The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/kern/subr_witness.c

Version: -  FREEBSD  -  FREEBSD-13-STABLE  -  FREEBSD-13-0  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  l41  -  OPENBSD  -  linux-2.6  -  MK84  -  PLAN9  -  xnu-8792 
SearchContext: -  none  -  3  -  10 

    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: releng/9.0/sys/kern/subr_witness.c 218909 2011-02-21 09:01:34Z brucec $");
   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/systm.h>
  110 
  111 #ifdef DDB
  112 #include <ddb/ddb.h>
  113 #endif
  114 
  115 #include <machine/stdarg.h>
  116 
  117 #if !defined(DDB) && !defined(STACK)
  118 #error "DDB or STACK options are required for WITNESS"
  119 #endif
  120 
  121 /* Note that these traces do not work with KTR_ALQ. */
  122 #if 0
  123 #define KTR_WITNESS     KTR_SUBSYS
  124 #else
  125 #define KTR_WITNESS     0
  126 #endif
  127 
  128 #define LI_RECURSEMASK  0x0000ffff      /* Recursion depth of lock instance. */
  129 #define LI_EXCLUSIVE    0x00010000      /* Exclusive lock instance. */
  130 #define LI_NORELEASE    0x00020000      /* Lock not allowed to be released. */
  131 
  132 /* Define this to check for blessed mutexes */
  133 #undef BLESSING
  134 
  135 #define WITNESS_COUNT           1024
  136 #define WITNESS_CHILDCOUNT      (WITNESS_COUNT * 4)
  137 #define WITNESS_HASH_SIZE       251     /* Prime, gives load factor < 2 */
  138 #define WITNESS_PENDLIST        768
  139 
  140 /* Allocate 256 KB of stack data space */
  141 #define WITNESS_LO_DATA_COUNT   2048
  142 
  143 /* Prime, gives load factor of ~2 at full load */
  144 #define WITNESS_LO_HASH_SIZE    1021
  145 
  146 /*
  147  * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
  148  * will hold LOCK_NCHILDREN locks.  We handle failure ok, and we should
  149  * probably be safe for the most part, but it's still a SWAG.
  150  */
  151 #define LOCK_NCHILDREN  5
  152 #define LOCK_CHILDCOUNT 2048
  153 
  154 #define MAX_W_NAME      64
  155 
  156 #define BADSTACK_SBUF_SIZE      (256 * WITNESS_COUNT)
  157 #define FULLGRAPH_SBUF_SIZE     512
  158 
  159 /*
  160  * These flags go in the witness relationship matrix and describe the
  161  * relationship between any two struct witness objects.
  162  */
  163 #define WITNESS_UNRELATED        0x00    /* No lock order relation. */
  164 #define WITNESS_PARENT           0x01    /* Parent, aka direct ancestor. */
  165 #define WITNESS_ANCESTOR         0x02    /* Direct or indirect ancestor. */
  166 #define WITNESS_CHILD            0x04    /* Child, aka direct descendant. */
  167 #define WITNESS_DESCENDANT       0x08    /* Direct or indirect descendant. */
  168 #define WITNESS_ANCESTOR_MASK    (WITNESS_PARENT | WITNESS_ANCESTOR)
  169 #define WITNESS_DESCENDANT_MASK  (WITNESS_CHILD | WITNESS_DESCENDANT)
  170 #define WITNESS_RELATED_MASK                                            \
  171         (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
  172 #define WITNESS_REVERSAL         0x10    /* A lock order reversal has been
  173                                           * observed. */
  174 #define WITNESS_RESERVED1        0x20    /* Unused flag, reserved. */
  175 #define WITNESS_RESERVED2        0x40    /* Unused flag, reserved. */
  176 #define WITNESS_LOCK_ORDER_KNOWN 0x80    /* This lock order is known. */
  177 
  178 /* Descendant to ancestor flags */
  179 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
  180 
  181 /* Ancestor to descendant flags */
  182 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
  183 
  184 #define WITNESS_INDEX_ASSERT(i)                                         \
  185         MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
  186 
  187 MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
  188 
  189 /*
  190  * Lock instances.  A lock instance is the data associated with a lock while
  191  * it is held by witness.  For example, a lock instance will hold the
  192  * recursion count of a lock.  Lock instances are held in lists.  Spin locks
  193  * are held in a per-cpu list while sleep locks are held in per-thread list.
  194  */
  195 struct lock_instance {
  196         struct lock_object      *li_lock;
  197         const char              *li_file;
  198         int                     li_line;
  199         u_int                   li_flags;
  200 };
  201 
  202 /*
  203  * A simple list type used to build the list of locks held by a thread
  204  * or CPU.  We can't simply embed the list in struct lock_object since a
  205  * lock may be held by more than one thread if it is a shared lock.  Locks
  206  * are added to the head of the list, so we fill up each list entry from
  207  * "the back" logically.  To ease some of the arithmetic, we actually fill
  208  * in each list entry the normal way (children[0] then children[1], etc.) but
  209  * when we traverse the list we read children[count-1] as the first entry
  210  * down to children[0] as the final entry.
  211  */
  212 struct lock_list_entry {
  213         struct lock_list_entry  *ll_next;
  214         struct lock_instance    ll_children[LOCK_NCHILDREN];
  215         u_int                   ll_count;
  216 };
  217 
  218 /*
  219  * The main witness structure. One of these per named lock type in the system
  220  * (for example, "vnode interlock").
  221  */
  222 struct witness {
  223         char                    w_name[MAX_W_NAME];
  224         uint32_t                w_index;  /* Index in the relationship matrix */
  225         struct lock_class       *w_class;
  226         STAILQ_ENTRY(witness)   w_list;         /* List of all witnesses. */
  227         STAILQ_ENTRY(witness)   w_typelist;     /* Witnesses of a type. */
  228         struct witness          *w_hash_next; /* Linked list in hash buckets. */
  229         const char              *w_file; /* File where last acquired */
  230         uint32_t                w_line; /* Line where last acquired */
  231         uint32_t                w_refcount;
  232         uint16_t                w_num_ancestors; /* direct/indirect
  233                                                   * ancestor count */
  234         uint16_t                w_num_descendants; /* direct/indirect
  235                                                     * descendant count */
  236         int16_t                 w_ddb_level;
  237         unsigned                w_displayed:1;
  238         unsigned                w_reversed:1;
  239 };
  240 
  241 STAILQ_HEAD(witness_list, witness);
  242 
  243 /*
  244  * The witness hash table. Keys are witness names (const char *), elements are
  245  * witness objects (struct witness *).
  246  */
  247 struct witness_hash {
  248         struct witness  *wh_array[WITNESS_HASH_SIZE];
  249         uint32_t        wh_size;
  250         uint32_t        wh_count;
  251 };
  252 
  253 /*
  254  * Key type for the lock order data hash table.
  255  */
  256 struct witness_lock_order_key {
  257         uint16_t        from;
  258         uint16_t        to;
  259 };
  260 
  261 struct witness_lock_order_data {
  262         struct stack                    wlod_stack;
  263         struct witness_lock_order_key   wlod_key;
  264         struct witness_lock_order_data  *wlod_next;
  265 };
  266 
  267 /*
  268  * The witness lock order data hash table. Keys are witness index tuples
  269  * (struct witness_lock_order_key), elements are lock order data objects
  270  * (struct witness_lock_order_data). 
  271  */
  272 struct witness_lock_order_hash {
  273         struct witness_lock_order_data  *wloh_array[WITNESS_LO_HASH_SIZE];
  274         u_int   wloh_size;
  275         u_int   wloh_count;
  276 };
  277 
  278 #ifdef BLESSING
  279 struct witness_blessed {
  280         const char      *b_lock1;
  281         const char      *b_lock2;
  282 };
  283 #endif
  284 
  285 struct witness_pendhelp {
  286         const char              *wh_type;
  287         struct lock_object      *wh_lock;
  288 };
  289 
  290 struct witness_order_list_entry {
  291         const char              *w_name;
  292         struct lock_class       *w_class;
  293 };
  294 
  295 /*
  296  * Returns 0 if one of the locks is a spin lock and the other is not.
  297  * Returns 1 otherwise.
  298  */
  299 static __inline int
  300 witness_lock_type_equal(struct witness *w1, struct witness *w2)
  301 {
  302 
  303         return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
  304                 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
  305 }
  306 
  307 static __inline int
  308 witness_lock_order_key_empty(const struct witness_lock_order_key *key)
  309 {
  310 
  311         return (key->from == 0 && key->to == 0);
  312 }
  313 
  314 static __inline int
  315 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
  316     const struct witness_lock_order_key *b)
  317 {
  318 
  319         return (a->from == b->from && a->to == b->to);
  320 }
  321 
  322 static int      _isitmyx(struct witness *w1, struct witness *w2, int rmask,
  323                     const char *fname);
  324 #ifdef KDB
  325 static void     _witness_debugger(int cond, const char *msg);
  326 #endif
  327 static void     adopt(struct witness *parent, struct witness *child);
  328 #ifdef BLESSING
  329 static int      blessed(struct witness *, struct witness *);
  330 #endif
  331 static void     depart(struct witness *w);
  332 static struct witness   *enroll(const char *description,
  333                             struct lock_class *lock_class);
  334 static struct lock_instance     *find_instance(struct lock_list_entry *list,
  335                                     struct lock_object *lock);
  336 static int      isitmychild(struct witness *parent, struct witness *child);
  337 static int      isitmydescendant(struct witness *parent, struct witness *child);
  338 static void     itismychild(struct witness *parent, struct witness *child);
  339 static int      sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
  340 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
  341 static int      sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
  342 static void     witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
  343 #ifdef DDB
  344 static void     witness_ddb_compute_levels(void);
  345 static void     witness_ddb_display(int(*)(const char *fmt, ...));
  346 static void     witness_ddb_display_descendants(int(*)(const char *fmt, ...),
  347                     struct witness *, int indent);
  348 static void     witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
  349                     struct witness_list *list);
  350 static void     witness_ddb_level_descendants(struct witness *parent, int l);
  351 static void     witness_ddb_list(struct thread *td);
  352 #endif
  353 static void     witness_free(struct witness *m);
  354 static struct witness   *witness_get(void);
  355 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
  356 static struct witness   *witness_hash_get(const char *key);
  357 static void     witness_hash_put(struct witness *w);
  358 static void     witness_init_hash_tables(void);
  359 static void     witness_increment_graph_generation(void);
  360 static void     witness_lock_list_free(struct lock_list_entry *lle);
  361 static struct lock_list_entry   *witness_lock_list_get(void);
  362 static int      witness_lock_order_add(struct witness *parent,
  363                     struct witness *child);
  364 static int      witness_lock_order_check(struct witness *parent,
  365                     struct witness *child);
  366 static struct witness_lock_order_data   *witness_lock_order_get(
  367                                             struct witness *parent,
  368                                             struct witness *child);
  369 static void     witness_list_lock(struct lock_instance *instance,
  370                     int (*prnt)(const char *fmt, ...));
  371 static void     witness_setflag(struct lock_object *lock, int flag, int set);
  372 
  373 #ifdef KDB
  374 #define witness_debugger(c)     _witness_debugger(c, __func__)
  375 #else
  376 #define witness_debugger(c)
  377 #endif
  378 
  379 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking");
  380 
  381 /*
  382  * If set to 0, lock order checking is disabled.  If set to -1,
  383  * witness is completely disabled.  Otherwise witness performs full
  384  * lock order checking for all locks.  At runtime, lock order checking
  385  * may be toggled.  However, witness cannot be reenabled once it is
  386  * completely disabled.
  387  */
  388 static int witness_watch = 1;
  389 TUNABLE_INT("debug.witness.watch", &witness_watch);
  390 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
  391     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
  392 
  393 #ifdef KDB
  394 /*
  395  * When KDB is enabled and witness_kdb is 1, it will cause the system
  396  * to drop into kdebug() when:
  397  *      - a lock hierarchy violation occurs
  398  *      - locks are held when going to sleep.
  399  */
  400 #ifdef WITNESS_KDB
  401 int     witness_kdb = 1;
  402 #else
  403 int     witness_kdb = 0;
  404 #endif
  405 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
  406 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
  407 
  408 /*
  409  * When KDB is enabled and witness_trace is 1, it will cause the system
  410  * to print a stack trace:
  411  *      - a lock hierarchy violation occurs
  412  *      - locks are held when going to sleep.
  413  */
  414 int     witness_trace = 1;
  415 TUNABLE_INT("debug.witness.trace", &witness_trace);
  416 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
  417 #endif /* KDB */
  418 
  419 #ifdef WITNESS_SKIPSPIN
  420 int     witness_skipspin = 1;
  421 #else
  422 int     witness_skipspin = 0;
  423 #endif
  424 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
  425 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
  426     0, "");
  427 
  428 /*
  429  * Call this to print out the relations between locks.
  430  */
  431 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
  432     NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
  433 
  434 /*
  435  * Call this to print out the witness faulty stacks.
  436  */
  437 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
  438     NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
  439 
  440 static struct mtx w_mtx;
  441 
  442 /* w_list */
  443 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
  444 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
  445 
  446 /* w_typelist */
  447 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
  448 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
  449 
  450 /* lock list */
  451 static struct lock_list_entry *w_lock_list_free = NULL;
  452 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
  453 static u_int pending_cnt;
  454 
  455 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
  456 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
  457 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
  458 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
  459     "");
  460 
  461 static struct witness *w_data;
  462 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
  463 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
  464 static struct witness_hash w_hash;      /* The witness hash table. */
  465 
  466 /* The lock order data hash */
  467 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
  468 static struct witness_lock_order_data *w_lofree = NULL;
  469 static struct witness_lock_order_hash w_lohash;
  470 static int w_max_used_index = 0;
  471 static unsigned int w_generation = 0;
  472 static const char w_notrunning[] = "Witness not running\n";
  473 static const char w_stillcold[] = "Witness is still cold\n";
  474 
  475 
  476 static struct witness_order_list_entry order_lists[] = {
  477         /*
  478          * sx locks
  479          */
  480         { "proctree", &lock_class_sx },
  481         { "allproc", &lock_class_sx },
  482         { "allprison", &lock_class_sx },
  483         { NULL, NULL },
  484         /*
  485          * Various mutexes
  486          */
  487         { "Giant", &lock_class_mtx_sleep },
  488         { "pipe mutex", &lock_class_mtx_sleep },
  489         { "sigio lock", &lock_class_mtx_sleep },
  490         { "process group", &lock_class_mtx_sleep },
  491         { "process lock", &lock_class_mtx_sleep },
  492         { "session", &lock_class_mtx_sleep },
  493         { "uidinfo hash", &lock_class_rw },
  494 #ifdef  HWPMC_HOOKS
  495         { "pmc-sleep", &lock_class_mtx_sleep },
  496 #endif
  497         { "time lock", &lock_class_mtx_sleep },
  498         { NULL, NULL },
  499         /*
  500          * Sockets
  501          */
  502         { "accept", &lock_class_mtx_sleep },
  503         { "so_snd", &lock_class_mtx_sleep },
  504         { "so_rcv", &lock_class_mtx_sleep },
  505         { "sellck", &lock_class_mtx_sleep },
  506         { NULL, NULL },
  507         /*
  508          * Routing
  509          */
  510         { "so_rcv", &lock_class_mtx_sleep },
  511         { "radix node head", &lock_class_rw },
  512         { "rtentry", &lock_class_mtx_sleep },
  513         { "ifaddr", &lock_class_mtx_sleep },
  514         { NULL, NULL },
  515         /*
  516          * IPv4 multicast:
  517          * protocol locks before interface locks, after UDP locks.
  518          */
  519         { "udpinp", &lock_class_rw },
  520         { "in_multi_mtx", &lock_class_mtx_sleep },
  521         { "igmp_mtx", &lock_class_mtx_sleep },
  522         { "if_addr_mtx", &lock_class_mtx_sleep },
  523         { NULL, NULL },
  524         /*
  525          * IPv6 multicast:
  526          * protocol locks before interface locks, after UDP locks.
  527          */
  528         { "udpinp", &lock_class_rw },
  529         { "in6_multi_mtx", &lock_class_mtx_sleep },
  530         { "mld_mtx", &lock_class_mtx_sleep },
  531         { "if_addr_mtx", &lock_class_mtx_sleep },
  532         { NULL, NULL },
  533         /*
  534          * UNIX Domain Sockets
  535          */
  536         { "unp_global_rwlock", &lock_class_rw },
  537         { "unp_list_lock", &lock_class_mtx_sleep },
  538         { "unp", &lock_class_mtx_sleep },
  539         { "so_snd", &lock_class_mtx_sleep },
  540         { NULL, NULL },
  541         /*
  542          * UDP/IP
  543          */
  544         { "udp", &lock_class_rw },
  545         { "udpinp", &lock_class_rw },
  546         { "so_snd", &lock_class_mtx_sleep },
  547         { NULL, NULL },
  548         /*
  549          * TCP/IP
  550          */
  551         { "tcp", &lock_class_rw },
  552         { "tcpinp", &lock_class_rw },
  553         { "so_snd", &lock_class_mtx_sleep },
  554         { NULL, NULL },
  555         /*
  556          * netatalk
  557          */
  558         { "ddp_list_mtx", &lock_class_mtx_sleep },
  559         { "ddp_mtx", &lock_class_mtx_sleep },
  560         { NULL, NULL },
  561         /*
  562          * BPF
  563          */
  564         { "bpf global lock", &lock_class_mtx_sleep },
  565         { "bpf interface lock", &lock_class_mtx_sleep },
  566         { "bpf cdev lock", &lock_class_mtx_sleep },
  567         { NULL, NULL },
  568         /*
  569          * NFS server
  570          */
  571         { "nfsd_mtx", &lock_class_mtx_sleep },
  572         { "so_snd", &lock_class_mtx_sleep },
  573         { NULL, NULL },
  574 
  575         /*
  576          * IEEE 802.11
  577          */
  578         { "802.11 com lock", &lock_class_mtx_sleep},
  579         { NULL, NULL },
  580         /*
  581          * Network drivers
  582          */
  583         { "network driver", &lock_class_mtx_sleep},
  584         { NULL, NULL },
  585 
  586         /*
  587          * Netgraph
  588          */
  589         { "ng_node", &lock_class_mtx_sleep },
  590         { "ng_worklist", &lock_class_mtx_sleep },
  591         { NULL, NULL },
  592         /*
  593          * CDEV
  594          */
  595         { "system map", &lock_class_mtx_sleep },
  596         { "vm page queue mutex", &lock_class_mtx_sleep },
  597         { "vnode interlock", &lock_class_mtx_sleep },
  598         { "cdev", &lock_class_mtx_sleep },
  599         { NULL, NULL },
  600         /*
  601          * VM
  602          * 
  603          */
  604         { "vm object", &lock_class_mtx_sleep },
  605         { "page lock", &lock_class_mtx_sleep },
  606         { "vm page queue mutex", &lock_class_mtx_sleep },
  607         { "pmap", &lock_class_mtx_sleep },
  608         { NULL, NULL },
  609         /*
  610          * kqueue/VFS interaction
  611          */
  612         { "kqueue", &lock_class_mtx_sleep },
  613         { "struct mount mtx", &lock_class_mtx_sleep },
  614         { "vnode interlock", &lock_class_mtx_sleep },
  615         { NULL, NULL },
  616         /*
  617          * ZFS locking
  618          */
  619         { "dn->dn_mtx", &lock_class_sx },
  620         { "dr->dt.di.dr_mtx", &lock_class_sx },
  621         { "db->db_mtx", &lock_class_sx },
  622         { NULL, NULL },
  623         /*
  624          * spin locks
  625          */
  626 #ifdef SMP
  627         { "ap boot", &lock_class_mtx_spin },
  628 #endif
  629         { "rm.mutex_mtx", &lock_class_mtx_spin },
  630         { "sio", &lock_class_mtx_spin },
  631         { "scrlock", &lock_class_mtx_spin },
  632 #ifdef __i386__
  633         { "cy", &lock_class_mtx_spin },
  634 #endif
  635 #ifdef __sparc64__
  636         { "pcib_mtx", &lock_class_mtx_spin },
  637         { "rtc_mtx", &lock_class_mtx_spin },
  638 #endif
  639         { "scc_hwmtx", &lock_class_mtx_spin },
  640         { "uart_hwmtx", &lock_class_mtx_spin },
  641         { "fast_taskqueue", &lock_class_mtx_spin },
  642         { "intr table", &lock_class_mtx_spin },
  643 #ifdef  HWPMC_HOOKS
  644         { "pmc-per-proc", &lock_class_mtx_spin },
  645 #endif
  646         { "process slock", &lock_class_mtx_spin },
  647         { "sleepq chain", &lock_class_mtx_spin },
  648         { "umtx lock", &lock_class_mtx_spin },
  649         { "rm_spinlock", &lock_class_mtx_spin },
  650         { "turnstile chain", &lock_class_mtx_spin },
  651         { "turnstile lock", &lock_class_mtx_spin },
  652         { "sched lock", &lock_class_mtx_spin },
  653         { "td_contested", &lock_class_mtx_spin },
  654         { "callout", &lock_class_mtx_spin },
  655         { "entropy harvest mutex", &lock_class_mtx_spin },
  656         { "syscons video lock", &lock_class_mtx_spin },
  657 #ifdef SMP
  658         { "smp rendezvous", &lock_class_mtx_spin },
  659 #endif
  660 #ifdef __powerpc__
  661         { "tlb0", &lock_class_mtx_spin },
  662 #endif
  663         /*
  664          * leaf locks
  665          */
  666         { "intrcnt", &lock_class_mtx_spin },
  667         { "icu", &lock_class_mtx_spin },
  668 #if defined(SMP) && defined(__sparc64__)
  669         { "ipi", &lock_class_mtx_spin },
  670 #endif
  671 #ifdef __i386__
  672         { "allpmaps", &lock_class_mtx_spin },
  673         { "descriptor tables", &lock_class_mtx_spin },
  674 #endif
  675         { "clk", &lock_class_mtx_spin },
  676         { "cpuset", &lock_class_mtx_spin },
  677         { "mprof lock", &lock_class_mtx_spin },
  678         { "zombie lock", &lock_class_mtx_spin },
  679         { "ALD Queue", &lock_class_mtx_spin },
  680 #ifdef __ia64__
  681         { "MCA spin lock", &lock_class_mtx_spin },
  682 #endif
  683 #if defined(__i386__) || defined(__amd64__)
  684         { "pcicfg", &lock_class_mtx_spin },
  685         { "NDIS thread lock", &lock_class_mtx_spin },
  686 #endif
  687         { "tw_osl_io_lock", &lock_class_mtx_spin },
  688         { "tw_osl_q_lock", &lock_class_mtx_spin },
  689         { "tw_cl_io_lock", &lock_class_mtx_spin },
  690         { "tw_cl_intr_lock", &lock_class_mtx_spin },
  691         { "tw_cl_gen_lock", &lock_class_mtx_spin },
  692 #ifdef  HWPMC_HOOKS
  693         { "pmc-leaf", &lock_class_mtx_spin },
  694 #endif
  695         { "blocked lock", &lock_class_mtx_spin },
  696         { NULL, NULL },
  697         { NULL, NULL }
  698 };
  699 
  700 #ifdef BLESSING
  701 /*
  702  * Pairs of locks which have been blessed
  703  * Don't complain about order problems with blessed locks
  704  */
  705 static struct witness_blessed blessed_list[] = {
  706 };
  707 static int blessed_count =
  708         sizeof(blessed_list) / sizeof(struct witness_blessed);
  709 #endif
  710 
  711 /*
  712  * This global is set to 0 once it becomes safe to use the witness code.
  713  */
  714 static int witness_cold = 1;
  715 
  716 /*
  717  * This global is set to 1 once the static lock orders have been enrolled
  718  * so that a warning can be issued for any spin locks enrolled later.
  719  */
  720 static int witness_spin_warn = 0;
  721 
  722 /*
  723  * The WITNESS-enabled diagnostic code.  Note that the witness code does
  724  * assume that the early boot is single-threaded at least until after this
  725  * routine is completed.
  726  */
  727 static void
  728 witness_initialize(void *dummy __unused)
  729 {
  730         struct lock_object *lock;
  731         struct witness_order_list_entry *order;
  732         struct witness *w, *w1;
  733         int i;
  734 
  735         w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
  736             M_NOWAIT | M_ZERO);
  737 
  738         /*
  739          * We have to release Giant before initializing its witness
  740          * structure so that WITNESS doesn't get confused.
  741          */
  742         mtx_unlock(&Giant);
  743         mtx_assert(&Giant, MA_NOTOWNED);
  744 
  745         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
  746         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
  747             MTX_NOWITNESS | MTX_NOPROFILE);
  748         for (i = WITNESS_COUNT - 1; i >= 0; i--) {
  749                 w = &w_data[i];
  750                 memset(w, 0, sizeof(*w));
  751                 w_data[i].w_index = i;  /* Witness index never changes. */
  752                 witness_free(w);
  753         }
  754         KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
  755             ("%s: Invalid list of free witness objects", __func__));
  756 
  757         /* Witness with index 0 is not used to aid in debugging. */
  758         STAILQ_REMOVE_HEAD(&w_free, w_list);
  759         w_free_cnt--;
  760 
  761         memset(w_rmatrix, 0,
  762             (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
  763 
  764         for (i = 0; i < LOCK_CHILDCOUNT; i++)
  765                 witness_lock_list_free(&w_locklistdata[i]);
  766         witness_init_hash_tables();
  767 
  768         /* First add in all the specified order lists. */
  769         for (order = order_lists; order->w_name != NULL; order++) {
  770                 w = enroll(order->w_name, order->w_class);
  771                 if (w == NULL)
  772                         continue;
  773                 w->w_file = "order list";
  774                 for (order++; order->w_name != NULL; order++) {
  775                         w1 = enroll(order->w_name, order->w_class);
  776                         if (w1 == NULL)
  777                                 continue;
  778                         w1->w_file = "order list";
  779                         itismychild(w, w1);
  780                         w = w1;
  781                 }
  782         }
  783         witness_spin_warn = 1;
  784 
  785         /* Iterate through all locks and add them to witness. */
  786         for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
  787                 lock = pending_locks[i].wh_lock;
  788                 KASSERT(lock->lo_flags & LO_WITNESS,
  789                     ("%s: lock %s is on pending list but not LO_WITNESS",
  790                     __func__, lock->lo_name));
  791                 lock->lo_witness = enroll(pending_locks[i].wh_type,
  792                     LOCK_CLASS(lock));
  793         }
  794 
  795         /* Mark the witness code as being ready for use. */
  796         witness_cold = 0;
  797 
  798         mtx_lock(&Giant);
  799 }
  800 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
  801     NULL);
  802 
  803 void
  804 witness_init(struct lock_object *lock, const char *type)
  805 {
  806         struct lock_class *class;
  807 
  808         /* Various sanity checks. */
  809         class = LOCK_CLASS(lock);
  810         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
  811             (class->lc_flags & LC_RECURSABLE) == 0)
  812                 panic("%s: lock (%s) %s can not be recursable", __func__,
  813                     class->lc_name, lock->lo_name);
  814         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  815             (class->lc_flags & LC_SLEEPABLE) == 0)
  816                 panic("%s: lock (%s) %s can not be sleepable", __func__,
  817                     class->lc_name, lock->lo_name);
  818         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
  819             (class->lc_flags & LC_UPGRADABLE) == 0)
  820                 panic("%s: lock (%s) %s can not be upgradable", __func__,
  821                     class->lc_name, lock->lo_name);
  822 
  823         /*
  824          * If we shouldn't watch this lock, then just clear lo_witness.
  825          * Otherwise, if witness_cold is set, then it is too early to
  826          * enroll this lock, so defer it to witness_initialize() by adding
  827          * it to the pending_locks list.  If it is not too early, then enroll
  828          * the lock now.
  829          */
  830         if (witness_watch < 1 || panicstr != NULL ||
  831             (lock->lo_flags & LO_WITNESS) == 0)
  832                 lock->lo_witness = NULL;
  833         else if (witness_cold) {
  834                 pending_locks[pending_cnt].wh_lock = lock;
  835                 pending_locks[pending_cnt++].wh_type = type;
  836                 if (pending_cnt > WITNESS_PENDLIST)
  837                         panic("%s: pending locks list is too small, bump it\n",
  838                             __func__);
  839         } else
  840                 lock->lo_witness = enroll(type, class);
  841 }
  842 
  843 void
  844 witness_destroy(struct lock_object *lock)
  845 {
  846         struct lock_class *class;
  847         struct witness *w;
  848 
  849         class = LOCK_CLASS(lock);
  850 
  851         if (witness_cold)
  852                 panic("lock (%s) %s destroyed while witness_cold",
  853                     class->lc_name, lock->lo_name);
  854 
  855         /* XXX: need to verify that no one holds the lock */
  856         if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
  857                 return;
  858         w = lock->lo_witness;
  859 
  860         mtx_lock_spin(&w_mtx);
  861         MPASS(w->w_refcount > 0);
  862         w->w_refcount--;
  863 
  864         if (w->w_refcount == 0)
  865                 depart(w);
  866         mtx_unlock_spin(&w_mtx);
  867 }
  868 
  869 #ifdef DDB
  870 static void
  871 witness_ddb_compute_levels(void)
  872 {
  873         struct witness *w;
  874 
  875         /*
  876          * First clear all levels.
  877          */
  878         STAILQ_FOREACH(w, &w_all, w_list)
  879                 w->w_ddb_level = -1;
  880 
  881         /*
  882          * Look for locks with no parents and level all their descendants.
  883          */
  884         STAILQ_FOREACH(w, &w_all, w_list) {
  885 
  886                 /* If the witness has ancestors (is not a root), skip it. */
  887                 if (w->w_num_ancestors > 0)
  888                         continue;
  889                 witness_ddb_level_descendants(w, 0);
  890         }
  891 }
  892 
  893 static void
  894 witness_ddb_level_descendants(struct witness *w, int l)
  895 {
  896         int i;
  897 
  898         if (w->w_ddb_level >= l)
  899                 return;
  900 
  901         w->w_ddb_level = l;
  902         l++;
  903 
  904         for (i = 1; i <= w_max_used_index; i++) {
  905                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
  906                         witness_ddb_level_descendants(&w_data[i], l);
  907         }
  908 }
  909 
  910 static void
  911 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
  912     struct witness *w, int indent)
  913 {
  914         int i;
  915 
  916         for (i = 0; i < indent; i++)
  917                 prnt(" ");
  918         prnt("%s (type: %s, depth: %d, active refs: %d)",
  919              w->w_name, w->w_class->lc_name,
  920              w->w_ddb_level, w->w_refcount);
  921         if (w->w_displayed) {
  922                 prnt(" -- (already displayed)\n");
  923                 return;
  924         }
  925         w->w_displayed = 1;
  926         if (w->w_file != NULL && w->w_line != 0)
  927                 prnt(" -- last acquired @ %s:%d\n", w->w_file,
  928                     w->w_line);
  929         else
  930                 prnt(" -- never acquired\n");
  931         indent++;
  932         WITNESS_INDEX_ASSERT(w->w_index);
  933         for (i = 1; i <= w_max_used_index; i++) {
  934                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
  935                         witness_ddb_display_descendants(prnt, &w_data[i],
  936                             indent);
  937         }
  938 }
  939 
  940 static void
  941 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
  942     struct witness_list *list)
  943 {
  944         struct witness *w;
  945 
  946         STAILQ_FOREACH(w, list, w_typelist) {
  947                 if (w->w_file == NULL || w->w_ddb_level > 0)
  948                         continue;
  949 
  950                 /* This lock has no anscestors - display its descendants. */
  951                 witness_ddb_display_descendants(prnt, w, 0);
  952         }
  953 }
  954         
  955 static void
  956 witness_ddb_display(int(*prnt)(const char *fmt, ...))
  957 {
  958         struct witness *w;
  959 
  960         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
  961         witness_ddb_compute_levels();
  962 
  963         /* Clear all the displayed flags. */
  964         STAILQ_FOREACH(w, &w_all, w_list)
  965                 w->w_displayed = 0;
  966 
  967         /*
  968          * First, handle sleep locks which have been acquired at least
  969          * once.
  970          */
  971         prnt("Sleep locks:\n");
  972         witness_ddb_display_list(prnt, &w_sleep);
  973         
  974         /*
  975          * Now do spin locks which have been acquired at least once.
  976          */
  977         prnt("\nSpin locks:\n");
  978         witness_ddb_display_list(prnt, &w_spin);
  979         
  980         /*
  981          * Finally, any locks which have not been acquired yet.
  982          */
  983         prnt("\nLocks which were never acquired:\n");
  984         STAILQ_FOREACH(w, &w_all, w_list) {
  985                 if (w->w_file != NULL || w->w_refcount == 0)
  986                         continue;
  987                 prnt("%s (type: %s, depth: %d)\n", w->w_name,
  988                     w->w_class->lc_name, w->w_ddb_level);
  989         }
  990 }
  991 #endif /* DDB */
  992 
  993 /* Trim useless garbage from filenames. */
  994 static const char *
  995 fixup_filename(const char *file)
  996 {
  997 
  998         if (file == NULL)
  999                 return (NULL);
 1000         while (strncmp(file, "../", 3) == 0)
 1001                 file += 3;
 1002         return (file);
 1003 }
 1004 
 1005 int
 1006 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
 1007 {
 1008 
 1009         if (witness_watch == -1 || panicstr != NULL)
 1010                 return (0);
 1011 
 1012         /* Require locks that witness knows about. */
 1013         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
 1014             lock2->lo_witness == NULL)
 1015                 return (EINVAL);
 1016 
 1017         mtx_assert(&w_mtx, MA_NOTOWNED);
 1018         mtx_lock_spin(&w_mtx);
 1019 
 1020         /*
 1021          * If we already have either an explicit or implied lock order that
 1022          * is the other way around, then return an error.
 1023          */
 1024         if (witness_watch &&
 1025             isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
 1026                 mtx_unlock_spin(&w_mtx);
 1027                 return (EDOOFUS);
 1028         }
 1029         
 1030         /* Try to add the new order. */
 1031         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 1032             lock2->lo_witness->w_name, lock1->lo_witness->w_name);
 1033         itismychild(lock1->lo_witness, lock2->lo_witness);
 1034         mtx_unlock_spin(&w_mtx);
 1035         return (0);
 1036 }
 1037 
 1038 void
 1039 witness_checkorder(struct lock_object *lock, int flags, const char *file,
 1040     int line, struct lock_object *interlock)
 1041 {
 1042         struct lock_list_entry *lock_list, *lle;
 1043         struct lock_instance *lock1, *lock2, *plock;
 1044         struct lock_class *class;
 1045         struct witness *w, *w1;
 1046         struct thread *td;
 1047         int i, j;
 1048 
 1049         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
 1050             panicstr != NULL)
 1051                 return;
 1052 
 1053         w = lock->lo_witness;
 1054         class = LOCK_CLASS(lock);
 1055         td = curthread;
 1056         file = fixup_filename(file);
 1057 
 1058         if (class->lc_flags & LC_SLEEPLOCK) {
 1059 
 1060                 /*
 1061                  * Since spin locks include a critical section, this check
 1062                  * implicitly enforces a lock order of all sleep locks before
 1063                  * all spin locks.
 1064                  */
 1065                 if (td->td_critnest != 0 && !kdb_active)
 1066                         panic("blockable sleep lock (%s) %s @ %s:%d",
 1067                             class->lc_name, lock->lo_name, file, line);
 1068 
 1069                 /*
 1070                  * If this is the first lock acquired then just return as
 1071                  * no order checking is needed.
 1072                  */
 1073                 lock_list = td->td_sleeplocks;
 1074                 if (lock_list == NULL || lock_list->ll_count == 0)
 1075                         return;
 1076         } else {
 1077 
 1078                 /*
 1079                  * If this is the first lock, just return as no order
 1080                  * checking is needed.  Avoid problems with thread
 1081                  * migration pinning the thread while checking if
 1082                  * spinlocks are held.  If at least one spinlock is held
 1083                  * the thread is in a safe path and it is allowed to
 1084                  * unpin it.
 1085                  */
 1086                 sched_pin();
 1087                 lock_list = PCPU_GET(spinlocks);
 1088                 if (lock_list == NULL || lock_list->ll_count == 0) {
 1089                         sched_unpin();
 1090                         return;
 1091                 }
 1092                 sched_unpin();
 1093         }
 1094 
 1095         /*
 1096          * Check to see if we are recursing on a lock we already own.  If
 1097          * so, make sure that we don't mismatch exclusive and shared lock
 1098          * acquires.
 1099          */
 1100         lock1 = find_instance(lock_list, lock);
 1101         if (lock1 != NULL) {
 1102                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
 1103                     (flags & LOP_EXCLUSIVE) == 0) {
 1104                         printf("shared lock of (%s) %s @ %s:%d\n",
 1105                             class->lc_name, lock->lo_name, file, line);
 1106                         printf("while exclusively locked from %s:%d\n",
 1107                             lock1->li_file, lock1->li_line);
 1108                         panic("share->excl");
 1109                 }
 1110                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
 1111                     (flags & LOP_EXCLUSIVE) != 0) {
 1112                         printf("exclusive lock of (%s) %s @ %s:%d\n",
 1113                             class->lc_name, lock->lo_name, file, line);
 1114                         printf("while share locked from %s:%d\n",
 1115                             lock1->li_file, lock1->li_line);
 1116                         panic("excl->share");
 1117                 }
 1118                 return;
 1119         }
 1120 
 1121         /*
 1122          * Find the previously acquired lock, but ignore interlocks.
 1123          */
 1124         plock = &lock_list->ll_children[lock_list->ll_count - 1];
 1125         if (interlock != NULL && plock->li_lock == interlock) {
 1126                 if (lock_list->ll_count > 1)
 1127                         plock =
 1128                             &lock_list->ll_children[lock_list->ll_count - 2];
 1129                 else {
 1130                         lle = lock_list->ll_next;
 1131 
 1132                         /*
 1133                          * The interlock is the only lock we hold, so
 1134                          * simply return.
 1135                          */
 1136                         if (lle == NULL)
 1137                                 return;
 1138                         plock = &lle->ll_children[lle->ll_count - 1];
 1139                 }
 1140         }
 1141         
 1142         /*
 1143          * Try to perform most checks without a lock.  If this succeeds we
 1144          * can skip acquiring the lock and return success.
 1145          */
 1146         w1 = plock->li_lock->lo_witness;
 1147         if (witness_lock_order_check(w1, w))
 1148                 return;
 1149 
 1150         /*
 1151          * Check for duplicate locks of the same type.  Note that we only
 1152          * have to check for this on the last lock we just acquired.  Any
 1153          * other cases will be caught as lock order violations.
 1154          */
 1155         mtx_lock_spin(&w_mtx);
 1156         witness_lock_order_add(w1, w);
 1157         if (w1 == w) {
 1158                 i = w->w_index;
 1159                 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
 1160                     !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
 1161                     w_rmatrix[i][i] |= WITNESS_REVERSAL;
 1162                         w->w_reversed = 1;
 1163                         mtx_unlock_spin(&w_mtx);
 1164                         printf(
 1165                             "acquiring duplicate lock of same type: \"%s\"\n", 
 1166                             w->w_name);
 1167                         printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
 1168                                plock->li_file, plock->li_line);
 1169                         printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
 1170                         witness_debugger(1);
 1171                     } else
 1172                             mtx_unlock_spin(&w_mtx);
 1173                 return;
 1174         }
 1175         mtx_assert(&w_mtx, MA_OWNED);
 1176 
 1177         /*
 1178          * If we know that the lock we are acquiring comes after
 1179          * the lock we most recently acquired in the lock order tree,
 1180          * then there is no need for any further checks.
 1181          */
 1182         if (isitmychild(w1, w))
 1183                 goto out;
 1184 
 1185         for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
 1186                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
 1187 
 1188                         MPASS(j < WITNESS_COUNT);
 1189                         lock1 = &lle->ll_children[i];
 1190 
 1191                         /*
 1192                          * Ignore the interlock the first time we see it.
 1193                          */
 1194                         if (interlock != NULL && interlock == lock1->li_lock) {
 1195                                 interlock = NULL;
 1196                                 continue;
 1197                         }
 1198 
 1199                         /*
 1200                          * If this lock doesn't undergo witness checking,
 1201                          * then skip it.
 1202                          */
 1203                         w1 = lock1->li_lock->lo_witness;
 1204                         if (w1 == NULL) {
 1205                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
 1206                                     ("lock missing witness structure"));
 1207                                 continue;
 1208                         }
 1209 
 1210                         /*
 1211                          * If we are locking Giant and this is a sleepable
 1212                          * lock, then skip it.
 1213                          */
 1214                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1215                             lock == &Giant.lock_object)
 1216                                 continue;
 1217 
 1218                         /*
 1219                          * If we are locking a sleepable lock and this lock
 1220                          * is Giant, then skip it.
 1221                          */
 1222                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1223                             lock1->li_lock == &Giant.lock_object)
 1224                                 continue;
 1225 
 1226                         /*
 1227                          * If we are locking a sleepable lock and this lock
 1228                          * isn't sleepable, we want to treat it as a lock
 1229                          * order violation to enfore a general lock order of
 1230                          * sleepable locks before non-sleepable locks.
 1231                          */
 1232                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1233                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
 1234                                 goto reversal;
 1235 
 1236                         /*
 1237                          * If we are locking Giant and this is a non-sleepable
 1238                          * lock, then treat it as a reversal.
 1239                          */
 1240                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
 1241                             lock == &Giant.lock_object)
 1242                                 goto reversal;
 1243 
 1244                         /*
 1245                          * Check the lock order hierarchy for a reveresal.
 1246                          */
 1247                         if (!isitmydescendant(w, w1))
 1248                                 continue;
 1249                 reversal:
 1250 
 1251                         /*
 1252                          * We have a lock order violation, check to see if it
 1253                          * is allowed or has already been yelled about.
 1254                          */
 1255 #ifdef BLESSING
 1256 
 1257                         /*
 1258                          * If the lock order is blessed, just bail.  We don't
 1259                          * look for other lock order violations though, which
 1260                          * may be a bug.
 1261                          */
 1262                         if (blessed(w, w1))
 1263                                 goto out;
 1264 #endif
 1265 
 1266                         /* Bail if this violation is known */
 1267                         if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
 1268                                 goto out;
 1269 
 1270                         /* Record this as a violation */
 1271                         w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
 1272                         w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
 1273                         w->w_reversed = w1->w_reversed = 1;
 1274                         witness_increment_graph_generation();
 1275                         mtx_unlock_spin(&w_mtx);
 1276                         
 1277                         /*
 1278                          * Ok, yell about it.
 1279                          */
 1280                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1281                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
 1282                                 printf(
 1283                 "lock order reversal: (sleepable after non-sleepable)\n");
 1284                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
 1285                             && lock == &Giant.lock_object)
 1286                                 printf(
 1287                 "lock order reversal: (Giant after non-sleepable)\n");
 1288                         else
 1289                                 printf("lock order reversal:\n");
 1290 
 1291                         /*
 1292                          * Try to locate an earlier lock with
 1293                          * witness w in our list.
 1294                          */
 1295                         do {
 1296                                 lock2 = &lle->ll_children[i];
 1297                                 MPASS(lock2->li_lock != NULL);
 1298                                 if (lock2->li_lock->lo_witness == w)
 1299                                         break;
 1300                                 if (i == 0 && lle->ll_next != NULL) {
 1301                                         lle = lle->ll_next;
 1302                                         i = lle->ll_count - 1;
 1303                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
 1304                                 } else
 1305                                         i--;
 1306                         } while (i >= 0);
 1307                         if (i < 0) {
 1308                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 1309                                     lock1->li_lock, lock1->li_lock->lo_name,
 1310                                     w1->w_name, lock1->li_file, lock1->li_line);
 1311                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
 1312                                     lock->lo_name, w->w_name, file, line);
 1313                         } else {
 1314                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 1315                                     lock2->li_lock, lock2->li_lock->lo_name,
 1316                                     lock2->li_lock->lo_witness->w_name,
 1317                                     lock2->li_file, lock2->li_line);
 1318                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
 1319                                     lock1->li_lock, lock1->li_lock->lo_name,
 1320                                     w1->w_name, lock1->li_file, lock1->li_line);
 1321                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
 1322                                     lock->lo_name, w->w_name, file, line);
 1323                         }
 1324                         witness_debugger(1);
 1325                         return;
 1326                 }
 1327         }
 1328 
 1329         /*
 1330          * If requested, build a new lock order.  However, don't build a new
 1331          * relationship between a sleepable lock and Giant if it is in the
 1332          * wrong direction.  The correct lock order is that sleepable locks
 1333          * always come before Giant.
 1334          */
 1335         if (flags & LOP_NEWORDER &&
 1336             !(plock->li_lock == &Giant.lock_object &&
 1337             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
 1338                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 1339                     w->w_name, plock->li_lock->lo_witness->w_name);
 1340                 itismychild(plock->li_lock->lo_witness, w);
 1341         }
 1342 out:
 1343         mtx_unlock_spin(&w_mtx);
 1344 }
 1345 
 1346 void
 1347 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
 1348 {
 1349         struct lock_list_entry **lock_list, *lle;
 1350         struct lock_instance *instance;
 1351         struct witness *w;
 1352         struct thread *td;
 1353 
 1354         if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
 1355             panicstr != NULL)
 1356                 return;
 1357         w = lock->lo_witness;
 1358         td = curthread;
 1359         file = fixup_filename(file);
 1360 
 1361         /* Determine lock list for this lock. */
 1362         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
 1363                 lock_list = &td->td_sleeplocks;
 1364         else
 1365                 lock_list = PCPU_PTR(spinlocks);
 1366 
 1367         /* Check to see if we are recursing on a lock we already own. */
 1368         instance = find_instance(*lock_list, lock);
 1369         if (instance != NULL) {
 1370                 instance->li_flags++;
 1371                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
 1372                     td->td_proc->p_pid, lock->lo_name,
 1373                     instance->li_flags & LI_RECURSEMASK);
 1374                 instance->li_file = file;
 1375                 instance->li_line = line;
 1376                 return;
 1377         }
 1378 
 1379         /* Update per-witness last file and line acquire. */
 1380         w->w_file = file;
 1381         w->w_line = line;
 1382 
 1383         /* Find the next open lock instance in the list and fill it. */
 1384         lle = *lock_list;
 1385         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
 1386                 lle = witness_lock_list_get();
 1387                 if (lle == NULL)
 1388                         return;
 1389                 lle->ll_next = *lock_list;
 1390                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
 1391                     td->td_proc->p_pid, lle);
 1392                 *lock_list = lle;
 1393         }
 1394         instance = &lle->ll_children[lle->ll_count++];
 1395         instance->li_lock = lock;
 1396         instance->li_line = line;
 1397         instance->li_file = file;
 1398         if ((flags & LOP_EXCLUSIVE) != 0)
 1399                 instance->li_flags = LI_EXCLUSIVE;
 1400         else
 1401                 instance->li_flags = 0;
 1402         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
 1403             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
 1404 }
 1405 
 1406 void
 1407 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
 1408 {
 1409         struct lock_instance *instance;
 1410         struct lock_class *class;
 1411 
 1412         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
 1413         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
 1414                 return;
 1415         class = LOCK_CLASS(lock);
 1416         file = fixup_filename(file);
 1417         if (witness_watch) {
 1418                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1419                         panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
 1420                             class->lc_name, lock->lo_name, file, line);
 1421                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
 1422                         panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
 1423                             class->lc_name, lock->lo_name, file, line);
 1424         }
 1425         instance = find_instance(curthread->td_sleeplocks, lock);
 1426         if (instance == NULL)
 1427                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
 1428                     class->lc_name, lock->lo_name, file, line);
 1429         if (witness_watch) {
 1430                 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
 1431                         panic("upgrade of exclusive lock (%s) %s @ %s:%d",
 1432                             class->lc_name, lock->lo_name, file, line);
 1433                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1434                         panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1435                             class->lc_name, lock->lo_name,
 1436                             instance->li_flags & LI_RECURSEMASK, file, line);
 1437         }
 1438         instance->li_flags |= LI_EXCLUSIVE;
 1439 }
 1440 
 1441 void
 1442 witness_downgrade(struct lock_object *lock, int flags, const char *file,
 1443     int line)
 1444 {
 1445         struct lock_instance *instance;
 1446         struct lock_class *class;
 1447 
 1448         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
 1449         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
 1450                 return;
 1451         class = LOCK_CLASS(lock);
 1452         file = fixup_filename(file);
 1453         if (witness_watch) {
 1454                 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1455                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
 1456                             class->lc_name, lock->lo_name, file, line);
 1457                 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
 1458                         panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
 1459                             class->lc_name, lock->lo_name, file, line);
 1460         }
 1461         instance = find_instance(curthread->td_sleeplocks, lock);
 1462         if (instance == NULL)
 1463                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
 1464                     class->lc_name, lock->lo_name, file, line);
 1465         if (witness_watch) {
 1466                 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
 1467                         panic("downgrade of shared lock (%s) %s @ %s:%d",
 1468                             class->lc_name, lock->lo_name, file, line);
 1469                 if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1470                         panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1471                             class->lc_name, lock->lo_name,
 1472                             instance->li_flags & LI_RECURSEMASK, file, line);
 1473         }
 1474         instance->li_flags &= ~LI_EXCLUSIVE;
 1475 }
 1476 
 1477 void
 1478 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
 1479 {
 1480         struct lock_list_entry **lock_list, *lle;
 1481         struct lock_instance *instance;
 1482         struct lock_class *class;
 1483         struct thread *td;
 1484         register_t s;
 1485         int i, j;
 1486 
 1487         if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
 1488                 return;
 1489         td = curthread;
 1490         class = LOCK_CLASS(lock);
 1491         file = fixup_filename(file);
 1492 
 1493         /* Find lock instance associated with this lock. */
 1494         if (class->lc_flags & LC_SLEEPLOCK)
 1495                 lock_list = &td->td_sleeplocks;
 1496         else
 1497                 lock_list = PCPU_PTR(spinlocks);
 1498         lle = *lock_list;
 1499         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
 1500                 for (i = 0; i < (*lock_list)->ll_count; i++) {
 1501                         instance = &(*lock_list)->ll_children[i];
 1502                         if (instance->li_lock == lock)
 1503                                 goto found;
 1504                 }
 1505 
 1506         /*
 1507          * When disabling WITNESS through witness_watch we could end up in
 1508          * having registered locks in the td_sleeplocks queue.
 1509          * We have to make sure we flush these queues, so just search for
 1510          * eventual register locks and remove them.
 1511          */
 1512         if (witness_watch > 0)
 1513                 panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
 1514                     lock->lo_name, file, line);
 1515         else
 1516                 return;
 1517 found:
 1518 
 1519         /* First, check for shared/exclusive mismatches. */
 1520         if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
 1521             (flags & LOP_EXCLUSIVE) == 0) {
 1522                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1523                     lock->lo_name, file, line);
 1524                 printf("while exclusively locked from %s:%d\n",
 1525                     instance->li_file, instance->li_line);
 1526                 panic("excl->ushare");
 1527         }
 1528         if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
 1529             (flags & LOP_EXCLUSIVE) != 0) {
 1530                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1531                     lock->lo_name, file, line);
 1532                 printf("while share locked from %s:%d\n", instance->li_file,
 1533                     instance->li_line);
 1534                 panic("share->uexcl");
 1535         }
 1536         /* If we are recursed, unrecurse. */
 1537         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
 1538                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
 1539                     td->td_proc->p_pid, instance->li_lock->lo_name,
 1540                     instance->li_flags);
 1541                 instance->li_flags--;
 1542                 return;
 1543         }
 1544         /* The lock is now being dropped, check for NORELEASE flag */
 1545         if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
 1546                 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1547                     lock->lo_name, file, line);
 1548                 panic("lock marked norelease");
 1549         }
 1550 
 1551         /* Otherwise, remove this item from the list. */
 1552         s = intr_disable();
 1553         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
 1554             td->td_proc->p_pid, instance->li_lock->lo_name,
 1555             (*lock_list)->ll_count - 1);
 1556         for (j = i; j < (*lock_list)->ll_count - 1; j++)
 1557                 (*lock_list)->ll_children[j] =
 1558                     (*lock_list)->ll_children[j + 1];
 1559         (*lock_list)->ll_count--;
 1560         intr_restore(s);
 1561 
 1562         /*
 1563          * In order to reduce contention on w_mtx, we want to keep always an
 1564          * head object into lists so that frequent allocation from the 
 1565          * free witness pool (and subsequent locking) is avoided.
 1566          * In order to maintain the current code simple, when the head
 1567          * object is totally unloaded it means also that we do not have
 1568          * further objects in the list, so the list ownership needs to be
 1569          * hand over to another object if the current head needs to be freed.
 1570          */
 1571         if ((*lock_list)->ll_count == 0) {
 1572                 if (*lock_list == lle) {
 1573                         if (lle->ll_next == NULL)
 1574                                 return;
 1575                 } else
 1576                         lle = *lock_list;
 1577                 *lock_list = lle->ll_next;
 1578                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
 1579                     td->td_proc->p_pid, lle);
 1580                 witness_lock_list_free(lle);
 1581         }
 1582 }
 1583 
 1584 void
 1585 witness_thread_exit(struct thread *td)
 1586 {
 1587         struct lock_list_entry *lle;
 1588         int i, n;
 1589 
 1590         lle = td->td_sleeplocks;
 1591         if (lle == NULL || panicstr != NULL)
 1592                 return;
 1593         if (lle->ll_count != 0) {
 1594                 for (n = 0; lle != NULL; lle = lle->ll_next)
 1595                         for (i = lle->ll_count - 1; i >= 0; i--) {
 1596                                 if (n == 0)
 1597                 printf("Thread %p exiting with the following locks held:\n",
 1598                                             td);
 1599                                 n++;
 1600                                 witness_list_lock(&lle->ll_children[i], printf);
 1601                                 
 1602                         }
 1603                 panic("Thread %p cannot exit while holding sleeplocks\n", td);
 1604         }
 1605         witness_lock_list_free(lle);
 1606 }
 1607 
 1608 /*
 1609  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
 1610  * exempt Giant and sleepable locks from the checks as well.  If any
 1611  * non-exempt locks are held, then a supplied message is printed to the
 1612  * console along with a list of the offending locks.  If indicated in the
 1613  * flags then a failure results in a panic as well.
 1614  */
 1615 int
 1616 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
 1617 {
 1618         struct lock_list_entry *lock_list, *lle;
 1619         struct lock_instance *lock1;
 1620         struct thread *td;
 1621         va_list ap;
 1622         int i, n;
 1623 
 1624         if (witness_cold || witness_watch < 1 || panicstr != NULL)
 1625                 return (0);
 1626         n = 0;
 1627         td = curthread;
 1628         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
 1629                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1630                         lock1 = &lle->ll_children[i];
 1631                         if (lock1->li_lock == lock)
 1632                                 continue;
 1633                         if (flags & WARN_GIANTOK &&
 1634                             lock1->li_lock == &Giant.lock_object)
 1635                                 continue;
 1636                         if (flags & WARN_SLEEPOK &&
 1637                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
 1638                                 continue;
 1639                         if (n == 0) {
 1640                                 va_start(ap, fmt);
 1641                                 vprintf(fmt, ap);
 1642                                 va_end(ap);
 1643                                 printf(" with the following");
 1644                                 if (flags & WARN_SLEEPOK)
 1645                                         printf(" non-sleepable");
 1646                                 printf(" locks held:\n");
 1647                         }
 1648                         n++;
 1649                         witness_list_lock(lock1, printf);
 1650                 }
 1651 
 1652         /*
 1653          * Pin the thread in order to avoid problems with thread migration.
 1654          * Once that all verifies are passed about spinlocks ownership,
 1655          * the thread is in a safe path and it can be unpinned.
 1656          */
 1657         sched_pin();
 1658         lock_list = PCPU_GET(spinlocks);
 1659         if (lock_list != NULL && lock_list->ll_count != 0) {
 1660                 sched_unpin();
 1661 
 1662                 /*
 1663                  * We should only have one spinlock and as long as
 1664                  * the flags cannot match for this locks class,
 1665                  * check if the first spinlock is the one curthread
 1666                  * should hold.
 1667                  */
 1668                 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
 1669                 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
 1670                     lock1->li_lock == lock && n == 0)
 1671                         return (0);
 1672 
 1673                 va_start(ap, fmt);
 1674                 vprintf(fmt, ap);
 1675                 va_end(ap);
 1676                 printf(" with the following");
 1677                 if (flags & WARN_SLEEPOK)
 1678                         printf(" non-sleepable");
 1679                 printf(" locks held:\n");
 1680                 n += witness_list_locks(&lock_list, printf);
 1681         } else
 1682                 sched_unpin();
 1683         if (flags & WARN_PANIC && n)
 1684                 panic("%s", __func__);
 1685         else
 1686                 witness_debugger(n);
 1687         return (n);
 1688 }
 1689 
 1690 const char *
 1691 witness_file(struct lock_object *lock)
 1692 {
 1693         struct witness *w;
 1694 
 1695         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
 1696                 return ("?");
 1697         w = lock->lo_witness;
 1698         return (w->w_file);
 1699 }
 1700 
 1701 int
 1702 witness_line(struct lock_object *lock)
 1703 {
 1704         struct witness *w;
 1705 
 1706         if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
 1707                 return (0);
 1708         w = lock->lo_witness;
 1709         return (w->w_line);
 1710 }
 1711 
 1712 static struct witness *
 1713 enroll(const char *description, struct lock_class *lock_class)
 1714 {
 1715         struct witness *w;
 1716         struct witness_list *typelist;
 1717 
 1718         MPASS(description != NULL);
 1719 
 1720         if (witness_watch == -1 || panicstr != NULL)
 1721                 return (NULL);
 1722         if ((lock_class->lc_flags & LC_SPINLOCK)) {
 1723                 if (witness_skipspin)
 1724                         return (NULL);
 1725                 else
 1726                         typelist = &w_spin;
 1727         } else if ((lock_class->lc_flags & LC_SLEEPLOCK))
 1728                 typelist = &w_sleep;
 1729         else
 1730                 panic("lock class %s is not sleep or spin",
 1731                     lock_class->lc_name);
 1732 
 1733         mtx_lock_spin(&w_mtx);
 1734         w = witness_hash_get(description);
 1735         if (w)
 1736                 goto found;
 1737         if ((w = witness_get()) == NULL)
 1738                 return (NULL);
 1739         MPASS(strlen(description) < MAX_W_NAME);
 1740         strcpy(w->w_name, description);
 1741         w->w_class = lock_class;
 1742         w->w_refcount = 1;
 1743         STAILQ_INSERT_HEAD(&w_all, w, w_list);
 1744         if (lock_class->lc_flags & LC_SPINLOCK) {
 1745                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
 1746                 w_spin_cnt++;
 1747         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
 1748                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
 1749                 w_sleep_cnt++;
 1750         }
 1751 
 1752         /* Insert new witness into the hash */
 1753         witness_hash_put(w);
 1754         witness_increment_graph_generation();
 1755         mtx_unlock_spin(&w_mtx);
 1756         return (w);
 1757 found:
 1758         w->w_refcount++;
 1759         mtx_unlock_spin(&w_mtx);
 1760         if (lock_class != w->w_class)
 1761                 panic(
 1762                         "lock (%s) %s does not match earlier (%s) lock",
 1763                         description, lock_class->lc_name,
 1764                         w->w_class->lc_name);
 1765         return (w);
 1766 }
 1767 
 1768 static void
 1769 depart(struct witness *w)
 1770 {
 1771         struct witness_list *list;
 1772 
 1773         MPASS(w->w_refcount == 0);
 1774         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
 1775                 list = &w_sleep;
 1776                 w_sleep_cnt--;
 1777         } else {
 1778                 list = &w_spin;
 1779                 w_spin_cnt--;
 1780         }
 1781         /*
 1782          * Set file to NULL as it may point into a loadable module.
 1783          */
 1784         w->w_file = NULL;
 1785         w->w_line = 0;
 1786         witness_increment_graph_generation();
 1787 }
 1788 
 1789 
 1790 static void
 1791 adopt(struct witness *parent, struct witness *child)
 1792 {
 1793         int pi, ci, i, j;
 1794 
 1795         if (witness_cold == 0)
 1796                 mtx_assert(&w_mtx, MA_OWNED);
 1797 
 1798         /* If the relationship is already known, there's no work to be done. */
 1799         if (isitmychild(parent, child))
 1800                 return;
 1801 
 1802         /* When the structure of the graph changes, bump up the generation. */
 1803         witness_increment_graph_generation();
 1804 
 1805         /*
 1806          * The hard part ... create the direct relationship, then propagate all
 1807          * indirect relationships.
 1808          */
 1809         pi = parent->w_index;
 1810         ci = child->w_index;
 1811         WITNESS_INDEX_ASSERT(pi);
 1812         WITNESS_INDEX_ASSERT(ci);
 1813         MPASS(pi != ci);
 1814         w_rmatrix[pi][ci] |= WITNESS_PARENT;
 1815         w_rmatrix[ci][pi] |= WITNESS_CHILD;
 1816 
 1817         /*
 1818          * If parent was not already an ancestor of child,
 1819          * then we increment the descendant and ancestor counters.
 1820          */
 1821         if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
 1822                 parent->w_num_descendants++;
 1823                 child->w_num_ancestors++;
 1824         }
 1825 
 1826         /* 
 1827          * Find each ancestor of 'pi'. Note that 'pi' itself is counted as 
 1828          * an ancestor of 'pi' during this loop.
 1829          */
 1830         for (i = 1; i <= w_max_used_index; i++) {
 1831                 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && 
 1832                     (i != pi))
 1833                         continue;
 1834 
 1835                 /* Find each descendant of 'i' and mark it as a descendant. */
 1836                 for (j = 1; j <= w_max_used_index; j++) {
 1837 
 1838                         /* 
 1839                          * Skip children that are already marked as
 1840                          * descendants of 'i'.
 1841                          */
 1842                         if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
 1843                                 continue;
 1844 
 1845                         /*
 1846                          * We are only interested in descendants of 'ci'. Note
 1847                          * that 'ci' itself is counted as a descendant of 'ci'.
 1848                          */
 1849                         if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && 
 1850                             (j != ci))
 1851                                 continue;
 1852                         w_rmatrix[i][j] |= WITNESS_ANCESTOR;
 1853                         w_rmatrix[j][i] |= WITNESS_DESCENDANT;
 1854                         w_data[i].w_num_descendants++;
 1855                         w_data[j].w_num_ancestors++;
 1856 
 1857                         /* 
 1858                          * Make sure we aren't marking a node as both an
 1859                          * ancestor and descendant. We should have caught 
 1860                          * this as a lock order reversal earlier.
 1861                          */
 1862                         if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
 1863                             (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
 1864                                 printf("witness rmatrix paradox! [%d][%d]=%d "
 1865                                     "both ancestor and descendant\n",
 1866                                     i, j, w_rmatrix[i][j]); 
 1867                                 kdb_backtrace();
 1868                                 printf("Witness disabled.\n");
 1869                                 witness_watch = -1;
 1870                         }
 1871                         if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
 1872                             (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
 1873                                 printf("witness rmatrix paradox! [%d][%d]=%d "
 1874                                     "both ancestor and descendant\n",
 1875                                     j, i, w_rmatrix[j][i]); 
 1876                                 kdb_backtrace();
 1877                                 printf("Witness disabled.\n");
 1878                                 witness_watch = -1;
 1879                         }
 1880                 }
 1881         }
 1882 }
 1883 
 1884 static void
 1885 itismychild(struct witness *parent, struct witness *child)
 1886 {
 1887 
 1888         MPASS(child != NULL && parent != NULL);
 1889         if (witness_cold == 0)
 1890                 mtx_assert(&w_mtx, MA_OWNED);
 1891 
 1892         if (!witness_lock_type_equal(parent, child)) {
 1893                 if (witness_cold == 0)
 1894                         mtx_unlock_spin(&w_mtx);
 1895                 panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
 1896                     "the same lock type", __func__, parent->w_name,
 1897                     parent->w_class->lc_name, child->w_name,
 1898                     child->w_class->lc_name);
 1899         }
 1900         adopt(parent, child);
 1901 }
 1902 
 1903 /*
 1904  * Generic code for the isitmy*() functions. The rmask parameter is the
 1905  * expected relationship of w1 to w2.
 1906  */
 1907 static int
 1908 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
 1909 {
 1910         unsigned char r1, r2;
 1911         int i1, i2;
 1912 
 1913         i1 = w1->w_index;
 1914         i2 = w2->w_index;
 1915         WITNESS_INDEX_ASSERT(i1);
 1916         WITNESS_INDEX_ASSERT(i2);
 1917         r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
 1918         r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
 1919 
 1920         /* The flags on one better be the inverse of the flags on the other */
 1921         if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
 1922                 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
 1923                 printf("%s: rmatrix mismatch between %s (index %d) and %s "
 1924                     "(index %d): w_rmatrix[%d][%d] == %hhx but "
 1925                     "w_rmatrix[%d][%d] == %hhx\n",
 1926                     fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
 1927                     i2, i1, r2);
 1928                 kdb_backtrace();
 1929                 printf("Witness disabled.\n");
 1930                 witness_watch = -1;
 1931         }
 1932         return (r1 & rmask);
 1933 }
 1934 
 1935 /*
 1936  * Checks if @child is a direct child of @parent.
 1937  */
 1938 static int
 1939 isitmychild(struct witness *parent, struct witness *child)
 1940 {
 1941 
 1942         return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
 1943 }
 1944 
 1945 /*
 1946  * Checks if @descendant is a direct or inderect descendant of @ancestor.
 1947  */
 1948 static int
 1949 isitmydescendant(struct witness *ancestor, struct witness *descendant)
 1950 {
 1951 
 1952         return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
 1953             __func__));
 1954 }
 1955 
 1956 #ifdef BLESSING
 1957 static int
 1958 blessed(struct witness *w1, struct witness *w2)
 1959 {
 1960         int i;
 1961         struct witness_blessed *b;
 1962 
 1963         for (i = 0; i < blessed_count; i++) {
 1964                 b = &blessed_list[i];
 1965                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
 1966                         if (strcmp(w2->w_name, b->b_lock2) == 0)
 1967                                 return (1);
 1968                         continue;
 1969                 }
 1970                 if (strcmp(w1->w_name, b->b_lock2) == 0)
 1971                         if (strcmp(w2->w_name, b->b_lock1) == 0)
 1972                                 return (1);
 1973         }
 1974         return (0);
 1975 }
 1976 #endif
 1977 
 1978 static struct witness *
 1979 witness_get(void)
 1980 {
 1981         struct witness *w;
 1982         int index;
 1983 
 1984         if (witness_cold == 0)
 1985                 mtx_assert(&w_mtx, MA_OWNED);
 1986 
 1987         if (witness_watch == -1) {
 1988                 mtx_unlock_spin(&w_mtx);
 1989                 return (NULL);
 1990         }
 1991         if (STAILQ_EMPTY(&w_free)) {
 1992                 witness_watch = -1;
 1993                 mtx_unlock_spin(&w_mtx);
 1994                 printf("WITNESS: unable to allocate a new witness object\n");
 1995                 return (NULL);
 1996         }
 1997         w = STAILQ_FIRST(&w_free);
 1998         STAILQ_REMOVE_HEAD(&w_free, w_list);
 1999         w_free_cnt--;
 2000         index = w->w_index;
 2001         MPASS(index > 0 && index == w_max_used_index+1 &&
 2002             index < WITNESS_COUNT);
 2003         bzero(w, sizeof(*w));
 2004         w->w_index = index;
 2005         if (index > w_max_used_index)
 2006                 w_max_used_index = index;
 2007         return (w);
 2008 }
 2009 
 2010 static void
 2011 witness_free(struct witness *w)
 2012 {
 2013 
 2014         STAILQ_INSERT_HEAD(&w_free, w, w_list);
 2015         w_free_cnt++;
 2016 }
 2017 
 2018 static struct lock_list_entry *
 2019 witness_lock_list_get(void)
 2020 {
 2021         struct lock_list_entry *lle;
 2022 
 2023         if (witness_watch == -1)
 2024                 return (NULL);
 2025         mtx_lock_spin(&w_mtx);
 2026         lle = w_lock_list_free;
 2027         if (lle == NULL) {
 2028                 witness_watch = -1;
 2029                 mtx_unlock_spin(&w_mtx);
 2030                 printf("%s: witness exhausted\n", __func__);
 2031                 return (NULL);
 2032         }
 2033         w_lock_list_free = lle->ll_next;
 2034         mtx_unlock_spin(&w_mtx);
 2035         bzero(lle, sizeof(*lle));
 2036         return (lle);
 2037 }
 2038                 
 2039 static void
 2040 witness_lock_list_free(struct lock_list_entry *lle)
 2041 {
 2042 
 2043         mtx_lock_spin(&w_mtx);
 2044         lle->ll_next = w_lock_list_free;
 2045         w_lock_list_free = lle;
 2046         mtx_unlock_spin(&w_mtx);
 2047 }
 2048 
 2049 static struct lock_instance *
 2050 find_instance(struct lock_list_entry *list, struct lock_object *lock)
 2051 {
 2052         struct lock_list_entry *lle;
 2053         struct lock_instance *instance;
 2054         int i;
 2055 
 2056         for (lle = list; lle != NULL; lle = lle->ll_next)
 2057                 for (i = lle->ll_count - 1; i >= 0; i--) {
 2058                         instance = &lle->ll_children[i];
 2059                         if (instance->li_lock == lock)
 2060                                 return (instance);
 2061                 }
 2062         return (NULL);
 2063 }
 2064 
 2065 static void
 2066 witness_list_lock(struct lock_instance *instance,
 2067     int (*prnt)(const char *fmt, ...))
 2068 {
 2069         struct lock_object *lock;
 2070 
 2071         lock = instance->li_lock;
 2072         prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
 2073             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
 2074         if (lock->lo_witness->w_name != lock->lo_name)
 2075                 prnt(" (%s)", lock->lo_witness->w_name);
 2076         prnt(" r = %d (%p) locked @ %s:%d\n",
 2077             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
 2078             instance->li_line);
 2079 }
 2080 
 2081 #ifdef DDB
 2082 static int
 2083 witness_thread_has_locks(struct thread *td)
 2084 {
 2085 
 2086         if (td->td_sleeplocks == NULL)
 2087                 return (0);
 2088         return (td->td_sleeplocks->ll_count != 0);
 2089 }
 2090 
 2091 static int
 2092 witness_proc_has_locks(struct proc *p)
 2093 {
 2094         struct thread *td;
 2095 
 2096         FOREACH_THREAD_IN_PROC(p, td) {
 2097                 if (witness_thread_has_locks(td))
 2098                         return (1);
 2099         }
 2100         return (0);
 2101 }
 2102 #endif
 2103 
 2104 int
 2105 witness_list_locks(struct lock_list_entry **lock_list,
 2106     int (*prnt)(const char *fmt, ...))
 2107 {
 2108         struct lock_list_entry *lle;
 2109         int i, nheld;
 2110 
 2111         nheld = 0;
 2112         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
 2113                 for (i = lle->ll_count - 1; i >= 0; i--) {
 2114                         witness_list_lock(&lle->ll_children[i], prnt);
 2115                         nheld++;
 2116                 }
 2117         return (nheld);
 2118 }
 2119 
 2120 /*
 2121  * This is a bit risky at best.  We call this function when we have timed
 2122  * out acquiring a spin lock, and we assume that the other CPU is stuck
 2123  * with this lock held.  So, we go groveling around in the other CPU's
 2124  * per-cpu data to try to find the lock instance for this spin lock to
 2125  * see when it was last acquired.
 2126  */
 2127 void
 2128 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
 2129     int (*prnt)(const char *fmt, ...))
 2130 {
 2131         struct lock_instance *instance;
 2132         struct pcpu *pc;
 2133 
 2134         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
 2135                 return;
 2136         pc = pcpu_find(owner->td_oncpu);
 2137         instance = find_instance(pc->pc_spinlocks, lock);
 2138         if (instance != NULL)
 2139                 witness_list_lock(instance, prnt);
 2140 }
 2141 
 2142 void
 2143 witness_save(struct lock_object *lock, const char **filep, int *linep)
 2144 {
 2145         struct lock_list_entry *lock_list;
 2146         struct lock_instance *instance;
 2147         struct lock_class *class;
 2148 
 2149         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
 2150         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
 2151                 return;
 2152         class = LOCK_CLASS(lock);
 2153         if (class->lc_flags & LC_SLEEPLOCK)
 2154                 lock_list = curthread->td_sleeplocks;
 2155         else {
 2156                 if (witness_skipspin)
 2157                         return;
 2158                 lock_list = PCPU_GET(spinlocks);
 2159         }
 2160         instance = find_instance(lock_list, lock);
 2161         if (instance == NULL)
 2162                 panic("%s: lock (%s) %s not locked", __func__,
 2163                     class->lc_name, lock->lo_name);
 2164         *filep = instance->li_file;
 2165         *linep = instance->li_line;
 2166 }
 2167 
 2168 void
 2169 witness_restore(struct lock_object *lock, const char *file, int line)
 2170 {
 2171         struct lock_list_entry *lock_list;
 2172         struct lock_instance *instance;
 2173         struct lock_class *class;
 2174 
 2175         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
 2176         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
 2177                 return;
 2178         class = LOCK_CLASS(lock);
 2179         if (class->lc_flags & LC_SLEEPLOCK)
 2180                 lock_list = curthread->td_sleeplocks;
 2181         else {
 2182                 if (witness_skipspin)
 2183                         return;
 2184                 lock_list = PCPU_GET(spinlocks);
 2185         }
 2186         instance = find_instance(lock_list, lock);
 2187         if (instance == NULL)
 2188                 panic("%s: lock (%s) %s not locked", __func__,
 2189                     class->lc_name, lock->lo_name);
 2190         lock->lo_witness->w_file = file;
 2191         lock->lo_witness->w_line = line;
 2192         instance->li_file = file;
 2193         instance->li_line = line;
 2194 }
 2195 
 2196 void
 2197 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
 2198 {
 2199 #ifdef INVARIANT_SUPPORT
 2200         struct lock_instance *instance;
 2201         struct lock_class *class;
 2202 
 2203         if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
 2204                 return;
 2205         class = LOCK_CLASS(lock);
 2206         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
 2207                 instance = find_instance(curthread->td_sleeplocks, lock);
 2208         else if ((class->lc_flags & LC_SPINLOCK) != 0)
 2209                 instance = find_instance(PCPU_GET(spinlocks), lock);
 2210         else {
 2211                 panic("Lock (%s) %s is not sleep or spin!",
 2212                     class->lc_name, lock->lo_name);
 2213         }
 2214         file = fixup_filename(file);
 2215         switch (flags) {
 2216         case LA_UNLOCKED:
 2217                 if (instance != NULL)
 2218                         panic("Lock (%s) %s locked @ %s:%d.",
 2219                             class->lc_name, lock->lo_name, file, line);
 2220                 break;
 2221         case LA_LOCKED:
 2222         case LA_LOCKED | LA_RECURSED:
 2223         case LA_LOCKED | LA_NOTRECURSED:
 2224         case LA_SLOCKED:
 2225         case LA_SLOCKED | LA_RECURSED:
 2226         case LA_SLOCKED | LA_NOTRECURSED:
 2227         case LA_XLOCKED:
 2228         case LA_XLOCKED | LA_RECURSED:
 2229         case LA_XLOCKED | LA_NOTRECURSED:
 2230                 if (instance == NULL) {
 2231                         panic("Lock (%s) %s not locked @ %s:%d.",
 2232                             class->lc_name, lock->lo_name, file, line);
 2233                         break;
 2234                 }
 2235                 if ((flags & LA_XLOCKED) != 0 &&
 2236                     (instance->li_flags & LI_EXCLUSIVE) == 0)
 2237                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
 2238                             class->lc_name, lock->lo_name, file, line);
 2239                 if ((flags & LA_SLOCKED) != 0 &&
 2240                     (instance->li_flags & LI_EXCLUSIVE) != 0)
 2241                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
 2242                             class->lc_name, lock->lo_name, file, line);
 2243                 if ((flags & LA_RECURSED) != 0 &&
 2244                     (instance->li_flags & LI_RECURSEMASK) == 0)
 2245                         panic("Lock (%s) %s not recursed @ %s:%d.",
 2246                             class->lc_name, lock->lo_name, file, line);
 2247                 if ((flags & LA_NOTRECURSED) != 0 &&
 2248                     (instance->li_flags & LI_RECURSEMASK) != 0)
 2249                         panic("Lock (%s) %s recursed @ %s:%d.",
 2250                             class->lc_name, lock->lo_name, file, line);
 2251                 break;
 2252         default:
 2253                 panic("Invalid lock assertion at %s:%d.", file, line);
 2254 
 2255         }
 2256 #endif  /* INVARIANT_SUPPORT */
 2257 }
 2258 
 2259 static void
 2260 witness_setflag(struct lock_object *lock, int flag, int set)
 2261 {
 2262         struct lock_list_entry *lock_list;
 2263         struct lock_instance *instance;
 2264         struct lock_class *class;
 2265 
 2266         if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
 2267                 return;
 2268         class = LOCK_CLASS(lock);
 2269         if (class->lc_flags & LC_SLEEPLOCK)
 2270                 lock_list = curthread->td_sleeplocks;
 2271         else {
 2272                 if (witness_skipspin)
 2273                         return;
 2274                 lock_list = PCPU_GET(spinlocks);
 2275         }
 2276         instance = find_instance(lock_list, lock);
 2277         if (instance == NULL)
 2278                 panic("%s: lock (%s) %s not locked", __func__,
 2279                     class->lc_name, lock->lo_name);
 2280 
 2281         if (set)
 2282                 instance->li_flags |= flag;
 2283         else
 2284                 instance->li_flags &= ~flag;
 2285 }
 2286 
 2287 void
 2288 witness_norelease(struct lock_object *lock)
 2289 {
 2290 
 2291         witness_setflag(lock, LI_NORELEASE, 1);
 2292 }
 2293 
 2294 void
 2295 witness_releaseok(struct lock_object *lock)
 2296 {
 2297 
 2298         witness_setflag(lock, LI_NORELEASE, 0);
 2299 }
 2300 
 2301 #ifdef DDB
 2302 static void
 2303 witness_ddb_list(struct thread *td)
 2304 {
 2305 
 2306         KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
 2307         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
 2308 
 2309         if (witness_watch < 1)
 2310                 return;
 2311 
 2312         witness_list_locks(&td->td_sleeplocks, db_printf);
 2313 
 2314         /*
 2315          * We only handle spinlocks if td == curthread.  This is somewhat broken
 2316          * if td is currently executing on some other CPU and holds spin locks
 2317          * as we won't display those locks.  If we had a MI way of getting
 2318          * the per-cpu data for a given cpu then we could use
 2319          * td->td_oncpu to get the list of spinlocks for this thread
 2320          * and "fix" this.
 2321          *
 2322          * That still wouldn't really fix this unless we locked the scheduler
 2323          * lock or stopped the other CPU to make sure it wasn't changing the
 2324          * list out from under us.  It is probably best to just not try to
 2325          * handle threads on other CPU's for now.
 2326          */
 2327         if (td == curthread && PCPU_GET(spinlocks) != NULL)
 2328                 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
 2329 }
 2330 
 2331 DB_SHOW_COMMAND(locks, db_witness_list)
 2332 {
 2333         struct thread *td;
 2334 
 2335         if (have_addr)
 2336                 td = db_lookup_thread(addr, TRUE);
 2337         else
 2338                 td = kdb_thread;
 2339         witness_ddb_list(td);
 2340 }
 2341 
 2342 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
 2343 {
 2344         struct thread *td;
 2345         struct proc *p;
 2346 
 2347         /*
 2348          * It would be nice to list only threads and processes that actually
 2349          * held sleep locks, but that information is currently not exported
 2350          * by WITNESS.
 2351          */
 2352         FOREACH_PROC_IN_SYSTEM(p) {
 2353                 if (!witness_proc_has_locks(p))
 2354                         continue;
 2355                 FOREACH_THREAD_IN_PROC(p, td) {
 2356                         if (!witness_thread_has_locks(td))
 2357                                 continue;
 2358                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
 2359                             p->p_comm, td, td->td_tid);
 2360                         witness_ddb_list(td);
 2361                 }
 2362         }
 2363 }
 2364 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
 2365 
 2366 DB_SHOW_COMMAND(witness, db_witness_display)
 2367 {
 2368 
 2369         witness_ddb_display(db_printf);
 2370 }
 2371 #endif
 2372 
 2373 static int
 2374 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
 2375 {
 2376         struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
 2377         struct witness *tmp_w1, *tmp_w2, *w1, *w2;
 2378         struct sbuf *sb;
 2379         u_int w_rmatrix1, w_rmatrix2;
 2380         int error, generation, i, j;
 2381 
 2382         tmp_data1 = NULL;
 2383         tmp_data2 = NULL;
 2384         tmp_w1 = NULL;
 2385         tmp_w2 = NULL;
 2386         if (witness_watch < 1) {
 2387                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
 2388                 return (error);
 2389         }
 2390         if (witness_cold) {
 2391                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
 2392                 return (error);
 2393         }
 2394         error = 0;
 2395         sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
 2396         if (sb == NULL)
 2397                 return (ENOMEM);
 2398 
 2399         /* Allocate and init temporary storage space. */
 2400         tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
 2401         tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
 2402         tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 
 2403             M_WAITOK | M_ZERO);
 2404         tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 
 2405             M_WAITOK | M_ZERO);
 2406         stack_zero(&tmp_data1->wlod_stack);
 2407         stack_zero(&tmp_data2->wlod_stack);
 2408 
 2409 restart:
 2410         mtx_lock_spin(&w_mtx);
 2411         generation = w_generation;
 2412         mtx_unlock_spin(&w_mtx);
 2413         sbuf_printf(sb, "Number of known direct relationships is %d\n",
 2414             w_lohash.wloh_count);
 2415         for (i = 1; i < w_max_used_index; i++) {
 2416                 mtx_lock_spin(&w_mtx);
 2417                 if (generation != w_generation) {
 2418                         mtx_unlock_spin(&w_mtx);
 2419 
 2420                         /* The graph has changed, try again. */
 2421                         req->oldidx = 0;
 2422                         sbuf_clear(sb);
 2423                         goto restart;
 2424                 }
 2425 
 2426                 w1 = &w_data[i];
 2427                 if (w1->w_reversed == 0) {
 2428                         mtx_unlock_spin(&w_mtx);
 2429                         continue;
 2430                 }
 2431 
 2432                 /* Copy w1 locally so we can release the spin lock. */
 2433                 *tmp_w1 = *w1;
 2434                 mtx_unlock_spin(&w_mtx);
 2435 
 2436                 if (tmp_w1->w_reversed == 0)
 2437                         continue;
 2438                 for (j = 1; j < w_max_used_index; j++) {
 2439                         if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
 2440                                 continue;
 2441 
 2442                         mtx_lock_spin(&w_mtx);
 2443                         if (generation != w_generation) {
 2444                                 mtx_unlock_spin(&w_mtx);
 2445 
 2446                                 /* The graph has changed, try again. */
 2447                                 req->oldidx = 0;
 2448                                 sbuf_clear(sb);
 2449                                 goto restart;
 2450                         }
 2451 
 2452                         w2 = &w_data[j];
 2453                         data1 = witness_lock_order_get(w1, w2);
 2454                         data2 = witness_lock_order_get(w2, w1);
 2455 
 2456                         /*
 2457                          * Copy information locally so we can release the
 2458                          * spin lock.
 2459                          */
 2460                         *tmp_w2 = *w2;
 2461                         w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
 2462                         w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
 2463 
 2464                         if (data1) {
 2465                                 stack_zero(&tmp_data1->wlod_stack);
 2466                                 stack_copy(&data1->wlod_stack,
 2467                                     &tmp_data1->wlod_stack);
 2468                         }
 2469                         if (data2 && data2 != data1) {
 2470                                 stack_zero(&tmp_data2->wlod_stack);
 2471                                 stack_copy(&data2->wlod_stack,
 2472                                     &tmp_data2->wlod_stack);
 2473                         }
 2474                         mtx_unlock_spin(&w_mtx);
 2475 
 2476                         sbuf_printf(sb,
 2477             "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
 2478                             tmp_w1->w_name, tmp_w1->w_class->lc_name, 
 2479                             tmp_w2->w_name, tmp_w2->w_class->lc_name);
 2480 #if 0
 2481                         sbuf_printf(sb,
 2482                         "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
 2483                             tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
 2484                             tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
 2485 #endif
 2486                         if (data1) {
 2487                                 sbuf_printf(sb,
 2488                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
 2489                                     tmp_w1->w_name, tmp_w1->w_class->lc_name, 
 2490                                     tmp_w2->w_name, tmp_w2->w_class->lc_name);
 2491                                 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
 2492                                 sbuf_printf(sb, "\n");
 2493                         }
 2494                         if (data2 && data2 != data1) {
 2495                                 sbuf_printf(sb,
 2496                         "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
 2497                                     tmp_w2->w_name, tmp_w2->w_class->lc_name, 
 2498                                     tmp_w1->w_name, tmp_w1->w_class->lc_name);
 2499                                 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
 2500                                 sbuf_printf(sb, "\n");
 2501                         }
 2502                 }
 2503         }
 2504         mtx_lock_spin(&w_mtx);
 2505         if (generation != w_generation) {
 2506                 mtx_unlock_spin(&w_mtx);
 2507 
 2508                 /*
 2509                  * The graph changed while we were printing stack data,
 2510                  * try again.
 2511                  */
 2512                 req->oldidx = 0;
 2513                 sbuf_clear(sb);
 2514                 goto restart;
 2515         }
 2516         mtx_unlock_spin(&w_mtx);
 2517 
 2518         /* Free temporary storage space. */
 2519         free(tmp_data1, M_TEMP);
 2520         free(tmp_data2, M_TEMP);
 2521         free(tmp_w1, M_TEMP);
 2522         free(tmp_w2, M_TEMP);
 2523 
 2524         sbuf_finish(sb);
 2525         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
 2526         sbuf_delete(sb);
 2527 
 2528         return (error);
 2529 }
 2530 
 2531 static int
 2532 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
 2533 {
 2534         struct witness *w;
 2535         struct sbuf *sb;
 2536         int error;
 2537 
 2538         if (witness_watch < 1) {
 2539                 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
 2540                 return (error);
 2541         }
 2542         if (witness_cold) {
 2543                 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
 2544                 return (error);
 2545         }
 2546         error = 0;
 2547 
 2548         error = sysctl_wire_old_buffer(req, 0);
 2549         if (error != 0)
 2550                 return (error);
 2551         sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
 2552         if (sb == NULL)
 2553                 return (ENOMEM);
 2554         sbuf_printf(sb, "\n");
 2555 
 2556         mtx_lock_spin(&w_mtx);
 2557         STAILQ_FOREACH(w, &w_all, w_list)
 2558                 w->w_displayed = 0;
 2559         STAILQ_FOREACH(w, &w_all, w_list)
 2560                 witness_add_fullgraph(sb, w);
 2561         mtx_unlock_spin(&w_mtx);
 2562 
 2563         /*
 2564          * Close the sbuf and return to userland.
 2565          */
 2566         error = sbuf_finish(sb);
 2567         sbuf_delete(sb);
 2568 
 2569         return (error);
 2570 }
 2571 
 2572 static int
 2573 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
 2574 {
 2575         int error, value;
 2576 
 2577         value = witness_watch;
 2578         error = sysctl_handle_int(oidp, &value, 0, req);
 2579         if (error != 0 || req->newptr == NULL)
 2580                 return (error);
 2581         if (value > 1 || value < -1 ||
 2582             (witness_watch == -1 && value != witness_watch))
 2583                 return (EINVAL);
 2584         witness_watch = value;
 2585         return (0);
 2586 }
 2587 
 2588 static void
 2589 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
 2590 {
 2591         int i;
 2592 
 2593         if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
 2594                 return;
 2595         w->w_displayed = 1;
 2596 
 2597         WITNESS_INDEX_ASSERT(w->w_index);
 2598         for (i = 1; i <= w_max_used_index; i++) {
 2599                 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
 2600                         sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
 2601                             w_data[i].w_name);
 2602                         witness_add_fullgraph(sb, &w_data[i]);
 2603                 }
 2604         }
 2605 }
 2606 
 2607 /*
 2608  * A simple hash function. Takes a key pointer and a key size. If size == 0,
 2609  * interprets the key as a string and reads until the null
 2610  * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
 2611  * hash value computed from the key.
 2612  */
 2613 static uint32_t
 2614 witness_hash_djb2(const uint8_t *key, uint32_t size)
 2615 {
 2616         unsigned int hash = 5381;
 2617         int i;
 2618 
 2619         /* hash = hash * 33 + key[i] */
 2620         if (size)
 2621                 for (i = 0; i < size; i++)
 2622                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
 2623         else
 2624                 for (i = 0; key[i] != 0; i++)
 2625                         hash = ((hash << 5) + hash) + (unsigned int)key[i];
 2626 
 2627         return (hash);
 2628 }
 2629 
 2630 
 2631 /*
 2632  * Initializes the two witness hash tables. Called exactly once from
 2633  * witness_initialize().
 2634  */
 2635 static void
 2636 witness_init_hash_tables(void)
 2637 {
 2638         int i;
 2639 
 2640         MPASS(witness_cold);
 2641 
 2642         /* Initialize the hash tables. */
 2643         for (i = 0; i < WITNESS_HASH_SIZE; i++)
 2644                 w_hash.wh_array[i] = NULL;
 2645 
 2646         w_hash.wh_size = WITNESS_HASH_SIZE;
 2647         w_hash.wh_count = 0;
 2648 
 2649         /* Initialize the lock order data hash. */
 2650         w_lofree = NULL;
 2651         for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
 2652                 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
 2653                 w_lodata[i].wlod_next = w_lofree;
 2654                 w_lofree = &w_lodata[i];
 2655         }
 2656         w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
 2657         w_lohash.wloh_count = 0;
 2658         for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
 2659                 w_lohash.wloh_array[i] = NULL;
 2660 }
 2661 
 2662 static struct witness *
 2663 witness_hash_get(const char *key)
 2664 {
 2665         struct witness *w;
 2666         uint32_t hash;
 2667         
 2668         MPASS(key != NULL);
 2669         if (witness_cold == 0)
 2670                 mtx_assert(&w_mtx, MA_OWNED);
 2671         hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
 2672         w = w_hash.wh_array[hash];
 2673         while (w != NULL) {
 2674                 if (strcmp(w->w_name, key) == 0)
 2675                         goto out;
 2676                 w = w->w_hash_next;
 2677         }
 2678 
 2679 out:
 2680         return (w);
 2681 }
 2682 
 2683 static void
 2684 witness_hash_put(struct witness *w)
 2685 {
 2686         uint32_t hash;
 2687 
 2688         MPASS(w != NULL);
 2689         MPASS(w->w_name != NULL);
 2690         if (witness_cold == 0)
 2691                 mtx_assert(&w_mtx, MA_OWNED);
 2692         KASSERT(witness_hash_get(w->w_name) == NULL,
 2693             ("%s: trying to add a hash entry that already exists!", __func__));
 2694         KASSERT(w->w_hash_next == NULL,
 2695             ("%s: w->w_hash_next != NULL", __func__));
 2696 
 2697         hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
 2698         w->w_hash_next = w_hash.wh_array[hash];
 2699         w_hash.wh_array[hash] = w;
 2700         w_hash.wh_count++;
 2701 }
 2702 
 2703 
 2704 static struct witness_lock_order_data *
 2705 witness_lock_order_get(struct witness *parent, struct witness *child)
 2706 {
 2707         struct witness_lock_order_data *data = NULL;
 2708         struct witness_lock_order_key key;
 2709         unsigned int hash;
 2710 
 2711         MPASS(parent != NULL && child != NULL);
 2712         key.from = parent->w_index;
 2713         key.to = child->w_index;
 2714         WITNESS_INDEX_ASSERT(key.from);
 2715         WITNESS_INDEX_ASSERT(key.to);
 2716         if ((w_rmatrix[parent->w_index][child->w_index]
 2717             & WITNESS_LOCK_ORDER_KNOWN) == 0)
 2718                 goto out;
 2719 
 2720         hash = witness_hash_djb2((const char*)&key,
 2721             sizeof(key)) % w_lohash.wloh_size;
 2722         data = w_lohash.wloh_array[hash];
 2723         while (data != NULL) {
 2724                 if (witness_lock_order_key_equal(&data->wlod_key, &key))
 2725                         break;
 2726                 data = data->wlod_next;
 2727         }
 2728 
 2729 out:
 2730         return (data);
 2731 }
 2732 
 2733 /*
 2734  * Verify that parent and child have a known relationship, are not the same,
 2735  * and child is actually a child of parent.  This is done without w_mtx
 2736  * to avoid contention in the common case.
 2737  */
 2738 static int
 2739 witness_lock_order_check(struct witness *parent, struct witness *child)
 2740 {
 2741 
 2742         if (parent != child &&
 2743             w_rmatrix[parent->w_index][child->w_index]
 2744             & WITNESS_LOCK_ORDER_KNOWN &&
 2745             isitmychild(parent, child))
 2746                 return (1);
 2747 
 2748         return (0);
 2749 }
 2750 
 2751 static int
 2752 witness_lock_order_add(struct witness *parent, struct witness *child)
 2753 {
 2754         struct witness_lock_order_data *data = NULL;
 2755         struct witness_lock_order_key key;
 2756         unsigned int hash;
 2757         
 2758         MPASS(parent != NULL && child != NULL);
 2759         key.from = parent->w_index;
 2760         key.to = child->w_index;
 2761         WITNESS_INDEX_ASSERT(key.from);
 2762         WITNESS_INDEX_ASSERT(key.to);
 2763         if (w_rmatrix[parent->w_index][child->w_index]
 2764             & WITNESS_LOCK_ORDER_KNOWN)
 2765                 return (1);
 2766 
 2767         hash = witness_hash_djb2((const char*)&key,
 2768             sizeof(key)) % w_lohash.wloh_size;
 2769         w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
 2770         data = w_lofree;
 2771         if (data == NULL)
 2772                 return (0);
 2773         w_lofree = data->wlod_next;
 2774         data->wlod_next = w_lohash.wloh_array[hash];
 2775         data->wlod_key = key;
 2776         w_lohash.wloh_array[hash] = data;
 2777         w_lohash.wloh_count++;
 2778         stack_zero(&data->wlod_stack);
 2779         stack_save(&data->wlod_stack);
 2780         return (1);
 2781 }
 2782 
 2783 /* Call this whenver the structure of the witness graph changes. */
 2784 static void
 2785 witness_increment_graph_generation(void)
 2786 {
 2787 
 2788         if (witness_cold == 0)
 2789                 mtx_assert(&w_mtx, MA_OWNED);
 2790         w_generation++;
 2791 }
 2792 
 2793 #ifdef KDB
 2794 static void
 2795 _witness_debugger(int cond, const char *msg)
 2796 {
 2797 
 2798         if (witness_trace && cond)
 2799                 kdb_backtrace();
 2800         if (witness_kdb && cond)
 2801                 kdb_enter(KDB_WHY_WITNESS, msg);
 2802 }
 2803 #endif

Cache object: d9a509032d4fbee89b09d03b0e667de3


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.