sys/kern/subr_witness.c

   1 /*-
   2  * Copyright (c) 2008 Isilon Systems, Inc.
   3  * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
   4  * Copyright (c) 1998 Berkeley Software Design, Inc.
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Berkeley Software Design Inc's name may not be used to endorse or
  16  *    promote products derived from this software without specific prior
  17  *    written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
  20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  22  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
  23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  29  * SUCH DAMAGE.
  30  *
  31  *      from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
  32  *      and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
  33  */
  34
  35 /*
  36  * Implementation of the `witness' lock verifier.  Originally implemented for
  37  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
  38  * classes in FreeBSD.
  39  */
  40
  41 /*
  42  *      Main Entry: witness
  43  *      Pronunciation: 'wit-n&s
  44  *      Function: noun
  45  *      Etymology: Middle English witnesse, from Old English witnes knowledge,
  46  *          testimony, witness, from 2wit
  47  *      Date: before 12th century
  48  *      1 : attestation of a fact or event : TESTIMONY
  49  *      2 : one that gives evidence; specifically : one who testifies in
  50  *          a cause or before a judicial tribunal
  51  *      3 : one asked to be present at a transaction so as to be able to
  52  *          testify to its having taken place
  53  *      4 : one who has personal knowledge of something
  54  *      5 a : something serving as evidence or proof : SIGN
  55  *        b : public affirmation by word or example of usually
  56  *            religious faith or conviction <the heroic witness to divine
  57  *            life -- Pilot>
  58  *      6 capitalized : a member of the Jehovah's Witnesses
  59  */
  60
  61 /*
  62  * Special rules concerning Giant and lock orders:
  63  *
  64  * 1) Giant must be acquired before any other mutexes.  Stated another way,
  65  *    no other mutex may be held when Giant is acquired.
  66  *
  67  * 2) Giant must be released when blocking on a sleepable lock.
  68  *
  69  * This rule is less obvious, but is a result of Giant providing the same
  70  * semantics as spl().  Basically, when a thread sleeps, it must release
  71  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
  72  * 2).
  73  *
  74  * 3) Giant may be acquired before or after sleepable locks.
  75  *
  76  * This rule is also not quite as obvious.  Giant may be acquired after
  77  * a sleepable lock because it is a non-sleepable lock and non-sleepable
  78  * locks may always be acquired while holding a sleepable lock.  The second
  79  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
  80  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
  81  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
  82  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
  83  * execute.  Thus, acquiring Giant both before and after a sleepable lock
  84  * will not result in a lock order reversal.
  85  */
  86
  87 #include <sys/cdefs.h>
  88 __FBSDID("$FreeBSD$");
  89
  90 #include "opt_ddb.h"
  91 #include "opt_hwpmc_hooks.h"
  92 #include "opt_stack.h"
  93 #include "opt_witness.h"
  94
  95 #include <sys/param.h>
  96 #include <sys/bus.h>
  97 #include <sys/kdb.h>
  98 #include <sys/kernel.h>
  99 #include <sys/ktr.h>
 100 #include <sys/lock.h>
 101 #include <sys/malloc.h>
 102 #include <sys/mutex.h>
 103 #include <sys/priv.h>
 104 #include <sys/proc.h>
 105 #include <sys/sbuf.h>
 106 #include <sys/sched.h>
 107 #include <sys/stack.h>
 108 #include <sys/sysctl.h>
 109 #include <sys/syslog.h>
 110 #include <sys/systm.h>
 111
 112 #ifdef DDB
 113 #include <ddb/ddb.h>
 114 #endif
 115
 116 #include <machine/stdarg.h>
 117
 118 #if !defined(DDB) && !defined(STACK)
 119 #error "DDB or STACK options are required for WITNESS"
 120 #endif
 121
 122 /* Note that these traces do not work with KTR_ALQ. */
 123 #if 0
 124 #define KTR_WITNESS     KTR_SUBSYS
 125 #else
 126 #define KTR_WITNESS     0
 127 #endif
 128
 129 #define LI_RECURSEMASK  0x0000ffff      /* Recursion depth of lock instance. */
 130 #define LI_EXCLUSIVE    0x00010000      /* Exclusive lock instance. */
 131 #define LI_NORELEASE    0x00020000      /* Lock not allowed to be released. */
 132
 133 /* Define this to check for blessed mutexes */
 134 #undef BLESSING
 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        (1024 + MAXCPU)
 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 #ifdef BLESSING
 280 struct witness_blessed {
 281         const char      *b_lock1;
 282         const char      *b_lock2;
 283 };
 284 #endif
 285
 286 struct witness_pendhelp {
 287         const char              *wh_type;
 288         struct lock_object      *wh_lock;
 289 };
 290
 291 struct witness_order_list_entry {
 292         const char              *w_name;
 293         struct lock_class       *w_class;
 294 };
 295
 296 /*
 297  * Returns 0 if one of the locks is a spin lock and the other is not.
 298  * Returns 1 otherwise.
 299  */
 300 static __inline int
 301 witness_lock_type_equal(struct witness *w1, struct witness *w2)
 302 {
 303
 304         return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
 305                 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
 306 }
 307
 308 static __inline int
 309 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
 310     const struct witness_lock_order_key *b)
 311 {
 312
 313         return (a->from == b->from && a->to == b->to);
 314 }
 315
 316 static int      _isitmyx(struct witness *w1, struct witness *w2, int rmask,
 317                     const char *fname);
 318 static void     adopt(struct witness *parent, struct witness *child);
 319 #ifdef BLESSING
 320 static int      blessed(struct witness *, struct witness *);
 321 #endif
 322 static void     depart(struct witness *w);
 323 static struct witness   *enroll(const char *description,
 324                             struct lock_class *lock_class);
 325 static struct lock_instance     *find_instance(struct lock_list_entry *list,
 326                                     const struct lock_object *lock);
 327 static int      isitmychild(struct witness *parent, struct witness *child);
 328 static int      isitmydescendant(struct witness *parent, struct witness *child);
 329 static void     itismychild(struct witness *parent, struct witness *child);
 330 static int      sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
 331 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
 332 static int      sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
 333 static int      sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
 334 static void     witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
 335 #ifdef DDB
 336 static void     witness_ddb_compute_levels(void);
 337 static void     witness_ddb_display(int(*)(const char *fmt, ...));
 338 static void     witness_ddb_display_descendants(int(*)(const char *fmt, ...),
 339                     struct witness *, int indent);
 340 static void     witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
 341                     struct witness_list *list);
 342 static void     witness_ddb_level_descendants(struct witness *parent, int l);
 343 static void     witness_ddb_list(struct thread *td);
 344 #endif
 345 static void     witness_debugger(int cond, const char *msg);
 346 static void     witness_free(struct witness *m);
 347 static struct witness   *witness_get(void);
 348 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
 349 static struct witness   *witness_hash_get(const char *key);
 350 static void     witness_hash_put(struct witness *w);
 351 static void     witness_init_hash_tables(void);
 352 static void     witness_increment_graph_generation(void);
 353 static void     witness_lock_list_free(struct lock_list_entry *lle);
 354 static struct lock_list_entry   *witness_lock_list_get(void);
 355 static int      witness_lock_order_add(struct witness *parent,
 356                     struct witness *child);
 357 static int      witness_lock_order_check(struct witness *parent,
 358                     struct witness *child);
 359 static struct witness_lock_order_data   *witness_lock_order_get(
 360                                             struct witness *parent,
 361                                             struct witness *child);
 362 static void     witness_list_lock(struct lock_instance *instance,
 363                     int (*prnt)(const char *fmt, ...));
 364 static int      witness_output(const char *fmt, ...) __printflike(1, 2);
 365 static int      witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
 366 static void     witness_setflag(struct lock_object *lock, int flag, int set);
 367
 368 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
 369     "Witness Locking");
 370
 371 /*
 372  * If set to 0, lock order checking is disabled.  If set to -1,
 373  * witness is completely disabled.  Otherwise witness performs full
 374  * lock order checking for all locks.  At runtime, lock order checking
 375  * may be toggled.  However, witness cannot be reenabled once it is
 376  * completely disabled.
 377  */
 378 static int witness_watch = 1;
 379 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
 380     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
 381
 382 #ifdef KDB
 383 /*
 384  * When KDB is enabled and witness_kdb is 1, it will cause the system
 385  * to drop into kdebug() when:
 386  *      - a lock hierarchy violation occurs
 387  *      - locks are held when going to sleep.
 388  */
 389 #ifdef WITNESS_KDB
 390 int     witness_kdb = 1;
 391 #else
 392 int     witness_kdb = 0;
 393 #endif
 394 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
 395 #endif /* KDB */
 396
 397 #if defined(DDB) || defined(KDB)
 398 /*
 399  * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
 400  * to print a stack trace:
 401  *      - a lock hierarchy violation occurs
 402  *      - locks are held when going to sleep.
 403  */
 404 int     witness_trace = 1;
 405 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
 406 #endif /* DDB || KDB */
 407
 408 #ifdef WITNESS_SKIPSPIN
 409 int     witness_skipspin = 1;
 410 #else
 411 int     witness_skipspin = 0;
 412 #endif
 413 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
 414
 415 int badstack_sbuf_size;
 416
 417 int witness_count = WITNESS_COUNT;
 418 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
 419     &witness_count, 0, "");
 420
 421 /*
 422  * Output channel for witness messages.  By default we print to the console.
 423  */
 424 enum witness_channel {
 425         WITNESS_CONSOLE,
 426         WITNESS_LOG,
 427         WITNESS_NONE,
 428 };
 429
 430 static enum witness_channel witness_channel = WITNESS_CONSOLE;
 431 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
 432     CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
 433     "Output channel for warnings");
 434
 435 /*
 436  * Call this to print out the relations between locks.
 437  */
 438 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
 439     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
 440
 441 /*
 442  * Call this to print out the witness faulty stacks.
 443  */
 444 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
 445     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
 446
 447 static struct mtx w_mtx;
 448
 449 /* w_list */
 450 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
 451 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
 452
 453 /* w_typelist */
 454 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
 455 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
 456
 457 /* lock list */
 458 static struct lock_list_entry *w_lock_list_free = NULL;
 459 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
 460 static u_int pending_cnt;
 461
 462 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
 463 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
 464 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
 465 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
 466     "");
 467
 468 static struct witness *w_data;
 469 static uint8_t **w_rmatrix;
 470 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
 471 static struct witness_hash w_hash;      /* The witness hash table. */
 472
 473 /* The lock order data hash */
 474 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
 475 static struct witness_lock_order_data *w_lofree = NULL;
 476 static struct witness_lock_order_hash w_lohash;
 477 static int w_max_used_index = 0;
 478 static unsigned int w_generation = 0;
 479 static const char w_notrunning[] = "Witness not running\n";
 480 static const char w_stillcold[] = "Witness is still cold\n";
 481
 482
 483 static struct witness_order_list_entry order_lists[] = {
 484         /*
 485          * sx locks
 486          */
 487         { "proctree", &lock_class_sx },
 488         { "allproc", &lock_class_sx },
 489         { "allprison", &lock_class_sx },
 490         { NULL, NULL },
 491         /*
 492          * Various mutexes
 493          */
 494         { "Giant", &lock_class_mtx_sleep },
 495         { "pipe mutex", &lock_class_mtx_sleep },
 496         { "sigio lock", &lock_class_mtx_sleep },
 497         { "process group", &lock_class_mtx_sleep },
 498         { "process lock", &lock_class_mtx_sleep },
 499         { "session", &lock_class_mtx_sleep },
 500         { "uidinfo hash", &lock_class_rw },
 501 #ifdef  HWPMC_HOOKS
 502         { "pmc-sleep", &lock_class_mtx_sleep },
 503 #endif
 504         { "time lock", &lock_class_mtx_sleep },
 505         { NULL, NULL },
 506         /*
 507          * umtx
 508          */
 509         { "umtx lock", &lock_class_mtx_sleep },
 510         { NULL, NULL },
 511         /*
 512          * Sockets
 513          */
 514         { "accept", &lock_class_mtx_sleep },
 515         { "so_snd", &lock_class_mtx_sleep },
 516         { "so_rcv", &lock_class_mtx_sleep },
 517         { "sellck", &lock_class_mtx_sleep },
 518         { NULL, NULL },
 519         /*
 520          * Routing
 521          */
 522         { "so_rcv", &lock_class_mtx_sleep },
 523         { "radix node head", &lock_class_rw },
 524         { "rtentry", &lock_class_mtx_sleep },
 525         { "ifaddr", &lock_class_mtx_sleep },
 526         { NULL, NULL },
 527         /*
 528          * IPv4 multicast:
 529          * protocol locks before interface locks, after UDP locks.
 530          */
 531         { "udpinp", &lock_class_rw },
 532         { "in_multi_mtx", &lock_class_mtx_sleep },
 533         { "igmp_mtx", &lock_class_mtx_sleep },
 534         { "if_addr_lock", &lock_class_rw },
 535         { NULL, NULL },
 536         /*
 537          * IPv6 multicast:
 538          * protocol locks before interface locks, after UDP locks.
 539          */
 540         { "udpinp", &lock_class_rw },
 541         { "in6_multi_mtx", &lock_class_mtx_sleep },
 542         { "mld_mtx", &lock_class_mtx_sleep },
 543         { "if_addr_lock", &lock_class_rw },
 544         { NULL, NULL },
 545         /*
 546          * UNIX Domain Sockets
 547          */
 548         { "unp_link_rwlock", &lock_class_rw },
 549         { "unp_list_lock", &lock_class_mtx_sleep },
 550         { "unp", &lock_class_mtx_sleep },
 551         { "so_snd", &lock_class_mtx_sleep },
 552         { NULL, NULL },
 553         /*
 554          * UDP/IP
 555          */
 556         { "udp", &lock_class_rw },
 557         { "udpinp", &lock_class_rw },
 558         { "so_snd", &lock_class_mtx_sleep },
 559         { NULL, NULL },
 560         /*
 561          * TCP/IP
 562          */
 563         { "tcp", &lock_class_rw },
 564         { "tcpinp", &lock_class_rw },
 565         { "so_snd", &lock_class_mtx_sleep },
 566         { NULL, NULL },
 567         /*
 568          * BPF
 569          */
 570         { "bpf global lock", &lock_class_mtx_sleep },
 571         { "bpf interface lock", &lock_class_rw },
 572         { "bpf cdev lock", &lock_class_mtx_sleep },
 573         { NULL, NULL },
 574         /*
 575          * NFS server
 576          */
 577         { "nfsd_mtx", &lock_class_mtx_sleep },
 578         { "so_snd", &lock_class_mtx_sleep },
 579         { NULL, NULL },
 580
 581         /*
 582          * IEEE 802.11
 583          */
 584         { "802.11 com lock", &lock_class_mtx_sleep},
 585         { NULL, NULL },
 586         /*
 587          * Network drivers
 588          */
 589         { "network driver", &lock_class_mtx_sleep},
 590         { NULL, NULL },
 591
 592         /*
 593          * Netgraph
 594          */
 595         { "ng_node", &lock_class_mtx_sleep },
 596         { "ng_worklist", &lock_class_mtx_sleep },
 597         { NULL, NULL },
 598         /*
 599          * CDEV
 600          */
 601         { "vm map (system)", &lock_class_mtx_sleep },
 602         { "vm page queue", &lock_class_mtx_sleep },
 603         { "vnode interlock", &lock_class_mtx_sleep },
 604         { "cdev", &lock_class_mtx_sleep },
 605         { NULL, NULL },
 606         /*
 607          * VM
 608          */
 609         { "vm map (user)", &lock_class_sx },
 610         { "vm object", &lock_class_rw },
 611         { "vm page", &lock_class_mtx_sleep },
 612         { "vm page queue", &lock_class_mtx_sleep },
 613         { "pmap pv global", &lock_class_rw },
 614         { "pmap", &lock_class_mtx_sleep },
 615         { "pmap pv list", &lock_class_rw },
 616         { "vm page free queue", &lock_class_mtx_sleep },
 617         { NULL, NULL },
 618         /*
 619          * kqueue/VFS interaction
 620          */
 621         { "kqueue", &lock_class_mtx_sleep },
 622         { "struct mount mtx", &lock_class_mtx_sleep },
 623         { "vnode interlock", &lock_class_mtx_sleep },
 624         { NULL, NULL },
 625         /*
 626          * ZFS locking
 627          */
 628         { "dn->dn_mtx", &lock_class_sx },
 629         { "dr->dt.di.dr_mtx", &lock_class_sx },
 630         { "db->db_mtx", &lock_class_sx },
 631         { NULL, NULL },
 632         /*
 633          * spin locks
 634          */
 635 #ifdef SMP
 636         { "ap boot", &lock_class_mtx_spin },
 637 #endif
 638         { "rm.mutex_mtx", &lock_class_mtx_spin },
 639         { "sio", &lock_class_mtx_spin },
 640         { "scrlock", &lock_class_mtx_spin },
 641 #ifdef __i386__
 642         { "cy", &lock_class_mtx_spin },
 643 #endif
 644 #ifdef __sparc64__
 645         { "pcib_mtx", &lock_class_mtx_spin },
 646         { "rtc_mtx", &lock_class_mtx_spin },
 647 #endif
 648         { "scc_hwmtx", &lock_class_mtx_spin },
 649         { "uart_hwmtx", &lock_class_mtx_spin },
 650         { "fast_taskqueue", &lock_class_mtx_spin },
 651         { "intr table", &lock_class_mtx_spin },
 652 #ifdef  HWPMC_HOOKS
 653         { "pmc-per-proc", &lock_class_mtx_spin },
 654 #endif
 655         { "process slock", &lock_class_mtx_spin },
 656         { "sleepq chain", &lock_class_mtx_spin },
 657         { "rm_spinlock", &lock_class_mtx_spin },
 658         { "turnstile chain", &lock_class_mtx_spin },
 659         { "turnstile lock", &lock_class_mtx_spin },
 660         { "sched lock", &lock_class_mtx_spin },
 661         { "td_contested", &lock_class_mtx_spin },
 662         { "callout", &lock_class_mtx_spin },
 663         { "entropy harvest mutex", &lock_class_mtx_spin },
 664         { "syscons video lock", &lock_class_mtx_spin },
 665 #ifdef SMP
 666         { "smp rendezvous", &lock_class_mtx_spin },
 667 #endif
 668 #ifdef __powerpc__
 669         { "tlb0", &lock_class_mtx_spin },
 670 #endif
 671         /*
 672          * leaf locks
 673          */
 674         { "intrcnt", &lock_class_mtx_spin },
 675         { "icu", &lock_class_mtx_spin },
 676 #if defined(SMP) && defined(__sparc64__)
 677         { "ipi", &lock_class_mtx_spin },
 678 #endif
 679 #ifdef __i386__
 680         { "allpmaps", &lock_class_mtx_spin },
 681         { "descriptor tables", &lock_class_mtx_spin },
 682 #endif
 683         { "clk", &lock_class_mtx_spin },
 684         { "cpuset", &lock_class_mtx_spin },
 685         { "mprof lock", &lock_class_mtx_spin },
 686         { "zombie lock", &lock_class_mtx_spin },
 687         { "ALD Queue", &lock_class_mtx_spin },
 688 #if defined(__i386__) || defined(__amd64__)
 689         { "pcicfg", &lock_class_mtx_spin },
 690         { "NDIS thread lock", &lock_class_mtx_spin },
 691 #endif
 692         { "tw_osl_io_lock", &lock_class_mtx_spin },
 693         { "tw_osl_q_lock", &lock_class_mtx_spin },
 694         { "tw_cl_io_lock", &lock_class_mtx_spin },
 695         { "tw_cl_intr_lock", &lock_class_mtx_spin },
 696         { "tw_cl_gen_lock", &lock_class_mtx_spin },
 697 #ifdef  HWPMC_HOOKS
 698         { "pmc-leaf", &lock_class_mtx_spin },
 699 #endif
 700         { "blocked lock", &lock_class_mtx_spin },
 701         { NULL, NULL },
 702         { NULL, NULL }
 703 };
 704
 705 #ifdef BLESSING
 706 /*
 707  * Pairs of locks which have been blessed
 708  * Don't complain about order problems with blessed locks
 709  */
 710 static struct witness_blessed blessed_list[] = {
 711 };
 712 static int blessed_count =
 713         sizeof(blessed_list) / sizeof(struct witness_blessed);
 714 #endif
 715
 716 /*
 717  * This global is set to 0 once it becomes safe to use the witness code.
 718  */
 719 static int witness_cold = 1;
 720
 721 /*
 722  * This global is set to 1 once the static lock orders have been enrolled
 723  * so that a warning can be issued for any spin locks enrolled later.
 724  */
 725 static int witness_spin_warn = 0;
 726
 727 /* Trim useless garbage from filenames. */
 728 static const char *
 729 fixup_filename(const char *file)
 730 {
 731
 732         if (file == NULL)
 733                 return (NULL);
 734         while (strncmp(file, "../", 3) == 0)
 735                 file += 3;
 736         return (file);
 737 }
 738
 739 /*
 740  * The WITNESS-enabled diagnostic code.  Note that the witness code does
 741  * assume that the early boot is single-threaded at least until after this
 742  * routine is completed.
 743  */
 744 static void
 745 witness_initialize(void *dummy __unused)
 746 {
 747         struct lock_object *lock;
 748         struct witness_order_list_entry *order;
 749         struct witness *w, *w1;
 750         int i;
 751
 752         w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS,
 753             M_WAITOK | M_ZERO);
 754
 755         w_rmatrix = malloc(sizeof(*w_rmatrix) * (witness_count + 1),
 756             M_WITNESS, M_WAITOK | M_ZERO);
 757
 758         for (i = 0; i < witness_count + 1; i++) {
 759                 w_rmatrix[i] = malloc(sizeof(*w_rmatrix[i]) *
 760                     (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO);
 761         }
 762         badstack_sbuf_size = witness_count * 256;
 763
 764         /*
 765          * We have to release Giant before initializing its witness
 766          * structure so that WITNESS doesn't get confused.
 767          */
 768         mtx_unlock(&Giant);
 769         mtx_assert(&Giant, MA_NOTOWNED);
 770
 771         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
 772         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
 773             MTX_NOWITNESS | MTX_NOPROFILE);
 774         for (i = witness_count - 1; i >= 0; i--) {
 775                 w = &w_data[i];
 776                 memset(w, 0, sizeof(*w));
 777                 w_data[i].w_index = i;  /* Witness index never changes. */
 778                 witness_free(w);
 779         }
 780         KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
 781             ("%s: Invalid list of free witness objects", __func__));
 782
 783         /* Witness with index 0 is not used to aid in debugging. */
 784         STAILQ_REMOVE_HEAD(&w_free, w_list);
 785         w_free_cnt--;
 786
 787         for (i = 0; i < witness_count; i++) {
 788                 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
 789                     (witness_count + 1));
 790         }
 791
 792         for (i = 0; i < LOCK_CHILDCOUNT; i++)
 793                 witness_lock_list_free(&w_locklistdata[i]);
 794         witness_init_hash_tables();
 795
 796         /* First add in all the specified order lists. */
 797         for (order = order_lists; order->w_name != NULL; order++) {
 798                 w = enroll(order->w_name, order->w_class);
 799                 if (w == NULL)
 800                         continue;
 801                 w->w_file = "order list";
 802                 for (order++; order->w_name != NULL; order++) {
 803                         w1 = enroll(order->w_name, order->w_class);
 804                         if (w1 == NULL)
 805                                 continue;
 806                         w1->w_file = "order list";
 807                         itismychild(w, w1);
 808                         w = w1;
 809                 }
 810         }
 811         witness_spin_warn = 1;
 812
 813         /* Iterate through all locks and add them to witness. */
 814         for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
 815                 lock = pending_locks[i].wh_lock;
 816                 KASSERT(lock->lo_flags & LO_WITNESS,
 817                     ("%s: lock %s is on pending list but not LO_WITNESS",
 818                     __func__, lock->lo_name));
 819                 lock->lo_witness = enroll(pending_locks[i].wh_type,
 820                     LOCK_CLASS(lock));
 821         }
 822
 823         /* Mark the witness code as being ready for use. */
 824         witness_cold = 0;
 825
 826         mtx_lock(&Giant);
 827 }
 828 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
 829     NULL);
 830
 831 void
 832 witness_init(struct lock_object *lock, const char *type)
 833 {
 834         struct lock_class *class;
 835
 836         /* Various sanity checks. */
 837         class = LOCK_CLASS(lock);
 838         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
 839             (class->lc_flags & LC_RECURSABLE) == 0)
 840                 kassert_panic("%s: lock (%s) %s can not be recursable",
 841                     __func__, class->lc_name, lock->lo_name);
 842         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 843             (class->lc_flags & LC_SLEEPABLE) == 0)
 844                 kassert_panic("%s: lock (%s) %s can not be sleepable",
 845                     __func__, class->lc_name, lock->lo_name);
 846         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
 847             (class->lc_flags & LC_UPGRADABLE) == 0)
 848                 kassert_panic("%s: lock (%s) %s can not be upgradable",
 849                     __func__, class->lc_name, lock->lo_name);
 850
 851         /*
 852          * If we shouldn't watch this lock, then just clear lo_witness.
 853          * Otherwise, if witness_cold is set, then it is too early to
 854          * enroll this lock, so defer it to witness_initialize() by adding
 855          * it to the pending_locks list.  If it is not too early, then enroll
 856          * the lock now.
 857          */
 858         if (witness_watch < 1 || panicstr != NULL ||
 859             (lock->lo_flags & LO_WITNESS) == 0)
 860                 lock->lo_witness = NULL;
 861         else if (witness_cold) {
 862                 pending_locks[pending_cnt].wh_lock = lock;
 863                 pending_locks[pending_cnt++].wh_type = type;
 864                 if (pending_cnt > WITNESS_PENDLIST)
 865                         panic("%s: pending locks list is too small, "
 866                             "increase WITNESS_PENDLIST\n",
 867                             __func__);
 868         } else
 869                 lock->lo_witness = enroll(type, class);
 870 }
 871
 872 void
 873 witness_destroy(struct lock_object *lock)
 874 {
 875         struct lock_class *class;
 876         struct witness *w;
 877
 878         class = LOCK_CLASS(lock);
 879
 880         if (witness_cold)
 881                 panic("lock (%s) %s destroyed while witness_cold",
 882                     class->lc_name, lock->lo_name);
 883
 884         /* XXX: need to verify that no one holds the lock */
 885         if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
 886                 return;
 887         w = lock->lo_witness;
 888
 889         mtx_lock_spin(&w_mtx);
 890         MPASS(w->w_refcount > 0);
 891         w->w_refcount--;
 892
 893         if (w->w_refcount == 0)
 894                 depart(w);
 895         mtx_unlock_spin(&w_mtx);
 896 }
 897
 898 #ifdef DDB
 899 static void
 900 witness_ddb_compute_levels(void)
 901 {
 902         struct witness *w;
 903
 904         /*
 905          * First clear all levels.
 906          */
 907         STAILQ_FOREACH(w, &w_all, w_list)
 908                 w->w_ddb_level = -1;
 909
 910         /*
 911          * Look for locks with no parents and level all their descendants.
 912          */
 913         STAILQ_FOREACH(w, &w_all, w_list) {
 914
 915                 /* If the witness has ancestors (is not a root), skip it. */
 916                 if (w->w_num_ancestors > 0)
 917                         continue;
 918                 witness_ddb_level_descendants(w, 0);
 919         }
 920 }
 921
 922 static void
 923 witness_ddb_level_descendants(struct witness *w, int l)
 924 {
 925         int i;
 926
 927         if (w->w_ddb_level >= l)
 928                 return;
 929
 930         w->w_ddb_level = l;
 931         l++;
 932
 933         for (i = 1; i <= w_max_used_index; i++) {
 934                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 935                         witness_ddb_level_descendants(&w_data[i], l);
 936         }
 937 }
 938
 939 static void
 940 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
 941     struct witness *w, int indent)
 942 {
 943         int i;
 944
 945         for (i = 0; i < indent; i++)
 946                 prnt(" ");
 947         prnt("%s (type: %s, depth: %d, active refs: %d)",
 948              w->w_name, w->w_class->lc_name,
 949              w->w_ddb_level, w->w_refcount);
 950         if (w->w_displayed) {
 951                 prnt(" -- (already displayed)\n");
 952                 return;
 953         }
 954         w->w_displayed = 1;
 955         if (w->w_file != NULL && w->w_line != 0)
 956                 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
 957                     w->w_line);
 958         else
 959                 prnt(" -- never acquired\n");
 960         indent++;
 961         WITNESS_INDEX_ASSERT(w->w_index);
 962         for (i = 1; i <= w_max_used_index; i++) {
 963                 if (db_pager_quit)
 964                         return;
 965                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
 966                         witness_ddb_display_descendants(prnt, &w_data[i],
 967                             indent);
 968         }
 969 }
 970
 971 static void
 972 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
 973     struct witness_list *list)
 974 {
 975         struct witness *w;
 976
 977         STAILQ_FOREACH(w, list, w_typelist) {
 978                 if (w->w_file == NULL || w->w_ddb_level > 0)
 979                         continue;
 980
 981                 /* This lock has no anscestors - display its descendants. */
 982                 witness_ddb_display_descendants(prnt, w, 0);
 983                 if (db_pager_quit)
 984                         return;
 985         }
 986 }
 987
 988 static void
 989 witness_ddb_display(int(*prnt)(const char *fmt, ...))
 990 {
 991         struct witness *w;
 992
 993         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
 994         witness_ddb_compute_levels();
 995
 996         /* Clear all the displayed flags. */
 997         STAILQ_FOREACH(w, &w_all, w_list)
 998                 w->w_displayed = 0;
 999
1000         /*
1001          * First, handle sleep locks which have been acquired at least
1002          * once.
1003          */
1004         prnt("Sleep locks:\n");
1005         witness_ddb_display_list(prnt, &w_sleep);
1006         if (db_pager_quit)
1007                 return;
1008
1009         /*
1010          * Now do spin locks which have been acquired at least once.
1011          */
1012         prnt("\nSpin locks:\n");
1013         witness_ddb_display_list(prnt, &w_spin);
1014         if (db_pager_quit)
1015                 return;
1016
1017         /*
1018          * Finally, any locks which have not been acquired yet.
1019          */
1020         prnt("\nLocks which were never acquired:\n");
1021         STAILQ_FOREACH(w, &w_all, w_list) {
1022                 if (w->w_file != NULL || w->w_refcount == 0)
1023                         continue;
1024                 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1025                     w->w_class->lc_name, w->w_ddb_level);
1026                 if (db_pager_quit)
1027                         return;
1028         }
1029 }
1030 #endif /* DDB */
1031
1032 int
1033 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1034 {
1035
1036         if (witness_watch == -1 || panicstr != NULL)
1037                 return (0);
1038
1039         /* Require locks that witness knows about. */
1040         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1041             lock2->lo_witness == NULL)
1042                 return (EINVAL);
1043
1044         mtx_assert(&w_mtx, MA_NOTOWNED);
1045         mtx_lock_spin(&w_mtx);
1046
1047         /*
1048          * If we already have either an explicit or implied lock order that
1049          * is the other way around, then return an error.
1050          */
1051         if (witness_watch &&
1052             isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1053                 mtx_unlock_spin(&w_mtx);
1054                 return (EDOOFUS);
1055         }
1056
1057         /* Try to add the new order. */
1058         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1059             lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1060         itismychild(lock1->lo_witness, lock2->lo_witness);
1061         mtx_unlock_spin(&w_mtx);
1062         return (0);
1063 }
1064
1065 void
1066 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1067     int line, struct lock_object *interlock)
1068 {
1069         struct lock_list_entry *lock_list, *lle;
1070         struct lock_instance *lock1, *lock2, *plock;
1071         struct lock_class *class, *iclass;
1072         struct witness *w, *w1;
1073         struct thread *td;
1074         int i, j;
1075
1076         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1077             panicstr != NULL)
1078                 return;
1079
1080         w = lock->lo_witness;
1081         class = LOCK_CLASS(lock);
1082         td = curthread;
1083
1084         if (class->lc_flags & LC_SLEEPLOCK) {
1085
1086                 /*
1087                  * Since spin locks include a critical section, this check
1088                  * implicitly enforces a lock order of all sleep locks before
1089                  * all spin locks.
1090                  */
1091                 if (td->td_critnest != 0 && !kdb_active)
1092                         kassert_panic("acquiring blockable sleep lock with "
1093                             "spinlock or critical section held (%s) %s @ %s:%d",
1094                             class->lc_name, lock->lo_name,
1095                             fixup_filename(file), line);
1096
1097                 /*
1098                  * If this is the first lock acquired then just return as
1099                  * no order checking is needed.
1100                  */
1101                 lock_list = td->td_sleeplocks;
1102                 if (lock_list == NULL || lock_list->ll_count == 0)
1103                         return;
1104         } else {
1105
1106                 /*
1107                  * If this is the first lock, just return as no order
1108                  * checking is needed.  Avoid problems with thread
1109                  * migration pinning the thread while checking if
1110                  * spinlocks are held.  If at least one spinlock is held
1111                  * the thread is in a safe path and it is allowed to
1112                  * unpin it.
1113                  */
1114                 sched_pin();
1115                 lock_list = PCPU_GET(spinlocks);
1116                 if (lock_list == NULL || lock_list->ll_count == 0) {
1117                         sched_unpin();
1118                         return;
1119                 }
1120                 sched_unpin();
1121         }
1122
1123         /*
1124          * Check to see if we are recursing on a lock we already own.  If
1125          * so, make sure that we don't mismatch exclusive and shared lock
1126          * acquires.
1127          */
1128         lock1 = find_instance(lock_list, lock);
1129         if (lock1 != NULL) {
1130                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1131                     (flags & LOP_EXCLUSIVE) == 0) {
1132                         witness_output("shared lock of (%s) %s @ %s:%d\n",
1133                             class->lc_name, lock->lo_name,
1134                             fixup_filename(file), line);
1135                         witness_output("while exclusively locked from %s:%d\n",
1136                             fixup_filename(lock1->li_file), lock1->li_line);
1137                         kassert_panic("excl->share");
1138                 }
1139                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1140                     (flags & LOP_EXCLUSIVE) != 0) {
1141                         witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1142                             class->lc_name, lock->lo_name,
1143                             fixup_filename(file), line);
1144                         witness_output("while share locked from %s:%d\n",
1145                             fixup_filename(lock1->li_file), lock1->li_line);
1146                         kassert_panic("share->excl");
1147                 }
1148                 return;
1149         }
1150
1151         /* Warn if the interlock is not locked exactly once. */
1152         if (interlock != NULL) {
1153                 iclass = LOCK_CLASS(interlock);
1154                 lock1 = find_instance(lock_list, interlock);
1155                 if (lock1 == NULL)
1156                         kassert_panic("interlock (%s) %s not locked @ %s:%d",
1157                             iclass->lc_name, interlock->lo_name,
1158                             fixup_filename(file), line);
1159                 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1160                         kassert_panic("interlock (%s) %s recursed @ %s:%d",
1161                             iclass->lc_name, interlock->lo_name,
1162                             fixup_filename(file), line);
1163         }
1164
1165         /*
1166          * Find the previously acquired lock, but ignore interlocks.
1167          */
1168         plock = &lock_list->ll_children[lock_list->ll_count - 1];
1169         if (interlock != NULL && plock->li_lock == interlock) {
1170                 if (lock_list->ll_count > 1)
1171                         plock =
1172                             &lock_list->ll_children[lock_list->ll_count - 2];
1173                 else {
1174                         lle = lock_list->ll_next;
1175
1176                         /*
1177                          * The interlock is the only lock we hold, so
1178                          * simply return.
1179                          */
1180                         if (lle == NULL)
1181                                 return;
1182                         plock = &lle->ll_children[lle->ll_count - 1];
1183                 }
1184         }
1185
1186         /*
1187          * Try to perform most checks without a lock.  If this succeeds we
1188          * can skip acquiring the lock and return success.  Otherwise we redo
1189          * the check with the lock held to handle races with concurrent updates.
1190          */
1191         w1 = plock->li_lock->lo_witness;
1192         if (witness_lock_order_check(w1, w))
1193                 return;
1194
1195         mtx_lock_spin(&w_mtx);
1196         if (witness_lock_order_check(w1, w)) {
1197                 mtx_unlock_spin(&w_mtx);
1198                 return;
1199         }
1200         witness_lock_order_add(w1, w);
1201
1202         /*
1203          * Check for duplicate locks of the same type.  Note that we only
1204          * have to check for this on the last lock we just acquired.  Any
1205          * other cases will be caught as lock order violations.
1206          */
1207         if (w1 == w) {
1208                 i = w->w_index;
1209                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1210                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1211                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
1212                         w->w_reversed = 1;
1213                         mtx_unlock_spin(&w_mtx);
1214                         witness_output(
1215                             "acquiring duplicate lock of same type: \"%s\"\n",
1216                             w->w_name);
1217                         witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1218                             fixup_filename(plock->li_file), plock->li_line);
1219                         witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1220                             fixup_filename(file), line);
1221                         witness_debugger(1, __func__);
1222                 } else
1223                         mtx_unlock_spin(&w_mtx);
1224                 return;
1225         }
1226         mtx_assert(&w_mtx, MA_OWNED);
1227
1228         /*
1229          * If we know that the lock we are acquiring comes after
1230          * the lock we most recently acquired in the lock order tree,
1231          * then there is no need for any further checks.
1232          */
1233         if (isitmychild(w1, w))
1234                 goto out;
1235
1236         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1237                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1238
1239                         MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1240                         lock1 = &lle->ll_children[i];
1241
1242                         /*
1243                          * Ignore the interlock.
1244                          */
1245                         if (interlock == lock1->li_lock)
1246                                 continue;
1247
1248                         /*
1249                          * If this lock doesn't undergo witness checking,
1250                          * then skip it.
1251                          */
1252                         w1 = lock1->li_lock->lo_witness;
1253                         if (w1 == NULL) {
1254                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1255                                     ("lock missing witness structure"));
1256                                 continue;
1257                         }
1258
1259                         /*
1260                          * If we are locking Giant and this is a sleepable
1261                          * lock, then skip it.
1262                          */
1263                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1264                             lock == &Giant.lock_object)
1265                                 continue;
1266
1267                         /*
1268                          * If we are locking a sleepable lock and this lock
1269                          * is Giant, then skip it.
1270                          */
1271                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1272                             lock1->li_lock == &Giant.lock_object)
1273                                 continue;
1274
1275                         /*
1276                          * If we are locking a sleepable lock and this lock
1277                          * isn't sleepable, we want to treat it as a lock
1278                          * order violation to enfore a general lock order of
1279                          * sleepable locks before non-sleepable locks.
1280                          */
1281                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1282                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1283                                 goto reversal;
1284
1285                         /*
1286                          * If we are locking Giant and this is a non-sleepable
1287                          * lock, then treat it as a reversal.
1288                          */
1289                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1290                             lock == &Giant.lock_object)
1291                                 goto reversal;
1292
1293                         /*
1294                          * Check the lock order hierarchy for a reveresal.
1295                          */
1296                         if (!isitmydescendant(w, w1))
1297                                 continue;
1298                 reversal:
1299
1300                         /*
1301                          * We have a lock order violation, check to see if it
1302                          * is allowed or has already been yelled about.
1303                          */
1304 #ifdef BLESSING
1305
1306                         /*
1307                          * If the lock order is blessed, just bail.  We don't
1308                          * look for other lock order violations though, which
1309                          * may be a bug.
1310                          */
1311                         if (blessed(w, w1))
1312                                 goto out;
1313 #endif
1314
1315                         /* Bail if this violation is known */
1316                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1317                                 goto out;
1318
1319                         /* Record this as a violation */
1320                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1321                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1322                         w->w_reversed = w1->w_reversed = 1;
1323                         witness_increment_graph_generation();
1324                         mtx_unlock_spin(&w_mtx);
1325
1326 #ifdef WITNESS_NO_VNODE
1327                         /*
1328                          * There are known LORs between VNODE locks. They are
1329                          * not an indication of a bug. VNODE locks are flagged
1330                          * as such (LO_IS_VNODE) and we don't yell if the LOR
1331                          * is between 2 VNODE locks.
1332                          */
1333                         if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1334                             (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1335                                 return;
1336 #endif
1337
1338                         /*
1339                          * Ok, yell about it.
1340                          */
1341                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1342                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1343                                 witness_output(
1344                 "lock order reversal: (sleepable after non-sleepable)\n");
1345                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1346                             && lock == &Giant.lock_object)
1347                                 witness_output(
1348                 "lock order reversal: (Giant after non-sleepable)\n");
1349                         else
1350                                 witness_output("lock order reversal:\n");
1351
1352                         /*
1353                          * Try to locate an earlier lock with
1354                          * witness w in our list.
1355                          */
1356                         do {
1357                                 lock2 = &lle->ll_children[i];
1358                                 MPASS(lock2->li_lock != NULL);
1359                                 if (lock2->li_lock->lo_witness == w)
1360                                         break;
1361                                 if (i == 0 && lle->ll_next != NULL) {
1362                                         lle = lle->ll_next;
1363                                         i = lle->ll_count - 1;
1364                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
1365                                 } else
1366                                         i--;
1367                         } while (i >= 0);
1368                         if (i < 0) {
1369                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1370                                     lock1->li_lock, lock1->li_lock->lo_name,
1371                                     w1->w_name, fixup_filename(lock1->li_file),
1372                                     lock1->li_line);
1373                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1374                                     lock->lo_name, w->w_name,
1375                                     fixup_filename(file), line);
1376                         } else {
1377                                 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1378                                     lock2->li_lock, lock2->li_lock->lo_name,
1379                                     lock2->li_lock->lo_witness->w_name,
1380                                     fixup_filename(lock2->li_file),
1381                                     lock2->li_line);
1382                                 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1383                                     lock1->li_lock, lock1->li_lock->lo_name,
1384                                     w1->w_name, fixup_filename(lock1->li_file),
1385                                     lock1->li_line);
1386                                 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1387                                     lock->lo_name, w->w_name,
1388                                     fixup_filename(file), line);
1389                         }
1390                         witness_debugger(1, __func__);
1391                         return;
1392                 }
1393         }
1394
1395         /*
1396          * If requested, build a new lock order.  However, don't build a new
1397          * relationship between a sleepable lock and Giant if it is in the
1398          * wrong direction.  The correct lock order is that sleepable locks
1399          * always come before Giant.
1400          */
1401         if (flags & LOP_NEWORDER &&
1402             !(plock->li_lock == &Giant.lock_object &&
1403             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1404                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1405                     w->w_name, plock->li_lock->lo_witness->w_name);
1406                 itismychild(plock->li_lock->lo_witness, w);
1407         }
1408 out:
1409         mtx_unlock_spin(&w_mtx);
1410 }
1411
1412 void
1413 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1414 {
1415         struct lock_list_entry **lock_list, *lle;
1416         struct lock_instance *instance;
1417         struct witness *w;
1418         struct thread *td;
1419
1420         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1421             panicstr != NULL)
1422                 return;
1423         w = lock->lo_witness;
1424         td = curthread;
1425
1426         /* Determine lock list for this lock. */
1427         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1428                 lock_list = &td->td_sleeplocks;
1429         else
1430                 lock_list = PCPU_PTR(spinlocks);
1431
1432         /* Check to see if we are recursing on a lock we already own. */
1433         instance = find_instance(*lock_list, lock);
1434         if (instance != NULL) {
1435                 instance->li_flags++;
1436                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1437                     td->td_proc->p_pid, lock->lo_name,
1438                     instance->li_flags & LI_RECURSEMASK);
1439                 instance->li_file = file;
1440                 instance->li_line = line;
1441                 return;
1442         }
1443
1444         /* Update per-witness last file and line acquire. */
1445         w->w_file = file;
1446         w->w_line = line;
1447
1448         /* Find the next open lock instance in the list and fill it. */
1449         lle = *lock_list;
1450         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1451                 lle = witness_lock_list_get();
1452                 if (lle == NULL)
1453                         return;
1454                 lle->ll_next = *lock_list;
1455                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1456                     td->td_proc->p_pid, lle);
1457                 *lock_list = lle;
1458         }
1459         instance = &lle->ll_children[lle->ll_count++];
1460         instance->li_lock = lock;
1461         instance->li_line = line;
1462         instance->li_file = file;
1463         if ((flags & LOP_EXCLUSIVE) != 0)
1464                 instance->li_flags = LI_EXCLUSIVE;
1465         else
1466                 instance->li_flags = 0;
1467         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1468             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1469 }
1470
1471 void
1472 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1473 {
1474         struct lock_instance *instance;
1475         struct lock_class *class;
1476
1477         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1478         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1479                 return;
1480         class = LOCK_CLASS(lock);
1481         if (witness_watch) {
1482                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1483                         kassert_panic(
1484                             "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1485                             class->lc_name, lock->lo_name,
1486                             fixup_filename(file), line);
1487                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1488                         kassert_panic(
1489                             "upgrade of non-sleep lock (%s) %s @ %s:%d",
1490                             class->lc_name, lock->lo_name,
1491                             fixup_filename(file), line);
1492         }
1493         instance = find_instance(curthread->td_sleeplocks, lock);
1494         if (instance == NULL) {
1495                 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1496                     class->lc_name, lock->lo_name,
1497                     fixup_filename(file), line);
1498                 return;
1499         }
1500         if (witness_watch) {
1501                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1502                         kassert_panic(
1503                             "upgrade of exclusive lock (%s) %s @ %s:%d",
1504                             class->lc_name, lock->lo_name,
1505                             fixup_filename(file), line);
1506                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1507                         kassert_panic(
1508                             "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1509                             class->lc_name, lock->lo_name,
1510                             instance->li_flags & LI_RECURSEMASK,
1511                             fixup_filename(file), line);
1512         }
1513         instance->li_flags |= LI_EXCLUSIVE;
1514 }
1515
1516 void
1517 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1518     int line)
1519 {
1520         struct lock_instance *instance;
1521         struct lock_class *class;
1522
1523         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1524         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1525                 return;
1526         class = LOCK_CLASS(lock);
1527         if (witness_watch) {
1528                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1529                         kassert_panic(
1530                             "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1531                             class->lc_name, lock->lo_name,
1532                             fixup_filename(file), line);
1533                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1534                         kassert_panic(
1535                             "downgrade of non-sleep lock (%s) %s @ %s:%d",
1536                             class->lc_name, lock->lo_name,
1537                             fixup_filename(file), line);
1538         }
1539         instance = find_instance(curthread->td_sleeplocks, lock);
1540         if (instance == NULL) {
1541                 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1542                     class->lc_name, lock->lo_name,
1543                     fixup_filename(file), line);
1544                 return;
1545         }
1546         if (witness_watch) {
1547                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1548                         kassert_panic(
1549                             "downgrade of shared lock (%s) %s @ %s:%d",
1550                             class->lc_name, lock->lo_name,
1551                             fixup_filename(file), line);
1552                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1553                         kassert_panic(
1554                             "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1555                             class->lc_name, lock->lo_name,
1556                             instance->li_flags & LI_RECURSEMASK,
1557                             fixup_filename(file), line);
1558         }
1559         instance->li_flags &= ~LI_EXCLUSIVE;
1560 }
1561
1562 void
1563 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1564 {
1565         struct lock_list_entry **lock_list, *lle;
1566         struct lock_instance *instance;
1567         struct lock_class *class;
1568         struct thread *td;
1569         register_t s;
1570         int i, j;
1571
1572         if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1573                 return;
1574         td = curthread;
1575         class = LOCK_CLASS(lock);
1576
1577         /* Find lock instance associated with this lock. */
1578         if (class->lc_flags & LC_SLEEPLOCK)
1579                 lock_list = &td->td_sleeplocks;
1580         else
1581                 lock_list = PCPU_PTR(spinlocks);
1582         lle = *lock_list;
1583         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1584                 for (i = 0; i < (*lock_list)->ll_count; i++) {
1585                         instance = &(*lock_list)->ll_children[i];
1586                         if (instance->li_lock == lock)
1587                                 goto found;
1588                 }
1589
1590         /*
1591          * When disabling WITNESS through witness_watch we could end up in
1592          * having registered locks in the td_sleeplocks queue.
1593          * We have to make sure we flush these queues, so just search for
1594          * eventual register locks and remove them.
1595          */
1596         if (witness_watch > 0) {
1597                 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1598                     lock->lo_name, fixup_filename(file), line);
1599                 return;
1600         } else {
1601                 return;
1602         }
1603 found:
1604
1605         /* First, check for shared/exclusive mismatches. */
1606         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1607             (flags & LOP_EXCLUSIVE) == 0) {
1608                 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1609                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1610                 witness_output("while exclusively locked from %s:%d\n",
1611                     fixup_filename(instance->li_file), instance->li_line);
1612                 kassert_panic("excl->ushare");
1613         }
1614         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1615             (flags & LOP_EXCLUSIVE) != 0) {
1616                 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1617                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1618                 witness_output("while share locked from %s:%d\n",
1619                     fixup_filename(instance->li_file),
1620                     instance->li_line);
1621                 kassert_panic("share->uexcl");
1622         }
1623         /* If we are recursed, unrecurse. */
1624         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1625                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1626                     td->td_proc->p_pid, instance->li_lock->lo_name,
1627                     instance->li_flags);
1628                 instance->li_flags--;
1629                 return;
1630         }
1631         /* The lock is now being dropped, check for NORELEASE flag */
1632         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1633                 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1634                     class->lc_name, lock->lo_name, fixup_filename(file), line);
1635                 kassert_panic("lock marked norelease");
1636         }
1637
1638         /* Otherwise, remove this item from the list. */
1639         s = intr_disable();
1640         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1641             td->td_proc->p_pid, instance->li_lock->lo_name,
1642             (*lock_list)->ll_count - 1);
1643         for (j = i; j < (*lock_list)->ll_count - 1; j++)
1644                 (*lock_list)->ll_children[j] =
1645                     (*lock_list)->ll_children[j + 1];
1646         (*lock_list)->ll_count--;
1647         intr_restore(s);
1648
1649         /*
1650          * In order to reduce contention on w_mtx, we want to keep always an
1651          * head object into lists so that frequent allocation from the
1652          * free witness pool (and subsequent locking) is avoided.
1653          * In order to maintain the current code simple, when the head
1654          * object is totally unloaded it means also that we do not have
1655          * further objects in the list, so the list ownership needs to be
1656          * hand over to another object if the current head needs to be freed.
1657          */
1658         if ((*lock_list)->ll_count == 0) {
1659                 if (*lock_list == lle) {
1660                         if (lle->ll_next == NULL)
1661                                 return;
1662                 } else
1663                         lle = *lock_list;
1664                 *lock_list = lle->ll_next;
1665                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1666                     td->td_proc->p_pid, lle);
1667                 witness_lock_list_free(lle);
1668         }
1669 }
1670
1671 void
1672 witness_thread_exit(struct thread *td)
1673 {
1674         struct lock_list_entry *lle;
1675         int i, n;
1676
1677         lle = td->td_sleeplocks;
1678         if (lle == NULL || panicstr != NULL)
1679                 return;
1680         if (lle->ll_count != 0) {
1681                 for (n = 0; lle != NULL; lle = lle->ll_next)
1682                         for (i = lle->ll_count - 1; i >= 0; i--) {
1683                                 if (n == 0)
1684                                         witness_output(
1685                     "Thread %p exiting with the following locks held:\n", td);
1686                                 n++;
1687                                 witness_list_lock(&lle->ll_children[i],
1688                                     witness_output);
1689
1690                         }
1691                 kassert_panic(
1692                     "Thread %p cannot exit while holding sleeplocks\n", td);
1693         }
1694         witness_lock_list_free(lle);
1695 }
1696
1697 /*
1698  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
1699  * exempt Giant and sleepable locks from the checks as well.  If any
1700  * non-exempt locks are held, then a supplied message is printed to the
1701  * output channel along with a list of the offending locks.  If indicated in the
1702  * flags then a failure results in a panic as well.
1703  */
1704 int
1705 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1706 {
1707         struct lock_list_entry *lock_list, *lle;
1708         struct lock_instance *lock1;
1709         struct thread *td;
1710         va_list ap;
1711         int i, n;
1712
1713         if (witness_cold || witness_watch < 1 || panicstr != NULL)
1714                 return (0);
1715         n = 0;
1716         td = curthread;
1717         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1718                 for (i = lle->ll_count - 1; i >= 0; i--) {
1719                         lock1 = &lle->ll_children[i];
1720                         if (lock1->li_lock == lock)
1721                                 continue;
1722                         if (flags & WARN_GIANTOK &&
1723                             lock1->li_lock == &Giant.lock_object)
1724                                 continue;
1725                         if (flags & WARN_SLEEPOK &&
1726                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1727                                 continue;
1728                         if (n == 0) {
1729                                 va_start(ap, fmt);
1730                                 witness_voutput(fmt, ap);
1731                                 va_end(ap);
1732                                 witness_output(
1733                                     " with the following %slocks held:\n",
1734                                     (flags & WARN_SLEEPOK) != 0 ?
1735                                     "non-sleepable " : "");
1736                         }
1737                         n++;
1738                         witness_list_lock(lock1, witness_output);
1739                 }
1740
1741         /*
1742          * Pin the thread in order to avoid problems with thread migration.
1743          * Once that all verifies are passed about spinlocks ownership,
1744          * the thread is in a safe path and it can be unpinned.
1745          */
1746         sched_pin();
1747         lock_list = PCPU_GET(spinlocks);
1748         if (lock_list != NULL && lock_list->ll_count != 0) {
1749                 sched_unpin();
1750
1751                 /*
1752                  * We should only have one spinlock and as long as
1753                  * the flags cannot match for this locks class,
1754                  * check if the first spinlock is the one curthread
1755                  * should hold.
1756                  */
1757                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1758                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1759                     lock1->li_lock == lock && n == 0)
1760                         return (0);
1761
1762                 va_start(ap, fmt);
1763                 witness_voutput(fmt, ap);
1764                 va_end(ap);
1765                 witness_output(" with the following %slocks held:\n",
1766                     (flags & WARN_SLEEPOK) != 0 ?  "non-sleepable " : "");
1767                 n += witness_list_locks(&lock_list, witness_output);
1768         } else
1769                 sched_unpin();
1770         if (flags & WARN_PANIC && n)
1771                 kassert_panic("%s", __func__);
1772         else
1773                 witness_debugger(n, __func__);
1774         return (n);
1775 }
1776
1777 const char *
1778 witness_file(struct lock_object *lock)
1779 {
1780         struct witness *w;
1781
1782         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1783                 return ("?");
1784         w = lock->lo_witness;
1785         return (w->w_file);
1786 }
1787
1788 int
1789 witness_line(struct lock_object *lock)
1790 {
1791         struct witness *w;
1792
1793         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1794                 return (0);
1795         w = lock->lo_witness;
1796         return (w->w_line);
1797 }
1798
1799 static struct witness *
1800 enroll(const char *description, struct lock_class *lock_class)
1801 {
1802         struct witness *w;
1803         struct witness_list *typelist;
1804
1805         MPASS(description != NULL);
1806
1807         if (witness_watch == -1 || panicstr != NULL)
1808                 return (NULL);
1809         if ((lock_class->lc_flags & LC_SPINLOCK)) {
1810                 if (witness_skipspin)
1811                         return (NULL);
1812                 else
1813                         typelist = &w_spin;
1814         } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1815                 typelist = &w_sleep;
1816         } else {
1817                 kassert_panic("lock class %s is not sleep or spin",
1818                     lock_class->lc_name);
1819                 return (NULL);
1820         }
1821
1822         mtx_lock_spin(&w_mtx);
1823         w = witness_hash_get(description);
1824         if (w)
1825                 goto found;
1826         if ((w = witness_get()) == NULL)
1827                 return (NULL);
1828         MPASS(strlen(description) < MAX_W_NAME);
1829         strcpy(w->w_name, description);
1830         w->w_class = lock_class;
1831         w->w_refcount = 1;
1832         STAILQ_INSERT_HEAD(&w_all, w, w_list);
1833         if (lock_class->lc_flags & LC_SPINLOCK) {
1834                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1835                 w_spin_cnt++;
1836         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1837                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1838                 w_sleep_cnt++;
1839         }
1840
1841         /* Insert new witness into the hash */
1842         witness_hash_put(w);
1843         witness_increment_graph_generation();
1844         mtx_unlock_spin(&w_mtx);
1845         return (w);
1846 found:
1847         w->w_refcount++;
1848         mtx_unlock_spin(&w_mtx);
1849         if (lock_class != w->w_class)
1850                 kassert_panic(
1851                         "lock (%s) %s does not match earlier (%s) lock",
1852                         description, lock_class->lc_name,
1853                         w->w_class->lc_name);
1854         return (w);
1855 }
1856
1857 static void
1858 depart(struct witness *w)
1859 {
1860         struct witness_list *list;
1861
1862         MPASS(w->w_refcount == 0);
1863         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1864                 list = &w_sleep;
1865                 w_sleep_cnt--;
1866         } else {
1867                 list = &w_spin;
1868                 w_spin_cnt--;
1869         }
1870         /*
1871          * Set file to NULL as it may point into a loadable module.
1872          */
1873         w->w_file = NULL;
1874         w->w_line = 0;
1875         witness_increment_graph_generation();
1876 }
1877
1878
1879 static void
1880 adopt(struct witness *parent, struct witness *child)
1881 {
1882         int pi, ci, i, j;
1883
1884         if (witness_cold == 0)
1885                 mtx_assert(&w_mtx, MA_OWNED);
1886
1887         /* If the relationship is already known, there's no work to be done. */
1888         if (isitmychild(parent, child))
1889                 return;
1890
1891         /* When the structure of the graph changes, bump up the generation. */
1892         witness_increment_graph_generation();
1893
1894         /*
1895          * The hard part ... create the direct relationship, then propagate all
1896          * indirect relationships.
1897          */
1898         pi = parent->w_index;
1899         ci = child->w_index;
1900         WITNESS_INDEX_ASSERT(pi);
1901         WITNESS_INDEX_ASSERT(ci);
1902         MPASS(pi != ci);
1903         w_rmatrix[pi][ci] |= WITNESS_PARENT;
1904         w_rmatrix[ci][pi] |= WITNESS_CHILD;
1905
1906         /*
1907          * If parent was not already an ancestor of child,
1908          * then we increment the descendant and ancestor counters.
1909          */
1910         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1911                 parent->w_num_descendants++;
1912                 child->w_num_ancestors++;
1913         }
1914
1915         /*
1916          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1917          * an ancestor of 'pi' during this loop.
1918          */
1919         for (i = 1; i <= w_max_used_index; i++) {
1920                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1921                     (i != pi))
1922                         continue;
1923
1924                 /* Find each descendant of 'i' and mark it as a descendant. */
1925                 for (j = 1; j <= w_max_used_index; j++) {
1926
1927                         /*
1928                          * Skip children that are already marked as
1929                          * descendants of 'i'.
1930                          */
1931                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1932                                 continue;
1933
1934                         /*
1935                          * We are only interested in descendants of 'ci'. Note
1936                          * that 'ci' itself is counted as a descendant of 'ci'.
1937                          */
1938                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1939                             (j != ci))
1940                                 continue;
1941                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1942                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1943                         w_data[i].w_num_descendants++;
1944                         w_data[j].w_num_ancestors++;
1945
1946                         /*
1947                          * Make sure we aren't marking a node as both an
1948                          * ancestor and descendant. We should have caught
1949                          * this as a lock order reversal earlier.
1950                          */
1951                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1952                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1953                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1954                                     "both ancestor and descendant\n",
1955                                     i, j, w_rmatrix[i][j]);
1956                                 kdb_backtrace();
1957                                 printf("Witness disabled.\n");
1958                                 witness_watch = -1;
1959                         }
1960                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1961                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1962                                 printf("witness rmatrix paradox! [%d][%d]=%d "
1963                                     "both ancestor and descendant\n",
1964                                     j, i, w_rmatrix[j][i]);
1965                                 kdb_backtrace();
1966                                 printf("Witness disabled.\n");
1967                                 witness_watch = -1;
1968                         }
1969                 }
1970         }
1971 }
1972
1973 static void
1974 itismychild(struct witness *parent, struct witness *child)
1975 {
1976         int unlocked;
1977
1978         MPASS(child != NULL && parent != NULL);
1979         if (witness_cold == 0)
1980                 mtx_assert(&w_mtx, MA_OWNED);
1981
1982         if (!witness_lock_type_equal(parent, child)) {
1983                 if (witness_cold == 0) {
1984                         unlocked = 1;
1985                         mtx_unlock_spin(&w_mtx);
1986                 } else {
1987                         unlocked = 0;
1988                 }
1989                 kassert_panic(
1990                     "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1991                     "the same lock type", __func__, parent->w_name,
1992                     parent->w_class->lc_name, child->w_name,
1993                     child->w_class->lc_name);
1994                 if (unlocked)
1995                         mtx_lock_spin(&w_mtx);
1996         }
1997         adopt(parent, child);
1998 }
1999
2000 /*
2001  * Generic code for the isitmy*() functions. The rmask parameter is the
2002  * expected relationship of w1 to w2.
2003  */
2004 static int
2005 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2006 {
2007         unsigned char r1, r2;
2008         int i1, i2;
2009
2010         i1 = w1->w_index;
2011         i2 = w2->w_index;
2012         WITNESS_INDEX_ASSERT(i1);
2013         WITNESS_INDEX_ASSERT(i2);
2014         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2015         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2016
2017         /* The flags on one better be the inverse of the flags on the other */
2018         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2019             (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2020                 /* Don't squawk if we're potentially racing with an update. */
2021                 if (!mtx_owned(&w_mtx))
2022                         return (0);
2023                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2024                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
2025                     "w_rmatrix[%d][%d] == %hhx\n",
2026                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2027                     i2, i1, r2);
2028                 kdb_backtrace();
2029                 printf("Witness disabled.\n");
2030                 witness_watch = -1;
2031         }
2032         return (r1 & rmask);
2033 }
2034
2035 /*
2036  * Checks if @child is a direct child of @parent.
2037  */
2038 static int
2039 isitmychild(struct witness *parent, struct witness *child)
2040 {
2041
2042         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2043 }
2044
2045 /*
2046  * Checks if @descendant is a direct or inderect descendant of @ancestor.
2047  */
2048 static int
2049 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2050 {
2051
2052         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2053             __func__));
2054 }
2055
2056 #ifdef BLESSING
2057 static int
2058 blessed(struct witness *w1, struct witness *w2)
2059 {
2060         int i;
2061         struct witness_blessed *b;
2062
2063         for (i = 0; i < blessed_count; i++) {
2064                 b = &blessed_list[i];
2065                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2066                         if (strcmp(w2->w_name, b->b_lock2) == 0)
2067                                 return (1);
2068                         continue;
2069                 }
2070                 if (strcmp(w1->w_name, b->b_lock2) == 0)
2071                         if (strcmp(w2->w_name, b->b_lock1) == 0)
2072                                 return (1);
2073         }
2074         return (0);
2075 }
2076 #endif
2077
2078 static struct witness *
2079 witness_get(void)
2080 {
2081         struct witness *w;
2082         int index;
2083
2084         if (witness_cold == 0)
2085                 mtx_assert(&w_mtx, MA_OWNED);
2086
2087         if (witness_watch == -1) {
2088                 mtx_unlock_spin(&w_mtx);
2089                 return (NULL);
2090         }
2091         if (STAILQ_EMPTY(&w_free)) {
2092                 witness_watch = -1;
2093                 mtx_unlock_spin(&w_mtx);
2094                 printf("WITNESS: unable to allocate a new witness object\n");
2095                 return (NULL);
2096         }
2097         w = STAILQ_FIRST(&w_free);
2098         STAILQ_REMOVE_HEAD(&w_free, w_list);
2099         w_free_cnt--;
2100         index = w->w_index;
2101         MPASS(index > 0 && index == w_max_used_index+1 &&
2102             index < witness_count);
2103         bzero(w, sizeof(*w));
2104         w->w_index = index;
2105         if (index > w_max_used_index)
2106                 w_max_used_index = index;
2107         return (w);
2108 }
2109
2110 static void
2111 witness_free(struct witness *w)
2112 {
2113
2114         STAILQ_INSERT_HEAD(&w_free, w, w_list);
2115         w_free_cnt++;
2116 }
2117
2118 static struct lock_list_entry *
2119 witness_lock_list_get(void)
2120 {
2121         struct lock_list_entry *lle;
2122
2123         if (witness_watch == -1)
2124                 return (NULL);
2125         mtx_lock_spin(&w_mtx);
2126         lle = w_lock_list_free;
2127         if (lle == NULL) {
2128                 witness_watch = -1;
2129                 mtx_unlock_spin(&w_mtx);
2130                 printf("%s: witness exhausted\n", __func__);
2131                 return (NULL);
2132         }
2133         w_lock_list_free = lle->ll_next;
2134         mtx_unlock_spin(&w_mtx);
2135         bzero(lle, sizeof(*lle));
2136         return (lle);
2137 }
2138
2139 static void
2140 witness_lock_list_free(struct lock_list_entry *lle)
2141 {
2142
2143         mtx_lock_spin(&w_mtx);
2144         lle->ll_next = w_lock_list_free;
2145         w_lock_list_free = lle;
2146         mtx_unlock_spin(&w_mtx);
2147 }
2148
2149 static struct lock_instance *
2150 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2151 {
2152         struct lock_list_entry *lle;
2153         struct lock_instance *instance;
2154         int i;
2155
2156         for (lle = list; lle != NULL; lle = lle->ll_next)
2157                 for (i = lle->ll_count - 1; i >= 0; i--) {
2158                         instance = &lle->ll_children[i];
2159                         if (instance->li_lock == lock)
2160                                 return (instance);
2161                 }
2162         return (NULL);
2163 }
2164
2165 static void
2166 witness_list_lock(struct lock_instance *instance,
2167     int (*prnt)(const char *fmt, ...))
2168 {
2169         struct lock_object *lock;
2170
2171         lock = instance->li_lock;
2172         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2173             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2174         if (lock->lo_witness->w_name != lock->lo_name)
2175                 prnt(" (%s)", lock->lo_witness->w_name);
2176         prnt(" r = %d (%p) locked @ %s:%d\n",
2177             instance->li_flags & LI_RECURSEMASK, lock,
2178             fixup_filename(instance->li_file), instance->li_line);
2179 }
2180
2181 static int
2182 witness_output(const char *fmt, ...)
2183 {
2184         va_list ap;
2185         int ret;
2186
2187         va_start(ap, fmt);
2188         ret = witness_voutput(fmt, ap);
2189         va_end(ap);
2190         return (ret);
2191 }
2192
2193 static int
2194 witness_voutput(const char *fmt, va_list ap)
2195 {
2196         int ret;
2197
2198         ret = 0;
2199         switch (witness_channel) {
2200         case WITNESS_CONSOLE:
2201                 ret = vprintf(fmt, ap);
2202                 break;
2203         case WITNESS_LOG:
2204                 vlog(LOG_NOTICE, fmt, ap);
2205                 break;
2206         case WITNESS_NONE:
2207                 break;
2208         }
2209         return (ret);
2210 }
2211
2212 #ifdef DDB
2213 static int
2214 witness_thread_has_locks(struct thread *td)
2215 {
2216
2217         if (td->td_sleeplocks == NULL)
2218                 return (0);
2219         return (td->td_sleeplocks->ll_count != 0);
2220 }
2221
2222 static int
2223 witness_proc_has_locks(struct proc *p)
2224 {
2225         struct thread *td;
2226
2227         FOREACH_THREAD_IN_PROC(p, td) {
2228                 if (witness_thread_has_locks(td))
2229                         return (1);
2230         }
2231         return (0);
2232 }
2233 #endif
2234
2235 int
2236 witness_list_locks(struct lock_list_entry **lock_list,
2237     int (*prnt)(const char *fmt, ...))
2238 {
2239         struct lock_list_entry *lle;
2240         int i, nheld;
2241
2242         nheld = 0;
2243         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2244                 for (i = lle->ll_count - 1; i >= 0; i--) {
2245                         witness_list_lock(&lle->ll_children[i], prnt);
2246                         nheld++;
2247                 }
2248         return (nheld);
2249 }
2250
2251 /*
2252  * This is a bit risky at best.  We call this function when we have timed
2253  * out acquiring a spin lock, and we assume that the other CPU is stuck
2254  * with this lock held.  So, we go groveling around in the other CPU's
2255  * per-cpu data to try to find the lock instance for this spin lock to
2256  * see when it was last acquired.
2257  */
2258 void
2259 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2260     int (*prnt)(const char *fmt, ...))
2261 {
2262         struct lock_instance *instance;
2263         struct pcpu *pc;
2264
2265         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2266                 return;
2267         pc = pcpu_find(owner->td_oncpu);
2268         instance = find_instance(pc->pc_spinlocks, lock);
2269         if (instance != NULL)
2270                 witness_list_lock(instance, prnt);
2271 }
2272
2273 void
2274 witness_save(struct lock_object *lock, const char **filep, int *linep)
2275 {
2276         struct lock_list_entry *lock_list;
2277         struct lock_instance *instance;
2278         struct lock_class *class;
2279
2280         /*
2281          * This function is used independently in locking code to deal with
2282          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2283          * is gone.
2284          */
2285         if (SCHEDULER_STOPPED())
2286                 return;
2287         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2288         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2289                 return;
2290         class = LOCK_CLASS(lock);
2291         if (class->lc_flags & LC_SLEEPLOCK)
2292                 lock_list = curthread->td_sleeplocks;
2293         else {
2294                 if (witness_skipspin)
2295                         return;
2296                 lock_list = PCPU_GET(spinlocks);
2297         }
2298         instance = find_instance(lock_list, lock);
2299         if (instance == NULL) {
2300                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2301                     class->lc_name, lock->lo_name);
2302                 return;
2303         }
2304         *filep = instance->li_file;
2305         *linep = instance->li_line;
2306 }
2307
2308 void
2309 witness_restore(struct lock_object *lock, const char *file, int line)
2310 {
2311         struct lock_list_entry *lock_list;
2312         struct lock_instance *instance;
2313         struct lock_class *class;
2314
2315         /*
2316          * This function is used independently in locking code to deal with
2317          * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2318          * is gone.
2319          */
2320         if (SCHEDULER_STOPPED())
2321                 return;
2322         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2323         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2324                 return;
2325         class = LOCK_CLASS(lock);
2326         if (class->lc_flags & LC_SLEEPLOCK)
2327                 lock_list = curthread->td_sleeplocks;
2328         else {
2329                 if (witness_skipspin)
2330                         return;
2331                 lock_list = PCPU_GET(spinlocks);
2332         }
2333         instance = find_instance(lock_list, lock);
2334         if (instance == NULL)
2335                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2336                     class->lc_name, lock->lo_name);
2337         lock->lo_witness->w_file = file;
2338         lock->lo_witness->w_line = line;
2339         if (instance == NULL)
2340                 return;
2341         instance->li_file = file;
2342         instance->li_line = line;
2343 }
2344
2345 void
2346 witness_assert(const struct lock_object *lock, int flags, const char *file,
2347     int line)
2348 {
2349 #ifdef INVARIANT_SUPPORT
2350         struct lock_instance *instance;
2351         struct lock_class *class;
2352
2353         if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2354                 return;
2355         class = LOCK_CLASS(lock);
2356         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2357                 instance = find_instance(curthread->td_sleeplocks, lock);
2358         else if ((class->lc_flags & LC_SPINLOCK) != 0)
2359                 instance = find_instance(PCPU_GET(spinlocks), lock);
2360         else {
2361                 kassert_panic("Lock (%s) %s is not sleep or spin!",
2362                     class->lc_name, lock->lo_name);
2363                 return;
2364         }
2365         switch (flags) {
2366         case LA_UNLOCKED:
2367                 if (instance != NULL)
2368                         kassert_panic("Lock (%s) %s locked @ %s:%d.",
2369                             class->lc_name, lock->lo_name,
2370                             fixup_filename(file), line);
2371                 break;
2372         case LA_LOCKED:
2373         case LA_LOCKED | LA_RECURSED:
2374         case LA_LOCKED | LA_NOTRECURSED:
2375         case LA_SLOCKED:
2376         case LA_SLOCKED | LA_RECURSED:
2377         case LA_SLOCKED | LA_NOTRECURSED:
2378         case LA_XLOCKED:
2379         case LA_XLOCKED | LA_RECURSED:
2380         case LA_XLOCKED | LA_NOTRECURSED:
2381                 if (instance == NULL) {
2382                         kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2383                             class->lc_name, lock->lo_name,
2384                             fixup_filename(file), line);
2385                         break;
2386                 }
2387                 if ((flags & LA_XLOCKED) != 0 &&
2388                     (instance->li_flags & LI_EXCLUSIVE) == 0)
2389                         kassert_panic(
2390                             "Lock (%s) %s not exclusively locked @ %s:%d.",
2391                             class->lc_name, lock->lo_name,
2392                             fixup_filename(file), line);
2393                 if ((flags & LA_SLOCKED) != 0 &&
2394                     (instance->li_flags & LI_EXCLUSIVE) != 0)
2395                         kassert_panic(
2396                             "Lock (%s) %s exclusively locked @ %s:%d.",
2397                             class->lc_name, lock->lo_name,
2398                             fixup_filename(file), line);
2399                 if ((flags & LA_RECURSED) != 0 &&
2400                     (instance->li_flags & LI_RECURSEMASK) == 0)
2401                         kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2402                             class->lc_name, lock->lo_name,
2403                             fixup_filename(file), line);
2404                 if ((flags & LA_NOTRECURSED) != 0 &&
2405                     (instance->li_flags & LI_RECURSEMASK) != 0)
2406                         kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2407                             class->lc_name, lock->lo_name,
2408                             fixup_filename(file), line);
2409                 break;
2410         default:
2411                 kassert_panic("Invalid lock assertion at %s:%d.",
2412                     fixup_filename(file), line);
2413
2414         }
2415 #endif  /* INVARIANT_SUPPORT */
2416 }
2417
2418 static void
2419 witness_setflag(struct lock_object *lock, int flag, int set)
2420 {
2421         struct lock_list_entry *lock_list;
2422         struct lock_instance *instance;
2423         struct lock_class *class;
2424
2425         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2426                 return;
2427         class = LOCK_CLASS(lock);
2428         if (class->lc_flags & LC_SLEEPLOCK)
2429                 lock_list = curthread->td_sleeplocks;
2430         else {
2431                 if (witness_skipspin)
2432                         return;
2433                 lock_list = PCPU_GET(spinlocks);
2434         }
2435         instance = find_instance(lock_list, lock);
2436         if (instance == NULL) {
2437                 kassert_panic("%s: lock (%s) %s not locked", __func__,
2438                     class->lc_name, lock->lo_name);
2439                 return;
2440         }
2441
2442         if (set)
2443                 instance->li_flags |= flag;
2444         else
2445                 instance->li_flags &= ~flag;
2446 }
2447
2448 void
2449 witness_norelease(struct lock_object *lock)
2450 {
2451
2452         witness_setflag(lock, LI_NORELEASE, 1);
2453 }
2454
2455 void
2456 witness_releaseok(struct lock_object *lock)
2457 {
2458
2459         witness_setflag(lock, LI_NORELEASE, 0);
2460 }
2461
2462 #ifdef DDB
2463 static void
2464 witness_ddb_list(struct thread *td)
2465 {
2466
2467         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2468         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2469
2470         if (witness_watch < 1)
2471                 return;
2472
2473         witness_list_locks(&td->td_sleeplocks, db_printf);
2474
2475         /*
2476          * We only handle spinlocks if td == curthread.  This is somewhat broken
2477          * if td is currently executing on some other CPU and holds spin locks
2478          * as we won't display those locks.  If we had a MI way of getting
2479          * the per-cpu data for a given cpu then we could use
2480          * td->td_oncpu to get the list of spinlocks for this thread
2481          * and "fix" this.
2482          *
2483          * That still wouldn't really fix this unless we locked the scheduler
2484          * lock or stopped the other CPU to make sure it wasn't changing the
2485          * list out from under us.  It is probably best to just not try to
2486          * handle threads on other CPU's for now.
2487          */
2488         if (td == curthread && PCPU_GET(spinlocks) != NULL)
2489                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2490 }
2491
2492 DB_SHOW_COMMAND(locks, db_witness_list)
2493 {
2494         struct thread *td;
2495
2496         if (have_addr)
2497                 td = db_lookup_thread(addr, true);
2498         else
2499                 td = kdb_thread;
2500         witness_ddb_list(td);
2501 }
2502
2503 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2504 {
2505         struct thread *td;
2506         struct proc *p;
2507
2508         /*
2509          * It would be nice to list only threads and processes that actually
2510          * held sleep locks, but that information is currently not exported
2511          * by WITNESS.
2512          */
2513         FOREACH_PROC_IN_SYSTEM(p) {
2514                 if (!witness_proc_has_locks(p))
2515                         continue;
2516                 FOREACH_THREAD_IN_PROC(p, td) {
2517                         if (!witness_thread_has_locks(td))
2518                                 continue;
2519                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2520                             p->p_comm, td, td->td_tid);
2521                         witness_ddb_list(td);
2522                         if (db_pager_quit)
2523                                 return;
2524                 }
2525         }
2526 }
2527 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2528
2529 DB_SHOW_COMMAND(witness, db_witness_display)
2530 {
2531
2532         witness_ddb_display(db_printf);
2533 }
2534 #endif
2535
2536 static int
2537 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2538 {
2539         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2540         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2541         struct sbuf *sb;
2542         u_int w_rmatrix1, w_rmatrix2;
2543         int error, generation, i, j;
2544
2545         tmp_data1 = NULL;
2546         tmp_data2 = NULL;
2547         tmp_w1 = NULL;
2548         tmp_w2 = NULL;
2549         if (witness_watch < 1) {
2550                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2551                 return (error);
2552         }
2553         if (witness_cold) {
2554                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2555                 return (error);
2556         }
2557         error = 0;
2558         sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2559         if (sb == NULL)
2560                 return (ENOMEM);
2561
2562         /* Allocate and init temporary storage space. */
2563         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2564         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2565         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2566             M_WAITOK | M_ZERO);
2567         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2568             M_WAITOK | M_ZERO);
2569         stack_zero(&tmp_data1->wlod_stack);
2570         stack_zero(&tmp_data2->wlod_stack);
2571
2572 restart:
2573         mtx_lock_spin(&w_mtx);
2574         generation = w_generation;
2575         mtx_unlock_spin(&w_mtx);
2576         sbuf_printf(sb, "Number of known direct relationships is %d\n",
2577             w_lohash.wloh_count);
2578         for (i = 1; i < w_max_used_index; i++) {
2579                 mtx_lock_spin(&w_mtx);
2580                 if (generation != w_generation) {
2581                         mtx_unlock_spin(&w_mtx);
2582
2583                         /* The graph has changed, try again. */
2584                         req->oldidx = 0;
2585                         sbuf_clear(sb);
2586                         goto restart;
2587                 }
2588
2589                 w1 = &w_data[i];
2590                 if (w1->w_reversed == 0) {
2591                         mtx_unlock_spin(&w_mtx);
2592                         continue;
2593                 }
2594
2595                 /* Copy w1 locally so we can release the spin lock. */
2596                 *tmp_w1 = *w1;
2597                 mtx_unlock_spin(&w_mtx);
2598
2599                 if (tmp_w1->w_reversed == 0)
2600                         continue;
2601                 for (j = 1; j < w_max_used_index; j++) {
2602                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2603                                 continue;
2604
2605                         mtx_lock_spin(&w_mtx);
2606                         if (generation != w_generation) {
2607                                 mtx_unlock_spin(&w_mtx);
2608
2609                                 /* The graph has changed, try again. */
2610                                 req->oldidx = 0;
2611                                 sbuf_clear(sb);
2612                                 goto restart;
2613                         }
2614
2615                         w2 = &w_data[j];
2616                         data1 = witness_lock_order_get(w1, w2);
2617                         data2 = witness_lock_order_get(w2, w1);
2618
2619                         /*
2620                          * Copy information locally so we can release the
2621                          * spin lock.
2622                          */
2623                         *tmp_w2 = *w2;
2624                         w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2625                         w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2626
2627                         if (data1) {
2628                                 stack_zero(&tmp_data1->wlod_stack);
2629                                 stack_copy(&data1->wlod_stack,
2630                                     &tmp_data1->wlod_stack);
2631                         }
2632                         if (data2 && data2 != data1) {
2633                                 stack_zero(&tmp_data2->wlod_stack);
2634                                 stack_copy(&data2->wlod_stack,
2635                                     &tmp_data2->wlod_stack);
2636                         }
2637                         mtx_unlock_spin(&w_mtx);
2638
2639                         sbuf_printf(sb,
2640             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2641                             tmp_w1->w_name, tmp_w1->w_class->lc_name,
2642                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
2643                         if (data1) {
2644                                 sbuf_printf(sb,
2645                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2646                                     tmp_w1->w_name, tmp_w1->w_class->lc_name,
2647                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
2648                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2649                                 sbuf_printf(sb, "\n");
2650                         }
2651                         if (data2 && data2 != data1) {
2652                                 sbuf_printf(sb,
2653                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2654                                     tmp_w2->w_name, tmp_w2->w_class->lc_name,
2655                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
2656                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2657                                 sbuf_printf(sb, "\n");
2658                         }
2659                 }
2660         }
2661         mtx_lock_spin(&w_mtx);
2662         if (generation != w_generation) {
2663                 mtx_unlock_spin(&w_mtx);
2664
2665                 /*
2666                  * The graph changed while we were printing stack data,
2667                  * try again.
2668                  */
2669                 req->oldidx = 0;
2670                 sbuf_clear(sb);
2671                 goto restart;
2672         }
2673         mtx_unlock_spin(&w_mtx);
2674
2675         /* Free temporary storage space. */
2676         free(tmp_data1, M_TEMP);
2677         free(tmp_data2, M_TEMP);
2678         free(tmp_w1, M_TEMP);
2679         free(tmp_w2, M_TEMP);
2680
2681         sbuf_finish(sb);
2682         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2683         sbuf_delete(sb);
2684
2685         return (error);
2686 }
2687
2688 static int
2689 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2690 {
2691         static const struct {
2692                 enum witness_channel channel;
2693                 const char *name;
2694         } channels[] = {
2695                 { WITNESS_CONSOLE, "console" },
2696                 { WITNESS_LOG, "log" },
2697                 { WITNESS_NONE, "none" },
2698         };
2699         char buf[16];
2700         u_int i;
2701         int error;
2702
2703         buf[0] = '\0';
2704         for (i = 0; i < nitems(channels); i++)
2705                 if (witness_channel == channels[i].channel) {
2706                         snprintf(buf, sizeof(buf), "%s", channels[i].name);
2707                         break;
2708                 }
2709
2710         error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2711         if (error != 0 || req->newptr == NULL)
2712                 return (error);
2713
2714         error = EINVAL;
2715         for (i = 0; i < nitems(channels); i++)
2716                 if (strcmp(channels[i].name, buf) == 0) {
2717                         witness_channel = channels[i].channel;
2718                         error = 0;
2719                         break;
2720                 }
2721         return (error);
2722 }
2723
2724 static int
2725 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2726 {
2727         struct witness *w;
2728         struct sbuf *sb;
2729         int error;
2730
2731         if (witness_watch < 1) {
2732                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2733                 return (error);
2734         }
2735         if (witness_cold) {
2736                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2737                 return (error);
2738         }
2739         error = 0;
2740
2741         error = sysctl_wire_old_buffer(req, 0);
2742         if (error != 0)
2743                 return (error);
2744         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2745         if (sb == NULL)
2746                 return (ENOMEM);
2747         sbuf_printf(sb, "\n");
2748
2749         mtx_lock_spin(&w_mtx);
2750         STAILQ_FOREACH(w, &w_all, w_list)
2751                 w->w_displayed = 0;
2752         STAILQ_FOREACH(w, &w_all, w_list)
2753                 witness_add_fullgraph(sb, w);
2754         mtx_unlock_spin(&w_mtx);
2755
2756         /*
2757          * Close the sbuf and return to userland.
2758          */
2759         error = sbuf_finish(sb);
2760         sbuf_delete(sb);
2761
2762         return (error);
2763 }
2764
2765 static int
2766 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2767 {
2768         int error, value;
2769
2770         value = witness_watch;
2771         error = sysctl_handle_int(oidp, &value, 0, req);
2772         if (error != 0 || req->newptr == NULL)
2773                 return (error);
2774         if (value > 1 || value < -1 ||
2775             (witness_watch == -1 && value != witness_watch))
2776                 return (EINVAL);
2777         witness_watch = value;
2778         return (0);
2779 }
2780
2781 static void
2782 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2783 {
2784         int i;
2785
2786         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2787                 return;
2788         w->w_displayed = 1;
2789
2790         WITNESS_INDEX_ASSERT(w->w_index);
2791         for (i = 1; i <= w_max_used_index; i++) {
2792                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2793                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2794                             w_data[i].w_name);
2795                         witness_add_fullgraph(sb, &w_data[i]);
2796                 }
2797         }
2798 }
2799
2800 /*
2801  * A simple hash function. Takes a key pointer and a key size. If size == 0,
2802  * interprets the key as a string and reads until the null
2803  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2804  * hash value computed from the key.
2805  */
2806 static uint32_t
2807 witness_hash_djb2(const uint8_t *key, uint32_t size)
2808 {
2809         unsigned int hash = 5381;
2810         int i;
2811
2812         /* hash = hash * 33 + key[i] */
2813         if (size)
2814                 for (i = 0; i < size; i++)
2815                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2816         else
2817                 for (i = 0; key[i] != 0; i++)
2818                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
2819
2820         return (hash);
2821 }
2822
2823
2824 /*
2825  * Initializes the two witness hash tables. Called exactly once from
2826  * witness_initialize().
2827  */
2828 static void
2829 witness_init_hash_tables(void)
2830 {
2831         int i;
2832
2833         MPASS(witness_cold);
2834
2835         /* Initialize the hash tables. */
2836         for (i = 0; i < WITNESS_HASH_SIZE; i++)
2837                 w_hash.wh_array[i] = NULL;
2838
2839         w_hash.wh_size = WITNESS_HASH_SIZE;
2840         w_hash.wh_count = 0;
2841
2842         /* Initialize the lock order data hash. */
2843         w_lofree = NULL;
2844         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2845                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2846                 w_lodata[i].wlod_next = w_lofree;
2847                 w_lofree = &w_lodata[i];
2848         }
2849         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2850         w_lohash.wloh_count = 0;
2851         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2852                 w_lohash.wloh_array[i] = NULL;
2853 }
2854
2855 static struct witness *
2856 witness_hash_get(const char *key)
2857 {
2858         struct witness *w;
2859         uint32_t hash;
2860
2861         MPASS(key != NULL);
2862         if (witness_cold == 0)
2863                 mtx_assert(&w_mtx, MA_OWNED);
2864         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2865         w = w_hash.wh_array[hash];
2866         while (w != NULL) {
2867                 if (strcmp(w->w_name, key) == 0)
2868                         goto out;
2869                 w = w->w_hash_next;
2870         }
2871
2872 out:
2873         return (w);
2874 }
2875
2876 static void
2877 witness_hash_put(struct witness *w)
2878 {
2879         uint32_t hash;
2880
2881         MPASS(w != NULL);
2882         MPASS(w->w_name != NULL);
2883         if (witness_cold == 0)
2884                 mtx_assert(&w_mtx, MA_OWNED);
2885         KASSERT(witness_hash_get(w->w_name) == NULL,
2886             ("%s: trying to add a hash entry that already exists!", __func__));
2887         KASSERT(w->w_hash_next == NULL,
2888             ("%s: w->w_hash_next != NULL", __func__));
2889
2890         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2891         w->w_hash_next = w_hash.wh_array[hash];
2892         w_hash.wh_array[hash] = w;
2893         w_hash.wh_count++;
2894 }
2895
2896
2897 static struct witness_lock_order_data *
2898 witness_lock_order_get(struct witness *parent, struct witness *child)
2899 {
2900         struct witness_lock_order_data *data = NULL;
2901         struct witness_lock_order_key key;
2902         unsigned int hash;
2903
2904         MPASS(parent != NULL && child != NULL);
2905         key.from = parent->w_index;
2906         key.to = child->w_index;
2907         WITNESS_INDEX_ASSERT(key.from);
2908         WITNESS_INDEX_ASSERT(key.to);
2909         if ((w_rmatrix[parent->w_index][child->w_index]
2910             & WITNESS_LOCK_ORDER_KNOWN) == 0)
2911                 goto out;
2912
2913         hash = witness_hash_djb2((const char*)&key,
2914             sizeof(key)) % w_lohash.wloh_size;
2915         data = w_lohash.wloh_array[hash];
2916         while (data != NULL) {
2917                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2918                         break;
2919                 data = data->wlod_next;
2920         }
2921
2922 out:
2923         return (data);
2924 }
2925
2926 /*
2927  * Verify that parent and child have a known relationship, are not the same,
2928  * and child is actually a child of parent.  This is done without w_mtx
2929  * to avoid contention in the common case.
2930  */
2931 static int
2932 witness_lock_order_check(struct witness *parent, struct witness *child)
2933 {
2934
2935         if (parent != child &&
2936             w_rmatrix[parent->w_index][child->w_index]
2937             & WITNESS_LOCK_ORDER_KNOWN &&
2938             isitmychild(parent, child))
2939                 return (1);
2940
2941         return (0);
2942 }
2943
2944 static int
2945 witness_lock_order_add(struct witness *parent, struct witness *child)
2946 {
2947         struct witness_lock_order_data *data = NULL;
2948         struct witness_lock_order_key key;
2949         unsigned int hash;
2950
2951         MPASS(parent != NULL && child != NULL);
2952         key.from = parent->w_index;
2953         key.to = child->w_index;
2954         WITNESS_INDEX_ASSERT(key.from);
2955         WITNESS_INDEX_ASSERT(key.to);
2956         if (w_rmatrix[parent->w_index][child->w_index]
2957             & WITNESS_LOCK_ORDER_KNOWN)
2958                 return (1);
2959
2960         hash = witness_hash_djb2((const char*)&key,
2961             sizeof(key)) % w_lohash.wloh_size;
2962         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2963         data = w_lofree;
2964         if (data == NULL)
2965                 return (0);
2966         w_lofree = data->wlod_next;
2967         data->wlod_next = w_lohash.wloh_array[hash];
2968         data->wlod_key = key;
2969         w_lohash.wloh_array[hash] = data;
2970         w_lohash.wloh_count++;
2971         stack_zero(&data->wlod_stack);
2972         stack_save(&data->wlod_stack);
2973         return (1);
2974 }
2975
2976 /* Call this whenver the structure of the witness graph changes. */
2977 static void
2978 witness_increment_graph_generation(void)
2979 {
2980
2981         if (witness_cold == 0)
2982                 mtx_assert(&w_mtx, MA_OWNED);
2983         w_generation++;
2984 }
2985
2986 static int
2987 witness_output_drain(void *arg __unused, const char *data, int len)
2988 {
2989
2990         witness_output("%.*s", len, data);
2991         return (len);
2992 }
2993
2994 static void
2995 witness_debugger(int cond, const char *msg)
2996 {
2997         char buf[32];
2998         struct sbuf sb;
2999         struct stack st;
3000
3001         if (!cond)
3002                 return;
3003
3004         if (witness_trace) {
3005                 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3006                 sbuf_set_drain(&sb, witness_output_drain, NULL);
3007
3008                 stack_zero(&st);
3009                 stack_save(&st);
3010                 witness_output("stack backtrace:\n");
3011                 stack_sbuf_print_ddb(&sb, &st);
3012
3013                 sbuf_finish(&sb);
3014         }
3015
3016 #ifdef KDB
3017         if (witness_kdb)
3018                 kdb_enter(KDB_WHY_WITNESS, msg);
3019 #endif
3020 }