sys/kern/subr_witness.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 2008 Isilon Systems, Inc.
   5  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
   6  * Copyright (c) 1998 Berkeley Software Design, Inc.
   7  * All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Berkeley Software Design Inc's name may not be used to endorse or
  18  *    promote products derived from this software without specific prior
  19  *    written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
  22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  24  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
  25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  *      from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
  34  *      and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
  35  */
  36
  37 /*
  38  * Implementation of the `witness' lock verifier.  Originally implemented for
  39  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
  40  * classes in FreeBSD.
  41  */
  42
  43 /*
  44  *      Main Entry: witness
  45  *      Pronunciation: 'wit-n&s
  46  *      Function: noun
  47  *      Etymology: Middle English witnesse, from Old English witnes knowledge,
  48  *          testimony, witness, from 2wit
  49  *      Date: before 12th century
  50  *      1 : attestation of a fact or event : TESTIMONY
  51  *      2 : one that gives evidence; specifically : one who testifies in
  52  *          a cause or before a judicial tribunal
  53  *      3 : one asked to be present at a transaction so as to be able to
  54  *          testify to its having taken place
  55  *      4 : one who has personal knowledge of something
  56  *      5 a : something serving as evidence or proof : SIGN
  57  *        b : public affirmation by word or example of usually
  58  *            religious faith or conviction <the heroic witness to divine
  59  *            life -- Pilot>
  60  *      6 capitalized : a member of the Jehovah's Witnesses
  61  */
  62
  63 /*
  64  * Special rules concerning Giant and lock orders:
  65  *
  66  * 1) Giant must be acquired before any other mutexes.  Stated another way,
  67  *    no other mutex may be held when Giant is acquired.
  68  *
  69  * 2) Giant must be released when blocking on a sleepable lock.
  70  *
  71  * This rule is less obvious, but is a result of Giant providing the same
  72  * semantics as spl().  Basically, when a thread sleeps, it must release
  73  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
  74  * 2).
  75  *
  76  * 3) Giant may be acquired before or after sleepable locks.
  77  *
  78  * This rule is also not quite as obvious.  Giant may be acquired after
  79  * a sleepable lock because it is a non-sleepable lock and non-sleepable
  80  * locks may always be acquired while holding a sleepable lock.  The second
  81  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
  82  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
  83  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
  84  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
  85  * execute.  Thus, acquiring Giant both before and after a sleepable lock
  86  * will not result in a lock order reversal.
  87  */
  88
  89 #include <sys/cdefs.h>
  90 __FBSDID("$FreeBSD$");
  91
  92 #include "opt_ddb.h"
  93 #include "opt_hwpmc_hooks.h"
  94 #include "opt_stack.h"
  95 #include "opt_witness.h"
  96
  97 #include <sys/param.h>
  98 #include <sys/bus.h>
  99 #include <sys/kdb.h>
 100 #include <sys/kernel.h>
 101 #include <sys/ktr.h>
 102 #include <sys/lock.h>
 103 #include <sys/malloc.h>
 104 #include <sys/mutex.h>
 105 #include <sys/priv.h>
 106 #include <sys/proc.h>
 107 #include <sys/sbuf.h>
 108 #include <sys/sched.h>
 109 #include <sys/stack.h>
 110 #include <sys/sysctl.h>
 111 #include <sys/syslog.h>
 112 #include <sys/systm.h>
 113
 114 #ifdef DDB
 115 #include <ddb/ddb.h>
 116 #endif
 117
 118 #include <machine/stdarg.h>
 119
 120 #if !defined(DDB) && !defined(STACK)
 121 #error "DDB or STACK options are required for WITNESS"
 122 #endif
 123
 124 /* Note that these traces do not work with KTR_ALQ. */
 125 #if 0
 126 #define KTR_WITNESS     KTR_SUBSYS
 127 #else
 128 #define KTR_WITNESS     0
 129 #endif
 130
 131 #define LI_RECURSEMASK  0x0000ffff      /* Recursion depth of lock instance. */
 132 #define LI_EXCLUSIVE    0x00010000      /* Exclusive lock instance. */
 133 #define LI_NORELEASE    0x00020000      /* Lock not allowed to be released. */
 134
 135 #ifndef WITNESS_COUNT
 136 #define WITNESS_COUNT           1536
 137 #endif
 138 #define WITNESS_HASH_SIZE       251     /* Prime, gives load factor < 2 */
 139 #define WITNESS_PENDLIST        (512 + (MAXCPU * 4))
 140
 141 /* Allocate 256 KB of stack data space */
 142 #define WITNESS_LO_DATA_COUNT   2048
 143
 144 /* Prime, gives load factor of ~2 at full load */
 145 #define WITNESS_LO_HASH_SIZE    1021
 146
 147 /*
 148  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
 149  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
 150  * probably be safe for the most part, but it's still a SWAG.
 151  */
 152 #define LOCK_NCHILDREN  5
 153 #define LOCK_CHILDCOUNT 2048
 154
 155 #define MAX_W_NAME      64
 156
 157 #define FULLGRAPH_SBUF_SIZE     512
 158
 159 /*
 160  * These flags go in the witness relationship matrix and describe the
 161  * relationship between any two struct witness objects.
 162  */
 163 #define WITNESS_UNRELATED        0x00    /* No lock order relation. */
 164 #define WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
 165 #define WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
 166 #define WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
 167 #define WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
 168 #define WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
 169 #define WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
 170 #define WITNESS_RELATED_MASK                                            \
 171         (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
 172 #define WITNESS_REVERSAL         0x10    /* A lock order reversal has been
 173                                           * observed. */
 174 #define WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
 175 #define WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
 176 #define WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
 177
 178 /* Descendant to ancestor flags */
 179 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
 180
 181 /* Ancestor to descendant flags */
 182 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
 183
 184 #define WITNESS_INDEX_ASSERT(i)                                         \
 185         MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
 186
 187 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
 188
 189 /*
 190  * Lock instances.  A lock instance is the data associated with a lock while
 191  * it is held by witness.  For example, a lock instance will hold the
 192  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
 193  * are held in a per-cpu list while sleep locks are held in per-thread list.
 194  */
 195 struct lock_instance {
 196         struct lock_object      *li_lock;
 197         const char              *li_file;
 198         int                     li_line;
 199         u_int                   li_flags;
 200 };
 201
 202 /*
 203  * A simple list type used to build the list of locks held by a thread
 204  * or CPU.  We can't simply embed the list in struct lock_object since a
 205  * lock may be held by more than one thread if it is a shared lock.  Locks
 206  * are added to the head of the list, so we fill up each list entry from
 207  * "the back" logically.  To ease some of the arithmetic, we actually fill
 208  * in each list entry the normal way (children[0] then children[1], etc.) but
 209  * when we traverse the list we read children[count-1] as the first entry
 210  * down to children[0] as the final entry.
 211  */
 212 struct lock_list_entry {
 213         struct lock_list_entry  *ll_next;
 214         struct lock_instance    ll_children[LOCK_NCHILDREN];
 215         u_int                   ll_count;
 216 };
 217
 218 /*
 219  * The main witness structure. One of these per named lock type in the system
 220  * (for example, "vnode interlock").
 221  */
 222 struct witness {
 223         char                    w_name[MAX_W_NAME];
 224         uint32_t                w_index;  /* Index in the relationship matrix */
 225         struct lock_class       *w_class;
 226         STAILQ_ENTRY(witness)   w_list;         /* List of all witnesses. */
 227         STAILQ_ENTRY(witness)   w_typelist;     /* Witnesses of a type. */
 228         struct witness          *w_hash_next; /* Linked list in hash buckets. */
 229         const char              *w_file; /* File where last acquired */
 230         uint32_t                w_line; /* Line where last acquired */
 231         uint32_t                w_refcount;
 232         uint16_t                w_num_ancestors; /* direct/indirect
 233                                                   * ancestor count */
 234         uint16_t                w_num_descendants; /* direct/indirect
 235                                                     * descendant count */
 236         int16_t                 w_ddb_level;
 237         unsigned                w_displayed:1;
 238         unsigned                w_reversed:1;
 239 };
 240
 241 STAILQ_HEAD(witness_list, witness);
 242
 243 /*
 244  * The witness hash table. Keys are witness names (const char *), elements are
 245  * witness objects (struct witness *).
 246  */
 247 struct witness_hash {
 248         struct witness  *wh_array[WITNESS_HASH_SIZE];
 249         uint32_t        wh_size;
 250         uint32_t        wh_count;
 251 };
 252
 253 /*
 254  * Key type for the lock order data hash table.
 255  */
 256 struct witness_lock_order_key {
 257         uint16_t        from;
 258         uint16_t        to;
 259 };
 260
 261 struct witness_lock_order_data {
 262         struct stack                    wlod_stack;
 263         struct witness_lock_order_key   wlod_key;
 264         struct witness_lock_order_data  *wlod_next;
 265 };
 266
 267 /*
 268  * The witness lock order data hash table. Keys are witness index tuples
 269  * (struct witness_lock_order_key), elements are lock order data objects
 270  * (struct witness_lock_order_data).
 271  */
 272 struct witness_lock_order_hash {
 273         struct witness_lock_order_data  *wloh_array[WITNESS_LO_HASH_SIZE];
 274         u_int   wloh_size;
 275         u_int   wloh_count;
 276 };
 277
 278 struct witness_blessed {
 279         const char      *b_lock1;
 280         const char      *b_lock2;
 281 };
 282
 283 struct witness_pendhelp {
 284         const char              *wh_type;
 285         struct lock_object      *wh_lock;
 286 };
 287
 288 struct witness_order_list_entry {
 289         const char              *w_name;
 290         struct lock_class       *w_class;
 291 };
 292
 293 /*
 294  * Returns 0 if one of the locks is a spin lock and the other is not.
 295  * Returns 1 otherwise.
 296  */
 297 static __inline int
 298 witness_lock_type_equal(struct witness *w1, struct witness *w2)
 299 {
 300
 301         return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
 302                 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
 303 }
 304
 305 static __inline int
 306 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
 307     const struct witness_lock_order_key *b)
 308 {
 309
 310         return (a->from == b->from && a->to == b->to);
 311 }
 312
 313 static int      _isitmyx(struct witness *w1, struct witness *w2, int rmask,
 314                     const char *fname);
 315 static void     adopt(struct witness *parent, struct witness *child);
 316 static int      blessed(struct witness *, struct witness *);
 317 static void     depart(struct witness *w);
 318 static struct witness   *enroll(const char *description,
 319                             struct lock_class *lock_class);
 320 static struct lock_instance     *find_instance(struct lock_list_entry *list,
 321                                     const struct lock_object *lock);
 322 static int      isitmychild(struct witness *parent, struct witness *child);
 323 static int      isitmydescendant(struct witness *parent, struct witness *child);
 324 static void     itismychild(struct witness *parent, struct witness *child);
 325 static int      sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
 326 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
 327 static int      sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
 328 static int      sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
 329 static void     witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
 330 #ifdef DDB
 331 static void     witness_ddb_compute_levels(void);
 332 static void     witness_ddb_display(int(*)(const char *fmt, ...));
 333 static void     witness_ddb_display_descendants(int(*)(const char *fmt, ...),
 334                     struct witness *, int indent);
 335 static void     witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
 336                     struct witness_list *list);
 337 static void     witness_ddb_level_descendants(struct witness *parent, int l);
 338 static void     witness_ddb_list(struct thread *td);
 339 #endif
 340 static void     witness_debugger(int cond, const char *msg);
 341 static void     witness_free(struct witness *m);
 342 static struct witness   *witness_get(void);
 343 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
 344 static struct witness   *witness_hash_get(const char *key);
 345 static void     witness_hash_put(struct witness *w);
 346 static void     witness_init_hash_tables(void);
 347 static void     witness_increment_graph_generation(void);
 348 static void     witness_lock_list_free(struct lock_list_entry *lle);
 349 static struct lock_list_entry   *witness_lock_list_get(void);
 350 static int      witness_lock_order_add(struct witness *parent,
 351                     struct witness *child);
 352 static int      witness_lock_order_check(struct witness *parent,
 353                     struct witness *child);
 354 static struct witness_lock_order_data   *witness_lock_order_get(
 355                                             struct witness *parent,
 356                                             struct witness *child);
 357 static void     witness_list_lock(struct lock_instance *instance,
 358                     int (*prnt)(const char *fmt, ...));
 359 static int      witness_output(const char *fmt, ...) __printflike(1, 2);
 360 static int      witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
 361 static void     witness_setflag(struct lock_object *lock, int flag, int set);
 362
 363 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
 364     "Witness Locking");
 365
 366 /*
 367  * If set to 0, lock order checking is disabled.  If set to -1,
 368  * witness is completely disabled.  Otherwise witness performs full
 369  * lock order checking for all locks.  At runtime, lock order checking
 370  * may be toggled.  However, witness cannot be reenabled once it is
 371  * completely disabled.
 372  */
 373 static int witness_watch = 1;
 374 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
 375     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
 376
 377 #ifdef KDB
 378 /*
 379  * When KDB is enabled and witness_kdb is 1, it will cause the system
 380  * to drop into kdebug() when:
 381  *      - a lock hierarchy violation occurs
 382  *      - locks are held when going to sleep.
 383  */
 384 #ifdef WITNESS_KDB
 385 int     witness_kdb = 1;
 386 #else
 387 int     witness_kdb = 0;
 388 #endif
 389 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
 390 #endif /* KDB */
 391
 392 #if defined(DDB) || defined(KDB)
 393 /*
 394  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
 395  * to print a stack trace:
 396  *      - a lock hierarchy violation occurs
 397  *      - locks are held when going to sleep.
 398  */
 399 int     witness_trace = 1;
 400 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
 401 #endif /* DDB || KDB */
 402
 403 #ifdef WITNESS_SKIPSPIN
 404 int     witness_skipspin = 1;
 405 #else
 406 int     witness_skipspin = 0;
 407 #endif
 408 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
 409
 410 int badstack_sbuf_size;
 411
 412 int witness_count = WITNESS_COUNT;
 413 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
 414     &witness_count, 0, "");
 415
 416 /*
 417  * Output channel for witness messages.  By default we print to the console.
 418  */
 419 enum witness_channel {
 420         WITNESS_CONSOLE,
 421         WITNESS_LOG,
 422         WITNESS_NONE,
 423 };
 424
 425 static enum witness_channel witness_channel = WITNESS_CONSOLE;
 426 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
 427     CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
 428     "Output channel for warnings");
 429
 430 /*
 431  * Call this to print out the relations between locks.
 432  */
 433 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
 434     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
 435
 436 /*
 437  * Call this to print out the witness faulty stacks.
 438  */
 439 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
 440     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
 441
 442 static struct mtx w_mtx;
 443
 444 /* w_list */
 445 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 446 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 447
 448 /* w_typelist */
 449 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 450 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 451
 452 /* lock list */
 453 static struct lock_list_entry *w_lock_list_free = NULL;
 454 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
 455 static u_int pending_cnt;
 456
 457 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
 458 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
 459 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
 460 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
 461     "");
 462
 463 static struct witness *w_data;
 464 static uint8_t **w_rmatrix;
 465 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 466 static struct witness_hash w_hash;      /* The witness hash table. */
 467
 468 /* The lock order data hash */
 469 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
 470 static struct witness_lock_order_data *w_lofree = NULL;
 471 static struct witness_lock_order_hash w_lohash;
 472 static int w_max_used_index = 0;
 473 static unsigned int w_generation = 0;
 474 static const char w_notrunning[] = "Witness not running\n";
 475 static const char w_stillcold[] = "Witness is still cold\n";
 476 #ifdef __i386__
 477 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
 478 #endif
 479
 480 static struct witness_order_list_entry order_lists[] = {
 481         /*
 482          * sx locks
 483          */
 484         { "proctree", &lock_class_sx },
 485         { "allproc", &lock_class_sx },
 486         { "allprison", &lock_class_sx },
 487         { NULL, NULL },
 488         /*
 489          * Various mutexes
 490          */
 491         { "Giant", &lock_class_mtx_sleep },
 492         { "pipe mutex", &lock_class_mtx_sleep },
 493         { "sigio lock", &lock_class_mtx_sleep },
 494         { "process group", &lock_class_mtx_sleep },
 495 #ifdef  HWPMC_HOOKS
 496         { "pmc-sleep", &lock_class_mtx_sleep },
 497 #endif
 498         { "process lock", &lock_class_mtx_sleep },
 499         { "session", &lock_class_mtx_sleep },
 500         { "uidinfo hash", &lock_class_rw },
 501         { "time lock", &lock_class_mtx_sleep },
 502         { NULL, NULL },
 503         /*
 504          * umtx
 505          */
 506         { "umtx lock", &lock_class_mtx_sleep },
 507         { NULL, NULL },
 508         /*
 509          * Sockets
 510          */
 511         { "accept", &lock_class_mtx_sleep },
 512         { "so_snd", &lock_class_mtx_sleep },
 513         { "so_rcv", &lock_class_mtx_sleep },
 514         { "sellck", &lock_class_mtx_sleep },
 515         { NULL, NULL },
 516         /*
 517          * Routing
 518          */
 519         { "so_rcv", &lock_class_mtx_sleep },
 520         { "radix node head", &lock_class_rm },
 521         { "rtentry", &lock_class_mtx_sleep },
 522         { "ifaddr", &lock_class_mtx_sleep },
 523         { NULL, NULL },
 524         /*
 525          * IPv4 multicast:
 526          * protocol locks before interface locks, after UDP locks.
 527          */
 528         { "in_multi_sx", &lock_class_sx },
 529         { "udpinp", &lock_class_rw },
 530         { "in_multi_list_mtx", &lock_class_mtx_sleep },
 531         { "igmp_mtx", &lock_class_mtx_sleep },
 532         { "ifnet_rw", &lock_class_rw },
 533         { "if_addr_lock", &lock_class_mtx_sleep },
 534         { NULL, NULL },
 535         /*
 536          * IPv6 multicast:
 537          * protocol locks before interface locks, after UDP locks.
 538          */
 539         { "in6_multi_sx", &lock_class_sx },
 540         { "udpinp", &lock_class_rw },
 541         { "in6_multi_list_mtx", &lock_class_mtx_sleep },
 542         { "mld_mtx", &lock_class_mtx_sleep },
 543         { "ifnet_rw", &lock_class_rw },
 544         { "if_addr_lock", &lock_class_mtx_sleep },
 545         { NULL, NULL },
 546         /*
 547          * UNIX Domain Sockets
 548          */
 549         { "unp_link_rwlock", &lock_class_rw },
 550         { "unp_list_lock", &lock_class_mtx_sleep },
 551         { "unp", &lock_class_mtx_sleep },
 552         { "so_snd", &lock_class_mtx_sleep },
 553         { NULL, NULL },
 554         /*
 555          * UDP/IP
 556          */
 557         { "udp", &lock_class_mtx_sleep },
 558         { "udpinp", &lock_class_rw },
 559         { "so_snd", &lock_class_mtx_sleep },
 560         { NULL, NULL },
 561         /*
 562          * TCP/IP
 563          */
 564         { "tcp", &lock_class_mtx_sleep },
 565         { "tcpinp", &lock_class_rw },
 566         { "so_snd", &lock_class_mtx_sleep },
 567         { NULL, NULL },
 568         /*
 569          * BPF
 570          */
 571         { "bpf global lock", &lock_class_sx },
 572         { "bpf interface lock", &lock_class_rw },
 573         { "bpf cdev lock", &lock_class_mtx_sleep },
 574         { NULL, NULL },
 575         /*
 576          * NFS server
 577          */
 578         { "nfsd_mtx", &lock_class_mtx_sleep },
 579         { "so_snd", &lock_class_mtx_sleep },
 580         { NULL, NULL },
 581
 582         /*
 583          * IEEE 802.11
 584          */
 585         { "802.11 com lock", &lock_class_mtx_sleep},
 586         { NULL, NULL },
 587         /*
 588          * Network drivers
 589          */
 590         { "network driver", &lock_class_mtx_sleep},
 591         { NULL, NULL },
 592
 593         /*
 594          * Netgraph
 595          */
 596         { "ng_node", &lock_class_mtx_sleep },
 597         { "ng_worklist", &lock_class_mtx_sleep },
 598         { NULL, NULL },
 599         /*
 600          * CDEV
 601          */
 602         { "vm map (system)", &lock_class_mtx_sleep },
 603         { "vnode interlock", &lock_class_mtx_sleep },
 604         { "cdev", &lock_class_mtx_sleep },
 605         { "devthrd", &lock_class_mtx_sleep },
 606         { NULL, NULL },
 607         /*
 608          * VM
 609          */
 610         { "vm map (user)", &lock_class_sx },
 611         { "vm object", &lock_class_rw },
 612         { "vm page", &lock_class_mtx_sleep },
 613         { "pmap pv global", &lock_class_rw },
 614         { "pmap", &lock_class_mtx_sleep },
 615         { "pmap pv list", &lock_class_rw },
 616         { "vm page free queue", &lock_class_mtx_sleep },
 617         { "vm pagequeue", &lock_class_mtx_sleep },
 618         { NULL, NULL },
 619         /*
 620          * kqueue/VFS interaction
 621          */
 622         { "kqueue", &lock_class_mtx_sleep },
 623         { "struct mount mtx", &lock_class_mtx_sleep },
 624         { "vnode interlock", &lock_class_mtx_sleep },
 625         { NULL, NULL },
 626         /*
 627          * VFS namecache
 628          */
 629         { "ncvn", &lock_class_mtx_sleep },
 630         { "ncbuc", &lock_class_rw },
 631         { "vnode interlock", &lock_class_mtx_sleep },
 632         { "ncneg", &lock_class_mtx_sleep },
 633         { NULL, NULL },
 634         /*
 635          * ZFS locking
 636          */
 637         { "dn->dn_mtx", &lock_class_sx },
 638         { "dr->dt.di.dr_mtx", &lock_class_sx },
 639         { "db->db_mtx", &lock_class_sx },
 640         { NULL, NULL },
 641         /*
 642          * TCP log locks
 643          */
 644         { "TCP ID tree", &lock_class_rw },
 645         { "tcp log id bucket", &lock_class_mtx_sleep },
 646         { "tcpinp", &lock_class_rw },
 647         { "TCP log expireq", &lock_class_mtx_sleep },
 648         { NULL, NULL },
 649         /*
 650          * spin locks
 651          */
 652 #ifdef SMP
 653         { "ap boot", &lock_class_mtx_spin },
 654 #endif
 655         { "rm.mutex_mtx", &lock_class_mtx_spin },
 656         { "sio", &lock_class_mtx_spin },
 657 #ifdef __i386__
 658         { "cy", &lock_class_mtx_spin },
 659 #endif
 660 #ifdef __sparc64__
 661         { "pcib_mtx", &lock_class_mtx_spin },
 662         { "rtc_mtx", &lock_class_mtx_spin },
 663 #endif
 664         { "scc_hwmtx", &lock_class_mtx_spin },
 665         { "uart_hwmtx", &lock_class_mtx_spin },
 666         { "fast_taskqueue", &lock_class_mtx_spin },
 667         { "intr table", &lock_class_mtx_spin },
 668         { "process slock", &lock_class_mtx_spin },
 669         { "syscons video lock", &lock_class_mtx_spin },
 670         { "sleepq chain", &lock_class_mtx_spin },
 671         { "rm_spinlock", &lock_class_mtx_spin },
 672         { "turnstile chain", &lock_class_mtx_spin },
 673         { "turnstile lock", &lock_class_mtx_spin },
 674         { "sched lock", &lock_class_mtx_spin },
 675         { "td_contested", &lock_class_mtx_spin },
 676         { "callout", &lock_class_mtx_spin },
 677         { "entropy harvest mutex", &lock_class_mtx_spin },
 678 #ifdef SMP
 679         { "smp rendezvous", &lock_class_mtx_spin },
 680 #endif
 681 #ifdef __powerpc__
 682         { "tlb0", &lock_class_mtx_spin },
 683 #endif
 684         { NULL, NULL },
 685         { "sched lock", &lock_class_mtx_spin },
 686 #ifdef  HWPMC_HOOKS
 687         { "pmc-per-proc", &lock_class_mtx_spin },
 688 #endif
 689         { NULL, NULL },
 690         /*
 691          * leaf locks
 692          */
 693         { "intrcnt", &lock_class_mtx_spin },
 694         { "icu", &lock_class_mtx_spin },
 695 #if defined(SMP) && defined(__sparc64__)
 696         { "ipi", &lock_class_mtx_spin },
 697 #endif
 698 #ifdef __i386__
 699         { "allpmaps", &lock_class_mtx_spin },
 700         { "descriptor tables", &lock_class_mtx_spin },
 701 #endif
 702         { "clk", &lock_class_mtx_spin },
 703         { "cpuset", &lock_class_mtx_spin },
 704         { "mprof lock", &lock_class_mtx_spin },
 705         { "zombie lock", &lock_class_mtx_spin },
 706         { "ALD Queue", &lock_class_mtx_spin },
 707 #if defined(__i386__) || defined(__amd64__)
 708         { "pcicfg", &lock_class_mtx_spin },
 709         { "NDIS thread lock", &lock_class_mtx_spin },
 710 #endif
 711         { "tw_osl_io_lock", &lock_class_mtx_spin },
 712         { "tw_osl_q_lock", &lock_class_mtx_spin },
 713         { "tw_cl_io_lock", &lock_class_mtx_spin },
 714         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 715         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 716 #ifdef  HWPMC_HOOKS
 717         { "pmc-leaf", &lock_class_mtx_spin },
 718 #endif
 719         { "blocked lock", &lock_class_mtx_spin },
 720         { NULL, NULL },
 721         { NULL, NULL }
 722 };
 723
 724 /*
 725  * Pairs of locks which have been blessed.  Witness does not complain about
 726  * order problems with blessed lock pairs.  Please do not add an entry to the
 727  * table without an explanatory comment.
 728  */
 729 static struct witness_blessed blessed_list[] = {
 730         /*
 731          * See the comment in ufs_dirhash.c.  Basically, a vnode lock serializes
 732          * both lock orders, so a deadlock cannot happen as a result of this
 733          * LOR.
 734          */
 735         { "dirhash",    "bufwait" },
 736
 737         /*
 738          * A UFS vnode may be locked in vget() while a buffer belonging to the
 739          * parent directory vnode is locked.
 740          */
 741         { "ufs",        "bufwait" },
 742 };
 743
 744 /*
 745  * This global is set to 0 once it becomes safe to use the witness code.
 746  */
 747 static int witness_cold = 1;
 748
 749 /*
 750  * This global is set to 1 once the static lock orders have been enrolled
 751  * so that a warning can be issued for any spin locks enrolled later.
 752  */
 753 static int witness_spin_warn = 0;
 754
 755 /* Trim useless garbage from filenames. */
 756 static const char *
 757 fixup_filename(const char *file)
 758 {
 759
 760         if (file == NULL)
 761                 return (NULL);
 762         while (strncmp(file, "../", 3) == 0)
 763                 file += 3;
 764         return (file);
 765 }
 766
 767 /*
 768  * Calculate the size of early witness structures.
 769  */
 770 int
 771 witness_startup_count(void)
 772 {
 773         int sz;
 774
 775         sz = sizeof(struct witness) * witness_count;
 776         sz += sizeof(*w_rmatrix) * (witness_count + 1);
 777         sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
 778             (witness_count + 1);
 779
 780         return (sz);
 781 }
 782
 783 /*
 784  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 785  * assume that the early boot is single-threaded at least until after this
 786  * routine is completed.
 787  */
 788 void
 789 witness_startup(void *mem)
 790 {
 791         struct lock_object *lock;
 792         struct witness_order_list_entry *order;
 793         struct witness *w, *w1;
 794         uintptr_t p;
 795         int i;
 796
 797         p = (uintptr_t)mem;
 798         w_data = (void *)p;
 799         p += sizeof(struct witness) * witness_count;
 800
 801         w_rmatrix = (void *)p;
 802         p += sizeof(*w_rmatrix) * (witness_count + 1);
 803
 804         for (i = 0; i < witness_count + 1; i++) {
 805                 w_rmatrix[i] = (void *)p;
 806                 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
 807         }
 808         badstack_sbuf_size = witness_count * 256;
 809
 810         /*
 811          * We have to release Giant before initializing its witness
 812          * structure so that WITNESS doesn't get confused.
 813          */
 814         mtx_unlock(&Giant);
 815         mtx_assert(&Giant, MA_NOTOWNED);
 816
 817         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 818         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 819             MTX_NOWITNESS | MTX_NOPROFILE);
 820         for (i = witness_count - 1; i >= 0; i--) {
 821                 w = &w_data[i];
 822                 memset(w, 0, sizeof(*w));
 823                 w_data[i].w_index = i;  /* Witness index never changes. */
 824                 witness_free(w);
 825         }
 826         KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
 827             ("%s: Invalid list of free witness objects", __func__));
 828
 829         /* Witness with index 0 is not used to aid in debugging. */
 830         STAILQ_REMOVE_HEAD(&w_free, w_list);
 831         w_free_cnt--;
 832
 833         for (i = 0; i < witness_count; i++) {
 834                 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
 835                     (witness_count + 1));
 836         }
 837
 838         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 839                 witness_lock_list_free(&w_locklistdata[i]);
 840         witness_init_hash_tables();
 841
 842         /* First add in all the specified order lists. */
 843         for (order = order_lists; order->w_name != NULL; order++) {
 844                 w = enroll(order->w_name, order->w_class);
 845                 if (w == NULL)
 846                         continue;
 847                 w->w_file = "order list";
 848                 for (order++; order->w_name != NULL; order++) {
 849                         w1 = enroll(order->w_name, order->w_class);
 850                         if (w1 == NULL)
 851                                 continue;
 852                         w1->w_file = "order list";
 853                         itismychild(w, w1);
 854                         w = w1;
 855                 }
 856         }
 857         witness_spin_warn = 1;
 858
 859         /* Iterate through all locks and add them to witness. */
 860         for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
 861                 lock = pending_locks[i].wh_lock;
 862                 KASSERT(lock->lo_flags & LO_WITNESS,
 863                     ("%s: lock %s is on pending list but not LO_WITNESS",
 864                     __func__, lock->lo_name));
 865                 lock->lo_witness = enroll(pending_locks[i].wh_type,
 866                     LOCK_CLASS(lock));
 867         }
 868
 869         /* Mark the witness code as being ready for use. */
 870         witness_cold = 0;
 871
 872         mtx_lock(&Giant);
 873 }
 874
 875 void
 876 witness_init(struct lock_object *lock, const char *type)
 877 {
 878         struct lock_class *class;
 879
 880         /* Various sanity checks. */
 881         class = LOCK_CLASS(lock);
 882         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 883             (class->lc_flags & LC_RECURSABLE) == 0)
 884                 kassert_panic("%s: lock (%s) %s can not be recursable",
 885                     __func__, class->lc_name, lock->lo_name);
 886         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 887             (class->lc_flags & LC_SLEEPABLE) == 0)
 888                 kassert_panic("%s: lock (%s) %s can not be sleepable",
 889                     __func__, class->lc_name, lock->lo_name);
 890         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 891             (class->lc_flags & LC_UPGRADABLE) == 0)
 892                 kassert_panic("%s: lock (%s) %s can not be upgradable",
 893                     __func__, class->lc_name, lock->lo_name);
 894
 895         /*
 896          * If we shouldn't watch this lock, then just clear lo_witness.
 897          * Otherwise, if witness_cold is set, then it is too early to
 898          * enroll this lock, so defer it to witness_initialize() by adding
 899          * it to the pending_locks list.  If it is not too early, then enroll
 900          * the lock now.
 901          */
 902         if (witness_watch < 1 || panicstr != NULL ||
 903             (lock->lo_flags & LO_WITNESS) == 0)
 904                 lock->lo_witness = NULL;
 905         else if (witness_cold) {
 906                 pending_locks[pending_cnt].wh_lock = lock;
 907                 pending_locks[pending_cnt++].wh_type = type;
 908                 if (pending_cnt > WITNESS_PENDLIST)
 909                         panic("%s: pending locks list is too small, "
 910                             "increase WITNESS_PENDLIST\n",
 911                             __func__);
 912         } else
 913                 lock->lo_witness = enroll(type, class);
 914 }
 915
 916 void
 917 witness_destroy(struct lock_object *lock)
 918 {
 919         struct lock_class *class;
 920         struct witness *w;
 921
 922         class = LOCK_CLASS(lock);
 923
 924         if (witness_cold)
 925                 panic("lock (%s) %s destroyed while witness_cold",
 926                     class->lc_name, lock->lo_name);
 927
 928         /* XXX: need to verify that no one holds the lock */
 929         if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
 930                 return;
 931         w = lock->lo_witness;
 932
 933         mtx_lock_spin(&w_mtx);
 934         MPASS(w->w_refcount > 0);
 935         w->w_refcount--;
 936
 937         if (w->w_refcount == 0)
 938                 depart(w);
 939         mtx_unlock_spin(&w_mtx);
 940 }
 941
 942 #ifdef DDB
 943 static void
 944 witness_ddb_compute_levels(void)
 945 {
 946         struct witness *w;
 947
 948         /*
 949          * First clear all levels.
 950          */
 951         STAILQ_FOREACH(w, &w_all, w_list)
 952                 w->w_ddb_level = -1;
 953
 954         /*
 955          * Look for locks with no parents and level all their descendants.
 956          */
 957         STAILQ_FOREACH(w, &w_all, w_list) {
 958
 959                 /* If the witness has ancestors (is not a root), skip it. */
 960                 if (w->w_num_ancestors > 0)
 961                         continue;
 962                 witness_ddb_level_descendants(w, 0);
 963         }
 964 }
 965
 966 static void
 967 witness_ddb_level_descendants(struct witness *w, int l)
 968 {
 969         int i;
 970
 971         if (w->w_ddb_level >= l)
 972                 return;
 973
 974         w->w_ddb_level = l;
 975         l++;
 976
 977         for (i = 1; i <= w_max_used_index; i++) {
 978                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 979                         witness_ddb_level_descendants(&w_data[i], l);
 980         }
 981 }
 982
 983 static void
 984 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
 985     struct witness *w, int indent)
 986 {
 987         int i;
 988
 989         for (i = 0; i < indent; i++)
 990                 prnt(" ");
 991         prnt("%s (type: %s, depth: %d, active refs: %d)",
 992              w->w_name, w->w_class->lc_name,
 993              w->w_ddb_level, w->w_refcount);
 994         if (w->w_displayed) {
 995                 prnt(" -- (already displayed)\n");
 996                 return;
 997         }
 998         w->w_displayed = 1;
 999         if (w->w_file != NULL && w->w_line != 0)
1000                 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1001                     w->w_line);
1002         else
1003                 prnt(" -- never acquired\n");
1004         indent++;
1005         WITNESS_INDEX_ASSERT(w->w_index);
1006         for (i = 1; i <= w_max_used_index; i++) {
1007                 if (db_pager_quit)
1008                         return;
1009                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1010                         witness_ddb_display_descendants(prnt, &w_data[i],
1011                             indent);
1012         }
1013 }
1014
1015 static void
1016 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1017     struct witness_list *list)
1018 {
1019         struct witness *w;
1020
1021         STAILQ_FOREACH(w, list, w_typelist) {
1022                 if (w->w_file == NULL || w->w_ddb_level > 0)
1023                         continue;
1024
1025                 /* This lock has no anscestors - display its descendants. */
1026                 witness_ddb_display_descendants(prnt, w, 0);
1027                 if (db_pager_quit)
1028                         return;
1029         }
1030 }
1031
1032 static void
1033 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1034 {
1035         struct witness *w;
1036
1037         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1038         witness_ddb_compute_levels();
1039
1040         /* Clear all the displayed flags. */
1041         STAILQ_FOREACH(w, &w_all, w_list)
1042                 w->w_displayed = 0;
1043
1044         /*
1045          * First, handle sleep locks which have been acquired at least
1046          * once.
1047          */
1048         prnt("Sleep locks:\n");
1049         witness_ddb_display_list(prnt, &w_sleep);
1050         if (db_pager_quit)
1051                 return;
1052
1053         /*
1054          * Now do spin locks which have been acquired at least once.
1055          */
1056         prnt("\nSpin locks:\n");
1057         witness_ddb_display_list(prnt, &w_spin);
1058         if (db_pager_quit)
1059                 return;
1060
1061         /*
1062          * Finally, any locks which have not been acquired yet.
1063          */
1064         prnt("\nLocks which were never acquired:\n");
1065         STAILQ_FOREACH(w, &w_all, w_list) {
1066                 if (w->w_file != NULL || w->w_refcount == 0)
1067                         continue;
1068                 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1069                     w->w_class->lc_name, w->w_ddb_level);
1070                 if (db_pager_quit)
1071                         return;
1072         }
1073 }
1074 #endif /* DDB */
1075
1076 int
1077 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1078 {
1079
1080         if (witness_watch == -1 || panicstr != NULL)
1081                 return (0);
1082
1083         /* Require locks that witness knows about. */
1084         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1085             lock2->lo_witness == NULL)
1086                 return (EINVAL);
1087
1088         mtx_assert(&w_mtx, MA_NOTOWNED);
1089         mtx_lock_spin(&w_mtx);
1090
1091         /*
1092          * If we already have either an explicit or implied lock order that
1093          * is the other way around, then return an error.
1094          */
1095         if (witness_watch &&
1096             isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1097                 mtx_unlock_spin(&w_mtx);
1098                 return (EDOOFUS);
1099         }
1100
1101         /* Try to add the new order. */
1102         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1103             lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1104         itismychild(lock1->lo_witness, lock2->lo_witness);
1105         mtx_unlock_spin(&w_mtx);
1106         return (0);
1107 }
1108
1109 void
1110 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1111     int line, struct lock_object *interlock)
1112 {
1113         struct lock_list_entry *lock_list, *lle;
1114         struct lock_instance *lock1, *lock2, *plock;
1115         struct lock_class *class, *iclass;
1116         struct witness *w, *w1;
1117         struct thread *td;
1118         int i, j;
1119
1120         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1121             panicstr != NULL)
1122                 return;
1123
1124         w = lock->lo_witness;
1125         class = LOCK_CLASS(lock);
1126         td = curthread;
1127
1128         if (class->lc_flags & LC_SLEEPLOCK) {
1129
1130                 /*
1131                  * Since spin locks include a critical section, this check
1132                  * implicitly enforces a lock order of all sleep locks before
1133                  * all spin locks.
1134                  */
1135                 if (td->td_critnest != 0 && !kdb_active)
1136                         kassert_panic("acquiring blockable sleep lock with "
1137                             "spinlock or critical section held (%s) %s @ %s:%d",
1138                             class->lc_name, lock->lo_name,
1139                             fixup_filename(file), line);
1140
1141                 /*
1142                  * If this is the first lock acquired then just return as
1143                  * no order checking is needed.
1144                  */
1145                 lock_list = td->td_sleeplocks;
1146                 if (lock_list == NULL || lock_list->ll_count == 0)
1147                         return;
1148         } else {
1149
1150                 /*
1151                  * If this is the first lock, just return as no order
1152                  * checking is needed.  Avoid problems with thread
1153                  * migration pinning the thread while checking if
1154                  * spinlocks are held.  If at least one spinlock is held
1155                  * the thread is in a safe path and it is allowed to
1156                  * unpin it.
1157                  */
1158                 sched_pin();
1159                 lock_list = PCPU_GET(spinlocks);
1160                 if (lock_list == NULL || lock_list->ll_count == 0) {
1161                         sched_unpin();
1162                         return;
1163                 }
1164                 sched_unpin();
1165         }
1166
1167         /*
1168          * Check to see if we are recursing on a lock we already own.  If
1169          * so, make sure that we don't mismatch exclusive and shared lock
1170          * acquires.
1171          */
1172         lock1 = find_instance(lock_list, lock);
1173         if (lock1 != NULL) {
1174                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1175                     (flags & LOP_EXCLUSIVE) == 0) {
1176                         witness_output("shared lock of (%s) %s @ %s:%d\n",
1177                             class->lc_name, lock->lo_name,
1178                             fixup_filename(file), line);
1179                         witness_output("while exclusively locked from %s:%d\n",
1180                             fixup_filename(lock1->li_file), lock1->li_line);
1181                         kassert_panic("excl->share");
1182                 }
1183                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1184                     (flags & LOP_EXCLUSIVE) != 0) {
1185                         witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1186                             class->lc_name, lock->lo_name,
1187                             fixup_filename(file), line);
1188                         witness_output("while share locked from %s:%d\n",
1189                             fixup_filename(lock1->li_file), lock1->li_line);
1190                         kassert_panic("share->excl");
1191                 }
1192                 return;
1193         }
1194
1195         /* Warn if the interlock is not locked exactly once. */
1196         if (interlock != NULL) {
1197                 iclass = LOCK_CLASS(interlock);
1198                 lock1 = find_instance(lock_list, interlock);
1199                 if (lock1 == NULL)
1200                         kassert_panic("interlock (%s) %s not locked @ %s:%d",
1201                             iclass->lc_name, interlock->lo_name,
1202                             fixup_filename(file), line);
1203                 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1204                         kassert_panic("interlock (%s) %s recursed @ %s:%d",
1205                             iclass->lc_name, interlock->lo_name,
1206                             fixup_filename(file), line);
1207         }
1208
1209         /*
1210          * Find the previously acquired lock, but ignore interlocks.
1211          */
1212         plock = &lock_list->ll_children[lock_list->ll_count - 1];
1213         if (interlock != NULL && plock->li_lock == interlock) {
1214                 if (lock_list->ll_count > 1)
1215                         plock =
1216                             &lock_list->ll_children[lock_list->ll_count - 2];
1217                 else {
1218                         lle = lock_list->ll_next;
1219
1220                         /*
1221                          * The interlock is the only lock we hold, so
1222                          * simply return.
1223                          */
1224                         if (lle == NULL)
1225                                 return;
1226                         plock = &lle->ll_children[lle->ll_count - 1];
1227                 }
1228         }
1229
1230         /*
1231          * Try to perform most checks without a lock.  If this succeeds we
1232          * can skip acquiring the lock and return success.  Otherwise we redo
1233          * the check with the lock held to handle races with concurrent updates.
1234          */
1235         w1 = plock->li_lock->lo_witness;
1236         if (witness_lock_order_check(w1, w))
1237                 return;
1238
1239         mtx_lock_spin(&w_mtx);
1240         if (witness_lock_order_check(w1, w)) {
1241                 mtx_unlock_spin(&w_mtx);
1242                 return;
1243         }
1244         witness_lock_order_add(w1, w);
1245
1246         /*
1247          * Check for duplicate locks of the same type.  Note that we only
1248          * have to check for this on the last lock we just acquired.  Any
1249          * other cases will be caught as lock order violations.
1250          */
1251         if (w1 == w) {
1252                 i = w->w_index;
1253                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1254                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1255                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
1256                         w->w_reversed = 1;
1257                         mtx_unlock_spin(&w_mtx);
1258                         witness_output(
1259                             "acquiring duplicate lock of same type: \"%s\"\n",
1260                             w->w_name);
1261                         witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1262                             fixup_filename(plock->li_file), plock->li_line);
1263                         witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1264                             fixup_filename(file), line);
1265                         witness_debugger(1, __func__);
1266                 } else
1267                         mtx_unlock_spin(&w_mtx);
1268                 return;
1269         }
1270         mtx_assert(&w_mtx, MA_OWNED);
1271
1272         /*
1273          * If we know that the lock we are acquiring comes after
1274          * the lock we most recently acquired in the lock order tree,
1275          * then there is no need for any further checks.
1276          */
1277         if (isitmychild(w1, w))
1278                 goto out;
1279
1280         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1281                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1282
1283                         MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1284                         lock1 = &lle->ll_children[i];
1285
1286                         /*
1287                          * Ignore the interlock.
1288                          */
1289                         if (interlock == lock1->li_lock)
1290                                 continue;
1291
1292                         /*
1293                          * If this lock doesn't undergo witness checking,
1294                          * then skip it.
1295                          */
1296                         w1 = lock1->li_lock->lo_witness;
1297                         if (w1 == NULL) {
1298                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1299                                     ("lock missing witness structure"));
1300                                 continue;
1301                         }
1302
1303                         /*
1304                          * If we are locking Giant and this is a sleepable
1305                          * lock, then skip it.
1306                          */
1307                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1308                             lock == &Giant.lock_object)
1309                                 continue;
1310
1311                         /*
1312                          * If we are locking a sleepable lock and this lock
1313                          * is Giant, then skip it.
1314                          */
1315                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1316                             lock1->li_lock == &Giant.lock_object)
1317                                 continue;
1318
1319                         /*
1320                          * If we are locking a sleepable lock and this lock
1321                          * isn't sleepable, we want to treat it as a lock
1322                          * order violation to enfore a general lock order of
1323                          * sleepable locks before non-sleepable locks.
1324                          */
1325                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1326                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1327                                 goto reversal;
1328
1329                         /*
1330                          * If we are locking Giant and this is a non-sleepable
1331                          * lock, then treat it as a reversal.
1332                          */
1333                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1334                             lock == &Giant.lock_object)
1335                                 goto reversal;
1336
1337                         /*
1338                          * Check the lock order hierarchy for a reveresal.
1339                          */
1340                         if (!isitmydescendant(w, w1))
1341                                 continue;
1342                 reversal:
1343
1344                         /*
1345                          * We have a lock order violation, check to see if it
1346                          * is allowed or has already been yelled about.
1347                          */
1348
1349                         /* Bail if this violation is known */
1350                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1351                                 goto out;
1352
1353                         /* Record this as a violation */
1354                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1355                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1356                         w->w_reversed = w1->w_reversed = 1;
1357                         witness_increment_graph_generation();
1358
1359                         /*
1360                          * If the lock order is blessed, bail before logging
1361                          * anything.  We don't look for other lock order
1362                          * violations though, which may be a bug.
1363                          */
1364                         if (blessed(w, w1))
1365                                 goto out;
1366                         mtx_unlock_spin(&w_mtx);
1367
1368 #ifdef WITNESS_NO_VNODE
1369                         /*
1370                          * There are known LORs between VNODE locks. They are
1371                          * not an indication of a bug. VNODE locks are flagged
1372                          * as such (LO_IS_VNODE) and we don't yell if the LOR
1373                          * is between 2 VNODE locks.
1374                          */
1375                         if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1376                             (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1377                                 return;
1378 #endif
1379
1380                         /*
1381                          * Ok, yell about it.
1382                          */
1383                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1384                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1385                                 witness_output(
1386                 "lock order reversal: (sleepable after non-sleepable)\n");
1387                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1388                             && lock == &Giant.lock_object)
1389                                 witness_output(
1390                 "lock order reversal: (Giant after non-sleepable)\n");
1391                         else
1392                                 witness_output("lock order reversal:\n");
1393
1394                         /*
1395                          * Try to locate an earlier lock with
1396                          * witness w in our list.
1397                          */
1398                         do {
1399                                 lock2 = &lle->ll_children[i];
1400                                 MPASS(lock2->li_lock != NULL);
1401                                 if (lock2->li_lock->lo_witness == w)
1402                                         break;
1403                                 if (i == 0 && lle->ll_next != NULL) {
1404                                         lle = lle->ll_next;
1405                                         i = lle->ll_count - 1;
1406                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1407                                 } else
1408                                         i--;
1409                         } while (i >= 0);
1410                         if (i < 0) {
1411                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1412                                     lock1->li_lock, lock1->li_lock->lo_name,
1413                                     w1->w_name, fixup_filename(lock1->li_file),
1414                                     lock1->li_line);
1415                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1416                                     lock->lo_name, w->w_name,
1417                                     fixup_filename(file), line);
1418                         } else {
1419                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1420                                     lock2->li_lock, lock2->li_lock->lo_name,
1421                                     lock2->li_lock->lo_witness->w_name,
1422                                     fixup_filename(lock2->li_file),
1423                                     lock2->li_line);
1424                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1425                                     lock1->li_lock, lock1->li_lock->lo_name,
1426                                     w1->w_name, fixup_filename(lock1->li_file),
1427                                     lock1->li_line);
1428                                 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1429                                     lock->lo_name, w->w_name,
1430                                     fixup_filename(file), line);
1431                         }
1432                         witness_debugger(1, __func__);
1433                         return;
1434                 }
1435         }
1436
1437         /*
1438          * If requested, build a new lock order.  However, don't build a new
1439          * relationship between a sleepable lock and Giant if it is in the
1440          * wrong direction.  The correct lock order is that sleepable locks
1441          * always come before Giant.
1442          */
1443         if (flags & LOP_NEWORDER &&
1444             !(plock->li_lock == &Giant.lock_object &&
1445             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1446                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1447                     w->w_name, plock->li_lock->lo_witness->w_name);
1448                 itismychild(plock->li_lock->lo_witness, w);
1449         }
1450 out:
1451         mtx_unlock_spin(&w_mtx);
1452 }
1453
1454 void
1455 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1456 {
1457         struct lock_list_entry **lock_list, *lle;
1458         struct lock_instance *instance;
1459         struct witness *w;
1460         struct thread *td;
1461
1462         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1463             panicstr != NULL)
1464                 return;
1465         w = lock->lo_witness;
1466         td = curthread;
1467
1468         /* Determine lock list for this lock. */
1469         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1470                 lock_list = &td->td_sleeplocks;
1471         else
1472                 lock_list = PCPU_PTR(spinlocks);
1473
1474         /* Check to see if we are recursing on a lock we already own. */
1475         instance = find_instance(*lock_list, lock);
1476         if (instance != NULL) {
1477                 instance->li_flags++;
1478                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1479                     td->td_proc->p_pid, lock->lo_name,
1480                     instance->li_flags & LI_RECURSEMASK);
1481                 instance->li_file = file;
1482                 instance->li_line = line;
1483                 return;
1484         }
1485
1486         /* Update per-witness last file and line acquire. */
1487         w->w_file = file;
1488         w->w_line = line;
1489
1490         /* Find the next open lock instance in the list and fill it. */
1491         lle = *lock_list;
1492         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1493                 lle = witness_lock_list_get();
1494                 if (lle == NULL)
1495                         return;
1496                 lle->ll_next = *lock_list;
1497                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1498                     td->td_proc->p_pid, lle);
1499                 *lock_list = lle;
1500         }
1501         instance = &lle->ll_children[lle->ll_count++];
1502         instance->li_lock = lock;
1503         instance->li_line = line;
1504         instance->li_file = file;
1505         if ((flags & LOP_EXCLUSIVE) != 0)
1506                 instance->li_flags = LI_EXCLUSIVE;
1507         else
1508                 instance->li_flags = 0;
1509         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1510             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1511 }
1512
1513 void
1514 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1515 {
1516         struct lock_instance *instance;
1517         struct lock_class *class;
1518
1519         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1520         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1521                 return;
1522         class = LOCK_CLASS(lock);
1523         if (witness_watch) {
1524                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1525                         kassert_panic(
1526                             "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1527                             class->lc_name, lock->lo_name,
1528                             fixup_filename(file), line);
1529                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1530                         kassert_panic(
1531                             "upgrade of non-sleep lock (%s) %s @ %s:%d",
1532                             class->lc_name, lock->lo_name,
1533                             fixup_filename(file), line);
1534         }
1535         instance = find_instance(curthread->td_sleeplocks, lock);
1536         if (instance == NULL) {
1537                 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1538                     class->lc_name, lock->lo_name,
1539                     fixup_filename(file), line);
1540                 return;
1541         }
1542         if (witness_watch) {
1543                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1544                         kassert_panic(
1545                             "upgrade of exclusive lock (%s) %s @ %s:%d",
1546                             class->lc_name, lock->lo_name,
1547                             fixup_filename(file), line);
1548                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1549                         kassert_panic(
1550                             "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1551                             class->lc_name, lock->lo_name,
1552                             instance->li_flags & LI_RECURSEMASK,
1553                             fixup_filename(file), line);
1554         }
1555         instance->li_flags |= LI_EXCLUSIVE;
1556 }
1557
1558 void
1559 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1560     int line)
1561 {
1562         struct lock_instance *instance;
1563         struct lock_class *class;
1564
1565         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1566         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1567                 return;
1568         class = LOCK_CLASS(lock);
1569         if (witness_watch) {
1570                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1571                         kassert_panic(
1572                             "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1573                             class->lc_name, lock->lo_name,
1574                             fixup_filename(file), line);
1575                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1576                         kassert_panic(
1577                             "downgrade of non-sleep lock (%s) %s @ %s:%d",
1578                             class->lc_name, lock->lo_name,
1579                             fixup_filename(file), line);
1580         }
1581         instance = find_instance(curthread->td_sleeplocks, lock);
1582         if (instance == NULL) {
1583                 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1584                     class->lc_name, lock->lo_name,
1585                     fixup_filename(file), line);
1586                 return;
1587         }
1588         if (witness_watch) {
1589                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1590                         kassert_panic(
1591                             "downgrade of shared lock (%s) %s @ %s:%d",
1592                             class->lc_name, lock->lo_name,
1593                             fixup_filename(file), line);
1594                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1595                         kassert_panic(
1596                             "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1597                             class->lc_name, lock->lo_name,
1598                             instance->li_flags & LI_RECURSEMASK,
1599                             fixup_filename(file), line);
1600         }
1601         instance->li_flags &= ~LI_EXCLUSIVE;
1602 }
1603
1604 void
1605 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1606 {
1607         struct lock_list_entry **lock_list, *lle;
1608         struct lock_instance *instance;
1609         struct lock_class *class;
1610         struct thread *td;
1611         register_t s;
1612         int i, j;
1613
1614         if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1615                 return;
1616         td = curthread;
1617         class = LOCK_CLASS(lock);
1618
1619         /* Find lock instance associated with this lock. */
1620         if (class->lc_flags & LC_SLEEPLOCK)
1621                 lock_list = &td->td_sleeplocks;
1622         else
1623                 lock_list = PCPU_PTR(spinlocks);
1624         lle = *lock_list;
1625         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1626                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1627                         instance = &(*lock_list)->ll_children[i];
1628                         if (instance->li_lock == lock)
1629                                 goto found;
1630                 }
1631
1632         /*
1633          * When disabling WITNESS through witness_watch we could end up in
1634          * having registered locks in the td_sleeplocks queue.
1635          * We have to make sure we flush these queues, so just search for
1636          * eventual register locks and remove them.
1637          */
1638         if (witness_watch > 0) {
1639                 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1640                     lock->lo_name, fixup_filename(file), line);
1641                 return;
1642         } else {
1643                 return;
1644         }
1645 found:
1646
1647         /* First, check for shared/exclusive mismatches. */
1648         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1649             (flags & LOP_EXCLUSIVE) == 0) {
1650                 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1651                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1652                 witness_output("while exclusively locked from %s:%d\n",
1653                     fixup_filename(instance->li_file), instance->li_line);
1654                 kassert_panic("excl->ushare");
1655         }
1656         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1657             (flags & LOP_EXCLUSIVE) != 0) {
1658                 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1659                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1660                 witness_output("while share locked from %s:%d\n",
1661                     fixup_filename(instance->li_file),
1662                     instance->li_line);
1663                 kassert_panic("share->uexcl");
1664         }
1665         /* If we are recursed, unrecurse. */
1666         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1667                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1668                     td->td_proc->p_pid, instance->li_lock->lo_name,
1669                     instance->li_flags);
1670                 instance->li_flags--;
1671                 return;
1672         }
1673         /* The lock is now being dropped, check for NORELEASE flag */
1674         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1675                 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1676                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1677                 kassert_panic("lock marked norelease");
1678         }
1679
1680         /* Otherwise, remove this item from the list. */
1681         s = intr_disable();
1682         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1683             td->td_proc->p_pid, instance->li_lock->lo_name,
1684             (*lock_list)->ll_count - 1);
1685         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1686                 (*lock_list)->ll_children[j] =
1687                     (*lock_list)->ll_children[j + 1];
1688         (*lock_list)->ll_count--;
1689         intr_restore(s);
1690
1691         /*
1692          * In order to reduce contention on w_mtx, we want to keep always an
1693          * head object into lists so that frequent allocation from the
1694          * free witness pool (and subsequent locking) is avoided.
1695          * In order to maintain the current code simple, when the head
1696          * object is totally unloaded it means also that we do not have
1697          * further objects in the list, so the list ownership needs to be
1698          * hand over to another object if the current head needs to be freed.
1699          */
1700         if ((*lock_list)->ll_count == 0) {
1701                 if (*lock_list == lle) {
1702                         if (lle->ll_next == NULL)
1703                                 return;
1704                 } else
1705                         lle = *lock_list;
1706                 *lock_list = lle->ll_next;
1707                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1708                     td->td_proc->p_pid, lle);
1709                 witness_lock_list_free(lle);
1710         }
1711 }
1712
1713 void
1714 witness_thread_exit(struct thread *td)
1715 {
1716         struct lock_list_entry *lle;
1717         int i, n;
1718
1719         lle = td->td_sleeplocks;
1720         if (lle == NULL || panicstr != NULL)
1721                 return;
1722         if (lle->ll_count != 0) {
1723                 for (n = 0; lle != NULL; lle = lle->ll_next)
1724                         for (i = lle->ll_count - 1; i >= 0; i--) {
1725                                 if (n == 0)
1726                                         witness_output(
1727                     "Thread %p exiting with the following locks held:\n", td);
1728                                 n++;
1729                                 witness_list_lock(&lle->ll_children[i],
1730                                     witness_output);
1731
1732                         }
1733                 kassert_panic(
1734                     "Thread %p cannot exit while holding sleeplocks\n", td);
1735         }
1736         witness_lock_list_free(lle);
1737 }
1738
1739 /*
1740  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1741  * exempt Giant and sleepable locks from the checks as well.  If any
1742  * non-exempt locks are held, then a supplied message is printed to the
1743  * output channel along with a list of the offending locks.  If indicated in the
1744  * flags then a failure results in a panic as well.
1745  */
1746 int
1747 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1748 {
1749         struct lock_list_entry *lock_list, *lle;
1750         struct lock_instance *lock1;
1751         struct thread *td;
1752         va_list ap;
1753         int i, n;
1754
1755         if (witness_cold || witness_watch < 1 || panicstr != NULL)
1756                 return (0);
1757         n = 0;
1758         td = curthread;
1759         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1760                 for (i = lle->ll_count - 1; i >= 0; i--) {
1761                         lock1 = &lle->ll_children[i];
1762                         if (lock1->li_lock == lock)
1763                                 continue;
1764                         if (flags & WARN_GIANTOK &&
1765                             lock1->li_lock == &Giant.lock_object)
1766                                 continue;
1767                         if (flags & WARN_SLEEPOK &&
1768                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1769                                 continue;
1770                         if (n == 0) {
1771                                 va_start(ap, fmt);
1772                                 vprintf(fmt, ap);
1773                                 va_end(ap);
1774                                 printf(" with the following %slocks held:\n",
1775                                     (flags & WARN_SLEEPOK) != 0 ?
1776                                     "non-sleepable " : "");
1777                         }
1778                         n++;
1779                         witness_list_lock(lock1, printf);
1780                 }
1781
1782         /*
1783          * Pin the thread in order to avoid problems with thread migration.
1784          * Once that all verifies are passed about spinlocks ownership,
1785          * the thread is in a safe path and it can be unpinned.
1786          */
1787         sched_pin();
1788         lock_list = PCPU_GET(spinlocks);
1789         if (lock_list != NULL && lock_list->ll_count != 0) {
1790                 sched_unpin();
1791
1792                 /*
1793                  * We should only have one spinlock and as long as
1794                  * the flags cannot match for this locks class,
1795                  * check if the first spinlock is the one curthread
1796                  * should hold.
1797                  */
1798                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1799                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1800                     lock1->li_lock == lock && n == 0)
1801                         return (0);
1802
1803                 va_start(ap, fmt);
1804                 vprintf(fmt, ap);
1805                 va_end(ap);
1806                 printf(" with the following %slocks held:\n",
1807                     (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1808                 n += witness_list_locks(&lock_list, printf);
1809         } else
1810                 sched_unpin();
1811         if (flags & WARN_PANIC && n)
1812                 kassert_panic("%s", __func__);
1813         else
1814                 witness_debugger(n, __func__);
1815         return (n);
1816 }
1817
1818 const char *
1819 witness_file(struct lock_object *lock)
1820 {
1821         struct witness *w;
1822
1823         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1824                 return ("?");
1825         w = lock->lo_witness;
1826         return (w->w_file);
1827 }
1828
1829 int
1830 witness_line(struct lock_object *lock)
1831 {
1832         struct witness *w;
1833
1834         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1835                 return (0);
1836         w = lock->lo_witness;
1837         return (w->w_line);
1838 }
1839
1840 static struct witness *
1841 enroll(const char *description, struct lock_class *lock_class)
1842 {
1843         struct witness *w;
1844
1845         MPASS(description != NULL);
1846
1847         if (witness_watch == -1 || panicstr != NULL)
1848                 return (NULL);
1849         if ((lock_class->lc_flags & LC_SPINLOCK)) {
1850                 if (witness_skipspin)
1851                         return (NULL);
1852         } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1853                 kassert_panic("lock class %s is not sleep or spin",
1854                     lock_class->lc_name);
1855                 return (NULL);
1856         }
1857
1858         mtx_lock_spin(&w_mtx);
1859         w = witness_hash_get(description);
1860         if (w)
1861                 goto found;
1862         if ((w = witness_get()) == NULL)
1863                 return (NULL);
1864         MPASS(strlen(description) < MAX_W_NAME);
1865         strcpy(w->w_name, description);
1866         w->w_class = lock_class;
1867         w->w_refcount = 1;
1868         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1869         if (lock_class->lc_flags & LC_SPINLOCK) {
1870                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1871                 w_spin_cnt++;
1872         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1873                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1874                 w_sleep_cnt++;
1875         }
1876
1877         /* Insert new witness into the hash */
1878         witness_hash_put(w);
1879         witness_increment_graph_generation();
1880         mtx_unlock_spin(&w_mtx);
1881         return (w);
1882 found:
1883         w->w_refcount++;
1884         if (w->w_refcount == 1)
1885                 w->w_class = lock_class;
1886         mtx_unlock_spin(&w_mtx);
1887         if (lock_class != w->w_class)
1888                 kassert_panic(
1889                     "lock (%s) %s does not match earlier (%s) lock",
1890                     description, lock_class->lc_name,
1891                     w->w_class->lc_name);
1892         return (w);
1893 }
1894
1895 static void
1896 depart(struct witness *w)
1897 {
1898
1899         MPASS(w->w_refcount == 0);
1900         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1901                 w_sleep_cnt--;
1902         } else {
1903                 w_spin_cnt--;
1904         }
1905         /*
1906          * Set file to NULL as it may point into a loadable module.
1907          */
1908         w->w_file = NULL;
1909         w->w_line = 0;
1910         witness_increment_graph_generation();
1911 }
1912
1913
1914 static void
1915 adopt(struct witness *parent, struct witness *child)
1916 {
1917         int pi, ci, i, j;
1918
1919         if (witness_cold == 0)
1920                 mtx_assert(&w_mtx, MA_OWNED);
1921
1922         /* If the relationship is already known, there's no work to be done. */
1923         if (isitmychild(parent, child))
1924                 return;
1925
1926         /* When the structure of the graph changes, bump up the generation. */
1927         witness_increment_graph_generation();
1928
1929         /*
1930          * The hard part ... create the direct relationship, then propagate all
1931          * indirect relationships.
1932          */
1933         pi = parent->w_index;
1934         ci = child->w_index;
1935         WITNESS_INDEX_ASSERT(pi);
1936         WITNESS_INDEX_ASSERT(ci);
1937         MPASS(pi != ci);
1938         w_rmatrix[pi][ci] |= WITNESS_PARENT;
1939         w_rmatrix[ci][pi] |= WITNESS_CHILD;
1940
1941         /*
1942          * If parent was not already an ancestor of child,
1943          * then we increment the descendant and ancestor counters.
1944          */
1945         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1946                 parent->w_num_descendants++;
1947                 child->w_num_ancestors++;
1948         }
1949
1950         /*
1951          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1952          * an ancestor of 'pi' during this loop.
1953          */
1954         for (i = 1; i <= w_max_used_index; i++) {
1955                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1956                     (i != pi))
1957                         continue;
1958
1959                 /* Find each descendant of 'i' and mark it as a descendant. */
1960                 for (j = 1; j <= w_max_used_index; j++) {
1961
1962                         /*
1963                          * Skip children that are already marked as
1964                          * descendants of 'i'.
1965                          */
1966                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1967                                 continue;
1968
1969                         /*
1970                          * We are only interested in descendants of 'ci'. Note
1971                          * that 'ci' itself is counted as a descendant of 'ci'.
1972                          */
1973                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1974                             (j != ci))
1975                                 continue;
1976                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1977                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1978                         w_data[i].w_num_descendants++;
1979                         w_data[j].w_num_ancestors++;
1980
1981                         /*
1982                          * Make sure we aren't marking a node as both an
1983                          * ancestor and descendant. We should have caught
1984                          * this as a lock order reversal earlier.
1985                          */
1986                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1987                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1988                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1989                                     "both ancestor and descendant\n",
1990                                     i, j, w_rmatrix[i][j]);
1991                                 kdb_backtrace();
1992                                 printf("Witness disabled.\n");
1993                                 witness_watch = -1;
1994                         }
1995                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1996                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1997                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1998                                     "both ancestor and descendant\n",
1999                                     j, i, w_rmatrix[j][i]);
2000                                 kdb_backtrace();
2001                                 printf("Witness disabled.\n");
2002                                 witness_watch = -1;
2003                         }
2004                 }
2005         }
2006 }
2007
2008 static void
2009 itismychild(struct witness *parent, struct witness *child)
2010 {
2011         int unlocked;
2012
2013         MPASS(child != NULL && parent != NULL);
2014         if (witness_cold == 0)
2015                 mtx_assert(&w_mtx, MA_OWNED);
2016
2017         if (!witness_lock_type_equal(parent, child)) {
2018                 if (witness_cold == 0) {
2019                         unlocked = 1;
2020                         mtx_unlock_spin(&w_mtx);
2021                 } else {
2022                         unlocked = 0;
2023                 }
2024                 kassert_panic(
2025                     "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2026                     "the same lock type", __func__, parent->w_name,
2027                     parent->w_class->lc_name, child->w_name,
2028                     child->w_class->lc_name);
2029                 if (unlocked)
2030                         mtx_lock_spin(&w_mtx);
2031         }
2032         adopt(parent, child);
2033 }
2034
2035 /*
2036  * Generic code for the isitmy*() functions. The rmask parameter is the
2037  * expected relationship of w1 to w2.
2038  */
2039 static int
2040 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2041 {
2042         unsigned char r1, r2;
2043         int i1, i2;
2044
2045         i1 = w1->w_index;
2046         i2 = w2->w_index;
2047         WITNESS_INDEX_ASSERT(i1);
2048         WITNESS_INDEX_ASSERT(i2);
2049         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2050         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2051
2052         /* The flags on one better be the inverse of the flags on the other */
2053         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2054             (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2055                 /* Don't squawk if we're potentially racing with an update. */
2056                 if (!mtx_owned(&w_mtx))
2057                         return (0);
2058                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2059                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
2060                     "w_rmatrix[%d][%d] == %hhx\n",
2061                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2062                     i2, i1, r2);
2063                 kdb_backtrace();
2064                 printf("Witness disabled.\n");
2065                 witness_watch = -1;
2066         }
2067         return (r1 & rmask);
2068 }
2069
2070 /*
2071  * Checks if @child is a direct child of @parent.
2072  */
2073 static int
2074 isitmychild(struct witness *parent, struct witness *child)
2075 {
2076
2077         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2078 }
2079
2080 /*
2081  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2082  */
2083 static int
2084 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2085 {
2086
2087         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2088             __func__));
2089 }
2090
2091 static int
2092 blessed(struct witness *w1, struct witness *w2)
2093 {
2094         int i;
2095         struct witness_blessed *b;
2096
2097         for (i = 0; i < nitems(blessed_list); i++) {
2098                 b = &blessed_list[i];
2099                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2100                         if (strcmp(w2->w_name, b->b_lock2) == 0)
2101                                 return (1);
2102                         continue;
2103                 }
2104                 if (strcmp(w1->w_name, b->b_lock2) == 0)
2105                         if (strcmp(w2->w_name, b->b_lock1) == 0)
2106                                 return (1);
2107         }
2108         return (0);
2109 }
2110
2111 static struct witness *
2112 witness_get(void)
2113 {
2114         struct witness *w;
2115         int index;
2116
2117         if (witness_cold == 0)
2118                 mtx_assert(&w_mtx, MA_OWNED);
2119
2120         if (witness_watch == -1) {
2121                 mtx_unlock_spin(&w_mtx);
2122                 return (NULL);
2123         }
2124         if (STAILQ_EMPTY(&w_free)) {
2125                 witness_watch = -1;
2126                 mtx_unlock_spin(&w_mtx);
2127                 printf("WITNESS: unable to allocate a new witness object\n");
2128                 return (NULL);
2129         }
2130         w = STAILQ_FIRST(&w_free);
2131         STAILQ_REMOVE_HEAD(&w_free, w_list);
2132         w_free_cnt--;
2133         index = w->w_index;
2134         MPASS(index > 0 && index == w_max_used_index+1 &&
2135             index < witness_count);
2136         bzero(w, sizeof(*w));
2137         w->w_index = index;
2138         if (index > w_max_used_index)
2139                 w_max_used_index = index;
2140         return (w);
2141 }
2142
2143 static void
2144 witness_free(struct witness *w)
2145 {
2146
2147         STAILQ_INSERT_HEAD(&w_free, w, w_list);
2148         w_free_cnt++;
2149 }
2150
2151 static struct lock_list_entry *
2152 witness_lock_list_get(void)
2153 {
2154         struct lock_list_entry *lle;
2155
2156         if (witness_watch == -1)
2157                 return (NULL);
2158         mtx_lock_spin(&w_mtx);
2159         lle = w_lock_list_free;
2160         if (lle == NULL) {
2161                 witness_watch = -1;
2162                 mtx_unlock_spin(&w_mtx);
2163                 printf("%s: witness exhausted\n", __func__);
2164                 return (NULL);
2165         }
2166         w_lock_list_free = lle->ll_next;
2167         mtx_unlock_spin(&w_mtx);
2168         bzero(lle, sizeof(*lle));
2169         return (lle);
2170 }
2171
2172 static void
2173 witness_lock_list_free(struct lock_list_entry *lle)
2174 {
2175
2176         mtx_lock_spin(&w_mtx);
2177         lle->ll_next = w_lock_list_free;
2178         w_lock_list_free = lle;
2179         mtx_unlock_spin(&w_mtx);
2180 }
2181
2182 static struct lock_instance *
2183 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2184 {
2185         struct lock_list_entry *lle;
2186         struct lock_instance *instance;
2187         int i;
2188
2189         for (lle = list; lle != NULL; lle = lle->ll_next)
2190                 for (i = lle->ll_count - 1; i >= 0; i--) {
2191                         instance = &lle->ll_children[i];
2192                         if (instance->li_lock == lock)
2193                                 return (instance);
2194                 }
2195         return (NULL);
2196 }
2197
2198 static void
2199 witness_list_lock(struct lock_instance *instance,
2200     int (*prnt)(const char *fmt, ...))
2201 {
2202         struct lock_object *lock;
2203
2204         lock = instance->li_lock;
2205         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2206             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2207         if (lock->lo_witness->w_name != lock->lo_name)
2208                 prnt(" (%s)", lock->lo_witness->w_name);
2209         prnt(" r = %d (%p) locked @ %s:%d\n",
2210             instance->li_flags & LI_RECURSEMASK, lock,
2211             fixup_filename(instance->li_file), instance->li_line);
2212 }
2213
2214 static int
2215 witness_output(const char *fmt, ...)
2216 {
2217         va_list ap;
2218         int ret;
2219
2220         va_start(ap, fmt);
2221         ret = witness_voutput(fmt, ap);
2222         va_end(ap);
2223         return (ret);
2224 }
2225
2226 static int
2227 witness_voutput(const char *fmt, va_list ap)
2228 {
2229         int ret;
2230
2231         ret = 0;
2232         switch (witness_channel) {
2233         case WITNESS_CONSOLE:
2234                 ret = vprintf(fmt, ap);
2235                 break;
2236         case WITNESS_LOG:
2237                 vlog(LOG_NOTICE, fmt, ap);
2238                 break;
2239         case WITNESS_NONE:
2240                 break;
2241         }
2242         return (ret);
2243 }
2244
2245 #ifdef DDB
2246 static int
2247 witness_thread_has_locks(struct thread *td)
2248 {
2249
2250         if (td->td_sleeplocks == NULL)
2251                 return (0);
2252         return (td->td_sleeplocks->ll_count != 0);
2253 }
2254
2255 static int
2256 witness_proc_has_locks(struct proc *p)
2257 {
2258         struct thread *td;
2259
2260         FOREACH_THREAD_IN_PROC(p, td) {
2261                 if (witness_thread_has_locks(td))
2262                         return (1);
2263         }
2264         return (0);
2265 }
2266 #endif
2267
2268 int
2269 witness_list_locks(struct lock_list_entry **lock_list,
2270     int (*prnt)(const char *fmt, ...))
2271 {
2272         struct lock_list_entry *lle;
2273         int i, nheld;
2274
2275         nheld = 0;
2276         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2277                 for (i = lle->ll_count - 1; i >= 0; i--) {
2278                         witness_list_lock(&lle->ll_children[i], prnt);
2279                         nheld++;
2280                 }
2281         return (nheld);
2282 }
2283
2284 /*
2285  * This is a bit risky at best.  We call this function when we have timed
2286  * out acquiring a spin lock, and we assume that the other CPU is stuck
2287  * with this lock held.  So, we go groveling around in the other CPU's
2288  * per-cpu data to try to find the lock instance for this spin lock to
2289  * see when it was last acquired.
2290  */
2291 void
2292 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2293     int (*prnt)(const char *fmt, ...))
2294 {
2295         struct lock_instance *instance;
2296         struct pcpu *pc;
2297
2298         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2299                 return;
2300         pc = pcpu_find(owner->td_oncpu);
2301         instance = find_instance(pc->pc_spinlocks, lock);
2302         if (instance != NULL)
2303                 witness_list_lock(instance, prnt);
2304 }
2305
2306 void
2307 witness_save(struct lock_object *lock, const char **filep, int *linep)
2308 {
2309         struct lock_list_entry *lock_list;
2310         struct lock_instance *instance;
2311         struct lock_class *class;
2312
2313         /*
2314          * This function is used independently in locking code to deal with
2315          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2316          * is gone.
2317          */
2318         if (SCHEDULER_STOPPED())
2319                 return;
2320         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2321         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2322                 return;
2323         class = LOCK_CLASS(lock);
2324         if (class->lc_flags & LC_SLEEPLOCK)
2325                 lock_list = curthread->td_sleeplocks;
2326         else {
2327                 if (witness_skipspin)
2328                         return;
2329                 lock_list = PCPU_GET(spinlocks);
2330         }
2331         instance = find_instance(lock_list, lock);
2332         if (instance == NULL) {
2333                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2334                     class->lc_name, lock->lo_name);
2335                 return;
2336         }
2337         *filep = instance->li_file;
2338         *linep = instance->li_line;
2339 }
2340
2341 void
2342 witness_restore(struct lock_object *lock, const char *file, int line)
2343 {
2344         struct lock_list_entry *lock_list;
2345         struct lock_instance *instance;
2346         struct lock_class *class;
2347
2348         /*
2349          * This function is used independently in locking code to deal with
2350          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2351          * is gone.
2352          */
2353         if (SCHEDULER_STOPPED())
2354                 return;
2355         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2356         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2357                 return;
2358         class = LOCK_CLASS(lock);
2359         if (class->lc_flags & LC_SLEEPLOCK)
2360                 lock_list = curthread->td_sleeplocks;
2361         else {
2362                 if (witness_skipspin)
2363                         return;
2364                 lock_list = PCPU_GET(spinlocks);
2365         }
2366         instance = find_instance(lock_list, lock);
2367         if (instance == NULL)
2368                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2369                     class->lc_name, lock->lo_name);
2370         lock->lo_witness->w_file = file;
2371         lock->lo_witness->w_line = line;
2372         if (instance == NULL)
2373                 return;
2374         instance->li_file = file;
2375         instance->li_line = line;
2376 }
2377
2378 void
2379 witness_assert(const struct lock_object *lock, int flags, const char *file,
2380     int line)
2381 {
2382 #ifdef INVARIANT_SUPPORT
2383         struct lock_instance *instance;
2384         struct lock_class *class;
2385
2386         if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2387                 return;
2388         class = LOCK_CLASS(lock);
2389         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2390                 instance = find_instance(curthread->td_sleeplocks, lock);
2391         else if ((class->lc_flags & LC_SPINLOCK) != 0)
2392                 instance = find_instance(PCPU_GET(spinlocks), lock);
2393         else {
2394                 kassert_panic("Lock (%s) %s is not sleep or spin!",
2395                     class->lc_name, lock->lo_name);
2396                 return;
2397         }
2398         switch (flags) {
2399         case LA_UNLOCKED:
2400                 if (instance != NULL)
2401                         kassert_panic("Lock (%s) %s locked @ %s:%d.",
2402                             class->lc_name, lock->lo_name,
2403                             fixup_filename(file), line);
2404                 break;
2405         case LA_LOCKED:
2406         case LA_LOCKED | LA_RECURSED:
2407         case LA_LOCKED | LA_NOTRECURSED:
2408         case LA_SLOCKED:
2409         case LA_SLOCKED | LA_RECURSED:
2410         case LA_SLOCKED | LA_NOTRECURSED:
2411         case LA_XLOCKED:
2412         case LA_XLOCKED | LA_RECURSED:
2413         case LA_XLOCKED | LA_NOTRECURSED:
2414                 if (instance == NULL) {
2415                         kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2416                             class->lc_name, lock->lo_name,
2417                             fixup_filename(file), line);
2418                         break;
2419                 }
2420                 if ((flags & LA_XLOCKED) != 0 &&
2421                     (instance->li_flags & LI_EXCLUSIVE) == 0)
2422                         kassert_panic(
2423                             "Lock (%s) %s not exclusively locked @ %s:%d.",
2424                             class->lc_name, lock->lo_name,
2425                             fixup_filename(file), line);
2426                 if ((flags & LA_SLOCKED) != 0 &&
2427                     (instance->li_flags & LI_EXCLUSIVE) != 0)
2428                         kassert_panic(
2429                             "Lock (%s) %s exclusively locked @ %s:%d.",
2430                             class->lc_name, lock->lo_name,
2431                             fixup_filename(file), line);
2432                 if ((flags & LA_RECURSED) != 0 &&
2433                     (instance->li_flags & LI_RECURSEMASK) == 0)
2434                         kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2435                             class->lc_name, lock->lo_name,
2436                             fixup_filename(file), line);
2437                 if ((flags & LA_NOTRECURSED) != 0 &&
2438                     (instance->li_flags & LI_RECURSEMASK) != 0)
2439                         kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2440                             class->lc_name, lock->lo_name,
2441                             fixup_filename(file), line);
2442                 break;
2443         default:
2444                 kassert_panic("Invalid lock assertion at %s:%d.",
2445                     fixup_filename(file), line);
2446
2447         }
2448 #endif  /* INVARIANT_SUPPORT */
2449 }
2450
2451 static void
2452 witness_setflag(struct lock_object *lock, int flag, int set)
2453 {
2454         struct lock_list_entry *lock_list;
2455         struct lock_instance *instance;
2456         struct lock_class *class;
2457
2458         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2459                 return;
2460         class = LOCK_CLASS(lock);
2461         if (class->lc_flags & LC_SLEEPLOCK)
2462                 lock_list = curthread->td_sleeplocks;
2463         else {
2464                 if (witness_skipspin)
2465                         return;
2466                 lock_list = PCPU_GET(spinlocks);
2467         }
2468         instance = find_instance(lock_list, lock);
2469         if (instance == NULL) {
2470                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2471                     class->lc_name, lock->lo_name);
2472                 return;
2473         }
2474
2475         if (set)
2476                 instance->li_flags |= flag;
2477         else
2478                 instance->li_flags &= ~flag;
2479 }
2480
2481 void
2482 witness_norelease(struct lock_object *lock)
2483 {
2484
2485         witness_setflag(lock, LI_NORELEASE, 1);
2486 }
2487
2488 void
2489 witness_releaseok(struct lock_object *lock)
2490 {
2491
2492         witness_setflag(lock, LI_NORELEASE, 0);
2493 }
2494
2495 #ifdef DDB
2496 static void
2497 witness_ddb_list(struct thread *td)
2498 {
2499
2500         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2501         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2502
2503         if (witness_watch < 1)
2504                 return;
2505
2506         witness_list_locks(&td->td_sleeplocks, db_printf);
2507
2508         /*
2509          * We only handle spinlocks if td == curthread.  This is somewhat broken
2510          * if td is currently executing on some other CPU and holds spin locks
2511          * as we won't display those locks.  If we had a MI way of getting
2512          * the per-cpu data for a given cpu then we could use
2513          * td->td_oncpu to get the list of spinlocks for this thread
2514          * and "fix" this.
2515          *
2516          * That still wouldn't really fix this unless we locked the scheduler
2517          * lock or stopped the other CPU to make sure it wasn't changing the
2518          * list out from under us.  It is probably best to just not try to
2519          * handle threads on other CPU's for now.
2520          */
2521         if (td == curthread && PCPU_GET(spinlocks) != NULL)
2522                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2523 }
2524
2525 DB_SHOW_COMMAND(locks, db_witness_list)
2526 {
2527         struct thread *td;
2528
2529         if (have_addr)
2530                 td = db_lookup_thread(addr, true);
2531         else
2532                 td = kdb_thread;
2533         witness_ddb_list(td);
2534 }
2535
2536 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2537 {
2538         struct thread *td;
2539         struct proc *p;
2540
2541         /*
2542          * It would be nice to list only threads and processes that actually
2543          * held sleep locks, but that information is currently not exported
2544          * by WITNESS.
2545          */
2546         FOREACH_PROC_IN_SYSTEM(p) {
2547                 if (!witness_proc_has_locks(p))
2548                         continue;
2549                 FOREACH_THREAD_IN_PROC(p, td) {
2550                         if (!witness_thread_has_locks(td))
2551                                 continue;
2552                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2553                             p->p_comm, td, td->td_tid);
2554                         witness_ddb_list(td);
2555                         if (db_pager_quit)
2556                                 return;
2557                 }
2558         }
2559 }
2560 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2561
2562 DB_SHOW_COMMAND(witness, db_witness_display)
2563 {
2564
2565         witness_ddb_display(db_printf);
2566 }
2567 #endif
2568
2569 static void
2570 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2571 {
2572         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2573         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2574         int generation, i, j;
2575
2576         tmp_data1 = NULL;
2577         tmp_data2 = NULL;
2578         tmp_w1 = NULL;
2579         tmp_w2 = NULL;
2580
2581         /* Allocate and init temporary storage space. */
2582         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2583         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2584         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2585             M_WAITOK | M_ZERO);
2586         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2587             M_WAITOK | M_ZERO);
2588         stack_zero(&tmp_data1->wlod_stack);
2589         stack_zero(&tmp_data2->wlod_stack);
2590
2591 restart:
2592         mtx_lock_spin(&w_mtx);
2593         generation = w_generation;
2594         mtx_unlock_spin(&w_mtx);
2595         sbuf_printf(sb, "Number of known direct relationships is %d\n",
2596             w_lohash.wloh_count);
2597         for (i = 1; i < w_max_used_index; i++) {
2598                 mtx_lock_spin(&w_mtx);
2599                 if (generation != w_generation) {
2600                         mtx_unlock_spin(&w_mtx);
2601
2602                         /* The graph has changed, try again. */
2603                         *oldidx = 0;
2604                         sbuf_clear(sb);
2605                         goto restart;
2606                 }
2607
2608                 w1 = &w_data[i];
2609                 if (w1->w_reversed == 0) {
2610                         mtx_unlock_spin(&w_mtx);
2611                         continue;
2612                 }
2613
2614                 /* Copy w1 locally so we can release the spin lock. */
2615                 *tmp_w1 = *w1;
2616                 mtx_unlock_spin(&w_mtx);
2617
2618                 if (tmp_w1->w_reversed == 0)
2619                         continue;
2620                 for (j = 1; j < w_max_used_index; j++) {
2621                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2622                                 continue;
2623
2624                         mtx_lock_spin(&w_mtx);
2625                         if (generation != w_generation) {
2626                                 mtx_unlock_spin(&w_mtx);
2627
2628                                 /* The graph has changed, try again. */
2629                                 *oldidx = 0;
2630                                 sbuf_clear(sb);
2631                                 goto restart;
2632                         }
2633
2634                         w2 = &w_data[j];
2635                         data1 = witness_lock_order_get(w1, w2);
2636                         data2 = witness_lock_order_get(w2, w1);
2637
2638                         /*
2639                          * Copy information locally so we can release the
2640                          * spin lock.
2641                          */
2642                         *tmp_w2 = *w2;
2643
2644                         if (data1) {
2645                                 stack_zero(&tmp_data1->wlod_stack);
2646                                 stack_copy(&data1->wlod_stack,
2647                                     &tmp_data1->wlod_stack);
2648                         }
2649                         if (data2 && data2 != data1) {
2650                                 stack_zero(&tmp_data2->wlod_stack);
2651                                 stack_copy(&data2->wlod_stack,
2652                                     &tmp_data2->wlod_stack);
2653                         }
2654                         mtx_unlock_spin(&w_mtx);
2655
2656                         if (blessed(tmp_w1, tmp_w2))
2657                                 continue;
2658
2659                         sbuf_printf(sb,
2660             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2661                             tmp_w1->w_name, tmp_w1->w_class->lc_name,
2662                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
2663                         if (data1) {
2664                                 sbuf_printf(sb,
2665                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2666                                     tmp_w1->w_name, tmp_w1->w_class->lc_name,
2667                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
2668                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2669                                 sbuf_printf(sb, "\n");
2670                         }
2671                         if (data2 && data2 != data1) {
2672                                 sbuf_printf(sb,
2673                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2674                                     tmp_w2->w_name, tmp_w2->w_class->lc_name,
2675                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
2676                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2677                                 sbuf_printf(sb, "\n");
2678                         }
2679                 }
2680         }
2681         mtx_lock_spin(&w_mtx);
2682         if (generation != w_generation) {
2683                 mtx_unlock_spin(&w_mtx);
2684
2685                 /*
2686                  * The graph changed while we were printing stack data,
2687                  * try again.
2688                  */
2689                 *oldidx = 0;
2690                 sbuf_clear(sb);
2691                 goto restart;
2692         }
2693         mtx_unlock_spin(&w_mtx);
2694
2695         /* Free temporary storage space. */
2696         free(tmp_data1, M_TEMP);
2697         free(tmp_data2, M_TEMP);
2698         free(tmp_w1, M_TEMP);
2699         free(tmp_w2, M_TEMP);
2700 }
2701
2702 static int
2703 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2704 {
2705         struct sbuf *sb;
2706         int error;
2707
2708         if (witness_watch < 1) {
2709                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2710                 return (error);
2711         }
2712         if (witness_cold) {
2713                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2714                 return (error);
2715         }
2716         error = 0;
2717         sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2718         if (sb == NULL)
2719                 return (ENOMEM);
2720
2721         sbuf_print_witness_badstacks(sb, &req->oldidx);
2722
2723         sbuf_finish(sb);
2724         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2725         sbuf_delete(sb);
2726
2727         return (error);
2728 }
2729
2730 #ifdef DDB
2731 static int
2732 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2733 {
2734
2735         return (db_printf("%.*s", len, data));
2736 }
2737
2738 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2739 {
2740         struct sbuf sb;
2741         char buffer[128];
2742         size_t dummy;
2743
2744         sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2745         sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2746         sbuf_print_witness_badstacks(&sb, &dummy);
2747         sbuf_finish(&sb);
2748 }
2749 #endif
2750
2751 static int
2752 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2753 {
2754         static const struct {
2755                 enum witness_channel channel;
2756                 const char *name;
2757         } channels[] = {
2758                 { WITNESS_CONSOLE, "console" },
2759                 { WITNESS_LOG, "log" },
2760                 { WITNESS_NONE, "none" },
2761         };
2762         char buf[16];
2763         u_int i;
2764         int error;
2765
2766         buf[0] = '\0';
2767         for (i = 0; i < nitems(channels); i++)
2768                 if (witness_channel == channels[i].channel) {
2769                         snprintf(buf, sizeof(buf), "%s", channels[i].name);
2770                         break;
2771                 }
2772
2773         error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2774         if (error != 0 || req->newptr == NULL)
2775                 return (error);
2776
2777         error = EINVAL;
2778         for (i = 0; i < nitems(channels); i++)
2779                 if (strcmp(channels[i].name, buf) == 0) {
2780                         witness_channel = channels[i].channel;
2781                         error = 0;
2782                         break;
2783                 }
2784         return (error);
2785 }
2786
2787 static int
2788 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2789 {
2790         struct witness *w;
2791         struct sbuf *sb;
2792         int error;
2793
2794 #ifdef __i386__
2795         error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2796         return (error);
2797 #endif
2798
2799         if (witness_watch < 1) {
2800                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2801                 return (error);
2802         }
2803         if (witness_cold) {
2804                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2805                 return (error);
2806         }
2807         error = 0;
2808
2809         error = sysctl_wire_old_buffer(req, 0);
2810         if (error != 0)
2811                 return (error);
2812         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2813         if (sb == NULL)
2814                 return (ENOMEM);
2815         sbuf_printf(sb, "\n");
2816
2817         mtx_lock_spin(&w_mtx);
2818         STAILQ_FOREACH(w, &w_all, w_list)
2819                 w->w_displayed = 0;
2820         STAILQ_FOREACH(w, &w_all, w_list)
2821                 witness_add_fullgraph(sb, w);
2822         mtx_unlock_spin(&w_mtx);
2823
2824         /*
2825          * Close the sbuf and return to userland.
2826          */
2827         error = sbuf_finish(sb);
2828         sbuf_delete(sb);
2829
2830         return (error);
2831 }
2832
2833 static int
2834 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2835 {
2836         int error, value;
2837
2838         value = witness_watch;
2839         error = sysctl_handle_int(oidp, &value, 0, req);
2840         if (error != 0 || req->newptr == NULL)
2841                 return (error);
2842         if (value > 1 || value < -1 ||
2843             (witness_watch == -1 && value != witness_watch))
2844                 return (EINVAL);
2845         witness_watch = value;
2846         return (0);
2847 }
2848
2849 static void
2850 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2851 {
2852         int i;
2853
2854         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2855                 return;
2856         w->w_displayed = 1;
2857
2858         WITNESS_INDEX_ASSERT(w->w_index);
2859         for (i = 1; i <= w_max_used_index; i++) {
2860                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2861                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2862                             w_data[i].w_name);
2863                         witness_add_fullgraph(sb, &w_data[i]);
2864                 }
2865         }
2866 }
2867
2868 /*
2869  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2870  * interprets the key as a string and reads until the null
2871  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2872  * hash value computed from the key.
2873  */
2874 static uint32_t
2875 witness_hash_djb2(const uint8_t *key, uint32_t size)
2876 {
2877         unsigned int hash = 5381;
2878         int i;
2879
2880         /* hash = hash * 33 + key[i] */
2881         if (size)
2882                 for (i = 0; i < size; i++)
2883                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2884         else
2885                 for (i = 0; key[i] != 0; i++)
2886                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2887
2888         return (hash);
2889 }
2890
2891
2892 /*
2893  * Initializes the two witness hash tables. Called exactly once from
2894  * witness_initialize().
2895  */
2896 static void
2897 witness_init_hash_tables(void)
2898 {
2899         int i;
2900
2901         MPASS(witness_cold);
2902
2903         /* Initialize the hash tables. */
2904         for (i = 0; i < WITNESS_HASH_SIZE; i++)
2905                 w_hash.wh_array[i] = NULL;
2906
2907         w_hash.wh_size = WITNESS_HASH_SIZE;
2908         w_hash.wh_count = 0;
2909
2910         /* Initialize the lock order data hash. */
2911         w_lofree = NULL;
2912         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2913                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2914                 w_lodata[i].wlod_next = w_lofree;
2915                 w_lofree = &w_lodata[i];
2916         }
2917         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2918         w_lohash.wloh_count = 0;
2919         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2920                 w_lohash.wloh_array[i] = NULL;
2921 }
2922
2923 static struct witness *
2924 witness_hash_get(const char *key)
2925 {
2926         struct witness *w;
2927         uint32_t hash;
2928
2929         MPASS(key != NULL);
2930         if (witness_cold == 0)
2931                 mtx_assert(&w_mtx, MA_OWNED);
2932         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2933         w = w_hash.wh_array[hash];
2934         while (w != NULL) {
2935                 if (strcmp(w->w_name, key) == 0)
2936                         goto out;
2937                 w = w->w_hash_next;
2938         }
2939
2940 out:
2941         return (w);
2942 }
2943
2944 static void
2945 witness_hash_put(struct witness *w)
2946 {
2947         uint32_t hash;
2948
2949         MPASS(w != NULL);
2950         MPASS(w->w_name != NULL);
2951         if (witness_cold == 0)
2952                 mtx_assert(&w_mtx, MA_OWNED);
2953         KASSERT(witness_hash_get(w->w_name) == NULL,
2954             ("%s: trying to add a hash entry that already exists!", __func__));
2955         KASSERT(w->w_hash_next == NULL,
2956             ("%s: w->w_hash_next != NULL", __func__));
2957
2958         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2959         w->w_hash_next = w_hash.wh_array[hash];
2960         w_hash.wh_array[hash] = w;
2961         w_hash.wh_count++;
2962 }
2963
2964
2965 static struct witness_lock_order_data *
2966 witness_lock_order_get(struct witness *parent, struct witness *child)
2967 {
2968         struct witness_lock_order_data *data = NULL;
2969         struct witness_lock_order_key key;
2970         unsigned int hash;
2971
2972         MPASS(parent != NULL && child != NULL);
2973         key.from = parent->w_index;
2974         key.to = child->w_index;
2975         WITNESS_INDEX_ASSERT(key.from);
2976         WITNESS_INDEX_ASSERT(key.to);
2977         if ((w_rmatrix[parent->w_index][child->w_index]
2978             & WITNESS_LOCK_ORDER_KNOWN) == 0)
2979                 goto out;
2980
2981         hash = witness_hash_djb2((const char*)&key,
2982             sizeof(key)) % w_lohash.wloh_size;
2983         data = w_lohash.wloh_array[hash];
2984         while (data != NULL) {
2985                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2986                         break;
2987                 data = data->wlod_next;
2988         }
2989
2990 out:
2991         return (data);
2992 }
2993
2994 /*
2995  * Verify that parent and child have a known relationship, are not the same,
2996  * and child is actually a child of parent.  This is done without w_mtx
2997  * to avoid contention in the common case.
2998  */
2999 static int
3000 witness_lock_order_check(struct witness *parent, struct witness *child)
3001 {
3002
3003         if (parent != child &&
3004             w_rmatrix[parent->w_index][child->w_index]
3005             & WITNESS_LOCK_ORDER_KNOWN &&
3006             isitmychild(parent, child))
3007                 return (1);
3008
3009         return (0);
3010 }
3011
3012 static int
3013 witness_lock_order_add(struct witness *parent, struct witness *child)
3014 {
3015         struct witness_lock_order_data *data = NULL;
3016         struct witness_lock_order_key key;
3017         unsigned int hash;
3018
3019         MPASS(parent != NULL && child != NULL);
3020         key.from = parent->w_index;
3021         key.to = child->w_index;
3022         WITNESS_INDEX_ASSERT(key.from);
3023         WITNESS_INDEX_ASSERT(key.to);
3024         if (w_rmatrix[parent->w_index][child->w_index]
3025             & WITNESS_LOCK_ORDER_KNOWN)
3026                 return (1);
3027
3028         hash = witness_hash_djb2((const char*)&key,
3029             sizeof(key)) % w_lohash.wloh_size;
3030         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3031         data = w_lofree;
3032         if (data == NULL)
3033                 return (0);
3034         w_lofree = data->wlod_next;
3035         data->wlod_next = w_lohash.wloh_array[hash];
3036         data->wlod_key = key;
3037         w_lohash.wloh_array[hash] = data;
3038         w_lohash.wloh_count++;
3039         stack_zero(&data->wlod_stack);
3040         stack_save(&data->wlod_stack);
3041         return (1);
3042 }
3043
3044 /* Call this whenever the structure of the witness graph changes. */
3045 static void
3046 witness_increment_graph_generation(void)
3047 {
3048
3049         if (witness_cold == 0)
3050                 mtx_assert(&w_mtx, MA_OWNED);
3051         w_generation++;
3052 }
3053
3054 static int
3055 witness_output_drain(void *arg __unused, const char *data, int len)
3056 {
3057
3058         witness_output("%.*s", len, data);
3059         return (len);
3060 }
3061
3062 static void
3063 witness_debugger(int cond, const char *msg)
3064 {
3065         char buf[32];
3066         struct sbuf sb;
3067         struct stack st;
3068
3069         if (!cond)
3070                 return;
3071
3072         if (witness_trace) {
3073                 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3074                 sbuf_set_drain(&sb, witness_output_drain, NULL);
3075
3076                 stack_zero(&st);
3077                 stack_save(&st);
3078                 witness_output("stack backtrace:\n");
3079                 stack_sbuf_print_ddb(&sb, &st);
3080
3081                 sbuf_finish(&sb);
3082         }
3083
3084 #ifdef KDB
3085         if (witness_kdb)
3086                 kdb_enter(KDB_WHY_WITNESS, msg);
3087 #endif
3088 }