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) 1998 Berkeley Software Design, Inc. All rights reserved.
    3  *
    4  * Redistribution and use in source and binary forms, with or without
    5  * modification, are permitted provided that the following conditions
    6  * are met:
    7  * 1. Redistributions of source code must retain the above copyright
    8  *    notice, this list of conditions and the following disclaimer.
    9  * 2. Redistributions in binary form must reproduce the above copyright
   10  *    notice, this list of conditions and the following disclaimer in the
   11  *    documentation and/or other materials provided with the distribution.
   12  * 3. Berkeley Software Design Inc's name may not be used to endorse or
   13  *    promote products derived from this software without specific prior
   14  *    written permission.
   15  *
   16  * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
   17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   19  * ARE DISCLAIMED.  IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
   20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   26  * SUCH DAMAGE.
   27  *
   28  *      from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
   29  *      and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
   30  */
   31 
   32 /*
   33  * Implementation of the `witness' lock verifier.  Originally implemented for
   34  * mutexes in BSD/OS.  Extended to handle generic lock objects and lock
   35  * classes in FreeBSD.
   36  */
   37 
   38 /*
   39  *      Main Entry: witness
   40  *      Pronunciation: 'wit-n&s
   41  *      Function: noun
   42  *      Etymology: Middle English witnesse, from Old English witnes knowledge,
   43  *          testimony, witness, from 2wit
   44  *      Date: before 12th century
   45  *      1 : attestation of a fact or event : TESTIMONY
   46  *      2 : one that gives evidence; specifically : one who testifies in
   47  *          a cause or before a judicial tribunal
   48  *      3 : one asked to be present at a transaction so as to be able to
   49  *          testify to its having taken place
   50  *      4 : one who has personal knowledge of something
   51  *      5 a : something serving as evidence or proof : SIGN
   52  *        b : public affirmation by word or example of usually
   53  *            religious faith or conviction <the heroic witness to divine
   54  *            life -- Pilot>
   55  *      6 capitalized : a member of the Jehovah's Witnesses 
   56  */
   57 
   58 /*
   59  * Special rules concerning Giant and lock orders:
   60  *
   61  * 1) Giant must be acquired before any other mutexes.  Stated another way,
   62  *    no other mutex may be held when Giant is acquired.
   63  *
   64  * 2) Giant must be released when blocking on a sleepable lock.
   65  *
   66  * This rule is less obvious, but is a result of Giant providing the same
   67  * semantics as spl().  Basically, when a thread sleeps, it must release
   68  * Giant.  When a thread blocks on a sleepable lock, it sleeps.  Hence rule
   69  * 2).
   70  *
   71  * 3) Giant may be acquired before or after sleepable locks.
   72  *
   73  * This rule is also not quite as obvious.  Giant may be acquired after
   74  * a sleepable lock because it is a non-sleepable lock and non-sleepable
   75  * locks may always be acquired while holding a sleepable lock.  The second
   76  * case, Giant before a sleepable lock, follows from rule 2) above.  Suppose
   77  * you have two threads T1 and T2 and a sleepable lock X.  Suppose that T1
   78  * acquires X and blocks on Giant.  Then suppose that T2 acquires Giant and
   79  * blocks on X.  When T2 blocks on X, T2 will release Giant allowing T1 to
   80  * execute.  Thus, acquiring Giant both before and after a sleepable lock
   81  * will not result in a lock order reversal.
   82  */
   83 
   84 #include <sys/cdefs.h>
   85 __FBSDID("$FreeBSD: releng/6.0/sys/kern/subr_witness.c 151667 2005-10-25 20:17:11Z jhb $");
   86 
   87 #include "opt_ddb.h"
   88 #include "opt_witness.h"
   89 
   90 #include <sys/param.h>
   91 #include <sys/bus.h>
   92 #include <sys/kdb.h>
   93 #include <sys/kernel.h>
   94 #include <sys/ktr.h>
   95 #include <sys/lock.h>
   96 #include <sys/malloc.h>
   97 #include <sys/mutex.h>
   98 #include <sys/proc.h>
   99 #include <sys/sysctl.h>
  100 #include <sys/systm.h>
  101 
  102 #include <ddb/ddb.h>
  103 
  104 #include <machine/stdarg.h>
  105 
  106 /* Define this to check for blessed mutexes */
  107 #undef BLESSING
  108 
  109 #define WITNESS_COUNT 1024
  110 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
  111 /*
  112  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
  113  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
  114  * probably be safe for the most part, but it's still a SWAG.
  115  */
  116 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
  117 
  118 #define WITNESS_NCHILDREN 6
  119 
  120 struct witness_child_list_entry;
  121 
  122 struct witness {
  123         const   char *w_name;
  124         struct  lock_class *w_class;
  125         STAILQ_ENTRY(witness) w_list;           /* List of all witnesses. */
  126         STAILQ_ENTRY(witness) w_typelist;       /* Witnesses of a type. */
  127         struct  witness_child_list_entry *w_children;   /* Great evilness... */
  128         const   char *w_file;
  129         int     w_line;
  130         u_int   w_level;
  131         u_int   w_refcount;
  132         u_char  w_Giant_squawked:1;
  133         u_char  w_other_squawked:1;
  134         u_char  w_same_squawked:1;
  135         u_char  w_displayed:1;
  136 };
  137 
  138 struct witness_child_list_entry {
  139         struct  witness_child_list_entry *wcl_next;
  140         struct  witness *wcl_children[WITNESS_NCHILDREN];
  141         u_int   wcl_count;
  142 };
  143 
  144 STAILQ_HEAD(witness_list, witness);
  145 
  146 #ifdef BLESSING
  147 struct witness_blessed {
  148         const   char *b_lock1;
  149         const   char *b_lock2;
  150 };
  151 #endif
  152 
  153 struct witness_order_list_entry {
  154         const   char *w_name;
  155         struct  lock_class *w_class;
  156 };
  157 
  158 #ifdef BLESSING
  159 static int      blessed(struct witness *, struct witness *);
  160 #endif
  161 static int      depart(struct witness *w);
  162 static struct   witness *enroll(const char *description,
  163                                 struct lock_class *lock_class);
  164 static int      insertchild(struct witness *parent, struct witness *child);
  165 static int      isitmychild(struct witness *parent, struct witness *child);
  166 static int      isitmydescendant(struct witness *parent, struct witness *child);
  167 static int      itismychild(struct witness *parent, struct witness *child);
  168 static int      rebalancetree(struct witness_list *list);
  169 static void     removechild(struct witness *parent, struct witness *child);
  170 static int      reparentchildren(struct witness *newparent,
  171                     struct witness *oldparent);
  172 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
  173 static void     witness_displaydescendants(void(*)(const char *fmt, ...),
  174                                            struct witness *, int indent);
  175 static const char *fixup_filename(const char *file);
  176 static void     witness_leveldescendents(struct witness *parent, int level);
  177 static void     witness_levelall(void);
  178 static struct   witness *witness_get(void);
  179 static void     witness_free(struct witness *m);
  180 static struct   witness_child_list_entry *witness_child_get(void);
  181 static void     witness_child_free(struct witness_child_list_entry *wcl);
  182 static struct   lock_list_entry *witness_lock_list_get(void);
  183 static void     witness_lock_list_free(struct lock_list_entry *lle);
  184 static struct   lock_instance *find_instance(struct lock_list_entry *lock_list,
  185                                              struct lock_object *lock);
  186 static void     witness_list_lock(struct lock_instance *instance);
  187 #ifdef DDB
  188 static void     witness_list(struct thread *td);
  189 static void     witness_display_list(void(*prnt)(const char *fmt, ...),
  190                                      struct witness_list *list);
  191 static void     witness_display(void(*)(const char *fmt, ...));
  192 #endif
  193 
  194 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
  195 
  196 /*
  197  * If set to 0, witness is disabled.  If set to 1, witness performs full lock
  198  * order checking for all locks.  If set to 2 or higher, then witness skips
  199  * the full lock order check if the lock being acquired is at a higher level
  200  * (i.e. farther down in the tree) than the current lock.  This last mode is
  201  * somewhat experimental and not considered fully safe.  At runtime, this
  202  * value may be set to 0 to turn off witness.  witness is not allowed be
  203  * turned on once it is turned off, however.
  204  */
  205 static int witness_watch = 1;
  206 TUNABLE_INT("debug.witness.watch", &witness_watch);
  207 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
  208     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
  209 
  210 #ifdef KDB
  211 /*
  212  * When KDB is enabled and witness_kdb is set to 1, it will cause the system
  213  * to drop into kdebug() when:
  214  *      - a lock heirarchy violation occurs
  215  *      - locks are held when going to sleep.
  216  */
  217 #ifdef WITNESS_KDB
  218 int     witness_kdb = 1;
  219 #else
  220 int     witness_kdb = 0;
  221 #endif
  222 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
  223 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
  224 
  225 /*
  226  * When KDB is enabled and witness_trace is set to 1, it will cause the system
  227  * to print a stack trace:
  228  *      - a lock heirarchy violation occurs
  229  *      - locks are held when going to sleep.
  230  */
  231 int     witness_trace = 1;
  232 TUNABLE_INT("debug.witness.trace", &witness_trace);
  233 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
  234 #endif /* KDB */
  235 
  236 #ifdef WITNESS_SKIPSPIN
  237 int     witness_skipspin = 1;
  238 #else
  239 int     witness_skipspin = 0;
  240 #endif
  241 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
  242 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
  243     &witness_skipspin, 0, "");
  244 
  245 static struct mtx w_mtx;
  246 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
  247 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
  248 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
  249 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
  250 static struct witness_child_list_entry *w_child_free = NULL;
  251 static struct lock_list_entry *w_lock_list_free = NULL;
  252 
  253 static struct witness w_data[WITNESS_COUNT];
  254 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
  255 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
  256 
  257 static struct witness_order_list_entry order_lists[] = {
  258         /*
  259          * sx locks
  260          */
  261         { "proctree", &lock_class_sx },
  262         { "allproc", &lock_class_sx },
  263         { NULL, NULL },
  264         /*
  265          * Various mutexes
  266          */
  267         { "Giant", &lock_class_mtx_sleep },
  268         { "filedesc structure", &lock_class_mtx_sleep },
  269         { "pipe mutex", &lock_class_mtx_sleep },
  270         { "sigio lock", &lock_class_mtx_sleep },
  271         { "process group", &lock_class_mtx_sleep },
  272         { "process lock", &lock_class_mtx_sleep },
  273         { "session", &lock_class_mtx_sleep },
  274         { "uidinfo hash", &lock_class_mtx_sleep },
  275         { "uidinfo struct", &lock_class_mtx_sleep },
  276         { "allprison", &lock_class_mtx_sleep },
  277         { NULL, NULL },
  278         /*
  279          * Sockets
  280          */
  281         { "filedesc structure", &lock_class_mtx_sleep },
  282         { "accept", &lock_class_mtx_sleep },
  283         { "so_snd", &lock_class_mtx_sleep },
  284         { "so_rcv", &lock_class_mtx_sleep },
  285         { "sellck", &lock_class_mtx_sleep },
  286         { NULL, NULL },
  287         /*
  288          * Routing
  289          */
  290         { "so_rcv", &lock_class_mtx_sleep },
  291         { "radix node head", &lock_class_mtx_sleep },
  292         { "rtentry", &lock_class_mtx_sleep },
  293         { "ifaddr", &lock_class_mtx_sleep },
  294         { NULL, NULL },
  295         /*
  296          * Multicast - protocol locks before interface locks.
  297          */
  298         { "in_multi_mtx", &lock_class_mtx_sleep },
  299         { "igmp_mtx", &lock_class_mtx_sleep },
  300         { "if_addr_mtx", &lock_class_mtx_sleep },
  301         { NULL, NULL },
  302         /*
  303          * UNIX Domain Sockets
  304          */
  305         { "unp", &lock_class_mtx_sleep },
  306         { "so_snd", &lock_class_mtx_sleep },
  307         { NULL, NULL },
  308         /*
  309          * UDP/IP
  310          */
  311         { "udp", &lock_class_mtx_sleep },
  312         { "udpinp", &lock_class_mtx_sleep },
  313         { "so_snd", &lock_class_mtx_sleep },
  314         { NULL, NULL },
  315         /*
  316          * TCP/IP
  317          */
  318         { "tcp", &lock_class_mtx_sleep },
  319         { "tcpinp", &lock_class_mtx_sleep },
  320         { "so_snd", &lock_class_mtx_sleep },
  321         { NULL, NULL },
  322         /*
  323          * SLIP
  324          */
  325         { "slip_mtx", &lock_class_mtx_sleep },
  326         { "slip sc_mtx", &lock_class_mtx_sleep },
  327         { NULL, NULL },
  328         /*
  329          * netatalk
  330          */
  331         { "ddp_list_mtx", &lock_class_mtx_sleep },
  332         { "ddp_mtx", &lock_class_mtx_sleep },
  333         { NULL, NULL },
  334         /*
  335          * BPF
  336          */
  337         { "bpf global lock", &lock_class_mtx_sleep },
  338         { "bpf interface lock", &lock_class_mtx_sleep },
  339         { "bpf cdev lock", &lock_class_mtx_sleep },
  340         { NULL, NULL },
  341         /*
  342          * NFS server
  343          */
  344         { "nfsd_mtx", &lock_class_mtx_sleep },
  345         { "so_snd", &lock_class_mtx_sleep },
  346         { NULL, NULL },
  347         /*
  348          * CDEV
  349          */
  350         { "system map", &lock_class_mtx_sleep },
  351         { "vm page queue mutex", &lock_class_mtx_sleep },
  352         { "vnode interlock", &lock_class_mtx_sleep },
  353         { "cdev", &lock_class_mtx_sleep },
  354         { NULL, NULL },
  355         /*
  356          * spin locks
  357          */
  358 #ifdef SMP
  359         { "ap boot", &lock_class_mtx_spin },
  360 #endif
  361         { "rm.mutex_mtx", &lock_class_mtx_spin },
  362         { "sio", &lock_class_mtx_spin },
  363 #ifdef __i386__
  364         { "cy", &lock_class_mtx_spin },
  365 #endif
  366         { "uart_hwmtx", &lock_class_mtx_spin },
  367         { "sabtty", &lock_class_mtx_spin },
  368         { "zstty", &lock_class_mtx_spin },
  369         { "ng_node", &lock_class_mtx_spin },
  370         { "ng_worklist", &lock_class_mtx_spin },
  371         { "taskqueue_fast", &lock_class_mtx_spin },
  372         { "intr table", &lock_class_mtx_spin },
  373         { "ithread table lock", &lock_class_mtx_spin },
  374         { "sleepq chain", &lock_class_mtx_spin },
  375         { "sched lock", &lock_class_mtx_spin },
  376         { "turnstile chain", &lock_class_mtx_spin },
  377         { "td_contested", &lock_class_mtx_spin },
  378         { "callout", &lock_class_mtx_spin },
  379         { "entropy harvest mutex", &lock_class_mtx_spin },
  380         /*
  381          * leaf locks
  382          */
  383         { "allpmaps", &lock_class_mtx_spin },
  384         { "vm page queue free mutex", &lock_class_mtx_spin },
  385         { "icu", &lock_class_mtx_spin },
  386 #ifdef SMP
  387         { "smp rendezvous", &lock_class_mtx_spin },
  388 #if defined(__i386__) || defined(__amd64__)
  389         { "tlb", &lock_class_mtx_spin },
  390 #endif
  391 #ifdef __sparc64__
  392         { "ipi", &lock_class_mtx_spin },
  393         { "rtc_mtx", &lock_class_mtx_spin },
  394 #endif
  395 #endif
  396         { "clk", &lock_class_mtx_spin },
  397         { "mutex profiling lock", &lock_class_mtx_spin },
  398         { "kse zombie lock", &lock_class_mtx_spin },
  399         { "ALD Queue", &lock_class_mtx_spin },
  400 #ifdef __ia64__
  401         { "MCA spin lock", &lock_class_mtx_spin },
  402 #endif
  403 #if defined(__i386__) || defined(__amd64__)
  404         { "pcicfg", &lock_class_mtx_spin },
  405         { "NDIS thread lock", &lock_class_mtx_spin },
  406 #endif
  407         { "tw_osl_io_lock", &lock_class_mtx_spin },
  408         { "tw_osl_q_lock", &lock_class_mtx_spin },
  409         { "tw_cl_io_lock", &lock_class_mtx_spin },
  410         { "tw_cl_intr_lock", &lock_class_mtx_spin },
  411         { "tw_cl_gen_lock", &lock_class_mtx_spin },
  412         { NULL, NULL },
  413         { NULL, NULL }
  414 };
  415 
  416 #ifdef BLESSING
  417 /*
  418  * Pairs of locks which have been blessed
  419  * Don't complain about order problems with blessed locks
  420  */
  421 static struct witness_blessed blessed_list[] = {
  422 };
  423 static int blessed_count =
  424         sizeof(blessed_list) / sizeof(struct witness_blessed);
  425 #endif
  426 
  427 /*
  428  * List of all locks in the system.
  429  */
  430 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
  431 
  432 static struct mtx all_mtx = {
  433         { &lock_class_mtx_sleep,        /* mtx_object.lo_class */
  434           "All locks list",             /* mtx_object.lo_name */
  435           "All locks list",             /* mtx_object.lo_type */
  436           LO_INITIALIZED,               /* mtx_object.lo_flags */
  437           { NULL, NULL },               /* mtx_object.lo_list */
  438           NULL },                       /* mtx_object.lo_witness */
  439         MTX_UNOWNED, 0                  /* mtx_lock, mtx_recurse */
  440 };
  441 
  442 /*
  443  * This global is set to 0 once it becomes safe to use the witness code.
  444  */
  445 static int witness_cold = 1;
  446 
  447 /*
  448  * This global is set to 1 once the static lock orders have been enrolled
  449  * so that a warning can be issued for any spin locks enrolled later.
  450  */
  451 static int witness_spin_warn = 0;
  452 
  453 /*
  454  * Global variables for book keeping.
  455  */
  456 static int lock_cur_cnt;
  457 static int lock_max_cnt;
  458 
  459 /*
  460  * The WITNESS-enabled diagnostic code.
  461  */
  462 static void
  463 witness_initialize(void *dummy __unused)
  464 {
  465         struct lock_object *lock;
  466         struct witness_order_list_entry *order;
  467         struct witness *w, *w1;
  468         int i;
  469 
  470         /*
  471          * We have to release Giant before initializing its witness
  472          * structure so that WITNESS doesn't get confused.
  473          */
  474         mtx_unlock(&Giant);
  475         mtx_assert(&Giant, MA_NOTOWNED);
  476 
  477         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
  478         TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
  479         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
  480             MTX_NOWITNESS);
  481         for (i = 0; i < WITNESS_COUNT; i++)
  482                 witness_free(&w_data[i]);
  483         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
  484                 witness_child_free(&w_childdata[i]);
  485         for (i = 0; i < LOCK_CHILDCOUNT; i++)
  486                 witness_lock_list_free(&w_locklistdata[i]);
  487 
  488         /* First add in all the specified order lists. */
  489         for (order = order_lists; order->w_name != NULL; order++) {
  490                 w = enroll(order->w_name, order->w_class);
  491                 if (w == NULL)
  492                         continue;
  493                 w->w_file = "order list";
  494                 for (order++; order->w_name != NULL; order++) {
  495                         w1 = enroll(order->w_name, order->w_class);
  496                         if (w1 == NULL)
  497                                 continue;
  498                         w1->w_file = "order list";
  499                         if (!itismychild(w, w1))
  500                                 panic("Not enough memory for static orders!");
  501                         w = w1;
  502                 }
  503         }
  504         witness_spin_warn = 1;
  505 
  506         /* Iterate through all locks and add them to witness. */
  507         mtx_lock(&all_mtx);
  508         TAILQ_FOREACH(lock, &all_locks, lo_list) {
  509                 if (lock->lo_flags & LO_WITNESS)
  510                         lock->lo_witness = enroll(lock->lo_type,
  511                             lock->lo_class);
  512                 else
  513                         lock->lo_witness = NULL;
  514         }
  515         mtx_unlock(&all_mtx);
  516 
  517         /* Mark the witness code as being ready for use. */
  518         witness_cold = 0;
  519 
  520         mtx_lock(&Giant);
  521 }
  522 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
  523 
  524 static int
  525 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
  526 {
  527         int error, value;
  528 
  529         value = witness_watch;
  530         error = sysctl_handle_int(oidp, &value, 0, req);
  531         if (error != 0 || req->newptr == NULL)
  532                 return (error);
  533         error = suser(req->td);
  534         if (error != 0)
  535                 return (error);
  536         if (value == witness_watch)
  537                 return (0);
  538         if (value != 0)
  539                 return (EINVAL);
  540         witness_watch = 0;
  541         return (0);
  542 }
  543 
  544 void
  545 witness_init(struct lock_object *lock)
  546 {
  547         struct lock_class *class;
  548 
  549         class = lock->lo_class;
  550         if (lock->lo_flags & LO_INITIALIZED)
  551                 panic("%s: lock (%s) %s is already initialized", __func__,
  552                     class->lc_name, lock->lo_name);
  553         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
  554             (class->lc_flags & LC_RECURSABLE) == 0)
  555                 panic("%s: lock (%s) %s can not be recursable", __func__,
  556                     class->lc_name, lock->lo_name);
  557         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  558             (class->lc_flags & LC_SLEEPABLE) == 0)
  559                 panic("%s: lock (%s) %s can not be sleepable", __func__,
  560                     class->lc_name, lock->lo_name);
  561         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
  562             (class->lc_flags & LC_UPGRADABLE) == 0)
  563                 panic("%s: lock (%s) %s can not be upgradable", __func__,
  564                     class->lc_name, lock->lo_name);
  565 
  566         mtx_lock(&all_mtx);
  567         TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
  568         lock->lo_flags |= LO_INITIALIZED;
  569         lock_cur_cnt++;
  570         if (lock_cur_cnt > lock_max_cnt)
  571                 lock_max_cnt = lock_cur_cnt;
  572         mtx_unlock(&all_mtx);
  573         if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
  574             (lock->lo_flags & LO_WITNESS) != 0)
  575                 lock->lo_witness = enroll(lock->lo_type, class);
  576         else
  577                 lock->lo_witness = NULL;
  578 }
  579 
  580 void
  581 witness_destroy(struct lock_object *lock)
  582 {
  583         struct witness *w;
  584 
  585         if (witness_cold)
  586                 panic("lock (%s) %s destroyed while witness_cold",
  587                     lock->lo_class->lc_name, lock->lo_name);
  588         if ((lock->lo_flags & LO_INITIALIZED) == 0)
  589                 panic("%s: lock (%s) %s is not initialized", __func__,
  590                     lock->lo_class->lc_name, lock->lo_name);
  591 
  592         /* XXX: need to verify that no one holds the lock */
  593         w = lock->lo_witness;
  594         if (w != NULL) {
  595                 mtx_lock_spin(&w_mtx);
  596                 MPASS(w->w_refcount > 0);
  597                 w->w_refcount--;
  598 
  599                 /*
  600                  * Lock is already released if we have an allocation failure
  601                  * and depart() fails.
  602                  */
  603                 if (w->w_refcount != 0 || depart(w))
  604                         mtx_unlock_spin(&w_mtx);
  605         }
  606 
  607         mtx_lock(&all_mtx);
  608         lock_cur_cnt--;
  609         TAILQ_REMOVE(&all_locks, lock, lo_list);
  610         lock->lo_flags &= ~LO_INITIALIZED;
  611         mtx_unlock(&all_mtx);
  612 }
  613 
  614 #ifdef DDB
  615 static void
  616 witness_display_list(void(*prnt)(const char *fmt, ...),
  617                      struct witness_list *list)
  618 {
  619         struct witness *w;
  620 
  621         STAILQ_FOREACH(w, list, w_typelist) {
  622                 if (w->w_file == NULL || w->w_level > 0)
  623                         continue;
  624                 /*
  625                  * This lock has no anscestors, display its descendants. 
  626                  */
  627                 witness_displaydescendants(prnt, w, 0);
  628         }
  629 }
  630         
  631 static void
  632 witness_display(void(*prnt)(const char *fmt, ...))
  633 {
  634         struct witness *w;
  635 
  636         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
  637         witness_levelall();
  638 
  639         /* Clear all the displayed flags. */
  640         STAILQ_FOREACH(w, &w_all, w_list) {
  641                 w->w_displayed = 0;
  642         }
  643 
  644         /*
  645          * First, handle sleep locks which have been acquired at least
  646          * once.
  647          */
  648         prnt("Sleep locks:\n");
  649         witness_display_list(prnt, &w_sleep);
  650         
  651         /*
  652          * Now do spin locks which have been acquired at least once.
  653          */
  654         prnt("\nSpin locks:\n");
  655         witness_display_list(prnt, &w_spin);
  656         
  657         /*
  658          * Finally, any locks which have not been acquired yet.
  659          */
  660         prnt("\nLocks which were never acquired:\n");
  661         STAILQ_FOREACH(w, &w_all, w_list) {
  662                 if (w->w_file != NULL || w->w_refcount == 0)
  663                         continue;
  664                 prnt("%s\n", w->w_name);
  665         }
  666 }
  667 #endif /* DDB */
  668 
  669 /* Trim useless garbage from filenames. */
  670 static const char *
  671 fixup_filename(const char *file)
  672 {
  673 
  674         if (file == NULL)
  675                 return (NULL);
  676         while (strncmp(file, "../", 3) == 0)
  677                 file += 3;
  678         return (file);
  679 }
  680 
  681 int
  682 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
  683 {
  684 
  685         if (witness_watch == 0 || panicstr != NULL)
  686                 return (0);
  687 
  688         /* Require locks that witness knows about. */
  689         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
  690             lock2->lo_witness == NULL)
  691                 return (EINVAL);
  692 
  693         MPASS(!mtx_owned(&w_mtx));
  694         mtx_lock_spin(&w_mtx);
  695 
  696         /*
  697          * If we already have either an explicit or implied lock order that
  698          * is the other way around, then return an error.
  699          */
  700         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
  701                 mtx_unlock_spin(&w_mtx);
  702                 return (EDOOFUS);
  703         }
  704         
  705         /* Try to add the new order. */
  706         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
  707             lock2->lo_type, lock1->lo_type);
  708         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
  709                 return (ENOMEM);
  710         mtx_unlock_spin(&w_mtx);
  711         return (0);
  712 }
  713 
  714 void
  715 witness_checkorder(struct lock_object *lock, int flags, const char *file,
  716     int line)
  717 {
  718         struct lock_list_entry **lock_list, *lle;
  719         struct lock_instance *lock1, *lock2;
  720         struct lock_class *class;
  721         struct witness *w, *w1;
  722         struct thread *td;
  723         int i, j;
  724 
  725         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
  726             panicstr != NULL)
  727                 return;
  728 
  729         /*
  730          * Try locks do not block if they fail to acquire the lock, thus
  731          * there is no danger of deadlocks or of switching while holding a
  732          * spin lock if we acquire a lock via a try operation.  This
  733          * function shouldn't even be called for try locks, so panic if
  734          * that happens.
  735          */
  736         if (flags & LOP_TRYLOCK)
  737                 panic("%s should not be called for try lock operations",
  738                     __func__);
  739 
  740         w = lock->lo_witness;
  741         class = lock->lo_class;
  742         td = curthread;
  743         file = fixup_filename(file);
  744 
  745         if (class->lc_flags & LC_SLEEPLOCK) {
  746                 /*
  747                  * Since spin locks include a critical section, this check
  748                  * implicitly enforces a lock order of all sleep locks before
  749                  * all spin locks.
  750                  */
  751                 if (td->td_critnest != 0 && !kdb_active)
  752                         panic("blockable sleep lock (%s) %s @ %s:%d",
  753                             class->lc_name, lock->lo_name, file, line);
  754 
  755                 /*
  756                  * If this is the first lock acquired then just return as
  757                  * no order checking is needed.
  758                  */
  759                 if (td->td_sleeplocks == NULL)
  760                         return;
  761                 lock_list = &td->td_sleeplocks;
  762         } else {
  763                 /*
  764                  * If this is the first lock, just return as no order
  765                  * checking is needed.  We check this in both if clauses
  766                  * here as unifying the check would require us to use a
  767                  * critical section to ensure we don't migrate while doing
  768                  * the check.  Note that if this is not the first lock, we
  769                  * are already in a critical section and are safe for the
  770                  * rest of the check.
  771                  */
  772                 if (PCPU_GET(spinlocks) == NULL)
  773                         return;
  774                 lock_list = PCPU_PTR(spinlocks);
  775         }
  776 
  777         /*
  778          * Check to see if we are recursing on a lock we already own.  If
  779          * so, make sure that we don't mismatch exclusive and shared lock
  780          * acquires.
  781          */
  782         lock1 = find_instance(*lock_list, lock);
  783         if (lock1 != NULL) {
  784                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
  785                     (flags & LOP_EXCLUSIVE) == 0) {
  786                         printf("shared lock of (%s) %s @ %s:%d\n",
  787                             class->lc_name, lock->lo_name, file, line);
  788                         printf("while exclusively locked from %s:%d\n",
  789                             lock1->li_file, lock1->li_line);
  790                         panic("share->excl");
  791                 }
  792                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
  793                     (flags & LOP_EXCLUSIVE) != 0) {
  794                         printf("exclusive lock of (%s) %s @ %s:%d\n",
  795                             class->lc_name, lock->lo_name, file, line);
  796                         printf("while share locked from %s:%d\n",
  797                             lock1->li_file, lock1->li_line);
  798                         panic("excl->share");
  799                 }
  800                 return;
  801         }
  802 
  803         /*
  804          * Try locks do not block if they fail to acquire the lock, thus
  805          * there is no danger of deadlocks or of switching while holding a
  806          * spin lock if we acquire a lock via a try operation.
  807          */
  808         if (flags & LOP_TRYLOCK)
  809                 return;
  810 
  811         /*
  812          * Check for duplicate locks of the same type.  Note that we only
  813          * have to check for this on the last lock we just acquired.  Any
  814          * other cases will be caught as lock order violations.
  815          */
  816         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
  817         w1 = lock1->li_lock->lo_witness;
  818         if (w1 == w) {
  819                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
  820                     (flags & LOP_DUPOK))
  821                         return;
  822                 w->w_same_squawked = 1;
  823                 printf("acquiring duplicate lock of same type: \"%s\"\n", 
  824                         lock->lo_type);
  825                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
  826                     lock1->li_file, lock1->li_line);
  827                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
  828 #ifdef KDB
  829                 goto debugger;
  830 #else
  831                 return;
  832 #endif
  833         }
  834         MPASS(!mtx_owned(&w_mtx));
  835         mtx_lock_spin(&w_mtx);
  836         /*
  837          * If we have a known higher number just say ok
  838          */
  839         if (witness_watch > 1 && w->w_level > w1->w_level) {
  840                 mtx_unlock_spin(&w_mtx);
  841                 return;
  842         }
  843         /*
  844          * If we know that the the lock we are acquiring comes after
  845          * the lock we most recently acquired in the lock order tree,
  846          * then there is no need for any further checks.
  847          */
  848         if (isitmydescendant(w1, w)) {
  849                 mtx_unlock_spin(&w_mtx);
  850                 return;
  851         }
  852         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
  853                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
  854 
  855                         MPASS(j < WITNESS_COUNT);
  856                         lock1 = &lle->ll_children[i];
  857                         w1 = lock1->li_lock->lo_witness;
  858 
  859                         /*
  860                          * If this lock doesn't undergo witness checking,
  861                          * then skip it.
  862                          */
  863                         if (w1 == NULL) {
  864                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
  865                                     ("lock missing witness structure"));
  866                                 continue;
  867                         }
  868                         /*
  869                          * If we are locking Giant and this is a sleepable
  870                          * lock, then skip it.
  871                          */
  872                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
  873                             lock == &Giant.mtx_object)
  874                                 continue;
  875                         /*
  876                          * If we are locking a sleepable lock and this lock
  877                          * is Giant, then skip it.
  878                          */
  879                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  880                             lock1->li_lock == &Giant.mtx_object)
  881                                 continue;
  882                         /*
  883                          * If we are locking a sleepable lock and this lock
  884                          * isn't sleepable, we want to treat it as a lock
  885                          * order violation to enfore a general lock order of
  886                          * sleepable locks before non-sleepable locks.
  887                          */
  888                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  889                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
  890                                 goto reversal;
  891                         /*
  892                          * Check the lock order hierarchy for a reveresal.
  893                          */
  894                         if (!isitmydescendant(w, w1))
  895                                 continue;
  896                 reversal:
  897                         /*
  898                          * We have a lock order violation, check to see if it
  899                          * is allowed or has already been yelled about.
  900                          */
  901                         mtx_unlock_spin(&w_mtx);
  902 #ifdef BLESSING
  903                         /*
  904                          * If the lock order is blessed, just bail.  We don't
  905                          * look for other lock order violations though, which
  906                          * may be a bug.
  907                          */
  908                         if (blessed(w, w1))
  909                                 return;
  910 #endif
  911                         if (lock1->li_lock == &Giant.mtx_object) {
  912                                 if (w1->w_Giant_squawked)
  913                                         return;
  914                                 else
  915                                         w1->w_Giant_squawked = 1;
  916                         } else {
  917                                 if (w1->w_other_squawked)
  918                                         return;
  919                                 else
  920                                         w1->w_other_squawked = 1;
  921                         }
  922                         /*
  923                          * Ok, yell about it.
  924                          */
  925                         printf("lock order reversal\n");
  926                         /*
  927                          * Try to locate an earlier lock with
  928                          * witness w in our list.
  929                          */
  930                         do {
  931                                 lock2 = &lle->ll_children[i];
  932                                 MPASS(lock2->li_lock != NULL);
  933                                 if (lock2->li_lock->lo_witness == w)
  934                                         break;
  935                                 if (i == 0 && lle->ll_next != NULL) {
  936                                         lle = lle->ll_next;
  937                                         i = lle->ll_count - 1;
  938                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
  939                                 } else
  940                                         i--;
  941                         } while (i >= 0);
  942                         if (i < 0) {
  943                                 printf(" 1st %p %s (%s) @ %s:%d\n",
  944                                     lock1->li_lock, lock1->li_lock->lo_name,
  945                                     lock1->li_lock->lo_type, lock1->li_file,
  946                                     lock1->li_line);
  947                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
  948                                     lock->lo_name, lock->lo_type, file, line);
  949                         } else {
  950                                 printf(" 1st %p %s (%s) @ %s:%d\n",
  951                                     lock2->li_lock, lock2->li_lock->lo_name,
  952                                     lock2->li_lock->lo_type, lock2->li_file,
  953                                     lock2->li_line);
  954                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
  955                                     lock1->li_lock, lock1->li_lock->lo_name,
  956                                     lock1->li_lock->lo_type, lock1->li_file,
  957                                     lock1->li_line);
  958                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
  959                                     lock->lo_name, lock->lo_type, file, line);
  960                         }
  961 #ifdef KDB
  962                         goto debugger;
  963 #else
  964                         return;
  965 #endif
  966                 }
  967         }
  968         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
  969         /*
  970          * If requested, build a new lock order.  However, don't build a new
  971          * relationship between a sleepable lock and Giant if it is in the
  972          * wrong direction.  The correct lock order is that sleepable locks
  973          * always come before Giant.
  974          */
  975         if (flags & LOP_NEWORDER &&
  976             !(lock1->li_lock == &Giant.mtx_object &&
  977             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
  978                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
  979                     lock->lo_type, lock1->li_lock->lo_type);
  980                 if (!itismychild(lock1->li_lock->lo_witness, w))
  981                         /* Witness is dead. */
  982                         return;
  983         } 
  984         mtx_unlock_spin(&w_mtx);
  985         return;
  986 
  987 #ifdef KDB
  988 debugger:
  989         if (witness_trace)
  990                 kdb_backtrace();
  991         if (witness_kdb)
  992                 kdb_enter(__func__);
  993 #endif
  994 }
  995 
  996 void
  997 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
  998 {
  999         struct lock_list_entry **lock_list, *lle;
 1000         struct lock_instance *instance;
 1001         struct witness *w;
 1002         struct thread *td;
 1003 
 1004         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 1005             panicstr != NULL)
 1006                 return;
 1007         w = lock->lo_witness;
 1008         td = curthread;
 1009         file = fixup_filename(file);
 1010 
 1011         /* Determine lock list for this lock. */
 1012         if (lock->lo_class->lc_flags & LC_SLEEPLOCK)
 1013                 lock_list = &td->td_sleeplocks;
 1014         else
 1015                 lock_list = PCPU_PTR(spinlocks);
 1016 
 1017         /* Check to see if we are recursing on a lock we already own. */
 1018         instance = find_instance(*lock_list, lock);
 1019         if (instance != NULL) {
 1020                 instance->li_flags++;
 1021                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
 1022                     td->td_proc->p_pid, lock->lo_name,
 1023                     instance->li_flags & LI_RECURSEMASK);
 1024                 instance->li_file = file;
 1025                 instance->li_line = line;
 1026                 return;
 1027         }
 1028 
 1029         /* Update per-witness last file and line acquire. */
 1030         w->w_file = file;
 1031         w->w_line = line;
 1032 
 1033         /* Find the next open lock instance in the list and fill it. */
 1034         lle = *lock_list;
 1035         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
 1036                 lle = witness_lock_list_get();
 1037                 if (lle == NULL)
 1038                         return;
 1039                 lle->ll_next = *lock_list;
 1040                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
 1041                     td->td_proc->p_pid, lle);
 1042                 *lock_list = lle;
 1043         }
 1044         instance = &lle->ll_children[lle->ll_count++];
 1045         instance->li_lock = lock;
 1046         instance->li_line = line;
 1047         instance->li_file = file;
 1048         if ((flags & LOP_EXCLUSIVE) != 0)
 1049                 instance->li_flags = LI_EXCLUSIVE;
 1050         else
 1051                 instance->li_flags = 0;
 1052         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
 1053             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
 1054 }
 1055 
 1056 void
 1057 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
 1058 {
 1059         struct lock_instance *instance;
 1060         struct lock_class *class;
 1061 
 1062         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1063         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1064                 return;
 1065         class = lock->lo_class;
 1066         file = fixup_filename(file);
 1067         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1068                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
 1069                     class->lc_name, lock->lo_name, file, line);
 1070         if ((flags & LOP_TRYLOCK) == 0)
 1071                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
 1072                     lock->lo_name, file, line);
 1073         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
 1074                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
 1075                     class->lc_name, lock->lo_name, file, line);
 1076         instance = find_instance(curthread->td_sleeplocks, lock);
 1077         if (instance == NULL)
 1078                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
 1079                     class->lc_name, lock->lo_name, file, line);
 1080         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
 1081                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
 1082                     class->lc_name, lock->lo_name, file, line);
 1083         if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1084                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1085                     class->lc_name, lock->lo_name,
 1086                     instance->li_flags & LI_RECURSEMASK, file, line);
 1087         instance->li_flags |= LI_EXCLUSIVE;
 1088 }
 1089 
 1090 void
 1091 witness_downgrade(struct lock_object *lock, int flags, const char *file,
 1092     int line)
 1093 {
 1094         struct lock_instance *instance;
 1095         struct lock_class *class;
 1096 
 1097         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1098         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1099                 return;
 1100         class = lock->lo_class;
 1101         file = fixup_filename(file);
 1102         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1103                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
 1104                     class->lc_name, lock->lo_name, file, line);
 1105         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
 1106                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
 1107                     class->lc_name, lock->lo_name, file, line);
 1108         instance = find_instance(curthread->td_sleeplocks, lock);
 1109         if (instance == NULL)
 1110                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
 1111                     class->lc_name, lock->lo_name, file, line);
 1112         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
 1113                 panic("downgrade of shared lock (%s) %s @ %s:%d",
 1114                     class->lc_name, lock->lo_name, file, line);
 1115         if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1116                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1117                     class->lc_name, lock->lo_name,
 1118                     instance->li_flags & LI_RECURSEMASK, file, line);
 1119         instance->li_flags &= ~LI_EXCLUSIVE;
 1120 }
 1121 
 1122 void
 1123 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
 1124 {
 1125         struct lock_list_entry **lock_list, *lle;
 1126         struct lock_instance *instance;
 1127         struct lock_class *class;
 1128         struct thread *td;
 1129         register_t s;
 1130         int i, j;
 1131 
 1132         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 1133             panicstr != NULL)
 1134                 return;
 1135         td = curthread;
 1136         class = lock->lo_class;
 1137         file = fixup_filename(file);
 1138 
 1139         /* Find lock instance associated with this lock. */
 1140         if (class->lc_flags & LC_SLEEPLOCK)
 1141                 lock_list = &td->td_sleeplocks;
 1142         else
 1143                 lock_list = PCPU_PTR(spinlocks);
 1144         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
 1145                 for (i = 0; i < (*lock_list)->ll_count; i++) {
 1146                         instance = &(*lock_list)->ll_children[i];
 1147                         if (instance->li_lock == lock)
 1148                                 goto found;
 1149                 }
 1150         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
 1151             file, line);
 1152 found:
 1153 
 1154         /* First, check for shared/exclusive mismatches. */
 1155         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
 1156             (flags & LOP_EXCLUSIVE) == 0) {
 1157                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1158                     lock->lo_name, file, line);
 1159                 printf("while exclusively locked from %s:%d\n",
 1160                     instance->li_file, instance->li_line);
 1161                 panic("excl->ushare");
 1162         }
 1163         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
 1164             (flags & LOP_EXCLUSIVE) != 0) {
 1165                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1166                     lock->lo_name, file, line);
 1167                 printf("while share locked from %s:%d\n", instance->li_file,
 1168                     instance->li_line);
 1169                 panic("share->uexcl");
 1170         }
 1171 
 1172         /* If we are recursed, unrecurse. */
 1173         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
 1174                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
 1175                     td->td_proc->p_pid, instance->li_lock->lo_name,
 1176                     instance->li_flags);
 1177                 instance->li_flags--;
 1178                 return;
 1179         }
 1180 
 1181         /* Otherwise, remove this item from the list. */
 1182         s = intr_disable();
 1183         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
 1184             td->td_proc->p_pid, instance->li_lock->lo_name,
 1185             (*lock_list)->ll_count - 1);
 1186         for (j = i; j < (*lock_list)->ll_count - 1; j++)
 1187                 (*lock_list)->ll_children[j] =
 1188                     (*lock_list)->ll_children[j + 1];
 1189         (*lock_list)->ll_count--;
 1190         intr_restore(s);
 1191 
 1192         /* If this lock list entry is now empty, free it. */
 1193         if ((*lock_list)->ll_count == 0) {
 1194                 lle = *lock_list;
 1195                 *lock_list = lle->ll_next;
 1196                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
 1197                     td->td_proc->p_pid, lle);
 1198                 witness_lock_list_free(lle);
 1199         }
 1200 }
 1201 
 1202 /*
 1203  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
 1204  * exempt Giant and sleepable locks from the checks as well.  If any
 1205  * non-exempt locks are held, then a supplied message is printed to the
 1206  * console along with a list of the offending locks.  If indicated in the
 1207  * flags then a failure results in a panic as well.
 1208  */
 1209 int
 1210 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
 1211 {
 1212         struct lock_list_entry *lle;
 1213         struct lock_instance *lock1;
 1214         struct thread *td;
 1215         va_list ap;
 1216         int i, n;
 1217 
 1218         if (witness_cold || witness_watch == 0 || panicstr != NULL)
 1219                 return (0);
 1220         n = 0;
 1221         td = curthread;
 1222         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
 1223                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1224                         lock1 = &lle->ll_children[i];
 1225                         if (lock1->li_lock == lock)
 1226                                 continue;
 1227                         if (flags & WARN_GIANTOK &&
 1228                             lock1->li_lock == &Giant.mtx_object)
 1229                                 continue;
 1230                         if (flags & WARN_SLEEPOK &&
 1231                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
 1232                                 continue;
 1233                         if (n == 0) {
 1234                                 va_start(ap, fmt);
 1235                                 vprintf(fmt, ap);
 1236                                 va_end(ap);
 1237                                 printf(" with the following");
 1238                                 if (flags & WARN_SLEEPOK)
 1239                                         printf(" non-sleepable");
 1240                                 printf(" locks held:\n");
 1241                         }
 1242                         n++;
 1243                         witness_list_lock(lock1);
 1244                 }
 1245         if (PCPU_GET(spinlocks) != NULL) {
 1246                 /*
 1247                  * Since we already hold a spinlock preemption is
 1248                  * already blocked.
 1249                  */
 1250                 if (n == 0) {
 1251                         va_start(ap, fmt);
 1252                         vprintf(fmt, ap);
 1253                         va_end(ap);
 1254                         printf(" with the following");
 1255                         if (flags & WARN_SLEEPOK)
 1256                                 printf(" non-sleepable");
 1257                         printf(" locks held:\n");
 1258                 }
 1259                 n += witness_list_locks(PCPU_PTR(spinlocks));
 1260         }
 1261         if (flags & WARN_PANIC && n)
 1262                 panic("witness_warn");
 1263 #ifdef KDB
 1264         else if (witness_kdb && n)
 1265                 kdb_enter(__func__);
 1266         else if (witness_trace && n)
 1267                 kdb_backtrace();
 1268 #endif
 1269         return (n);
 1270 }
 1271 
 1272 const char *
 1273 witness_file(struct lock_object *lock)
 1274 {
 1275         struct witness *w;
 1276 
 1277         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
 1278                 return ("?");
 1279         w = lock->lo_witness;
 1280         return (w->w_file);
 1281 }
 1282 
 1283 int
 1284 witness_line(struct lock_object *lock)
 1285 {
 1286         struct witness *w;
 1287 
 1288         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
 1289                 return (0);
 1290         w = lock->lo_witness;
 1291         return (w->w_line);
 1292 }
 1293 
 1294 static struct witness *
 1295 enroll(const char *description, struct lock_class *lock_class)
 1296 {
 1297         struct witness *w;
 1298 
 1299         if (witness_watch == 0 || panicstr != NULL)
 1300                 return (NULL);
 1301         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
 1302                 return (NULL);
 1303         mtx_lock_spin(&w_mtx);
 1304         STAILQ_FOREACH(w, &w_all, w_list) {
 1305                 if (w->w_name == description || (w->w_refcount > 0 &&
 1306                     strcmp(description, w->w_name) == 0)) {
 1307                         w->w_refcount++;
 1308                         mtx_unlock_spin(&w_mtx);
 1309                         if (lock_class != w->w_class)
 1310                                 panic(
 1311                                 "lock (%s) %s does not match earlier (%s) lock",
 1312                                     description, lock_class->lc_name,
 1313                                     w->w_class->lc_name);
 1314                         return (w);
 1315                 }
 1316         }
 1317         /*
 1318          * We issue a warning for any spin locks not defined in the static
 1319          * order list as a way to discourage their use (folks should really
 1320          * be using non-spin mutexes most of the time).  However, several
 1321          * 3rd part device drivers use spin locks because that is all they
 1322          * have available on Windows and Linux and they think that normal
 1323          * mutexes are insufficient.
 1324          */
 1325         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
 1326                 printf("WITNESS: spin lock %s not in order list", description);
 1327         if ((w = witness_get()) == NULL)
 1328                 return (NULL);
 1329         w->w_name = description;
 1330         w->w_class = lock_class;
 1331         w->w_refcount = 1;
 1332         STAILQ_INSERT_HEAD(&w_all, w, w_list);
 1333         if (lock_class->lc_flags & LC_SPINLOCK)
 1334                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
 1335         else if (lock_class->lc_flags & LC_SLEEPLOCK)
 1336                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
 1337         else {
 1338                 mtx_unlock_spin(&w_mtx);
 1339                 panic("lock class %s is not sleep or spin",
 1340                     lock_class->lc_name);
 1341         }
 1342         mtx_unlock_spin(&w_mtx);
 1343         return (w);
 1344 }
 1345 
 1346 /* Don't let the door bang you on the way out... */
 1347 static int
 1348 depart(struct witness *w)
 1349 {
 1350         struct witness_child_list_entry *wcl, *nwcl;
 1351         struct witness_list *list;
 1352         struct witness *parent;
 1353 
 1354         MPASS(w->w_refcount == 0);
 1355         if (w->w_class->lc_flags & LC_SLEEPLOCK)
 1356                 list = &w_sleep;
 1357         else
 1358                 list = &w_spin;
 1359         /*
 1360          * First, we run through the entire tree looking for any
 1361          * witnesses that the outgoing witness is a child of.  For
 1362          * each parent that we find, we reparent all the direct
 1363          * children of the outgoing witness to its parent.
 1364          */
 1365         STAILQ_FOREACH(parent, list, w_typelist) {
 1366                 if (!isitmychild(parent, w))
 1367                         continue;
 1368                 removechild(parent, w);
 1369                 if (!reparentchildren(parent, w))
 1370                         return (0);
 1371         }
 1372 
 1373         /*
 1374          * Now we go through and free up the child list of the
 1375          * outgoing witness.
 1376          */
 1377         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
 1378                 nwcl = wcl->wcl_next;
 1379                 witness_child_free(wcl);
 1380         }
 1381 
 1382         /*
 1383          * Detach from various lists and free.
 1384          */
 1385         STAILQ_REMOVE(list, w, witness, w_typelist);
 1386         STAILQ_REMOVE(&w_all, w, witness, w_list);
 1387         witness_free(w);
 1388 
 1389         /* Finally, fixup the tree. */
 1390         return (rebalancetree(list));
 1391 }
 1392 
 1393 /*
 1394  * Prune an entire lock order tree.  We look for cases where a lock
 1395  * is now both a descendant and a direct child of a given lock.  In
 1396  * that case, we want to remove the direct child link from the tree.
 1397  *
 1398  * Returns false if insertchild() fails.
 1399  */
 1400 static int
 1401 rebalancetree(struct witness_list *list)
 1402 {
 1403         struct witness *child, *parent;
 1404 
 1405         STAILQ_FOREACH(child, list, w_typelist) {
 1406                 STAILQ_FOREACH(parent, list, w_typelist) {
 1407                         if (!isitmychild(parent, child))
 1408                                 continue;
 1409                         removechild(parent, child);
 1410                         if (isitmydescendant(parent, child))
 1411                                 continue;
 1412                         if (!insertchild(parent, child))
 1413                                 return (0);
 1414                 }
 1415         }
 1416         witness_levelall();
 1417         return (1);
 1418 }
 1419 
 1420 /*
 1421  * Add "child" as a direct child of "parent".  Returns false if
 1422  * we fail due to out of memory.
 1423  */
 1424 static int
 1425 insertchild(struct witness *parent, struct witness *child)
 1426 {
 1427         struct witness_child_list_entry **wcl;
 1428 
 1429         MPASS(child != NULL && parent != NULL);
 1430 
 1431         /*
 1432          * Insert "child" after "parent"
 1433          */
 1434         wcl = &parent->w_children;
 1435         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
 1436                 wcl = &(*wcl)->wcl_next;
 1437         if (*wcl == NULL) {
 1438                 *wcl = witness_child_get();
 1439                 if (*wcl == NULL)
 1440                         return (0);
 1441         }
 1442         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
 1443 
 1444         return (1);
 1445 }
 1446 
 1447 /*
 1448  * Make all the direct descendants of oldparent be direct descendants
 1449  * of newparent.
 1450  */
 1451 static int
 1452 reparentchildren(struct witness *newparent, struct witness *oldparent)
 1453 {
 1454         struct witness_child_list_entry *wcl;
 1455         int i;
 1456 
 1457         /* Avoid making a witness a child of itself. */
 1458         MPASS(!isitmychild(oldparent, newparent));
 1459         
 1460         for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 1461                 for (i = 0; i < wcl->wcl_count; i++)
 1462                         if (!insertchild(newparent, wcl->wcl_children[i]))
 1463                                 return (0);
 1464         return (1);
 1465 }
 1466 
 1467 static int
 1468 itismychild(struct witness *parent, struct witness *child)
 1469 {
 1470         struct witness_list *list;
 1471 
 1472         MPASS(child != NULL && parent != NULL);
 1473         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
 1474             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
 1475                 panic(
 1476                 "%s: parent (%s) and child (%s) are not the same lock type",
 1477                     __func__, parent->w_class->lc_name,
 1478                     child->w_class->lc_name);
 1479 
 1480         if (!insertchild(parent, child))
 1481                 return (0);
 1482 
 1483         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
 1484                 list = &w_sleep;
 1485         else
 1486                 list = &w_spin;
 1487         return (rebalancetree(list));
 1488 }
 1489 
 1490 static void
 1491 removechild(struct witness *parent, struct witness *child)
 1492 {
 1493         struct witness_child_list_entry **wcl, *wcl1;
 1494         int i;
 1495 
 1496         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
 1497                 for (i = 0; i < (*wcl)->wcl_count; i++)
 1498                         if ((*wcl)->wcl_children[i] == child)
 1499                                 goto found;
 1500         return;
 1501 found:
 1502         (*wcl)->wcl_count--;
 1503         if ((*wcl)->wcl_count > i)
 1504                 (*wcl)->wcl_children[i] =
 1505                     (*wcl)->wcl_children[(*wcl)->wcl_count];
 1506         MPASS((*wcl)->wcl_children[i] != NULL);
 1507         if ((*wcl)->wcl_count != 0)
 1508                 return;
 1509         wcl1 = *wcl;
 1510         *wcl = wcl1->wcl_next;
 1511         witness_child_free(wcl1);
 1512 }
 1513 
 1514 static int
 1515 isitmychild(struct witness *parent, struct witness *child)
 1516 {
 1517         struct witness_child_list_entry *wcl;
 1518         int i;
 1519 
 1520         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
 1521                 for (i = 0; i < wcl->wcl_count; i++) {
 1522                         if (wcl->wcl_children[i] == child)
 1523                                 return (1);
 1524                 }
 1525         }
 1526         return (0);
 1527 }
 1528 
 1529 static int
 1530 isitmydescendant(struct witness *parent, struct witness *child)
 1531 {
 1532         struct witness_child_list_entry *wcl;
 1533         int i, j;
 1534 
 1535         if (isitmychild(parent, child))
 1536                 return (1);
 1537         j = 0;
 1538         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
 1539                 MPASS(j < 1000);
 1540                 for (i = 0; i < wcl->wcl_count; i++) {
 1541                         if (isitmydescendant(wcl->wcl_children[i], child))
 1542                                 return (1);
 1543                 }
 1544                 j++;
 1545         }
 1546         return (0);
 1547 }
 1548 
 1549 static void
 1550 witness_levelall (void)
 1551 {
 1552         struct witness_list *list;
 1553         struct witness *w, *w1;
 1554 
 1555         /*
 1556          * First clear all levels.
 1557          */
 1558         STAILQ_FOREACH(w, &w_all, w_list) {
 1559                 w->w_level = 0;
 1560         }
 1561 
 1562         /*
 1563          * Look for locks with no parent and level all their descendants.
 1564          */
 1565         STAILQ_FOREACH(w, &w_all, w_list) {
 1566                 /*
 1567                  * This is just an optimization, technically we could get
 1568                  * away just walking the all list each time.
 1569                  */
 1570                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
 1571                         list = &w_sleep;
 1572                 else
 1573                         list = &w_spin;
 1574                 STAILQ_FOREACH(w1, list, w_typelist) {
 1575                         if (isitmychild(w1, w))
 1576                                 goto skip;
 1577                 }
 1578                 witness_leveldescendents(w, 0);
 1579         skip:
 1580                 ;       /* silence GCC 3.x */
 1581         }
 1582 }
 1583 
 1584 static void
 1585 witness_leveldescendents(struct witness *parent, int level)
 1586 {
 1587         struct witness_child_list_entry *wcl;
 1588         int i;
 1589 
 1590         if (parent->w_level < level)
 1591                 parent->w_level = level;
 1592         level++;
 1593         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 1594                 for (i = 0; i < wcl->wcl_count; i++)
 1595                         witness_leveldescendents(wcl->wcl_children[i], level);
 1596 }
 1597 
 1598 static void
 1599 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
 1600                            struct witness *parent, int indent)
 1601 {
 1602         struct witness_child_list_entry *wcl;
 1603         int i, level;
 1604 
 1605         level = parent->w_level;
 1606         prnt("%-2d", level);
 1607         for (i = 0; i < indent; i++)
 1608                 prnt(" ");
 1609         if (parent->w_refcount > 0)
 1610                 prnt("%s", parent->w_name);
 1611         else
 1612                 prnt("(dead)");
 1613         if (parent->w_displayed) {
 1614                 prnt(" -- (already displayed)\n");
 1615                 return;
 1616         }
 1617         parent->w_displayed = 1;
 1618         if (parent->w_refcount > 0) {
 1619                 if (parent->w_file != NULL)
 1620                         prnt(" -- last acquired @ %s:%d", parent->w_file,
 1621                             parent->w_line);
 1622         }
 1623         prnt("\n");
 1624         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
 1625                 for (i = 0; i < wcl->wcl_count; i++)
 1626                             witness_displaydescendants(prnt,
 1627                                 wcl->wcl_children[i], indent + 1);
 1628 }
 1629 
 1630 #ifdef BLESSING
 1631 static int
 1632 blessed(struct witness *w1, struct witness *w2)
 1633 {
 1634         int i;
 1635         struct witness_blessed *b;
 1636 
 1637         for (i = 0; i < blessed_count; i++) {
 1638                 b = &blessed_list[i];
 1639                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
 1640                         if (strcmp(w2->w_name, b->b_lock2) == 0)
 1641                                 return (1);
 1642                         continue;
 1643                 }
 1644                 if (strcmp(w1->w_name, b->b_lock2) == 0)
 1645                         if (strcmp(w2->w_name, b->b_lock1) == 0)
 1646                                 return (1);
 1647         }
 1648         return (0);
 1649 }
 1650 #endif
 1651 
 1652 static struct witness *
 1653 witness_get(void)
 1654 {
 1655         struct witness *w;
 1656 
 1657         if (witness_watch == 0) {
 1658                 mtx_unlock_spin(&w_mtx);
 1659                 return (NULL);
 1660         }
 1661         if (STAILQ_EMPTY(&w_free)) {
 1662                 witness_watch = 0;
 1663                 mtx_unlock_spin(&w_mtx);
 1664                 printf("%s: witness exhausted\n", __func__);
 1665                 return (NULL);
 1666         }
 1667         w = STAILQ_FIRST(&w_free);
 1668         STAILQ_REMOVE_HEAD(&w_free, w_list);
 1669         bzero(w, sizeof(*w));
 1670         return (w);
 1671 }
 1672 
 1673 static void
 1674 witness_free(struct witness *w)
 1675 {
 1676 
 1677         STAILQ_INSERT_HEAD(&w_free, w, w_list);
 1678 }
 1679 
 1680 static struct witness_child_list_entry *
 1681 witness_child_get(void)
 1682 {
 1683         struct witness_child_list_entry *wcl;
 1684 
 1685         if (witness_watch == 0) {
 1686                 mtx_unlock_spin(&w_mtx);
 1687                 return (NULL);
 1688         }
 1689         wcl = w_child_free;
 1690         if (wcl == NULL) {
 1691                 witness_watch = 0;
 1692                 mtx_unlock_spin(&w_mtx);
 1693                 printf("%s: witness exhausted\n", __func__);
 1694                 return (NULL);
 1695         }
 1696         w_child_free = wcl->wcl_next;
 1697         bzero(wcl, sizeof(*wcl));
 1698         return (wcl);
 1699 }
 1700 
 1701 static void
 1702 witness_child_free(struct witness_child_list_entry *wcl)
 1703 {
 1704 
 1705         wcl->wcl_next = w_child_free;
 1706         w_child_free = wcl;
 1707 }
 1708 
 1709 static struct lock_list_entry *
 1710 witness_lock_list_get(void)
 1711 {
 1712         struct lock_list_entry *lle;
 1713 
 1714         if (witness_watch == 0)
 1715                 return (NULL);
 1716         mtx_lock_spin(&w_mtx);
 1717         lle = w_lock_list_free;
 1718         if (lle == NULL) {
 1719                 witness_watch = 0;
 1720                 mtx_unlock_spin(&w_mtx);
 1721                 printf("%s: witness exhausted\n", __func__);
 1722                 return (NULL);
 1723         }
 1724         w_lock_list_free = lle->ll_next;
 1725         mtx_unlock_spin(&w_mtx);
 1726         bzero(lle, sizeof(*lle));
 1727         return (lle);
 1728 }
 1729                 
 1730 static void
 1731 witness_lock_list_free(struct lock_list_entry *lle)
 1732 {
 1733 
 1734         mtx_lock_spin(&w_mtx);
 1735         lle->ll_next = w_lock_list_free;
 1736         w_lock_list_free = lle;
 1737         mtx_unlock_spin(&w_mtx);
 1738 }
 1739 
 1740 static struct lock_instance *
 1741 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
 1742 {
 1743         struct lock_list_entry *lle;
 1744         struct lock_instance *instance;
 1745         int i;
 1746 
 1747         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
 1748                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1749                         instance = &lle->ll_children[i];
 1750                         if (instance->li_lock == lock)
 1751                                 return (instance);
 1752                 }
 1753         return (NULL);
 1754 }
 1755 
 1756 static void
 1757 witness_list_lock(struct lock_instance *instance)
 1758 {
 1759         struct lock_object *lock;
 1760 
 1761         lock = instance->li_lock;
 1762         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
 1763             "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name);
 1764         if (lock->lo_type != lock->lo_name)
 1765                 printf(" (%s)", lock->lo_type);
 1766         printf(" r = %d (%p) locked @ %s:%d\n",
 1767             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
 1768             instance->li_line);
 1769 }
 1770 
 1771 #ifdef DDB
 1772 static int
 1773 witness_thread_has_locks(struct thread *td)
 1774 {
 1775 
 1776         return (td->td_sleeplocks != NULL);
 1777 }
 1778 
 1779 static int
 1780 witness_proc_has_locks(struct proc *p)
 1781 {
 1782         struct thread *td;
 1783 
 1784         FOREACH_THREAD_IN_PROC(p, td) {
 1785                 if (witness_thread_has_locks(td))
 1786                         return (1);
 1787         }
 1788         return (0);
 1789 }
 1790 #endif
 1791 
 1792 int
 1793 witness_list_locks(struct lock_list_entry **lock_list)
 1794 {
 1795         struct lock_list_entry *lle;
 1796         int i, nheld;
 1797 
 1798         nheld = 0;
 1799         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
 1800                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1801                         witness_list_lock(&lle->ll_children[i]);
 1802                         nheld++;
 1803                 }
 1804         return (nheld);
 1805 }
 1806 
 1807 /*
 1808  * This is a bit risky at best.  We call this function when we have timed
 1809  * out acquiring a spin lock, and we assume that the other CPU is stuck
 1810  * with this lock held.  So, we go groveling around in the other CPU's
 1811  * per-cpu data to try to find the lock instance for this spin lock to
 1812  * see when it was last acquired.
 1813  */
 1814 void
 1815 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
 1816 {
 1817         struct lock_instance *instance;
 1818         struct pcpu *pc;
 1819 
 1820         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
 1821                 return;
 1822         pc = pcpu_find(owner->td_oncpu);
 1823         instance = find_instance(pc->pc_spinlocks, lock);
 1824         if (instance != NULL)
 1825                 witness_list_lock(instance);
 1826 }
 1827 
 1828 void
 1829 witness_save(struct lock_object *lock, const char **filep, int *linep)
 1830 {
 1831         struct lock_instance *instance;
 1832 
 1833         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1834         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1835                 return;
 1836         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
 1837                 panic("%s: lock (%s) %s is not a sleep lock", __func__,
 1838                     lock->lo_class->lc_name, lock->lo_name);
 1839         instance = find_instance(curthread->td_sleeplocks, lock);
 1840         if (instance == NULL)
 1841                 panic("%s: lock (%s) %s not locked", __func__,
 1842                     lock->lo_class->lc_name, lock->lo_name);
 1843         *filep = instance->li_file;
 1844         *linep = instance->li_line;
 1845 }
 1846 
 1847 void
 1848 witness_restore(struct lock_object *lock, const char *file, int line)
 1849 {
 1850         struct lock_instance *instance;
 1851 
 1852         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1853         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1854                 return;
 1855         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0)
 1856                 panic("%s: lock (%s) %s is not a sleep lock", __func__,
 1857                     lock->lo_class->lc_name, lock->lo_name);
 1858         instance = find_instance(curthread->td_sleeplocks, lock);
 1859         if (instance == NULL)
 1860                 panic("%s: lock (%s) %s not locked", __func__,
 1861                     lock->lo_class->lc_name, lock->lo_name);
 1862         lock->lo_witness->w_file = file;
 1863         lock->lo_witness->w_line = line;
 1864         instance->li_file = file;
 1865         instance->li_line = line;
 1866 }
 1867 
 1868 void
 1869 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
 1870 {
 1871 #ifdef INVARIANT_SUPPORT
 1872         struct lock_instance *instance;
 1873 
 1874         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1875                 return;
 1876         if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0)
 1877                 instance = find_instance(curthread->td_sleeplocks, lock);
 1878         else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0)
 1879                 instance = find_instance(PCPU_GET(spinlocks), lock);
 1880         else {
 1881                 panic("Lock (%s) %s is not sleep or spin!",
 1882                     lock->lo_class->lc_name, lock->lo_name);
 1883         }
 1884         file = fixup_filename(file);
 1885         switch (flags) {
 1886         case LA_UNLOCKED:
 1887                 if (instance != NULL)
 1888                         panic("Lock (%s) %s locked @ %s:%d.",
 1889                             lock->lo_class->lc_name, lock->lo_name, file, line);
 1890                 break;
 1891         case LA_LOCKED:
 1892         case LA_LOCKED | LA_RECURSED:
 1893         case LA_LOCKED | LA_NOTRECURSED:
 1894         case LA_SLOCKED:
 1895         case LA_SLOCKED | LA_RECURSED:
 1896         case LA_SLOCKED | LA_NOTRECURSED:
 1897         case LA_XLOCKED:
 1898         case LA_XLOCKED | LA_RECURSED:
 1899         case LA_XLOCKED | LA_NOTRECURSED:
 1900                 if (instance == NULL) {
 1901                         panic("Lock (%s) %s not locked @ %s:%d.",
 1902                             lock->lo_class->lc_name, lock->lo_name, file, line);
 1903                         break;
 1904                 }
 1905                 if ((flags & LA_XLOCKED) != 0 &&
 1906                     (instance->li_flags & LI_EXCLUSIVE) == 0)
 1907                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
 1908                             lock->lo_class->lc_name, lock->lo_name, file, line);
 1909                 if ((flags & LA_SLOCKED) != 0 &&
 1910                     (instance->li_flags & LI_EXCLUSIVE) != 0)
 1911                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
 1912                             lock->lo_class->lc_name, lock->lo_name, file, line);
 1913                 if ((flags & LA_RECURSED) != 0 &&
 1914                     (instance->li_flags & LI_RECURSEMASK) == 0)
 1915                         panic("Lock (%s) %s not recursed @ %s:%d.",
 1916                             lock->lo_class->lc_name, lock->lo_name, file, line);
 1917                 if ((flags & LA_NOTRECURSED) != 0 &&
 1918                     (instance->li_flags & LI_RECURSEMASK) != 0)
 1919                         panic("Lock (%s) %s recursed @ %s:%d.",
 1920                             lock->lo_class->lc_name, lock->lo_name, file, line);
 1921                 break;
 1922         default:
 1923                 panic("Invalid lock assertion at %s:%d.", file, line);
 1924 
 1925         }
 1926 #endif  /* INVARIANT_SUPPORT */
 1927 }
 1928 
 1929 #ifdef DDB
 1930 static void
 1931 witness_list(struct thread *td)
 1932 {
 1933 
 1934         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1935         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
 1936 
 1937         if (witness_watch == 0)
 1938                 return;
 1939 
 1940         witness_list_locks(&td->td_sleeplocks);
 1941 
 1942         /*
 1943          * We only handle spinlocks if td == curthread.  This is somewhat broken
 1944          * if td is currently executing on some other CPU and holds spin locks
 1945          * as we won't display those locks.  If we had a MI way of getting
 1946          * the per-cpu data for a given cpu then we could use
 1947          * td->td_oncpu to get the list of spinlocks for this thread
 1948          * and "fix" this.
 1949          *
 1950          * That still wouldn't really fix this unless we locked sched_lock
 1951          * or stopped the other CPU to make sure it wasn't changing the list
 1952          * out from under us.  It is probably best to just not try to handle
 1953          * threads on other CPU's for now.
 1954          */
 1955         if (td == curthread && PCPU_GET(spinlocks) != NULL)
 1956                 witness_list_locks(PCPU_PTR(spinlocks));
 1957 }
 1958 
 1959 DB_SHOW_COMMAND(locks, db_witness_list)
 1960 {
 1961         struct thread *td;
 1962         pid_t pid;
 1963         struct proc *p;
 1964 
 1965         if (have_addr) {
 1966                 pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
 1967                     ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
 1968                     ((addr >> 16) % 16) * 10000;
 1969                 /* sx_slock(&allproc_lock); */
 1970                 FOREACH_PROC_IN_SYSTEM(p) {
 1971                         if (p->p_pid == pid)
 1972                                 break;
 1973                 }
 1974                 /* sx_sunlock(&allproc_lock); */
 1975                 if (p == NULL) {
 1976                         db_printf("pid %d not found\n", pid);
 1977                         return;
 1978                 }
 1979                 FOREACH_THREAD_IN_PROC(p, td) {
 1980                         witness_list(td);
 1981                 }
 1982         } else {
 1983                 td = curthread;
 1984                 witness_list(td);
 1985         }
 1986 }
 1987 
 1988 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
 1989 {
 1990         struct thread *td;
 1991         struct proc *p;
 1992 
 1993         /*
 1994          * It would be nice to list only threads and processes that actually
 1995          * held sleep locks, but that information is currently not exported
 1996          * by WITNESS.
 1997          */
 1998         FOREACH_PROC_IN_SYSTEM(p) {
 1999                 if (!witness_proc_has_locks(p))
 2000                         continue;
 2001                 FOREACH_THREAD_IN_PROC(p, td) {
 2002                         if (!witness_thread_has_locks(td))
 2003                                 continue;
 2004                         printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
 2005                             p->p_comm, td, td->td_tid);
 2006                         witness_list(td);
 2007                 }
 2008         }
 2009 }
 2010 
 2011 DB_SHOW_COMMAND(witness, db_witness_display)
 2012 {
 2013 
 2014         witness_display(db_printf);
 2015 }
 2016 #endif

Cache object: 0807e224bdc9a44589cdb01929694c4f


[ 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.