(struct witness *parent, struct witness *child)
3018 {
3019         struct witness_lock_order_data *data = NULL;
3020         struct witness_lock_order_key key;
3021         unsigned int hash;
3022
3023         MPASS(parent != NULL && child != NULL);
3024         key.from = parent->w_index;
3025         key.to = child->w_index;
3026         WITNESS_INDEX_ASSERT(key.from);
3027         WITNESS_INDEX_ASSERT(key.to);
3028         if (w_rmatrix[parent->w_index][child->w_index]
3029             & WITNESS_LOCK_ORDER_KNOWN)
3030                 return (1);
3031
3032         hash = witness_hash_djb2((const char*)&key,
3033             sizeof(key)) % w_lohash.wloh_size;
3034         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3035         data = w_lofree;
3036         if (data == NULL)
3037                 return (0);
3038         w_lofree = data->wlod_next;
3039         data->wlod_next = w_lohash.wloh_array[hash];
3040         data->wlod_key = key;
3041         w_lohash.wloh_array[hash] = data;
3042         w_lohash.wloh_count++;
3043         stack_zero(&data->wlod_stack);
3044         stack_save(&data->wlod_stack);
3045         return (1);
3046 }
3047
3048 /* Call this whenever the structure of the witness graph changes. */
3049 static void
3050 witness_increment_graph_generation(void)
3051 {
3052
3053         if (witness_cold == 0)
3054                 mtx_assert(&w_mtx, MA_OWNED);
3055         w_generation++;
3056 }
3057
3058 static int
3059 witness_output_drain(void *arg __unused, const char *data, int len)
3060 {
3061
3062         witness_output("%.*s", len, data);
3063         return (len);
3064 }
3065
3066 static void
3067 witness_debugger(int cond, const char *msg)
3068 {
3069         char buf[32];
3070         struct sbuf sb;
3071         struct stack st;
3072
3073         if (!cond)
3074                 return;
3075
3076         if (witness_trace) {
3077                 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3078                 sbuf_set_drain(&sb, witness_output_drain, NULL);
3079
3080                 stack_zero(&st);
3081                 stack_save(&st);
3082                 witness_output("stack backtrace:\n");
3083                 stack_sbuf_print_ddb(&sb, &st);
3084
3085                 sbuf_finish(&sb);
3086         }
3087
3088 #ifdef KDB
3089         if (witness_kdb)
3090                 kdb_enter(KDB_WHY_WITNESS, msg);
3091 #endif
3092 }