sys/kern/subr_witness.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 2008 Isilon Systems, Inc.
   5  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
   6  * Copyright (c) 1998 Berkeley Software Design, Inc.
   7  * All rights reserved.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  * 3. Berkeley Software Design Inc's name may not be used to endorse or
  18  *    promote products derived from this software without specific prior
  19  *    written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
  22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  24  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
  25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  *      from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
  34  *      and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
  35  */
  36
  37 /*
  38  * Implementation of the `witness' lock verifier.  Originally implemented for
  39  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
  40  * classes in FreeBSD.
  41  */
  42
  43 /*
  44  *      Main Entry: witness
  45  *      Pronunciation: 'wit-n&s
  46  *      Function: noun
  47  *      Etymology: Middle English witnesse, from Old English witnes knowledge,
  48  *          testimony, witness, from 2wit
  49  *      Date: before 12th century
  50  *      1 : attestation of a fact or event : TESTIMONY
  51  *      2 : one that gives evidence; specifically : one who testifies in
  52  *          a cause or before a judicial tribunal
  53  *      3 : one asked to be present at a transaction so as to be able to
  54  *          testify to its having taken place
  55  *      4 : one who has personal knowledge of something
  56  *      5 a : something serving as evidence or proof : SIGN
  57  *        b : public affirmation by word or example of usually
  58  *            religious faith or conviction <the heroic witness to divine
  59  *            life -- Pilot>
  60  *      6 capitalized : a member of the Jehovah's Witnesses
  61  */
  62
  63 /*
  64  * Special rules concerning Giant and lock orders:
  65  *
  66  * 1) Giant must be acquired before any other mutexes.  Stated another way,
  67  *    no other mutex may be held when Giant is acquired.
  68  *
  69  * 2) Giant must be released when blocking on a sleepable lock.
  70  *
  71  * This rule is less obvious, but is a result of Giant providing the same
  72  * semantics as spl().  Basically, when a thread sleeps, it must release
  73  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
  74  * 2).
  75  *
  76  * 3) Giant may be acquired before or after sleepable locks.
  77  *
  78  * This rule is also not quite as obvious.  Giant may be acquired after
  79  * a sleepable lock because it is a non-sleepable lock and non-sleepable
  80  * locks may always be acquired while holding a sleepable lock.  The second
  81  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
  82  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
  83  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
  84  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
  85  * execute.  Thus, acquiring Giant both before and after a sleepable lock
  86  * will not result in a lock order reversal.
  87  */
  88
  89 #include <sys/cdefs.h>
  90 __FBSDID("$FreeBSD$");
  91
  92 #include "opt_ddb.h"
  93 #include "opt_hwpmc_hooks.h"
  94 #include "opt_stack.h"
  95 #include "opt_witness.h"
  96
  97 #include <sys/param.h>
  98 #include <sys/bus.h>
  99 #include <sys/kdb.h>
 100 #include <sys/kernel.h>
 101 #include <sys/ktr.h>
 102 #include <sys/lock.h>
 103 #include <sys/malloc.h>
 104 #include <sys/mutex.h>
 105 #include <sys/priv.h>
 106 #include <sys/proc.h>
 107 #include <sys/sbuf.h>
 108 #include <sys/sched.h>
 109 #include <sys/stack.h>
 110 #include <sys/sysctl.h>
 111 #include <sys/syslog.h>
 112 #include <sys/systm.h>
 113
 114 #ifdef DDB
 115 #include <ddb/ddb.h>
 116 #endif
 117
 118 #include <machine/stdarg.h>
 119
 120 #if !defined(DDB) && !defined(STACK)
 121 #error "DDB or STACK options are required for WITNESS"
 122 #endif
 123
 124 /* Note that these traces do not work with KTR_ALQ. */
 125 #if 0
 126 #define KTR_WITNESS     KTR_SUBSYS
 127 #else
 128 #define KTR_WITNESS     0
 129 #endif
 130
 131 #define LI_RECURSEMASK  0x0000ffff      /* Recursion depth of lock instance. */
 132 #define LI_EXCLUSIVE    0x00010000      /* Exclusive lock instance. */
 133 #define LI_NORELEASE    0x00020000      /* Lock not allowed to be released. */
 134 #define LI_SLEEPABLE    0x00040000      /* Lock may be held while sleeping. */
 135
 136 #ifndef WITNESS_COUNT
 137 #define WITNESS_COUNT           1536
 138 #endif
 139 #define WITNESS_HASH_SIZE       251     /* Prime, gives load factor < 2 */
 140 #define WITNESS_PENDLIST        (512 + (MAXCPU * 4))
 141
 142 /* Allocate 256 KB of stack data space */
 143 #define WITNESS_LO_DATA_COUNT   2048
 144
 145 /* Prime, gives load factor of ~2 at full load */
 146 #define WITNESS_LO_HASH_SIZE    1021
 147
 148 /*
 149  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
 150  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
 151  * probably be safe for the most part, but it's still a SWAG.
 152  */
 153 #define LOCK_NCHILDREN  5
 154 #define LOCK_CHILDCOUNT 2048
 155
 156 #define MAX_W_NAME      64
 157
 158 #define FULLGRAPH_SBUF_SIZE     512
 159
 160 /*
 161  * These flags go in the witness relationship matrix and describe the
 162  * relationship between any two struct witness objects.
 163  */
 164 #define WITNESS_UNRELATED        0x00    /* No lock order relation. */
 165 #define WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
 166 #define WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
 167 #define WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
 168 #define WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
 169 #define WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
 170 #define WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
 171 #define WITNESS_RELATED_MASK                                            \
 172         (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
 173 #define WITNESS_REVERSAL         0x10    /* A lock order reversal has been
 174                                           * observed. */
 175 #define WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
 176 #define WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
 177 #define WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
 178
 179 /* Descendant to ancestor flags */
 180 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
 181
 182 /* Ancestor to descendant flags */
 183 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
 184
 185 #define WITNESS_INDEX_ASSERT(i)                                         \
 186         MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
 187
 188 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
 189
 190 /*
 191  * Lock instances.  A lock instance is the data associated with a lock while
 192  * it is held by witness.  For example, a lock instance will hold the
 193  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
 194  * are held in a per-cpu list while sleep locks are held in per-thread list.
 195  */
 196 struct lock_instance {
 197         struct lock_object      *li_lock;
 198         const char              *li_file;
 199         int                     li_line;
 200         u_int                   li_flags;
 201 };
 202
 203 /*
 204  * A simple list type used to build the list of locks held by a thread
 205  * or CPU.  We can't simply embed the list in struct lock_object since a
 206  * lock may be held by more than one thread if it is a shared lock.  Locks
 207  * are added to the head of the list, so we fill up each list entry from
 208  * "the back" logically.  To ease some of the arithmetic, we actually fill
 209  * in each list entry the normal way (children[0] then children[1], etc.) but
 210  * when we traverse the list we read children[count-1] as the first entry
 211  * down to children[0] as the final entry.
 212  */
 213 struct lock_list_entry {
 214         struct lock_list_entry  *ll_next;
 215         struct lock_instance    ll_children[LOCK_NCHILDREN];
 216         u_int                   ll_count;
 217 };
 218
 219 /*
 220  * The main witness structure. One of these per named lock type in the system
 221  * (for example, "vnode interlock").
 222  */
 223 struct witness {
 224         char                    w_name[MAX_W_NAME];
 225         uint32_t                w_index;  /* Index in the relationship matrix */
 226         struct lock_class       *w_class;
 227         STAILQ_ENTRY(witness)   w_list;         /* List of all witnesses. */
 228         STAILQ_ENTRY(witness)   w_typelist;     /* Witnesses of a type. */
 229         struct witness          *w_hash_next; /* Linked list in hash buckets. */
 230         const char              *w_file; /* File where last acquired */
 231         uint32_t                w_line; /* Line where last acquired */
 232         uint32_t                w_refcount;
 233         uint16_t                w_num_ancestors; /* direct/indirect
 234                                                   * ancestor count */
 235         uint16_t                w_num_descendants; /* direct/indirect
 236                                                     * descendant count */
 237         int16_t                 w_ddb_level;
 238         unsigned                w_displayed:1;
 239         unsigned                w_reversed:1;
 240 };
 241
 242 STAILQ_HEAD(witness_list, witness);
 243
 244 /*
 245  * The witness hash table. Keys are witness names (const char *), elements are
 246  * witness objects (struct witness *).
 247  */
 248 struct witness_hash {
 249         struct witness  *wh_array[WITNESS_HASH_SIZE];
 250         uint32_t        wh_size;
 251         uint32_t        wh_count;
 252 };
 253
 254 /*
 255  * Key type for the lock order data hash table.
 256  */
 257 struct witness_lock_order_key {
 258         uint16_t        from;
 259         uint16_t        to;
 260 };
 261
 262 struct witness_lock_order_data {
 263         struct stack                    wlod_stack;
 264         struct witness_lock_order_key   wlod_key;
 265         struct witness_lock_order_data  *wlod_next;
 266 };
 267
 268 /*
 269  * The witness lock order data hash table. Keys are witness index tuples
 270  * (struct witness_lock_order_key), elements are lock order data objects
 271  * (struct witness_lock_order_data).
 272  */
 273 struct witness_lock_order_hash {
 274         struct witness_lock_order_data  *wloh_array[WITNESS_LO_HASH_SIZE];
 275         u_int   wloh_size;
 276         u_int   wloh_count;
 277 };
 278
 279 struct witness_blessed {
 280         const char      *b_lock1;
 281         const char      *b_lock2;
 282 };
 283
 284 struct witness_pendhelp {
 285         const char              *wh_type;
 286         struct lock_object      *wh_lock;
 287 };
 288
 289 struct witness_order_list_entry {
 290         const char              *w_name;
 291         struct lock_class       *w_class;
 292 };
 293
 294 /*
 295  * Returns 0 if one of the locks is a spin lock and the other is not.
 296  * Returns 1 otherwise.
 297  */
 298 static __inline int
 299 witness_lock_type_equal(struct witness *w1, struct witness *w2)
 300 {
 301
 302         return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
 303                 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
 304 }
 305
 306 static __inline int
 307 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
 308     const struct witness_lock_order_key *b)
 309 {
 310
 311         return (a->from == b->from && a->to == b->to);
 312 }
 313
 314 static int      _isitmyx(struct witness *w1, struct witness *w2, int rmask,
 315                     const char *fname);
 316 static void     adopt(struct witness *parent, struct witness *child);
 317 static int      blessed(struct witness *, struct witness *);
 318 static void     depart(struct witness *w);
 319 static struct witness   *enroll(const char *description,
 320                             struct lock_class *lock_class);
 321 static struct lock_instance     *find_instance(struct lock_list_entry *list,
 322                                     const struct lock_object *lock);
 323 static int      isitmychild(struct witness *parent, struct witness *child);
 324 static int      isitmydescendant(struct witness *parent, struct witness *child);
 325 static void     itismychild(struct witness *parent, struct witness *child);
 326 static int      sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
 327 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
 328 static int      sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
 329 static int      sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
 330 static void     witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
 331 #ifdef DDB
 332 static void     witness_ddb_compute_levels(void);
 333 static void     witness_ddb_display(int(*)(const char *fmt, ...));
 334 static void     witness_ddb_display_descendants(int(*)(const char *fmt, ...),
 335                     struct witness *, int indent);
 336 static void     witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
 337                     struct witness_list *list);
 338 static void     witness_ddb_level_descendants(struct witness *parent, int l);
 339 static void     witness_ddb_list(struct thread *td);
 340 #endif
 341 static void     witness_debugger(int cond, const char *msg);
 342 static void     witness_free(struct witness *m);
 343 static struct witness   *witness_get(void);
 344 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
 345 static struct witness   *witness_hash_get(const char *key);
 346 static void     witness_hash_put(struct witness *w);
 347 static void     witness_init_hash_tables(void);
 348 static void     witness_increment_graph_generation(void);
 349 static void     witness_lock_list_free(struct lock_list_entry *lle);
 350 static struct lock_list_entry   *witness_lock_list_get(void);
 351 static int      witness_lock_order_add(struct witness *parent,
 352                     struct witness *child);
 353 static int      witness_lock_order_check(struct witness *parent,
 354                     struct witness *child);
 355 static struct witness_lock_order_data   *witness_lock_order_get(
 356                                             struct witness *parent,
 357                                             struct witness *child);
 358 static void     witness_list_lock(struct lock_instance *instance,
 359                     int (*prnt)(const char *fmt, ...));
 360 static int      witness_output(const char *fmt, ...) __printflike(1, 2);
 361 static int      witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
 362 static void     witness_setflag(struct lock_object *lock, int flag, int set);
 363
 364 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
 365     "Witness Locking");
 366
 367 /*
 368  * If set to 0, lock order checking is disabled.  If set to -1,
 369  * witness is completely disabled.  Otherwise witness performs full
 370  * lock order checking for all locks.  At runtime, lock order checking
 371  * may be toggled.  However, witness cannot be reenabled once it is
 372  * completely disabled.
 373  */
 374 static int witness_watch = 1;
 375 SYSCTL_PROC(_debug_witness, OID_AUTO, watch,
 376     CTLFLAG_RWTUN | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0,
 377     sysctl_debug_witness_watch, "I",
 378     "witness is watching lock operations");
 379
 380 #ifdef KDB
 381 /*
 382  * When KDB is enabled and witness_kdb is 1, it will cause the system
 383  * to drop into kdebug() when:
 384  *      - a lock hierarchy violation occurs
 385  *      - locks are held when going to sleep.
 386  */
 387 #ifdef WITNESS_KDB
 388 int     witness_kdb = 1;
 389 #else
 390 int     witness_kdb = 0;
 391 #endif
 392 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
 393 #endif /* KDB */
 394
 395 #if defined(DDB) || defined(KDB)
 396 /*
 397  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
 398  * to print a stack trace:
 399  *      - a lock hierarchy violation occurs
 400  *      - locks are held when going to sleep.
 401  */
 402 int     witness_trace = 1;
 403 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
 404 #endif /* DDB || KDB */
 405
 406 #ifdef WITNESS_SKIPSPIN
 407 int     witness_skipspin = 1;
 408 #else
 409 int     witness_skipspin = 0;
 410 #endif
 411 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
 412
 413 int badstack_sbuf_size;
 414
 415 int witness_count = WITNESS_COUNT;
 416 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
 417     &witness_count, 0, "");
 418
 419 /*
 420  * Output channel for witness messages.  By default we print to the console.
 421  */
 422 enum witness_channel {
 423         WITNESS_CONSOLE,
 424         WITNESS_LOG,
 425         WITNESS_NONE,
 426 };
 427
 428 static enum witness_channel witness_channel = WITNESS_CONSOLE;
 429 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel,
 430     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, NULL, 0,
 431     sysctl_debug_witness_channel, "A",
 432     "Output channel for warnings");
 433
 434 /*
 435  * Call this to print out the relations between locks.
 436  */
 437 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph,
 438     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
 439     sysctl_debug_witness_fullgraph, "A",
 440     "Show locks relation graphs");
 441
 442 /*
 443  * Call this to print out the witness faulty stacks.
 444  */
 445 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks,
 446     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
 447     sysctl_debug_witness_badstacks, "A",
 448     "Show bad witness stacks");
 449
 450 static struct mtx w_mtx;
 451
 452 /* w_list */
 453 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 454 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 455
 456 /* w_typelist */
 457 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 458 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 459
 460 /* lock list */
 461 static struct lock_list_entry *w_lock_list_free = NULL;
 462 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
 463 static u_int pending_cnt;
 464
 465 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
 466 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
 467 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
 468 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
 469     "");
 470
 471 static struct witness *w_data;
 472 static uint8_t **w_rmatrix;
 473 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 474 static struct witness_hash w_hash;      /* The witness hash table. */
 475
 476 /* The lock order data hash */
 477 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
 478 static struct witness_lock_order_data *w_lofree = NULL;
 479 static struct witness_lock_order_hash w_lohash;
 480 static int w_max_used_index = 0;
 481 static unsigned int w_generation = 0;
 482 static const char w_notrunning[] = "Witness not running\n";
 483 static const char w_stillcold[] = "Witness is still cold\n";
 484 #ifdef __i386__
 485 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
 486 #endif
 487
 488 static struct witness_order_list_entry order_lists[] = {
 489         /*
 490          * sx locks
 491          */
 492         { "proctree", &lock_class_sx },
 493         { "allproc", &lock_class_sx },
 494         { "allprison", &lock_class_sx },
 495         { NULL, NULL },
 496         /*
 497          * Various mutexes
 498          */
 499         { "Giant", &lock_class_mtx_sleep },
 500         { "pipe mutex", &lock_class_mtx_sleep },
 501         { "sigio lock", &lock_class_mtx_sleep },
 502         { "process group", &lock_class_mtx_sleep },
 503 #ifdef  HWPMC_HOOKS
 504         { "pmc-sleep", &lock_class_mtx_sleep },
 505 #endif
 506         { "process lock", &lock_class_mtx_sleep },
 507         { "session", &lock_class_mtx_sleep },
 508         { "uidinfo hash", &lock_class_rw },
 509         { "time lock", &lock_class_mtx_sleep },
 510         { NULL, NULL },
 511         /*
 512          * umtx
 513          */
 514         { "umtx lock", &lock_class_mtx_sleep },
 515         { NULL, NULL },
 516         /*
 517          * Sockets
 518          */
 519         { "accept", &lock_class_mtx_sleep },
 520         { "so_snd", &lock_class_mtx_sleep },
 521         { "so_rcv", &lock_class_mtx_sleep },
 522         { "sellck", &lock_class_mtx_sleep },
 523         { NULL, NULL },
 524         /*
 525          * Routing
 526          */
 527         { "so_rcv", &lock_class_mtx_sleep },
 528         { "radix node head", &lock_class_rm },
 529         { "rtentry", &lock_class_mtx_sleep },
 530         { "ifaddr", &lock_class_mtx_sleep },
 531         { NULL, NULL },
 532         /*
 533          * IPv4 multicast:
 534          * protocol locks before interface locks, after UDP locks.
 535          */
 536         { "in_multi_sx", &lock_class_sx },
 537         { "udpinp", &lock_class_rw },
 538         { "in_multi_list_mtx", &lock_class_mtx_sleep },
 539         { "igmp_mtx", &lock_class_mtx_sleep },
 540         { "ifnet_rw", &lock_class_rw },
 541         { "if_addr_lock", &lock_class_mtx_sleep },
 542         { NULL, NULL },
 543         /*
 544          * IPv6 multicast:
 545          * protocol locks before interface locks, after UDP locks.
 546          */
 547         { "in6_multi_sx", &lock_class_sx },
 548         { "udpinp", &lock_class_rw },
 549         { "in6_multi_list_mtx", &lock_class_mtx_sleep },
 550         { "mld_mtx", &lock_class_mtx_sleep },
 551         { "ifnet_rw", &lock_class_rw },
 552         { "if_addr_lock", &lock_class_mtx_sleep },
 553         { NULL, NULL },
 554         /*
 555          * UNIX Domain Sockets
 556          */
 557         { "unp_link_rwlock", &lock_class_rw },
 558         { "unp_list_lock", &lock_class_mtx_sleep },
 559         { "unp", &lock_class_mtx_sleep },
 560         { "so_snd", &lock_class_mtx_sleep },
 561         { NULL, NULL },
 562         /*
 563          * UDP/IP
 564          */
 565         { "udp", &lock_class_mtx_sleep },
 566         { "udpinp", &lock_class_rw },
 567         { "so_snd", &lock_class_mtx_sleep },
 568         { NULL, NULL },
 569         /*
 570          * TCP/IP
 571          */
 572         { "tcp", &lock_class_mtx_sleep },
 573         { "tcpinp", &lock_class_rw },
 574         { "so_snd", &lock_class_mtx_sleep },
 575         { NULL, NULL },
 576         /*
 577          * BPF
 578          */
 579         { "bpf global lock", &lock_class_sx },
 580         { "bpf cdev lock", &lock_class_mtx_sleep },
 581         { NULL, NULL },
 582         /*
 583          * NFS server
 584          */
 585         { "nfsd_mtx", &lock_class_mtx_sleep },
 586         { "so_snd", &lock_class_mtx_sleep },
 587         { NULL, NULL },
 588
 589         /*
 590          * IEEE 802.11
 591          */
 592         { "802.11 com lock", &lock_class_mtx_sleep},
 593         { NULL, NULL },
 594         /*
 595          * Network drivers
 596          */
 597         { "network driver", &lock_class_mtx_sleep},
 598         { NULL, NULL },
 599
 600         /*
 601          * Netgraph
 602          */
 603         { "ng_node", &lock_class_mtx_sleep },
 604         { "ng_worklist", &lock_class_mtx_sleep },
 605         { NULL, NULL },
 606         /*
 607          * CDEV
 608          */
 609         { "vm map (system)", &lock_class_mtx_sleep },
 610         { "vnode interlock", &lock_class_mtx_sleep },
 611         { "cdev", &lock_class_mtx_sleep },
 612         { "devthrd", &lock_class_mtx_sleep },
 613         { NULL, NULL },
 614         /*
 615          * VM
 616          */
 617         { "vm map (user)", &lock_class_sx },
 618         { "vm object", &lock_class_rw },
 619         { "vm page", &lock_class_mtx_sleep },
 620         { "pmap pv global", &lock_class_rw },
 621         { "pmap", &lock_class_mtx_sleep },
 622         { "pmap pv list", &lock_class_rw },
 623         { "vm page free queue", &lock_class_mtx_sleep },
 624         { "vm pagequeue", &lock_class_mtx_sleep },
 625         { NULL, NULL },
 626         /*
 627          * kqueue/VFS interaction
 628          */
 629         { "kqueue", &lock_class_mtx_sleep },
 630         { "struct mount mtx", &lock_class_mtx_sleep },
 631         { "vnode interlock", &lock_class_mtx_sleep },
 632         { NULL, NULL },
 633         /*
 634          * VFS namecache
 635          */
 636         { "ncvn", &lock_class_mtx_sleep },
 637         { "ncbuc", &lock_class_rw },
 638         { "vnode interlock", &lock_class_mtx_sleep },
 639         { "ncneg", &lock_class_mtx_sleep },
 640         { NULL, NULL },
 641         /*
 642          * ZFS locking
 643          */
 644         { "dn->dn_mtx", &lock_class_sx },
 645         { "dr->dt.di.dr_mtx", &lock_class_sx },
 646         { "db->db_mtx", &lock_class_sx },
 647         { NULL, NULL },
 648         /*
 649          * TCP log locks
 650          */
 651         { "TCP ID tree", &lock_class_rw },
 652         { "tcp log id bucket", &lock_class_mtx_sleep },
 653         { "tcpinp", &lock_class_rw },
 654         { "TCP log expireq", &lock_class_mtx_sleep },
 655         { NULL, NULL },
 656         /*
 657          * spin locks
 658          */
 659 #ifdef SMP
 660         { "ap boot", &lock_class_mtx_spin },
 661 #endif
 662         { "rm.mutex_mtx", &lock_class_mtx_spin },
 663         { "sio", &lock_class_mtx_spin },
 664 #ifdef __i386__
 665         { "cy", &lock_class_mtx_spin },
 666 #endif
 667         { "scc_hwmtx", &lock_class_mtx_spin },
 668         { "uart_hwmtx", &lock_class_mtx_spin },
 669         { "fast_taskqueue", &lock_class_mtx_spin },
 670         { "intr table", &lock_class_mtx_spin },
 671         { "process slock", &lock_class_mtx_spin },
 672         { "syscons video lock", &lock_class_mtx_spin },
 673         { "sleepq chain", &lock_class_mtx_spin },
 674         { "rm_spinlock", &lock_class_mtx_spin },
 675         { "turnstile chain", &lock_class_mtx_spin },
 676         { "turnstile lock", &lock_class_mtx_spin },
 677         { "sched lock", &lock_class_mtx_spin },
 678         { "td_contested", &lock_class_mtx_spin },
 679         { "callout", &lock_class_mtx_spin },
 680         { "entropy harvest mutex", &lock_class_mtx_spin },
 681 #ifdef SMP
 682         { "smp rendezvous", &lock_class_mtx_spin },
 683 #endif
 684 #ifdef __powerpc__
 685         { "tlb0", &lock_class_mtx_spin },
 686 #endif
 687         { NULL, NULL },
 688         { "sched lock", &lock_class_mtx_spin },
 689 #ifdef  HWPMC_HOOKS
 690         { "pmc-per-proc", &lock_class_mtx_spin },
 691 #endif
 692         { NULL, NULL },
 693         /*
 694          * leaf locks
 695          */
 696         { "intrcnt", &lock_class_mtx_spin },
 697         { "icu", &lock_class_mtx_spin },
 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 || KERNEL_PANICKED() ||
 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 || KERNEL_PANICKED())
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             KERNEL_PANICKED())
1122                 return;
1123
1124         w = lock->lo_witness;
1125         class = LOCK_CLASS(lock);
1126         td = curthread;
1127
1128         if (class->lc_flags & LC_SLEEPLOCK) {
1129
1130                 /*
1131                  * Since spin locks include a critical section, this check
1132                  * implicitly enforces a lock order of all sleep locks before
1133                  * all spin locks.
1134                  */
1135                 if (td->td_critnest != 0 && !kdb_active)
1136                         kassert_panic("acquiring blockable sleep lock with "
1137                             "spinlock or critical section held (%s) %s @ %s:%d",
1138                             class->lc_name, lock->lo_name,
1139                             fixup_filename(file), line);
1140
1141                 /*
1142                  * If this is the first lock acquired then just return as
1143                  * no order checking is needed.
1144                  */
1145                 lock_list = td->td_sleeplocks;
1146                 if (lock_list == NULL || lock_list->ll_count == 0)
1147                         return;
1148         } else {
1149
1150                 /*
1151                  * If this is the first lock, just return as no order
1152                  * checking is needed.  Avoid problems with thread
1153                  * migration pinning the thread while checking if
1154                  * spinlocks are held.  If at least one spinlock is held
1155                  * the thread is in a safe path and it is allowed to
1156                  * unpin it.
1157                  */
1158                 sched_pin();
1159                 lock_list = PCPU_GET(spinlocks);
1160                 if (lock_list == NULL || lock_list->ll_count == 0) {
1161                         sched_unpin();
1162                         return;
1163                 }
1164                 sched_unpin();
1165         }
1166
1167         /*
1168          * Check to see if we are recursing on a lock we already own.  If
1169          * so, make sure that we don't mismatch exclusive and shared lock
1170          * acquires.
1171          */
1172         lock1 = find_instance(lock_list, lock);
1173         if (lock1 != NULL) {
1174                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1175                     (flags & LOP_EXCLUSIVE) == 0) {
1176                         witness_output("shared lock of (%s) %s @ %s:%d\n",
1177                             class->lc_name, lock->lo_name,
1178                             fixup_filename(file), line);
1179                         witness_output("while exclusively locked from %s:%d\n",
1180                             fixup_filename(lock1->li_file), lock1->li_line);
1181                         kassert_panic("excl->share");
1182                 }
1183                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1184                     (flags & LOP_EXCLUSIVE) != 0) {
1185                         witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1186                             class->lc_name, lock->lo_name,
1187                             fixup_filename(file), line);
1188                         witness_output("while share locked from %s:%d\n",
1189                             fixup_filename(lock1->li_file), lock1->li_line);
1190                         kassert_panic("share->excl");
1191                 }
1192                 return;
1193         }
1194
1195         /* Warn if the interlock is not locked exactly once. */
1196         if (interlock != NULL) {
1197                 iclass = LOCK_CLASS(interlock);
1198                 lock1 = find_instance(lock_list, interlock);
1199                 if (lock1 == NULL)
1200                         kassert_panic("interlock (%s) %s not locked @ %s:%d",
1201                             iclass->lc_name, interlock->lo_name,
1202                             fixup_filename(file), line);
1203                 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1204                         kassert_panic("interlock (%s) %s recursed @ %s:%d",
1205                             iclass->lc_name, interlock->lo_name,
1206                             fixup_filename(file), line);
1207         }
1208
1209         /*
1210          * Find the previously acquired lock, but ignore interlocks.
1211          */
1212         plock = &lock_list->ll_children[lock_list->ll_count - 1];
1213         if (interlock != NULL && plock->li_lock == interlock) {
1214                 if (lock_list->ll_count > 1)
1215                         plock =
1216                             &lock_list->ll_children[lock_list->ll_count - 2];
1217                 else {
1218                         lle = lock_list->ll_next;
1219
1220                         /*
1221                          * The interlock is the only lock we hold, so
1222                          * simply return.
1223                          */
1224                         if (lle == NULL)
1225                                 return;
1226                         plock = &lle->ll_children[lle->ll_count - 1];
1227                 }
1228         }
1229
1230         /*
1231          * Try to perform most checks without a lock.  If this succeeds we
1232          * can skip acquiring the lock and return success.  Otherwise we redo
1233          * the check with the lock held to handle races with concurrent updates.
1234          */
1235         w1 = plock->li_lock->lo_witness;
1236         if (witness_lock_order_check(w1, w))
1237                 return;
1238
1239         mtx_lock_spin(&w_mtx);
1240         if (witness_lock_order_check(w1, w)) {
1241                 mtx_unlock_spin(&w_mtx);
1242                 return;
1243         }
1244         witness_lock_order_add(w1, w);
1245
1246         /*
1247          * Check for duplicate locks of the same type.  Note that we only
1248          * have to check for this on the last lock we just acquired.  Any
1249          * other cases will be caught as lock order violations.
1250          */
1251         if (w1 == w) {
1252                 i = w->w_index;
1253                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1254                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1255                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
1256                         w->w_reversed = 1;
1257                         mtx_unlock_spin(&w_mtx);
1258                         witness_output(
1259                             "acquiring duplicate lock of same type: \"%s\"\n",
1260                             w->w_name);
1261                         witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1262                             fixup_filename(plock->li_file), plock->li_line);
1263                         witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1264                             fixup_filename(file), line);
1265                         witness_debugger(1, __func__);
1266                 } else
1267                         mtx_unlock_spin(&w_mtx);
1268                 return;
1269         }
1270         mtx_assert(&w_mtx, MA_OWNED);
1271
1272         /*
1273          * If we know that the lock we are acquiring comes after
1274          * the lock we most recently acquired in the lock order tree,
1275          * then there is no need for any further checks.
1276          */
1277         if (isitmychild(w1, w))
1278                 goto out;
1279
1280         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1281                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1282
1283                         MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1284                         lock1 = &lle->ll_children[i];
1285
1286                         /*
1287                          * Ignore the interlock.
1288                          */
1289                         if (interlock == lock1->li_lock)
1290                                 continue;
1291
1292                         /*
1293                          * If this lock doesn't undergo witness checking,
1294                          * then skip it.
1295                          */
1296                         w1 = lock1->li_lock->lo_witness;
1297                         if (w1 == NULL) {
1298                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1299                                     ("lock missing witness structure"));
1300                                 continue;
1301                         }
1302
1303                         /*
1304                          * If we are locking Giant and this is a sleepable
1305                          * lock, then skip it.
1306                          */
1307                         if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1308                             lock == &Giant.lock_object)
1309                                 continue;
1310
1311                         /*
1312                          * If we are locking a sleepable lock and this lock
1313                          * is Giant, then skip it.
1314                          */
1315                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1316                             (flags & LOP_NOSLEEP) == 0 &&
1317                             lock1->li_lock == &Giant.lock_object)
1318                                 continue;
1319
1320                         /*
1321                          * If we are locking a sleepable lock and this lock
1322                          * isn't sleepable, we want to treat it as a lock
1323                          * order violation to enfore a general lock order of
1324                          * sleepable locks before non-sleepable locks.
1325                          */
1326                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1327                             (flags & LOP_NOSLEEP) == 0 &&
1328                             (lock1->li_flags & LI_SLEEPABLE) == 0)
1329                                 goto reversal;
1330
1331                         /*
1332                          * If we are locking Giant and this is a non-sleepable
1333                          * lock, then treat it as a reversal.
1334                          */
1335                         if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1336                             lock == &Giant.lock_object)
1337                                 goto reversal;
1338
1339                         /*
1340                          * Check the lock order hierarchy for a reveresal.
1341                          */
1342                         if (!isitmydescendant(w, w1))
1343                                 continue;
1344                 reversal:
1345
1346                         /*
1347                          * We have a lock order violation, check to see if it
1348                          * is allowed or has already been yelled about.
1349                          */
1350
1351                         /* Bail if this violation is known */
1352                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1353                                 goto out;
1354
1355                         /* Record this as a violation */
1356                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1357                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1358                         w->w_reversed = w1->w_reversed = 1;
1359                         witness_increment_graph_generation();
1360
1361                         /*
1362                          * If the lock order is blessed, bail before logging
1363                          * anything.  We don't look for other lock order
1364                          * violations though, which may be a bug.
1365                          */
1366                         if (blessed(w, w1))
1367                                 goto out;
1368                         mtx_unlock_spin(&w_mtx);
1369
1370 #ifdef WITNESS_NO_VNODE
1371                         /*
1372                          * There are known LORs between VNODE locks. They are
1373                          * not an indication of a bug. VNODE locks are flagged
1374                          * as such (LO_IS_VNODE) and we don't yell if the LOR
1375                          * is between 2 VNODE locks.
1376                          */
1377                         if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1378                             (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1379                                 return;
1380 #endif
1381
1382                         /*
1383                          * Ok, yell about it.
1384                          */
1385                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1386                             (flags & LOP_NOSLEEP) == 0 &&
1387                             (lock1->li_flags & LI_SLEEPABLE) == 0)
1388                                 witness_output(
1389                 "lock order reversal: (sleepable after non-sleepable)\n");
1390                         else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1391                             && lock == &Giant.lock_object)
1392                                 witness_output(
1393                 "lock order reversal: (Giant after non-sleepable)\n");
1394                         else
1395                                 witness_output("lock order reversal:\n");
1396
1397                         /*
1398                          * Try to locate an earlier lock with
1399                          * witness w in our list.
1400                          */
1401                         do {
1402                                 lock2 = &lle->ll_children[i];
1403                                 MPASS(lock2->li_lock != NULL);
1404                                 if (lock2->li_lock->lo_witness == w)
1405                                         break;
1406                                 if (i == 0 && lle->ll_next != NULL) {
1407                                         lle = lle->ll_next;
1408                                         i = lle->ll_count - 1;
1409                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1410                                 } else
1411                                         i--;
1412                         } while (i >= 0);
1413                         if (i < 0) {
1414                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1415                                     lock1->li_lock, lock1->li_lock->lo_name,
1416                                     w1->w_name, fixup_filename(lock1->li_file),
1417                                     lock1->li_line);
1418                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1419                                     lock->lo_name, w->w_name,
1420                                     fixup_filename(file), line);
1421                         } else {
1422                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1423                                     lock2->li_lock, lock2->li_lock->lo_name,
1424                                     lock2->li_lock->lo_witness->w_name,
1425                                     fixup_filename(lock2->li_file),
1426                                     lock2->li_line);
1427                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1428                                     lock1->li_lock, lock1->li_lock->lo_name,
1429                                     w1->w_name, fixup_filename(lock1->li_file),
1430                                     lock1->li_line);
1431                                 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1432                                     lock->lo_name, w->w_name,
1433                                     fixup_filename(file), line);
1434                         }
1435                         witness_debugger(1, __func__);
1436                         return;
1437                 }
1438         }
1439
1440         /*
1441          * If requested, build a new lock order.  However, don't build a new
1442          * relationship between a sleepable lock and Giant if it is in the
1443          * wrong direction.  The correct lock order is that sleepable locks
1444          * always come before Giant.
1445          */
1446         if (flags & LOP_NEWORDER &&
1447             !(plock->li_lock == &Giant.lock_object &&
1448             (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1449             (flags & LOP_NOSLEEP) == 0)) {
1450                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1451                     w->w_name, plock->li_lock->lo_witness->w_name);
1452                 itismychild(plock->li_lock->lo_witness, w);
1453         }
1454 out:
1455         mtx_unlock_spin(&w_mtx);
1456 }
1457
1458 void
1459 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1460 {
1461         struct lock_list_entry **lock_list, *lle;
1462         struct lock_instance *instance;
1463         struct witness *w;
1464         struct thread *td;
1465
1466         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1467             KERNEL_PANICKED())
1468                 return;
1469         w = lock->lo_witness;
1470         td = curthread;
1471
1472         /* Determine lock list for this lock. */
1473         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1474                 lock_list = &td->td_sleeplocks;
1475         else
1476                 lock_list = PCPU_PTR(spinlocks);
1477
1478         /* Check to see if we are recursing on a lock we already own. */
1479         instance = find_instance(*lock_list, lock);
1480         if (instance != NULL) {
1481                 instance->li_flags++;
1482                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1483                     td->td_proc->p_pid, lock->lo_name,
1484                     instance->li_flags & LI_RECURSEMASK);
1485                 instance->li_file = file;
1486                 instance->li_line = line;
1487                 return;
1488         }
1489
1490         /* Update per-witness last file and line acquire. */
1491         w->w_file = file;
1492         w->w_line = line;
1493
1494         /* Find the next open lock instance in the list and fill it. */
1495         lle = *lock_list;
1496         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1497                 lle = witness_lock_list_get();
1498                 if (lle == NULL)
1499                         return;
1500                 lle->ll_next = *lock_list;
1501                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1502                     td->td_proc->p_pid, lle);
1503                 *lock_list = lle;
1504         }
1505         instance = &lle->ll_children[lle->ll_count++];
1506         instance->li_lock = lock;
1507         instance->li_line = line;
1508         instance->li_file = file;
1509         instance->li_flags = 0;
1510         if ((flags & LOP_EXCLUSIVE) != 0)
1511                 instance->li_flags |= LI_EXCLUSIVE;
1512         if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1513                 instance->li_flags |= LI_SLEEPABLE;
1514         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1515             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1516 }
1517
1518 void
1519 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1520 {
1521         struct lock_instance *instance;
1522         struct lock_class *class;
1523
1524         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1525         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1526                 return;
1527         class = LOCK_CLASS(lock);
1528         if (witness_watch) {
1529                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1530                         kassert_panic(
1531                             "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1532                             class->lc_name, lock->lo_name,
1533                             fixup_filename(file), line);
1534                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1535                         kassert_panic(
1536                             "upgrade of non-sleep lock (%s) %s @ %s:%d",
1537                             class->lc_name, lock->lo_name,
1538                             fixup_filename(file), line);
1539         }
1540         instance = find_instance(curthread->td_sleeplocks, lock);
1541         if (instance == NULL) {
1542                 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1543                     class->lc_name, lock->lo_name,
1544                     fixup_filename(file), line);
1545                 return;
1546         }
1547         if (witness_watch) {
1548                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1549                         kassert_panic(
1550                             "upgrade of exclusive lock (%s) %s @ %s:%d",
1551                             class->lc_name, lock->lo_name,
1552                             fixup_filename(file), line);
1553                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1554                         kassert_panic(
1555                             "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1556                             class->lc_name, lock->lo_name,
1557                             instance->li_flags & LI_RECURSEMASK,
1558                             fixup_filename(file), line);
1559         }
1560         instance->li_flags |= LI_EXCLUSIVE;
1561 }
1562
1563 void
1564 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1565     int line)
1566 {
1567         struct lock_instance *instance;
1568         struct lock_class *class;
1569
1570         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1571         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1572                 return;
1573         class = LOCK_CLASS(lock);
1574         if (witness_watch) {
1575                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1576                         kassert_panic(
1577                             "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1578                             class->lc_name, lock->lo_name,
1579                             fixup_filename(file), line);
1580                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1581                         kassert_panic(
1582                             "downgrade of non-sleep lock (%s) %s @ %s:%d",
1583                             class->lc_name, lock->lo_name,
1584                             fixup_filename(file), line);
1585         }
1586         instance = find_instance(curthread->td_sleeplocks, lock);
1587         if (instance == NULL) {
1588                 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1589                     class->lc_name, lock->lo_name,
1590                     fixup_filename(file), line);
1591                 return;
1592         }
1593         if (witness_watch) {
1594                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1595                         kassert_panic(
1596                             "downgrade of shared lock (%s) %s @ %s:%d",
1597                             class->lc_name, lock->lo_name,
1598                             fixup_filename(file), line);
1599                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1600                         kassert_panic(
1601                             "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1602                             class->lc_name, lock->lo_name,
1603                             instance->li_flags & LI_RECURSEMASK,
1604                             fixup_filename(file), line);
1605         }
1606         instance->li_flags &= ~LI_EXCLUSIVE;
1607 }
1608
1609 void
1610 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1611 {
1612         struct lock_list_entry **lock_list, *lle;
1613         struct lock_instance *instance;
1614         struct lock_class *class;
1615         struct thread *td;
1616         register_t s;
1617         int i, j;
1618
1619         if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1620                 return;
1621         td = curthread;
1622         class = LOCK_CLASS(lock);
1623
1624         /* Find lock instance associated with this lock. */
1625         if (class->lc_flags & LC_SLEEPLOCK)
1626                 lock_list = &td->td_sleeplocks;
1627         else
1628                 lock_list = PCPU_PTR(spinlocks);
1629         lle = *lock_list;
1630         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1631                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1632                         instance = &(*lock_list)->ll_children[i];
1633                         if (instance->li_lock == lock)
1634                                 goto found;
1635                 }
1636
1637         /*
1638          * When disabling WITNESS through witness_watch we could end up in
1639          * having registered locks in the td_sleeplocks queue.
1640          * We have to make sure we flush these queues, so just search for
1641          * eventual register locks and remove them.
1642          */
1643         if (witness_watch > 0) {
1644                 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1645                     lock->lo_name, fixup_filename(file), line);
1646                 return;
1647         } else {
1648                 return;
1649         }
1650 found:
1651
1652         /* First, check for shared/exclusive mismatches. */
1653         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1654             (flags & LOP_EXCLUSIVE) == 0) {
1655                 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1656                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1657                 witness_output("while exclusively locked from %s:%d\n",
1658                     fixup_filename(instance->li_file), instance->li_line);
1659                 kassert_panic("excl->ushare");
1660         }
1661         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1662             (flags & LOP_EXCLUSIVE) != 0) {
1663                 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1664                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1665                 witness_output("while share locked from %s:%d\n",
1666                     fixup_filename(instance->li_file),
1667                     instance->li_line);
1668                 kassert_panic("share->uexcl");
1669         }
1670         /* If we are recursed, unrecurse. */
1671         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1672                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1673                     td->td_proc->p_pid, instance->li_lock->lo_name,
1674                     instance->li_flags);
1675                 instance->li_flags--;
1676                 return;
1677         }
1678         /* The lock is now being dropped, check for NORELEASE flag */
1679         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1680                 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1681                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1682                 kassert_panic("lock marked norelease");
1683         }
1684
1685         /* Otherwise, remove this item from the list. */
1686         s = intr_disable();
1687         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1688             td->td_proc->p_pid, instance->li_lock->lo_name,
1689             (*lock_list)->ll_count - 1);
1690         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1691                 (*lock_list)->ll_children[j] =
1692                     (*lock_list)->ll_children[j + 1];
1693         (*lock_list)->ll_count--;
1694         intr_restore(s);
1695
1696         /*
1697          * In order to reduce contention on w_mtx, we want to keep always an
1698          * head object into lists so that frequent allocation from the
1699          * free witness pool (and subsequent locking) is avoided.
1700          * In order to maintain the current code simple, when the head
1701          * object is totally unloaded it means also that we do not have
1702          * further objects in the list, so the list ownership needs to be
1703          * hand over to another object if the current head needs to be freed.
1704          */
1705         if ((*lock_list)->ll_count == 0) {
1706                 if (*lock_list == lle) {
1707                         if (lle->ll_next == NULL)
1708                                 return;
1709                 } else
1710                         lle = *lock_list;
1711                 *lock_list = lle->ll_next;
1712                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1713                     td->td_proc->p_pid, lle);
1714                 witness_lock_list_free(lle);
1715         }
1716 }
1717
1718 void
1719 witness_thread_exit(struct thread *td)
1720 {
1721         struct lock_list_entry *lle;
1722         int i, n;
1723
1724         lle = td->td_sleeplocks;
1725         if (lle == NULL || KERNEL_PANICKED())
1726                 return;
1727         if (lle->ll_count != 0) {
1728                 for (n = 0; lle != NULL; lle = lle->ll_next)
1729                         for (i = lle->ll_count - 1; i >= 0; i--) {
1730                                 if (n == 0)
1731                                         witness_output(
1732                     "Thread %p exiting with the following locks held:\n", td);
1733                                 n++;
1734                                 witness_list_lock(&lle->ll_children[i],
1735                                     witness_output);
1736
1737                         }
1738                 kassert_panic(
1739                     "Thread %p cannot exit while holding sleeplocks\n", td);
1740         }
1741         witness_lock_list_free(lle);
1742 }
1743
1744 /*
1745  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1746  * exempt Giant and sleepable locks from the checks as well.  If any
1747  * non-exempt locks are held, then a supplied message is printed to the
1748  * output channel along with a list of the offending locks.  If indicated in the
1749  * flags then a failure results in a panic as well.
1750  */
1751 int
1752 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1753 {
1754         struct lock_list_entry *lock_list, *lle;
1755         struct lock_instance *lock1;
1756         struct thread *td;
1757         va_list ap;
1758         int i, n;
1759
1760         if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1761                 return (0);
1762         n = 0;
1763         td = curthread;
1764         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1765                 for (i = lle->ll_count - 1; i >= 0; i--) {
1766                         lock1 = &lle->ll_children[i];
1767                         if (lock1->li_lock == lock)
1768                                 continue;
1769                         if (flags & WARN_GIANTOK &&
1770                             lock1->li_lock == &Giant.lock_object)
1771                                 continue;
1772                         if (flags & WARN_SLEEPOK &&
1773                             (lock1->li_flags & LI_SLEEPABLE) != 0)
1774                                 continue;
1775                         if (n == 0) {
1776                                 va_start(ap, fmt);
1777                                 vprintf(fmt, ap);
1778                                 va_end(ap);
1779                                 printf(" with the following %slocks held:\n",
1780                                     (flags & WARN_SLEEPOK) != 0 ?
1781                                     "non-sleepable " : "");
1782                         }
1783                         n++;
1784                         witness_list_lock(lock1, printf);
1785                 }
1786
1787         /*
1788          * Pin the thread in order to avoid problems with thread migration.
1789          * Once that all verifies are passed about spinlocks ownership,
1790          * the thread is in a safe path and it can be unpinned.
1791          */
1792         sched_pin();
1793         lock_list = PCPU_GET(spinlocks);
1794         if (lock_list != NULL && lock_list->ll_count != 0) {
1795                 sched_unpin();
1796
1797                 /*
1798                  * We should only have one spinlock and as long as
1799                  * the flags cannot match for this locks class,
1800                  * check if the first spinlock is the one curthread
1801                  * should hold.
1802                  */
1803                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1804                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1805                     lock1->li_lock == lock && n == 0)
1806                         return (0);
1807
1808                 va_start(ap, fmt);
1809                 vprintf(fmt, ap);
1810                 va_end(ap);
1811                 printf(" with the following %slocks held:\n",
1812                     (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1813                 n += witness_list_locks(&lock_list, printf);
1814         } else
1815                 sched_unpin();
1816         if (flags & WARN_PANIC && n)
1817                 kassert_panic("%s", __func__);
1818         else
1819                 witness_debugger(n, __func__);
1820         return (n);
1821 }
1822
1823 const char *
1824 witness_file(struct lock_object *lock)
1825 {
1826         struct witness *w;
1827
1828         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1829                 return ("?");
1830         w = lock->lo_witness;
1831         return (w->w_file);
1832 }
1833
1834 int
1835 witness_line(struct lock_object *lock)
1836 {
1837         struct witness *w;
1838
1839         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1840                 return (0);
1841         w = lock->lo_witness;
1842         return (w->w_line);
1843 }
1844
1845 static struct witness *
1846 enroll(const char *description, struct lock_class *lock_class)
1847 {
1848         struct witness *w;
1849
1850         MPASS(description != NULL);
1851
1852         if (witness_watch == -1 || KERNEL_PANICKED())
1853                 return (NULL);
1854         if ((lock_class->lc_flags & LC_SPINLOCK)) {
1855                 if (witness_skipspin)
1856                         return (NULL);
1857         } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1858                 kassert_panic("lock class %s is not sleep or spin",
1859                     lock_class->lc_name);
1860                 return (NULL);
1861         }
1862
1863         mtx_lock_spin(&w_mtx);
1864         w = witness_hash_get(description);
1865         if (w)
1866                 goto found;
1867         if ((w = witness_get()) == NULL)
1868                 return (NULL);
1869         MPASS(strlen(description) < MAX_W_NAME);
1870         strcpy(w->w_name, description);
1871         w->w_class = lock_class;
1872         w->w_refcount = 1;
1873         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1874         if (lock_class->lc_flags & LC_SPINLOCK) {
1875                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1876                 w_spin_cnt++;
1877         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1878                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1879                 w_sleep_cnt++;
1880         }
1881
1882         /* Insert new witness into the hash */
1883         witness_hash_put(w);
1884         witness_increment_graph_generation();
1885         mtx_unlock_spin(&w_mtx);
1886         return (w);
1887 found:
1888         w->w_refcount++;
1889         if (w->w_refcount == 1)
1890                 w->w_class = lock_class;
1891         mtx_unlock_spin(&w_mtx);
1892         if (lock_class != w->w_class)
1893                 kassert_panic(
1894                     "lock (%s) %s does not match earlier (%s) lock",
1895                     description, lock_class->lc_name,
1896                     w->w_class->lc_name);
1897         return (w);
1898 }
1899
1900 static void
1901 depart(struct witness *w)
1902 {
1903
1904         MPASS(w->w_refcount == 0);
1905         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1906                 w_sleep_cnt--;
1907         } else {
1908                 w_spin_cnt--;
1909         }
1910         /*
1911          * Set file to NULL as it may point into a loadable module.
1912          */
1913         w->w_file = NULL;
1914         w->w_line = 0;
1915         witness_increment_graph_generation();
1916 }
1917
1918 static void
1919 adopt(struct witness *parent, struct witness *child)
1920 {
1921         int pi, ci, i, j;
1922
1923         if (witness_cold == 0)
1924                 mtx_assert(&w_mtx, MA_OWNED);
1925
1926         /* If the relationship is already known, there's no work to be done. */
1927         if (isitmychild(parent, child))
1928                 return;
1929
1930         /* When the structure of the graph changes, bump up the generation. */
1931         witness_increment_graph_generation();
1932
1933         /*
1934          * The hard part ... create the direct relationship, then propagate all
1935          * indirect relationships.
1936          */
1937         pi = parent->w_index;
1938         ci = child->w_index;
1939         WITNESS_INDEX_ASSERT(pi);
1940         WITNESS_INDEX_ASSERT(ci);
1941         MPASS(pi != ci);
1942         w_rmatrix[pi][ci] |= WITNESS_PARENT;
1943         w_rmatrix[ci][pi] |= WITNESS_CHILD;
1944
1945         /*
1946          * If parent was not already an ancestor of child,
1947          * then we increment the descendant and ancestor counters.
1948          */
1949         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1950                 parent->w_num_descendants++;
1951                 child->w_num_ancestors++;
1952         }
1953
1954         /*
1955          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1956          * an ancestor of 'pi' during this loop.
1957          */
1958         for (i = 1; i <= w_max_used_index; i++) {
1959                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1960                     (i != pi))
1961                         continue;
1962
1963                 /* Find each descendant of 'i' and mark it as a descendant. */
1964                 for (j = 1; j <= w_max_used_index; j++) {
1965
1966                         /*
1967                          * Skip children that are already marked as
1968                          * descendants of 'i'.
1969                          */
1970                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1971                                 continue;
1972
1973                         /*
1974                          * We are only interested in descendants of 'ci'. Note
1975                          * that 'ci' itself is counted as a descendant of 'ci'.
1976                          */
1977                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1978                             (j != ci))
1979                                 continue;
1980                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1981                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1982                         w_data[i].w_num_descendants++;
1983                         w_data[j].w_num_ancestors++;
1984
1985                         /*
1986                          * Make sure we aren't marking a node as both an
1987                          * ancestor and descendant. We should have caught
1988                          * this as a lock order reversal earlier.
1989                          */
1990                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1991                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1992                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1993                                     "both ancestor and descendant\n",
1994                                     i, j, w_rmatrix[i][j]);
1995                                 kdb_backtrace();
1996                                 printf("Witness disabled.\n");
1997                                 witness_watch = -1;
1998                         }
1999                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2000                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2001                                 printf("witness rmatrix paradox! [%d][%d]=%d "
2002                                     "both ancestor and descendant\n",
2003                                     j, i, w_rmatrix[j][i]);
2004                                 kdb_backtrace();
2005                                 printf("Witness disabled.\n");
2006                                 witness_watch = -1;
2007                         }
2008                 }
2009         }
2010 }
2011
2012 static void
2013 itismychild(struct witness *parent, struct witness *child)
2014 {
2015         int unlocked;
2016
2017         MPASS(child != NULL && parent != NULL);
2018         if (witness_cold == 0)
2019                 mtx_assert(&w_mtx, MA_OWNED);
2020
2021         if (!witness_lock_type_equal(parent, child)) {
2022                 if (witness_cold == 0) {
2023                         unlocked = 1;
2024                         mtx_unlock_spin(&w_mtx);
2025                 } else {
2026                         unlocked = 0;
2027                 }
2028                 kassert_panic(
2029                     "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2030                     "the same lock type", __func__, parent->w_name,
2031                     parent->w_class->lc_name, child->w_name,
2032                     child->w_class->lc_name);
2033                 if (unlocked)
2034                         mtx_lock_spin(&w_mtx);
2035         }
2036         adopt(parent, child);
2037 }
2038
2039 /*
2040  * Generic code for the isitmy*() functions. The rmask parameter is the
2041  * expected relationship of w1 to w2.
2042  */
2043 static int
2044 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2045 {
2046         unsigned char r1, r2;
2047         int i1, i2;
2048
2049         i1 = w1->w_index;
2050         i2 = w2->w_index;
2051         WITNESS_INDEX_ASSERT(i1);
2052         WITNESS_INDEX_ASSERT(i2);
2053         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2054         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2055
2056         /* The flags on one better be the inverse of the flags on the other */
2057         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2058             (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2059                 /* Don't squawk if we're potentially racing with an update. */
2060                 if (!mtx_owned(&w_mtx))
2061                         return (0);
2062                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2063                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
2064                     "w_rmatrix[%d][%d] == %hhx\n",
2065                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2066                     i2, i1, r2);
2067                 kdb_backtrace();
2068                 printf("Witness disabled.\n");
2069                 witness_watch = -1;
2070         }
2071         return (r1 & rmask);
2072 }
2073
2074 /*
2075  * Checks if @child is a direct child of @parent.
2076  */
2077 static int
2078 isitmychild(struct witness *parent, struct witness *child)
2079 {
2080
2081         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2082 }
2083
2084 /*
2085  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2086  */
2087 static int
2088 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2089 {
2090
2091         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2092             __func__));
2093 }
2094
2095 static int
2096 blessed(struct witness *w1, struct witness *w2)
2097 {
2098         int i;
2099         struct witness_blessed *b;
2100
2101         for (i = 0; i < nitems(blessed_list); i++) {
2102                 b = &blessed_list[i];
2103                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2104                         if (strcmp(w2->w_name, b->b_lock2) == 0)
2105                                 return (1);
2106                         continue;
2107                 }
2108                 if (strcmp(w1->w_name, b->b_lock2) == 0)
2109                         if (strcmp(w2->w_name, b->b_lock1) == 0)
2110                                 return (1);
2111         }
2112         return (0);
2113 }
2114
2115 static struct witness *
2116 witness_get(void)
2117 {
2118         struct witness *w;
2119         int index;
2120
2121         if (witness_cold == 0)
2122                 mtx_assert(&w_mtx, MA_OWNED);
2123
2124         if (witness_watch == -1) {
2125                 mtx_unlock_spin(&w_mtx);
2126                 return (NULL);
2127         }
2128         if (STAILQ_EMPTY(&w_free)) {
2129                 witness_watch = -1;
2130                 mtx_unlock_spin(&w_mtx);
2131                 printf("WITNESS: unable to allocate a new witness object\n");
2132                 return (NULL);
2133         }
2134         w = STAILQ_FIRST(&w_free);
2135         STAILQ_REMOVE_HEAD(&w_free, w_list);
2136         w_free_cnt--;
2137         index = w->w_index;
2138         MPASS(index > 0 && index == w_max_used_index+1 &&
2139             index < witness_count);
2140         bzero(w, sizeof(*w));
2141         w->w_index = index;
2142         if (index > w_max_used_index)
2143                 w_max_used_index = index;
2144         return (w);
2145 }
2146
2147 static void
2148 witness_free(struct witness *w)
2149 {
2150
2151         STAILQ_INSERT_HEAD(&w_free, w, w_list);
2152         w_free_cnt++;
2153 }
2154
2155 static struct lock_list_entry *
2156 witness_lock_list_get(void)
2157 {
2158         struct lock_list_entry *lle;
2159
2160         if (witness_watch == -1)
2161                 return (NULL);
2162         mtx_lock_spin(&w_mtx);
2163         lle = w_lock_list_free;
2164         if (lle == NULL) {
2165                 witness_watch = -1;
2166                 mtx_unlock_spin(&w_mtx);
2167                 printf("%s: witness exhausted\n", __func__);
2168                 return (NULL);
2169         }
2170         w_lock_list_free = lle->ll_next;
2171         mtx_unlock_spin(&w_mtx);
2172         bzero(lle, sizeof(*lle));
2173         return (lle);
2174 }
2175
2176 static void
2177 witness_lock_list_free(struct lock_list_entry *lle)
2178 {
2179
2180         mtx_lock_spin(&w_mtx);
2181         lle->ll_next = w_lock_list_free;
2182         w_lock_list_free = lle;
2183         mtx_unlock_spin(&w_mtx);
2184 }
2185
2186 static struct lock_instance *
2187 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2188 {
2189         struct lock_list_entry *lle;
2190         struct lock_instance *instance;
2191         int i;
2192
2193         for (lle = list; lle != NULL; lle = lle->ll_next)
2194                 for (i = lle->ll_count - 1; i >= 0; i--) {
2195                         instance = &lle->ll_children[i];
2196                         if (instance->li_lock == lock)
2197                                 return (instance);
2198                 }
2199         return (NULL);
2200 }
2201
2202 static void
2203 witness_list_lock(struct lock_instance *instance,
2204     int (*prnt)(const char *fmt, ...))
2205 {
2206         struct lock_object *lock;
2207
2208         lock = instance->li_lock;
2209         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2210             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2211         if (lock->lo_witness->w_name != lock->lo_name)
2212                 prnt(" (%s)", lock->lo_witness->w_name);
2213         prnt(" r = %d (%p) locked @ %s:%d\n",
2214             instance->li_flags & LI_RECURSEMASK, lock,
2215             fixup_filename(instance->li_file), instance->li_line);
2216 }
2217
2218 static int
2219 witness_output(const char *fmt, ...)
2220 {
2221         va_list ap;
2222         int ret;
2223
2224         va_start(ap, fmt);
2225         ret = witness_voutput(fmt, ap);
2226         va_end(ap);
2227         return (ret);
2228 }
2229
2230 static int
2231 witness_voutput(const char *fmt, va_list ap)
2232 {
2233         int ret;
2234
2235         ret = 0;
2236         switch (witness_channel) {
2237         case WITNESS_CONSOLE:
2238                 ret = vprintf(fmt, ap);
2239                 break;
2240         case WITNESS_LOG:
2241                 vlog(LOG_NOTICE, fmt, ap);
2242                 break;
2243         case WITNESS_NONE:
2244                 break;
2245         }
2246         return (ret);
2247 }
2248
2249 #ifdef DDB
2250 static int
2251 witness_thread_has_locks(struct thread *td)
2252 {
2253
2254         if (td->td_sleeplocks == NULL)
2255                 return (0);
2256         return (td->td_sleeplocks->ll_count != 0);
2257 }
2258
2259 static int
2260 witness_proc_has_locks(struct proc *p)
2261 {
2262         struct thread *td;
2263
2264         FOREACH_THREAD_IN_PROC(p, td) {
2265                 if (witness_thread_has_locks(td))
2266                         return (1);
2267         }
2268         return (0);
2269 }
2270 #endif
2271
2272 int
2273 witness_list_locks(struct lock_list_entry **lock_list,
2274     int (*prnt)(const char *fmt, ...))
2275 {
2276         struct lock_list_entry *lle;
2277         int i, nheld;
2278
2279         nheld = 0;
2280         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2281                 for (i = lle->ll_count - 1; i >= 0; i--) {
2282                         witness_list_lock(&lle->ll_children[i], prnt);
2283                         nheld++;
2284                 }
2285         return (nheld);
2286 }
2287
2288 /*
2289  * This is a bit risky at best.  We call this function when we have timed
2290  * out acquiring a spin lock, and we assume that the other CPU is stuck
2291  * with this lock held.  So, we go groveling around in the other CPU's
2292  * per-cpu data to try to find the lock instance for this spin lock to
2293  * see when it was last acquired.
2294  */
2295 void
2296 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2297     int (*prnt)(const char *fmt, ...))
2298 {
2299         struct lock_instance *instance;
2300         struct pcpu *pc;
2301
2302         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2303                 return;
2304         pc = pcpu_find(owner->td_oncpu);
2305         instance = find_instance(pc->pc_spinlocks, lock);
2306         if (instance != NULL)
2307                 witness_list_lock(instance, prnt);
2308 }
2309
2310 void
2311 witness_save(struct lock_object *lock, const char **filep, int *linep)
2312 {
2313         struct lock_list_entry *lock_list;
2314         struct lock_instance *instance;
2315         struct lock_class *class;
2316
2317         /*
2318          * This function is used independently in locking code to deal with
2319          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2320          * is gone.
2321          */
2322         if (SCHEDULER_STOPPED())
2323                 return;
2324         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2325         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2326                 return;
2327         class = LOCK_CLASS(lock);
2328         if (class->lc_flags & LC_SLEEPLOCK)
2329                 lock_list = curthread->td_sleeplocks;
2330         else {
2331                 if (witness_skipspin)
2332                         return;
2333                 lock_list = PCPU_GET(spinlocks);
2334         }
2335         instance = find_instance(lock_list, lock);
2336         if (instance == NULL) {
2337                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2338                     class->lc_name, lock->lo_name);
2339                 return;
2340         }
2341         *filep = instance->li_file;
2342         *linep = instance->li_line;
2343 }
2344
2345 void
2346 witness_restore(struct lock_object *lock, const char *file, int line)
2347 {
2348         struct lock_list_entry *lock_list;
2349         struct lock_instance *instance;
2350         struct lock_class *class;
2351
2352         /*
2353          * This function is used independently in locking code to deal with
2354          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2355          * is gone.
2356          */
2357         if (SCHEDULER_STOPPED())
2358                 return;
2359         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2360         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2361                 return;
2362         class = LOCK_CLASS(lock);
2363         if (class->lc_flags & LC_SLEEPLOCK)
2364                 lock_list = curthread->td_sleeplocks;
2365         else {
2366                 if (witness_skipspin)
2367                         return;
2368                 lock_list = PCPU_GET(spinlocks);
2369         }
2370         instance = find_instance(lock_list, lock);
2371         if (instance == NULL)
2372                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2373                     class->lc_name, lock->lo_name);
2374         lock->lo_witness->w_file = file;
2375         lock->lo_witness->w_line = line;
2376         if (instance == NULL)
2377                 return;
2378         instance->li_file = file;
2379         instance->li_line = line;
2380 }
2381
2382 void
2383 witness_assert(const struct lock_object *lock, int flags, const char *file,
2384     int line)
2385 {
2386 #ifdef INVARIANT_SUPPORT
2387         struct lock_instance *instance;
2388         struct lock_class *class;
2389
2390         if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2391                 return;
2392         class = LOCK_CLASS(lock);
2393         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2394                 instance = find_instance(curthread->td_sleeplocks, lock);
2395         else if ((class->lc_flags & LC_SPINLOCK) != 0)
2396                 instance = find_instance(PCPU_GET(spinlocks), lock);
2397         else {
2398                 kassert_panic("Lock (%s) %s is not sleep or spin!",
2399                     class->lc_name, lock->lo_name);
2400                 return;
2401         }
2402         switch (flags) {
2403         case LA_UNLOCKED:
2404                 if (instance != NULL)
2405                         kassert_panic("Lock (%s) %s locked @ %s:%d.",
2406                             class->lc_name, lock->lo_name,
2407                             fixup_filename(file), line);
2408                 break;
2409         case LA_LOCKED:
2410         case LA_LOCKED | LA_RECURSED:
2411         case LA_LOCKED | LA_NOTRECURSED:
2412         case LA_SLOCKED:
2413         case LA_SLOCKED | LA_RECURSED:
2414         case LA_SLOCKED | LA_NOTRECURSED:
2415         case LA_XLOCKED:
2416         case LA_XLOCKED | LA_RECURSED:
2417         case LA_XLOCKED | LA_NOTRECURSED:
2418                 if (instance == NULL) {
2419                         kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2420                             class->lc_name, lock->lo_name,
2421                             fixup_filename(file), line);
2422                         break;
2423                 }
2424                 if ((flags & LA_XLOCKED) != 0 &&
2425                     (instance->li_flags & LI_EXCLUSIVE) == 0)
2426                         kassert_panic(
2427                             "Lock (%s) %s not exclusively locked @ %s:%d.",
2428                             class->lc_name, lock->lo_name,
2429                             fixup_filename(file), line);
2430                 if ((flags & LA_SLOCKED) != 0 &&
2431                     (instance->li_flags & LI_EXCLUSIVE) != 0)
2432                         kassert_panic(
2433                             "Lock (%s) %s exclusively locked @ %s:%d.",
2434                             class->lc_name, lock->lo_name,
2435                             fixup_filename(file), line);
2436                 if ((flags & LA_RECURSED) != 0 &&
2437                     (instance->li_flags & LI_RECURSEMASK) == 0)
2438                         kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2439                             class->lc_name, lock->lo_name,
2440                             fixup_filename(file), line);
2441                 if ((flags & LA_NOTRECURSED) != 0 &&
2442                     (instance->li_flags & LI_RECURSEMASK) != 0)
2443                         kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2444                             class->lc_name, lock->lo_name,
2445                             fixup_filename(file), line);
2446                 break;
2447         default:
2448                 kassert_panic("Invalid lock assertion at %s:%d.",
2449                     fixup_filename(file), line);
2450
2451         }
2452 #endif  /* INVARIANT_SUPPORT */
2453 }
2454
2455 static void
2456 witness_setflag(struct lock_object *lock, int flag, int set)
2457 {
2458         struct lock_list_entry *lock_list;
2459         struct lock_instance *instance;
2460         struct lock_class *class;
2461
2462         if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2463                 return;
2464         class = LOCK_CLASS(lock);
2465         if (class->lc_flags & LC_SLEEPLOCK)
2466                 lock_list = curthread->td_sleeplocks;
2467         else {
2468                 if (witness_skipspin)
2469                         return;
2470                 lock_list = PCPU_GET(spinlocks);
2471         }
2472         instance = find_instance(lock_list, lock);
2473         if (instance == NULL) {
2474                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2475                     class->lc_name, lock->lo_name);
2476                 return;
2477         }
2478
2479         if (set)
2480                 instance->li_flags |= flag;
2481         else
2482                 instance->li_flags &= ~flag;
2483 }
2484
2485 void
2486 witness_norelease(struct lock_object *lock)
2487 {
2488
2489         witness_setflag(lock, LI_NORELEASE, 1);
2490 }
2491
2492 void
2493 witness_releaseok(struct lock_object *lock)
2494 {
2495
2496         witness_setflag(lock, LI_NORELEASE, 0);
2497 }
2498
2499 #ifdef DDB
2500 static void
2501 witness_ddb_list(struct thread *td)
2502 {
2503
2504         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2505         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2506
2507         if (witness_watch < 1)
2508                 return;
2509
2510         witness_list_locks(&td->td_sleeplocks, db_printf);
2511
2512         /*
2513          * We only handle spinlocks if td == curthread.  This is somewhat broken
2514          * if td is currently executing on some other CPU and holds spin locks
2515          * as we won't display those locks.  If we had a MI way of getting
2516          * the per-cpu data for a given cpu then we could use
2517          * td->td_oncpu to get the list of spinlocks for this thread
2518          * and "fix" this.
2519          *
2520          * That still wouldn't really fix this unless we locked the scheduler
2521          * lock or stopped the other CPU to make sure it wasn't changing the
2522          * list out from under us.  It is probably best to just not try to
2523          * handle threads on other CPU's for now.
2524          */
2525         if (td == curthread && PCPU_GET(spinlocks) != NULL)
2526                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2527 }
2528
2529 DB_SHOW_COMMAND(locks, db_witness_list)
2530 {
2531         struct thread *td;
2532
2533         if (have_addr)
2534                 td = db_lookup_thread(addr, true);
2535         else
2536                 td = kdb_thread;
2537         witness_ddb_list(td);
2538 }
2539
2540 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2541 {
2542         struct thread *td;
2543         struct proc *p;
2544
2545         /*
2546          * It would be nice to list only threads and processes that actually
2547          * held sleep locks, but that information is currently not exported
2548          * by WITNESS.
2549          */
2550         FOREACH_PROC_IN_SYSTEM(p) {
2551                 if (!witness_proc_has_locks(p))
2552                         continue;
2553                 FOREACH_THREAD_IN_PROC(p, td) {
2554                         if (!witness_thread_has_locks(td))
2555                                 continue;
2556                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2557                             p->p_comm, td, td->td_tid);
2558                         witness_ddb_list(td);
2559                         if (db_pager_quit)
2560                                 return;
2561                 }
2562         }
2563 }
2564 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2565
2566 DB_SHOW_COMMAND(witness, db_witness_display)
2567 {
2568
2569         witness_ddb_display(db_printf);
2570 }
2571 #endif
2572
2573 static void
2574 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2575 {
2576         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2577         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2578         int generation, i, j;
2579
2580         tmp_data1 = NULL;
2581         tmp_data2 = NULL;
2582         tmp_w1 = NULL;
2583         tmp_w2 = NULL;
2584
2585         /* Allocate and init temporary storage space. */
2586         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2587         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2588         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2589             M_WAITOK | M_ZERO);
2590         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2591             M_WAITOK | M_ZERO);
2592         stack_zero(&tmp_data1->wlod_stack);
2593         stack_zero(&tmp_data2->wlod_stack);
2594
2595 restart:
2596         mtx_lock_spin(&w_mtx);
2597         generation = w_generation;
2598         mtx_unlock_spin(&w_mtx);
2599         sbuf_printf(sb, "Number of known direct relationships is %d\n",
2600             w_lohash.wloh_count);
2601         for (i = 1; i < w_max_used_index; i++) {
2602                 mtx_lock_spin(&w_mtx);
2603                 if (generation != w_generation) {
2604                         mtx_unlock_spin(&w_mtx);
2605
2606                         /* The graph has changed, try again. */
2607                         *oldidx = 0;
2608                         sbuf_clear(sb);
2609                         goto restart;
2610                 }
2611
2612                 w1 = &w_data[i];
2613                 if (w1->w_reversed == 0) {
2614                         mtx_unlock_spin(&w_mtx);
2615                         continue;
2616                 }
2617
2618                 /* Copy w1 locally so we can release the spin lock. */
2619                 *tmp_w1 = *w1;
2620                 mtx_unlock_spin(&w_mtx);
2621
2622                 if (tmp_w1->w_reversed == 0)
2623                         continue;
2624                 for (j = 1; j < w_max_used_index; j++) {
2625                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2626                                 continue;
2627
2628                         mtx_lock_spin(&w_mtx);
2629                         if (generation != w_generation) {
2630                                 mtx_unlock_spin(&w_mtx);
2631
2632                                 /* The graph has changed, try again. */
2633                                 *oldidx = 0;
2634                                 sbuf_clear(sb);
2635                                 goto restart;
2636                         }
2637
2638                         w2 = &w_data[j];
2639                         data1 = witness_lock_order_get(w1, w2);
2640                         data2 = witness_lock_order_get(w2, w1);
2641
2642                         /*
2643                          * Copy information locally so we can release the
2644                          * spin lock.
2645                          */
2646                         *tmp_w2 = *w2;
2647
2648                         if (data1) {
2649                                 stack_zero(&tmp_data1->wlod_stack);
2650                                 stack_copy(&data1->wlod_stack,
2651                                     &tmp_data1->wlod_stack);
2652                         }
2653                         if (data2 && data2 != data1) {
2654                                 stack_zero(&tmp_data2->wlod_stack);
2655                                 stack_copy(&data2->wlod_stack,
2656                                     &tmp_data2->wlod_stack);
2657                         }
2658                         mtx_unlock_spin(&w_mtx);
2659
2660                         if (blessed(tmp_w1, tmp_w2))
2661                                 continue;
2662
2663                         sbuf_printf(sb,
2664             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2665                             tmp_w1->w_name, tmp_w1->w_class->lc_name,
2666                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
2667                         if (data1) {
2668                                 sbuf_printf(sb,
2669                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2670                                     tmp_w1->w_name, tmp_w1->w_class->lc_name,
2671                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
2672                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2673                                 sbuf_printf(sb, "\n");
2674                         }
2675                         if (data2 && data2 != data1) {
2676                                 sbuf_printf(sb,
2677                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2678                                     tmp_w2->w_name, tmp_w2->w_class->lc_name,
2679                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
2680                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2681                                 sbuf_printf(sb, "\n");
2682                         }
2683                 }
2684         }
2685         mtx_lock_spin(&w_mtx);
2686         if (generation != w_generation) {
2687                 mtx_unlock_spin(&w_mtx);
2688
2689                 /*
2690                  * The graph changed while we were printing stack data,
2691                  * try again.
2692                  */
2693                 *oldidx = 0;
2694                 sbuf_clear(sb);
2695                 goto restart;
2696         }
2697         mtx_unlock_spin(&w_mtx);
2698
2699         /* Free temporary storage space. */
2700         free(tmp_data1, M_TEMP);
2701         free(tmp_data2, M_TEMP);
2702         free(tmp_w1, M_TEMP);
2703         free(tmp_w2, M_TEMP);
2704 }
2705
2706 static int
2707 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2708 {
2709         struct sbuf *sb;
2710         int error;
2711
2712         if (witness_watch < 1) {
2713                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2714                 return (error);
2715         }
2716         if (witness_cold) {
2717                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2718                 return (error);
2719         }
2720         error = 0;
2721         sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2722         if (sb == NULL)
2723                 return (ENOMEM);
2724
2725         sbuf_print_witness_badstacks(sb, &req->oldidx);
2726
2727         sbuf_finish(sb);
2728         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2729         sbuf_delete(sb);
2730
2731         return (error);
2732 }
2733
2734 #ifdef DDB
2735 static int
2736 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2737 {
2738
2739         return (db_printf("%.*s", len, data));
2740 }
2741
2742 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2743 {
2744         struct sbuf sb;
2745         char buffer[128];
2746         size_t dummy;
2747
2748         sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2749         sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2750         sbuf_print_witness_badstacks(&sb, &dummy);
2751         sbuf_finish(&sb);
2752 }
2753 #endif
2754
2755 static int
2756 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2757 {
2758         static const struct {
2759                 enum witness_channel channel;
2760                 const char *name;
2761         } channels[] = {
2762                 { WITNESS_CONSOLE, "console" },
2763                 { WITNESS_LOG, "log" },
2764                 { WITNESS_NONE, "none" },
2765         };
2766         char buf[16];
2767         u_int i;
2768         int error;
2769
2770         buf[0] = '\0';
2771         for (i = 0; i < nitems(channels); i++)
2772                 if (witness_channel == channels[i].channel) {
2773                         snprintf(buf, sizeof(buf), "%s", channels[i].name);
2774                         break;
2775                 }
2776
2777         error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2778         if (error != 0 || req->newptr == NULL)
2779                 return (error);
2780
2781         error = EINVAL;
2782         for (i = 0; i < nitems(channels); i++)
2783                 if (strcmp(channels[i].name, buf) == 0) {
2784                         witness_channel = channels[i].channel;
2785                         error = 0;
2786                         break;
2787                 }
2788         return (error);
2789 }
2790
2791 static int
2792 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2793 {
2794         struct witness *w;
2795         struct sbuf *sb;
2796         int error;
2797
2798 #ifdef __i386__
2799         error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2800         return (error);
2801 #endif
2802
2803         if (witness_watch < 1) {
2804                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2805                 return (error);
2806         }
2807         if (witness_cold) {
2808                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2809                 return (error);
2810         }
2811         error = 0;
2812
2813         error = sysctl_wire_old_buffer(req, 0);
2814         if (error != 0)
2815                 return (error);
2816         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2817         if (sb == NULL)
2818                 return (ENOMEM);
2819         sbuf_printf(sb, "\n");
2820
2821         mtx_lock_spin(&w_mtx);
2822         STAILQ_FOREACH(w, &w_all, w_list)
2823                 w->w_displayed = 0;
2824         STAILQ_FOREACH(w, &w_all, w_list)
2825                 witness_add_fullgraph(sb, w);
2826         mtx_unlock_spin(&w_mtx);
2827
2828         /*
2829          * Close the sbuf and return to userland.
2830          */
2831         error = sbuf_finish(sb);
2832         sbuf_delete(sb);
2833
2834         return (error);
2835 }
2836
2837 static int
2838 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2839 {
2840         int error, value;
2841
2842         value = witness_watch;
2843         error = sysctl_handle_int(oidp, &value, 0, req);
2844         if (error != 0 || req->newptr == NULL)
2845                 return (error);
2846         if (value > 1 || value < -1 ||
2847             (witness_watch == -1 && value != witness_watch))
2848                 return (EINVAL);
2849         witness_watch = value;
2850         return (0);
2851 }
2852
2853 static void
2854 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2855 {
2856         int i;
2857
2858         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2859                 return;
2860         w->w_displayed = 1;
2861
2862         WITNESS_INDEX_ASSERT(w->w_index);
2863         for (i = 1; i <= w_max_used_index; i++) {
2864                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2865                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2866                             w_data[i].w_name);
2867                         witness_add_fullgraph(sb, &w_data[i]);
2868                 }
2869         }
2870 }
2871
2872 /*
2873  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2874  * interprets the key as a string and reads until the null
2875  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2876  * hash value computed from the key.
2877  */
2878 static uint32_t
2879 witness_hash_djb2(const uint8_t *key, uint32_t size)
2880 {
2881         unsigned int hash = 5381;
2882         int i;
2883
2884         /* hash = hash * 33 + key[i] */
2885         if (size)
2886                 for (i = 0; i < size; i++)
2887                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2888         else
2889                 for (i = 0; key[i] != 0; i++)
2890                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2891
2892         return (hash);
2893 }
2894
2895 /*
2896  * Initializes the two witness hash tables. Called exactly once from
2897  * witness_initialize().
2898  */
2899 static void
2900 witness_init_hash_tables(void)
2901 {
2902         int i;
2903
2904         MPASS(witness_cold);
2905
2906         /* Initialize the hash tables. */
2907         for (i = 0; i < WITNESS_HASH_SIZE; i++)
2908                 w_hash.wh_array[i] = NULL;
2909
2910         w_hash.wh_size = WITNESS_HASH_SIZE;
2911         w_hash.wh_count = 0;
2912
2913         /* Initialize the lock order data hash. */
2914         w_lofree = NULL;
2915         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2916                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2917                 w_lodata[i].wlod_next = w_lofree;
2918                 w_lofree = &w_lodata[i];
2919         }
2920         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2921         w_lohash.wloh_count = 0;
2922         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2923                 w_lohash.wloh_array[i] = NULL;
2924 }
2925
2926 static struct witness *
2927 witness_hash_get(const char *key)
2928 {
2929         struct witness *w;
2930         uint32_t hash;
2931
2932         MPASS(key != NULL);
2933         if (witness_cold == 0)
2934                 mtx_assert(&w_mtx, MA_OWNED);
2935         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2936         w = w_hash.wh_array[hash];
2937         while (w != NULL) {
2938                 if (strcmp(w->w_name, key) == 0)
2939                         goto out;
2940                 w = w->w_hash_next;
2941         }
2942
2943 out:
2944         return (w);
2945 }
2946
2947 static void
2948 witness_hash_put(struct witness *w)
2949 {
2950         uint32_t hash;
2951
2952         MPASS(w != NULL);
2953         MPASS(w->w_name != NULL);
2954         if (witness_cold == 0)
2955                 mtx_assert(&w_mtx, MA_OWNED);
2956         KASSERT(witness_hash_get(w->w_name) == NULL,
2957             ("%s: trying to add a hash entry that already exists!", __func__));
2958         KASSERT(w->w_hash_next == NULL,
2959             ("%s: w->w_hash_next != NULL", __func__));
2960
2961         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2962         w->w_hash_next = w_hash.wh_array[hash];
2963         w_hash.wh_array[hash] = w;
2964         w_hash.wh_count++;
2965 }
2966
2967 static struct witness_lock_order_data *
2968 witness_lock_order_get(struct witness *parent, struct witness *child)
2969 {
2970         struct witness_lock_order_data *data = NULL;
2971         struct witness_lock_order_key key;
2972         unsigned int hash;
2973
2974         MPASS(parent != NULL && child != NULL);
2975         key.from = parent->w_index;
2976         key.to = child->w_index;
2977         WITNESS_INDEX_ASSERT(key.from);
2978         WITNESS_INDEX_ASSERT(key.to);
2979         if ((w_rmatrix[parent->w_index][child->w_index]
2980             & WITNESS_LOCK_ORDER_KNOWN) == 0)
2981                 goto out;
2982
2983         hash = witness_hash_djb2((const char*)&key,
2984             sizeof(key)) % w_lohash.wloh_size;
2985         data = w_lohash.wloh_array[hash];
2986         while (data != NULL) {
2987                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2988                         break;
2989                 data = data->wlod_next;
2990         }
2991
2992 out:
2993         return (data);
2994 }
2995
2996 /*
2997  * Verify that parent and child have a known relationship, are not the same,
2998  * and child is actually a child of parent.  This is done without w_mtx
2999  * to avoid contention in the common case.
3000  */
3001 static int
3002 witness_lock_order_check(struct witness *parent, struct witness *child)
3003 {
3004
3005         if (parent != child &&
3006             w_rmatrix[parent->w_index][child->w_index]
3007             & WITNESS_LOCK_ORDER_KNOWN &&
3008             isitmychild(parent, child))
3009                 return (1);
3010
3011         return (0);
3012 }
3013
3014 static int
3015 witness_lock_order_add(struct witness *parent, struct witness *child)
3016 {
3017         struct witness_lock_order_data *data = NULL;
3018         struct witness_lock_order_key key;
3019         unsigned int hash;
3020
3021         MPASS(parent != NULL && child != NULL);
3022         key.from = parent->w_index;
3023         key.to = child->w_index;
3024         WITNESS_INDEX_ASSERT(key.from);
3025         WITNESS_INDEX_ASSERT(key.to);
3026         if (w_rmatrix[parent->w_index][child->w_index]
3027             & WITNESS_LOCK_ORDER_KNOWN)
3028                 return (1);
3029
3030         hash = witness_hash_djb2((const char*)&key,
3031             sizeof(key)) % w_lohash.wloh_size;
3032         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3033         data = w_lofree;
3034         if (data == NULL)
3035                 return (0);
3036         w_lofree = data->wlod_next;
3037         data->wlod_next = w_lohash.wloh_array[hash];
3038         data->wlod_key = key;
3039         w_lohash.wloh_array[hash] = data;
3040         w_lohash.wloh_count++;
3041         stack_zero(&data->wlod_stack);
3042         stack_save(&data->wlod_stack);
3043         return (1);
3044 }
3045
3046 /* Call this whenever the structure of the witness graph changes. */
3047 static void
3048 witness_increment_graph_generation(void)
3049 {
3050
3051         if (witness_cold == 0)
3052                 mtx_assert(&w_mtx, MA_OWNED);
3053         w_generation++;
3054 }
3055
3056 static int
3057 witness_output_drain(void *arg __unused, const char *data, int len)
3058 {
3059
3060         witness_output("%.*s", len, data);
3061         return (len);
3062 }
3063
3064 static void
3065 witness_debugger(int cond, const char *msg)
3066 {
3067         char buf[32];
3068         struct sbuf sb;
3069         struct stack st;
3070
3071         if (!cond)
3072                 return;
3073
3074         if (witness_trace) {
3075                 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3076                 sbuf_set_drain(&sb, witness_output_drain, NULL);
3077
3078                 stack_zero(&st);
3079                 stack_save(&st);
3080                 witness_output("stack backtrace:\n");
3081                 stack_sbuf_print_ddb(&sb, &st);
3082
3083                 sbuf_finish(&sb);
3084         }
3085
3086 #ifdef KDB
3087         if (witness_kdb)
3088                 kdb_enter(KDB_WHY_WITNESS, msg);
3089 #endif
3090 }