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/6.3/sys/kern/subr_witness.c 173886 2007-11-24 19:45:58Z cvs2svn $");
   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 void     removechild(struct witness *parent, struct witness *child);
  169 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
  170 static const char *fixup_filename(const char *file);
  171 static struct   witness *witness_get(void);
  172 static void     witness_free(struct witness *m);
  173 static struct   witness_child_list_entry *witness_child_get(void);
  174 static void     witness_child_free(struct witness_child_list_entry *wcl);
  175 static struct   lock_list_entry *witness_lock_list_get(void);
  176 static void     witness_lock_list_free(struct lock_list_entry *lle);
  177 static struct   lock_instance *find_instance(struct lock_list_entry *lock_list,
  178                                              struct lock_object *lock);
  179 static void     witness_list_lock(struct lock_instance *instance);
  180 #ifdef DDB
  181 static void     witness_leveldescendents(struct witness *parent, int level);
  182 static void     witness_levelall(void);
  183 static void     witness_displaydescendants(void(*)(const char *fmt, ...),
  184                                            struct witness *, int indent);
  185 static void     witness_display_list(void(*prnt)(const char *fmt, ...),
  186                                      struct witness_list *list);
  187 static void     witness_display(void(*)(const char *fmt, ...));
  188 static void     witness_list(struct thread *td);
  189 #endif
  190 
  191 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
  192 
  193 /*
  194  * If set to 0, witness is disabled.  If set to a non-zero value, witness
  195  * performs full lock order checking for all locks.  At runtime, this
  196  * value may be set to 0 to turn off witness.  witness is not allowed be
  197  * turned on once it is turned off, however.
  198  */
  199 static int witness_watch = 1;
  200 TUNABLE_INT("debug.witness.watch", &witness_watch);
  201 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
  202     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
  203 
  204 #ifdef KDB
  205 /*
  206  * When KDB is enabled and witness_kdb is set to 1, it will cause the system
  207  * to drop into kdebug() when:
  208  *      - a lock heirarchy violation occurs
  209  *      - locks are held when going to sleep.
  210  */
  211 #ifdef WITNESS_KDB
  212 int     witness_kdb = 1;
  213 #else
  214 int     witness_kdb = 0;
  215 #endif
  216 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
  217 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
  218 
  219 /*
  220  * When KDB is enabled and witness_trace is set to 1, it will cause the system
  221  * to print a stack trace:
  222  *      - a lock heirarchy violation occurs
  223  *      - locks are held when going to sleep.
  224  */
  225 int     witness_trace = 1;
  226 TUNABLE_INT("debug.witness.trace", &witness_trace);
  227 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
  228 #endif /* KDB */
  229 
  230 #ifdef WITNESS_SKIPSPIN
  231 int     witness_skipspin = 1;
  232 #else
  233 int     witness_skipspin = 0;
  234 #endif
  235 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
  236 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
  237     &witness_skipspin, 0, "");
  238 
  239 static struct mtx w_mtx;
  240 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
  241 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
  242 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
  243 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
  244 static struct witness_child_list_entry *w_child_free = NULL;
  245 static struct lock_list_entry *w_lock_list_free = NULL;
  246 
  247 static int w_free_cnt, w_spin_cnt, w_sleep_cnt, w_child_free_cnt, w_child_cnt;
  248 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
  249 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
  250 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
  251     "");
  252 SYSCTL_INT(_debug_witness, OID_AUTO, child_free_cnt, CTLFLAG_RD,
  253     &w_child_free_cnt, 0, "");
  254 SYSCTL_INT(_debug_witness, OID_AUTO, child_cnt, CTLFLAG_RD, &w_child_cnt, 0,
  255     "");
  256 
  257 static struct witness w_data[WITNESS_COUNT];
  258 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
  259 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
  260 
  261 static struct witness_order_list_entry order_lists[] = {
  262         /*
  263          * sx locks
  264          */
  265         { "proctree", &lock_class_sx },
  266         { "allproc", &lock_class_sx },
  267         { NULL, NULL },
  268         /*
  269          * Various mutexes
  270          */
  271         { "Giant", &lock_class_mtx_sleep },
  272         { "filedesc structure", &lock_class_mtx_sleep },
  273         { "pipe mutex", &lock_class_mtx_sleep },
  274         { "sigio lock", &lock_class_mtx_sleep },
  275         { "process group", &lock_class_mtx_sleep },
  276         { "process lock", &lock_class_mtx_sleep },
  277         { "session", &lock_class_mtx_sleep },
  278         { "uidinfo hash", &lock_class_mtx_sleep },
  279         { "uidinfo struct", &lock_class_mtx_sleep },
  280         { "allprison", &lock_class_mtx_sleep },
  281         { NULL, NULL },
  282         /*
  283          * Sockets
  284          */
  285         { "filedesc structure", &lock_class_mtx_sleep },
  286         { "accept", &lock_class_mtx_sleep },
  287         { "so_snd", &lock_class_mtx_sleep },
  288         { "so_rcv", &lock_class_mtx_sleep },
  289         { "sellck", &lock_class_mtx_sleep },
  290         { NULL, NULL },
  291         /*
  292          * Routing
  293          */
  294         { "so_rcv", &lock_class_mtx_sleep },
  295         { "radix node head", &lock_class_mtx_sleep },
  296         { "rtentry", &lock_class_mtx_sleep },
  297         { "ifaddr", &lock_class_mtx_sleep },
  298         { NULL, NULL },
  299         /*
  300          * Multicast - protocol locks before interface locks.
  301          */
  302         { "in_multi_mtx", &lock_class_mtx_sleep },
  303         { "igmp_mtx", &lock_class_mtx_sleep },
  304         { "if_addr_mtx", &lock_class_mtx_sleep },
  305         { NULL, NULL },
  306         /*
  307          * UNIX Domain Sockets
  308          */
  309         { "unp", &lock_class_mtx_sleep },
  310         { "so_snd", &lock_class_mtx_sleep },
  311         { NULL, NULL },
  312         /*
  313          * UDP/IP
  314          */
  315         { "udp", &lock_class_mtx_sleep },
  316         { "udpinp", &lock_class_mtx_sleep },
  317         { "so_snd", &lock_class_mtx_sleep },
  318         { NULL, NULL },
  319         /*
  320          * TCP/IP
  321          */
  322         { "tcp", &lock_class_mtx_sleep },
  323         { "tcpinp", &lock_class_mtx_sleep },
  324         { "so_snd", &lock_class_mtx_sleep },
  325         { NULL, NULL },
  326         /*
  327          * SLIP
  328          */
  329         { "slip_mtx", &lock_class_mtx_sleep },
  330         { "slip sc_mtx", &lock_class_mtx_sleep },
  331         { NULL, NULL },
  332         /*
  333          * netatalk
  334          */
  335         { "ddp_list_mtx", &lock_class_mtx_sleep },
  336         { "ddp_mtx", &lock_class_mtx_sleep },
  337         { NULL, NULL },
  338         /*
  339          * BPF
  340          */
  341         { "bpf global lock", &lock_class_mtx_sleep },
  342         { "bpf interface lock", &lock_class_mtx_sleep },
  343         { "bpf cdev lock", &lock_class_mtx_sleep },
  344         { NULL, NULL },
  345         /*
  346          * NFS server
  347          */
  348         { "nfsd_mtx", &lock_class_mtx_sleep },
  349         { "so_snd", &lock_class_mtx_sleep },
  350         { NULL, NULL },
  351         /*
  352          * CDEV
  353          */
  354         { "system map", &lock_class_mtx_sleep },
  355         { "vm page queue mutex", &lock_class_mtx_sleep },
  356         { "vnode interlock", &lock_class_mtx_sleep },
  357         { "cdev", &lock_class_mtx_sleep },
  358         { NULL, NULL },
  359         /*
  360          * spin locks
  361          */
  362 #ifdef SMP
  363         { "ap boot", &lock_class_mtx_spin },
  364 #endif
  365         { "rm.mutex_mtx", &lock_class_mtx_spin },
  366         { "sio", &lock_class_mtx_spin },
  367 #ifdef __i386__
  368         { "cy", &lock_class_mtx_spin },
  369 #endif
  370         { "uart_hwmtx", &lock_class_mtx_spin },
  371         { "sabtty", &lock_class_mtx_spin },
  372         { "zstty", &lock_class_mtx_spin },
  373         { "ng_node", &lock_class_mtx_spin },
  374         { "ng_worklist", &lock_class_mtx_spin },
  375         { "fast_taskqueue", &lock_class_mtx_spin },
  376         { "intr table", &lock_class_mtx_spin },
  377         { "sleepq chain", &lock_class_mtx_spin },
  378         { "sched lock", &lock_class_mtx_spin },
  379         { "turnstile chain", &lock_class_mtx_spin },
  380         { "td_contested", &lock_class_mtx_spin },
  381         { "callout", &lock_class_mtx_spin },
  382         { "entropy harvest mutex", &lock_class_mtx_spin },
  383         /*
  384          * leaf locks
  385          */
  386         { "allpmaps", &lock_class_mtx_spin },
  387         { "vm page queue free mutex", &lock_class_mtx_spin },
  388         { "icu", &lock_class_mtx_spin },
  389 #ifdef SMP
  390         { "smp rendezvous", &lock_class_mtx_spin },
  391 #if defined(__i386__) || defined(__amd64__)
  392         { "tlb", &lock_class_mtx_spin },
  393 #endif
  394 #ifdef __sparc64__
  395         { "ipi", &lock_class_mtx_spin },
  396         { "rtc_mtx", &lock_class_mtx_spin },
  397 #endif
  398 #endif
  399         { "clk", &lock_class_mtx_spin },
  400         { "mutex profiling lock", &lock_class_mtx_spin },
  401         { "kse zombie lock", &lock_class_mtx_spin },
  402         { "ALD Queue", &lock_class_mtx_spin },
  403 #ifdef __ia64__
  404         { "MCA spin lock", &lock_class_mtx_spin },
  405 #endif
  406 #if defined(__i386__) || defined(__amd64__)
  407         { "pcicfg", &lock_class_mtx_spin },
  408         { "NDIS thread lock", &lock_class_mtx_spin },
  409 #endif
  410         { "tw_osl_io_lock", &lock_class_mtx_spin },
  411         { "tw_osl_q_lock", &lock_class_mtx_spin },
  412         { "tw_cl_io_lock", &lock_class_mtx_spin },
  413         { "tw_cl_intr_lock", &lock_class_mtx_spin },
  414         { "tw_cl_gen_lock", &lock_class_mtx_spin },
  415         { NULL, NULL },
  416         { NULL, NULL }
  417 };
  418 
  419 #ifdef BLESSING
  420 /*
  421  * Pairs of locks which have been blessed
  422  * Don't complain about order problems with blessed locks
  423  */
  424 static struct witness_blessed blessed_list[] = {
  425 };
  426 static int blessed_count =
  427         sizeof(blessed_list) / sizeof(struct witness_blessed);
  428 #endif
  429 
  430 /*
  431  * List of all locks in the system.
  432  */
  433 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
  434 
  435 static struct mtx all_mtx = {
  436         { &lock_class_mtx_sleep,        /* mtx_object.lo_class */
  437           "All locks list",             /* mtx_object.lo_name */
  438           "All locks list",             /* mtx_object.lo_type */
  439           LO_INITIALIZED,               /* mtx_object.lo_flags */
  440           { NULL, NULL },               /* mtx_object.lo_list */
  441           NULL },                       /* mtx_object.lo_witness */
  442         MTX_UNOWNED, 0                  /* mtx_lock, mtx_recurse */
  443 };
  444 
  445 /*
  446  * This global is set to 0 once it becomes safe to use the witness code.
  447  */
  448 static int witness_cold = 1;
  449 
  450 /*
  451  * This global is set to 1 once the static lock orders have been enrolled
  452  * so that a warning can be issued for any spin locks enrolled later.
  453  */
  454 static int witness_spin_warn = 0;
  455 
  456 /*
  457  * Global variables for book keeping.
  458  */
  459 static int lock_cur_cnt;
  460 static int lock_max_cnt;
  461 
  462 /*
  463  * The WITNESS-enabled diagnostic code.
  464  */
  465 static void
  466 witness_initialize(void *dummy __unused)
  467 {
  468         struct lock_object *lock;
  469         struct witness_order_list_entry *order;
  470         struct witness *w, *w1;
  471         int i;
  472 
  473         /*
  474          * We have to release Giant before initializing its witness
  475          * structure so that WITNESS doesn't get confused.
  476          */
  477         mtx_unlock(&Giant);
  478         mtx_assert(&Giant, MA_NOTOWNED);
  479 
  480         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
  481         TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
  482         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
  483             MTX_NOWITNESS);
  484         for (i = 0; i < WITNESS_COUNT; i++)
  485                 witness_free(&w_data[i]);
  486         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
  487                 witness_child_free(&w_childdata[i]);
  488         for (i = 0; i < LOCK_CHILDCOUNT; i++)
  489                 witness_lock_list_free(&w_locklistdata[i]);
  490 
  491         /* First add in all the specified order lists. */
  492         for (order = order_lists; order->w_name != NULL; order++) {
  493                 w = enroll(order->w_name, order->w_class);
  494                 if (w == NULL)
  495                         continue;
  496                 w->w_file = "order list";
  497                 for (order++; order->w_name != NULL; order++) {
  498                         w1 = enroll(order->w_name, order->w_class);
  499                         if (w1 == NULL)
  500                                 continue;
  501                         w1->w_file = "order list";
  502                         if (!itismychild(w, w1))
  503                                 panic("Not enough memory for static orders!");
  504                         w = w1;
  505                 }
  506         }
  507         witness_spin_warn = 1;
  508 
  509         /* Iterate through all locks and add them to witness. */
  510         mtx_lock(&all_mtx);
  511         TAILQ_FOREACH(lock, &all_locks, lo_list) {
  512                 if (lock->lo_flags & LO_WITNESS)
  513                         lock->lo_witness = enroll(lock->lo_type,
  514                             LOCK_CLASS(lock));
  515                 else
  516                         lock->lo_witness = NULL;
  517         }
  518         mtx_unlock(&all_mtx);
  519 
  520         /* Mark the witness code as being ready for use. */
  521         witness_cold = 0;
  522 
  523         mtx_lock(&Giant);
  524 }
  525 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
  526 
  527 static int
  528 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
  529 {
  530         int error, value;
  531 
  532         value = witness_watch;
  533         error = sysctl_handle_int(oidp, &value, 0, req);
  534         if (error != 0 || req->newptr == NULL)
  535                 return (error);
  536         error = suser(req->td);
  537         if (error != 0)
  538                 return (error);
  539         if (value == witness_watch)
  540                 return (0);
  541         if (value != 0)
  542                 return (EINVAL);
  543         witness_watch = 0;
  544         return (0);
  545 }
  546 
  547 void
  548 witness_init(struct lock_object *lock)
  549 {
  550         struct lock_class *class;
  551 
  552         class = LOCK_CLASS(lock);
  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_cur_cnt++;
  569         if (lock_cur_cnt > lock_max_cnt)
  570                 lock_max_cnt = lock_cur_cnt;
  571         mtx_unlock(&all_mtx);
  572         if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
  573             (lock->lo_flags & LO_WITNESS) != 0)
  574                 lock->lo_witness = enroll(lock->lo_type, class);
  575         else
  576                 lock->lo_witness = NULL;
  577 }
  578 
  579 void
  580 witness_destroy(struct lock_object *lock)
  581 {
  582         struct lock_class *class;
  583         struct witness *w;
  584 
  585         class = LOCK_CLASS(lock);
  586         if (witness_cold)
  587                 panic("lock (%s) %s destroyed while witness_cold",
  588                     class->lc_name, lock->lo_name);
  589 
  590         /* XXX: need to verify that no one holds the lock */
  591         w = lock->lo_witness;
  592         if (w != NULL) {
  593                 mtx_lock_spin(&w_mtx);
  594                 MPASS(w->w_refcount > 0);
  595                 w->w_refcount--;
  596 
  597                 /*
  598                  * Lock is already released if we have an allocation failure
  599                  * and depart() fails.
  600                  */
  601                 if (w->w_refcount != 0 || depart(w))
  602                         mtx_unlock_spin(&w_mtx);
  603         }
  604 
  605         mtx_lock(&all_mtx);
  606         lock_cur_cnt--;
  607         TAILQ_REMOVE(&all_locks, lock, lo_list);
  608         mtx_unlock(&all_mtx);
  609 }
  610 
  611 #ifdef DDB
  612 static void
  613 witness_levelall (void)
  614 {
  615         struct witness_list *list;
  616         struct witness *w, *w1;
  617 
  618         /*
  619          * First clear all levels.
  620          */
  621         STAILQ_FOREACH(w, &w_all, w_list) {
  622                 w->w_level = 0;
  623         }
  624 
  625         /*
  626          * Look for locks with no parent and level all their descendants.
  627          */
  628         STAILQ_FOREACH(w, &w_all, w_list) {
  629                 /*
  630                  * This is just an optimization, technically we could get
  631                  * away just walking the all list each time.
  632                  */
  633                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
  634                         list = &w_sleep;
  635                 else
  636                         list = &w_spin;
  637                 STAILQ_FOREACH(w1, list, w_typelist) {
  638                         if (isitmychild(w1, w))
  639                                 goto skip;
  640                 }
  641                 witness_leveldescendents(w, 0);
  642         skip:
  643                 ;       /* silence GCC 3.x */
  644         }
  645 }
  646 
  647 static void
  648 witness_leveldescendents(struct witness *parent, int level)
  649 {
  650         struct witness_child_list_entry *wcl;
  651         int i;
  652 
  653         if (parent->w_level < level)
  654                 parent->w_level = level;
  655         level++;
  656         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
  657                 for (i = 0; i < wcl->wcl_count; i++)
  658                         witness_leveldescendents(wcl->wcl_children[i], level);
  659 }
  660 
  661 static void
  662 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
  663                            struct witness *parent, int indent)
  664 {
  665         struct witness_child_list_entry *wcl;
  666         int i, level;
  667 
  668         level = parent->w_level;
  669         prnt("%-2d", level);
  670         for (i = 0; i < indent; i++)
  671                 prnt(" ");
  672         if (parent->w_refcount > 0)
  673                 prnt("%s", parent->w_name);
  674         else
  675                 prnt("(dead)");
  676         if (parent->w_displayed) {
  677                 prnt(" -- (already displayed)\n");
  678                 return;
  679         }
  680         parent->w_displayed = 1;
  681         if (parent->w_refcount > 0) {
  682                 if (parent->w_file != NULL)
  683                         prnt(" -- last acquired @ %s:%d", parent->w_file,
  684                             parent->w_line);
  685         }
  686         prnt("\n");
  687         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
  688                 for (i = 0; i < wcl->wcl_count; i++)
  689                             witness_displaydescendants(prnt,
  690                                 wcl->wcl_children[i], indent + 1);
  691 }
  692 
  693 static void
  694 witness_display_list(void(*prnt)(const char *fmt, ...),
  695                      struct witness_list *list)
  696 {
  697         struct witness *w;
  698 
  699         STAILQ_FOREACH(w, list, w_typelist) {
  700                 if (w->w_file == NULL || w->w_level > 0)
  701                         continue;
  702                 /*
  703                  * This lock has no anscestors, display its descendants. 
  704                  */
  705                 witness_displaydescendants(prnt, w, 0);
  706         }
  707 }
  708         
  709 static void
  710 witness_display(void(*prnt)(const char *fmt, ...))
  711 {
  712         struct witness *w;
  713 
  714         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
  715         witness_levelall();
  716 
  717         /* Clear all the displayed flags. */
  718         STAILQ_FOREACH(w, &w_all, w_list) {
  719                 w->w_displayed = 0;
  720         }
  721 
  722         /*
  723          * First, handle sleep locks which have been acquired at least
  724          * once.
  725          */
  726         prnt("Sleep locks:\n");
  727         witness_display_list(prnt, &w_sleep);
  728         
  729         /*
  730          * Now do spin locks which have been acquired at least once.
  731          */
  732         prnt("\nSpin locks:\n");
  733         witness_display_list(prnt, &w_spin);
  734         
  735         /*
  736          * Finally, any locks which have not been acquired yet.
  737          */
  738         prnt("\nLocks which were never acquired:\n");
  739         STAILQ_FOREACH(w, &w_all, w_list) {
  740                 if (w->w_file != NULL || w->w_refcount == 0)
  741                         continue;
  742                 prnt("%s\n", w->w_name);
  743         }
  744 }
  745 #endif /* DDB */
  746 
  747 /* Trim useless garbage from filenames. */
  748 static const char *
  749 fixup_filename(const char *file)
  750 {
  751 
  752         if (file == NULL)
  753                 return (NULL);
  754         while (strncmp(file, "../", 3) == 0)
  755                 file += 3;
  756         return (file);
  757 }
  758 
  759 int
  760 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
  761 {
  762 
  763         if (witness_watch == 0 || panicstr != NULL)
  764                 return (0);
  765 
  766         /* Require locks that witness knows about. */
  767         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
  768             lock2->lo_witness == NULL)
  769                 return (EINVAL);
  770 
  771         MPASS(!mtx_owned(&w_mtx));
  772         mtx_lock_spin(&w_mtx);
  773 
  774         /*
  775          * If we already have either an explicit or implied lock order that
  776          * is the other way around, then return an error.
  777          */
  778         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
  779                 mtx_unlock_spin(&w_mtx);
  780                 return (EDOOFUS);
  781         }
  782         
  783         /* Try to add the new order. */
  784         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
  785             lock2->lo_type, lock1->lo_type);
  786         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
  787                 return (ENOMEM);
  788         mtx_unlock_spin(&w_mtx);
  789         return (0);
  790 }
  791 
  792 void
  793 witness_checkorder(struct lock_object *lock, int flags, const char *file,
  794     int line)
  795 {
  796         struct lock_list_entry **lock_list, *lle;
  797         struct lock_instance *lock1, *lock2;
  798         struct lock_class *class;
  799         struct witness *w, *w1;
  800         struct thread *td;
  801         int i, j;
  802 
  803         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
  804             panicstr != NULL)
  805                 return;
  806 
  807         /*
  808          * Try locks do not block if they fail to acquire the lock, thus
  809          * there is no danger of deadlocks or of switching while holding a
  810          * spin lock if we acquire a lock via a try operation.  This
  811          * function shouldn't even be called for try locks, so panic if
  812          * that happens.
  813          */
  814         if (flags & LOP_TRYLOCK)
  815                 panic("%s should not be called for try lock operations",
  816                     __func__);
  817 
  818         w = lock->lo_witness;
  819         class = LOCK_CLASS(lock);
  820         td = curthread;
  821         file = fixup_filename(file);
  822 
  823         if (class->lc_flags & LC_SLEEPLOCK) {
  824                 /*
  825                  * Since spin locks include a critical section, this check
  826                  * implicitly enforces a lock order of all sleep locks before
  827                  * all spin locks.
  828                  */
  829                 if (td->td_critnest != 0 && !kdb_active)
  830                         panic("blockable sleep lock (%s) %s @ %s:%d",
  831                             class->lc_name, lock->lo_name, file, line);
  832 
  833                 /*
  834                  * If this is the first lock acquired then just return as
  835                  * no order checking is needed.
  836                  */
  837                 if (td->td_sleeplocks == NULL)
  838                         return;
  839                 lock_list = &td->td_sleeplocks;
  840         } else {
  841                 /*
  842                  * If this is the first lock, just return as no order
  843                  * checking is needed.  We check this in both if clauses
  844                  * here as unifying the check would require us to use a
  845                  * critical section to ensure we don't migrate while doing
  846                  * the check.  Note that if this is not the first lock, we
  847                  * are already in a critical section and are safe for the
  848                  * rest of the check.
  849                  */
  850                 if (PCPU_GET(spinlocks) == NULL)
  851                         return;
  852                 lock_list = PCPU_PTR(spinlocks);
  853         }
  854 
  855         /*
  856          * Check to see if we are recursing on a lock we already own.  If
  857          * so, make sure that we don't mismatch exclusive and shared lock
  858          * acquires.
  859          */
  860         lock1 = find_instance(*lock_list, lock);
  861         if (lock1 != NULL) {
  862                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
  863                     (flags & LOP_EXCLUSIVE) == 0) {
  864                         printf("shared lock of (%s) %s @ %s:%d\n",
  865                             class->lc_name, lock->lo_name, file, line);
  866                         printf("while exclusively locked from %s:%d\n",
  867                             lock1->li_file, lock1->li_line);
  868                         panic("share->excl");
  869                 }
  870                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
  871                     (flags & LOP_EXCLUSIVE) != 0) {
  872                         printf("exclusive lock of (%s) %s @ %s:%d\n",
  873                             class->lc_name, lock->lo_name, file, line);
  874                         printf("while share locked from %s:%d\n",
  875                             lock1->li_file, lock1->li_line);
  876                         panic("excl->share");
  877                 }
  878                 return;
  879         }
  880 
  881         /*
  882          * Try locks do not block if they fail to acquire the lock, thus
  883          * there is no danger of deadlocks or of switching while holding a
  884          * spin lock if we acquire a lock via a try operation.
  885          */
  886         if (flags & LOP_TRYLOCK)
  887                 return;
  888 
  889         /*
  890          * Check for duplicate locks of the same type.  Note that we only
  891          * have to check for this on the last lock we just acquired.  Any
  892          * other cases will be caught as lock order violations.
  893          */
  894         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
  895         w1 = lock1->li_lock->lo_witness;
  896         if (w1 == w) {
  897                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
  898                     (flags & LOP_DUPOK))
  899                         return;
  900                 w->w_same_squawked = 1;
  901                 printf("acquiring duplicate lock of same type: \"%s\"\n", 
  902                         lock->lo_type);
  903                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
  904                     lock1->li_file, lock1->li_line);
  905                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
  906 #ifdef KDB
  907                 goto debugger;
  908 #else
  909                 return;
  910 #endif
  911         }
  912         MPASS(!mtx_owned(&w_mtx));
  913         mtx_lock_spin(&w_mtx);
  914         /*
  915          * If we know that the the lock we are acquiring comes after
  916          * the lock we most recently acquired in the lock order tree,
  917          * then there is no need for any further checks.
  918          */
  919         if (isitmychild(w1, w)) {
  920                 mtx_unlock_spin(&w_mtx);
  921                 return;
  922         }
  923         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
  924                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
  925 
  926                         MPASS(j < WITNESS_COUNT);
  927                         lock1 = &lle->ll_children[i];
  928                         w1 = lock1->li_lock->lo_witness;
  929 
  930                         /*
  931                          * If this lock doesn't undergo witness checking,
  932                          * then skip it.
  933                          */
  934                         if (w1 == NULL) {
  935                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
  936                                     ("lock missing witness structure"));
  937                                 continue;
  938                         }
  939                         /*
  940                          * If we are locking Giant and this is a sleepable
  941                          * lock, then skip it.
  942                          */
  943                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
  944                             lock == &Giant.mtx_object)
  945                                 continue;
  946                         /*
  947                          * If we are locking a sleepable lock and this lock
  948                          * is Giant, then skip it.
  949                          */
  950                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  951                             lock1->li_lock == &Giant.mtx_object)
  952                                 continue;
  953                         /*
  954                          * If we are locking a sleepable lock and this lock
  955                          * isn't sleepable, we want to treat it as a lock
  956                          * order violation to enfore a general lock order of
  957                          * sleepable locks before non-sleepable locks.
  958                          */
  959                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  960                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
  961                                 goto reversal;
  962                         /*
  963                          * If we are locking Giant and this is a non-sleepable
  964                          * lock, then treat it as a reversal.
  965                          */
  966                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
  967                             lock == &Giant.mtx_object)
  968                                 goto reversal;
  969                         /*
  970                          * Check the lock order hierarchy for a reveresal.
  971                          */
  972                         if (!isitmydescendant(w, w1))
  973                                 continue;
  974                 reversal:
  975                         /*
  976                          * We have a lock order violation, check to see if it
  977                          * is allowed or has already been yelled about.
  978                          */
  979                         mtx_unlock_spin(&w_mtx);
  980 #ifdef BLESSING
  981                         /*
  982                          * If the lock order is blessed, just bail.  We don't
  983                          * look for other lock order violations though, which
  984                          * may be a bug.
  985                          */
  986                         if (blessed(w, w1))
  987                                 return;
  988 #endif
  989                         if (lock1->li_lock == &Giant.mtx_object) {
  990                                 if (w1->w_Giant_squawked)
  991                                         return;
  992                                 else
  993                                         w1->w_Giant_squawked = 1;
  994                         } else {
  995                                 if (w1->w_other_squawked)
  996                                         return;
  997                                 else
  998                                         w1->w_other_squawked = 1;
  999                         }
 1000                         /*
 1001                          * Ok, yell about it.
 1002                          */
 1003                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1004                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
 1005                                 printf(
 1006                 "lock order reversal: (sleepable after non-sleepable)\n");
 1007                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
 1008                             && lock == &Giant.mtx_object)
 1009                                 printf(
 1010                 "lock order reversal: (Giant after non-sleepable)\n");
 1011                         else
 1012                                 printf("lock order reversal:\n");
 1013                         /*
 1014                          * Try to locate an earlier lock with
 1015                          * witness w in our list.
 1016                          */
 1017                         do {
 1018                                 lock2 = &lle->ll_children[i];
 1019                                 MPASS(lock2->li_lock != NULL);
 1020                                 if (lock2->li_lock->lo_witness == w)
 1021                                         break;
 1022                                 if (i == 0 && lle->ll_next != NULL) {
 1023                                         lle = lle->ll_next;
 1024                                         i = lle->ll_count - 1;
 1025                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
 1026                                 } else
 1027                                         i--;
 1028                         } while (i >= 0);
 1029                         if (i < 0) {
 1030                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 1031                                     lock1->li_lock, lock1->li_lock->lo_name,
 1032                                     lock1->li_lock->lo_type, lock1->li_file,
 1033                                     lock1->li_line);
 1034                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
 1035                                     lock->lo_name, lock->lo_type, file, line);
 1036                         } else {
 1037                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 1038                                     lock2->li_lock, lock2->li_lock->lo_name,
 1039                                     lock2->li_lock->lo_type, lock2->li_file,
 1040                                     lock2->li_line);
 1041                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
 1042                                     lock1->li_lock, lock1->li_lock->lo_name,
 1043                                     lock1->li_lock->lo_type, lock1->li_file,
 1044                                     lock1->li_line);
 1045                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
 1046                                     lock->lo_name, lock->lo_type, file, line);
 1047                         }
 1048 #ifdef KDB
 1049                         goto debugger;
 1050 #else
 1051                         return;
 1052 #endif
 1053                 }
 1054         }
 1055         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 1056         /*
 1057          * If requested, build a new lock order.  However, don't build a new
 1058          * relationship between a sleepable lock and Giant if it is in the
 1059          * wrong direction.  The correct lock order is that sleepable locks
 1060          * always come before Giant.
 1061          */
 1062         if (flags & LOP_NEWORDER &&
 1063             !(lock1->li_lock == &Giant.mtx_object &&
 1064             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
 1065                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 1066                     lock->lo_type, lock1->li_lock->lo_type);
 1067                 if (!itismychild(lock1->li_lock->lo_witness, w))
 1068                         /* Witness is dead. */
 1069                         return;
 1070         } 
 1071         mtx_unlock_spin(&w_mtx);
 1072         return;
 1073 
 1074 #ifdef KDB
 1075 debugger:
 1076         if (witness_trace)
 1077                 kdb_backtrace();
 1078         if (witness_kdb)
 1079                 kdb_enter(__func__);
 1080 #endif
 1081 }
 1082 
 1083 void
 1084 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
 1085 {
 1086         struct lock_list_entry **lock_list, *lle;
 1087         struct lock_instance *instance;
 1088         struct witness *w;
 1089         struct thread *td;
 1090 
 1091         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 1092             panicstr != NULL)
 1093                 return;
 1094         w = lock->lo_witness;
 1095         td = curthread;
 1096         file = fixup_filename(file);
 1097 
 1098         /* Determine lock list for this lock. */
 1099         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
 1100                 lock_list = &td->td_sleeplocks;
 1101         else
 1102                 lock_list = PCPU_PTR(spinlocks);
 1103 
 1104         /* Check to see if we are recursing on a lock we already own. */
 1105         instance = find_instance(*lock_list, lock);
 1106         if (instance != NULL) {
 1107                 instance->li_flags++;
 1108                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
 1109                     td->td_proc->p_pid, lock->lo_name,
 1110                     instance->li_flags & LI_RECURSEMASK);
 1111                 instance->li_file = file;
 1112                 instance->li_line = line;
 1113                 return;
 1114         }
 1115 
 1116         /* Update per-witness last file and line acquire. */
 1117         w->w_file = file;
 1118         w->w_line = line;
 1119 
 1120         /* Find the next open lock instance in the list and fill it. */
 1121         lle = *lock_list;
 1122         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
 1123                 lle = witness_lock_list_get();
 1124                 if (lle == NULL)
 1125                         return;
 1126                 lle->ll_next = *lock_list;
 1127                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
 1128                     td->td_proc->p_pid, lle);
 1129                 *lock_list = lle;
 1130         }
 1131         instance = &lle->ll_children[lle->ll_count++];
 1132         instance->li_lock = lock;
 1133         instance->li_line = line;
 1134         instance->li_file = file;
 1135         if ((flags & LOP_EXCLUSIVE) != 0)
 1136                 instance->li_flags = LI_EXCLUSIVE;
 1137         else
 1138                 instance->li_flags = 0;
 1139         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
 1140             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
 1141 }
 1142 
 1143 void
 1144 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
 1145 {
 1146         struct lock_instance *instance;
 1147         struct lock_class *class;
 1148 
 1149         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1150         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1151                 return;
 1152         class = LOCK_CLASS(lock);
 1153         file = fixup_filename(file);
 1154         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1155                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
 1156                     class->lc_name, lock->lo_name, file, line);
 1157         if ((flags & LOP_TRYLOCK) == 0)
 1158                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
 1159                     lock->lo_name, file, line);
 1160         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
 1161                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
 1162                     class->lc_name, lock->lo_name, file, line);
 1163         instance = find_instance(curthread->td_sleeplocks, lock);
 1164         if (instance == NULL)
 1165                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
 1166                     class->lc_name, lock->lo_name, file, line);
 1167         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
 1168                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
 1169                     class->lc_name, lock->lo_name, file, line);
 1170         if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1171                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1172                     class->lc_name, lock->lo_name,
 1173                     instance->li_flags & LI_RECURSEMASK, file, line);
 1174         instance->li_flags |= LI_EXCLUSIVE;
 1175 }
 1176 
 1177 void
 1178 witness_downgrade(struct lock_object *lock, int flags, const char *file,
 1179     int line)
 1180 {
 1181         struct lock_instance *instance;
 1182         struct lock_class *class;
 1183 
 1184         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1185         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1186                 return;
 1187         class = LOCK_CLASS(lock);
 1188         file = fixup_filename(file);
 1189         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1190                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
 1191                     class->lc_name, lock->lo_name, file, line);
 1192         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
 1193                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
 1194                     class->lc_name, lock->lo_name, file, line);
 1195         instance = find_instance(curthread->td_sleeplocks, lock);
 1196         if (instance == NULL)
 1197                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
 1198                     class->lc_name, lock->lo_name, file, line);
 1199         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
 1200                 panic("downgrade of shared lock (%s) %s @ %s:%d",
 1201                     class->lc_name, lock->lo_name, file, line);
 1202         if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1203                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1204                     class->lc_name, lock->lo_name,
 1205                     instance->li_flags & LI_RECURSEMASK, file, line);
 1206         instance->li_flags &= ~LI_EXCLUSIVE;
 1207 }
 1208 
 1209 void
 1210 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
 1211 {
 1212         struct lock_list_entry **lock_list, *lle;
 1213         struct lock_instance *instance;
 1214         struct lock_class *class;
 1215         struct thread *td;
 1216         register_t s;
 1217         int i, j;
 1218 
 1219         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 1220             panicstr != NULL)
 1221                 return;
 1222         td = curthread;
 1223         class = LOCK_CLASS(lock);
 1224         file = fixup_filename(file);
 1225 
 1226         /* Find lock instance associated with this lock. */
 1227         if (class->lc_flags & LC_SLEEPLOCK)
 1228                 lock_list = &td->td_sleeplocks;
 1229         else
 1230                 lock_list = PCPU_PTR(spinlocks);
 1231         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
 1232                 for (i = 0; i < (*lock_list)->ll_count; i++) {
 1233                         instance = &(*lock_list)->ll_children[i];
 1234                         if (instance->li_lock == lock)
 1235                                 goto found;
 1236                 }
 1237         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
 1238             file, line);
 1239 found:
 1240 
 1241         /* First, check for shared/exclusive mismatches. */
 1242         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
 1243             (flags & LOP_EXCLUSIVE) == 0) {
 1244                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1245                     lock->lo_name, file, line);
 1246                 printf("while exclusively locked from %s:%d\n",
 1247                     instance->li_file, instance->li_line);
 1248                 panic("excl->ushare");
 1249         }
 1250         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
 1251             (flags & LOP_EXCLUSIVE) != 0) {
 1252                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1253                     lock->lo_name, file, line);
 1254                 printf("while share locked from %s:%d\n", instance->li_file,
 1255                     instance->li_line);
 1256                 panic("share->uexcl");
 1257         }
 1258 
 1259         /* If we are recursed, unrecurse. */
 1260         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
 1261                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
 1262                     td->td_proc->p_pid, instance->li_lock->lo_name,
 1263                     instance->li_flags);
 1264                 instance->li_flags--;
 1265                 return;
 1266         }
 1267 
 1268         /* Otherwise, remove this item from the list. */
 1269         s = intr_disable();
 1270         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
 1271             td->td_proc->p_pid, instance->li_lock->lo_name,
 1272             (*lock_list)->ll_count - 1);
 1273         for (j = i; j < (*lock_list)->ll_count - 1; j++)
 1274                 (*lock_list)->ll_children[j] =
 1275                     (*lock_list)->ll_children[j + 1];
 1276         (*lock_list)->ll_count--;
 1277         intr_restore(s);
 1278 
 1279         /* If this lock list entry is now empty, free it. */
 1280         if ((*lock_list)->ll_count == 0) {
 1281                 lle = *lock_list;
 1282                 *lock_list = lle->ll_next;
 1283                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
 1284                     td->td_proc->p_pid, lle);
 1285                 witness_lock_list_free(lle);
 1286         }
 1287 }
 1288 
 1289 /*
 1290  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
 1291  * exempt Giant and sleepable locks from the checks as well.  If any
 1292  * non-exempt locks are held, then a supplied message is printed to the
 1293  * console along with a list of the offending locks.  If indicated in the
 1294  * flags then a failure results in a panic as well.
 1295  */
 1296 int
 1297 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
 1298 {
 1299         struct lock_list_entry *lle;
 1300         struct lock_instance *lock1;
 1301         struct thread *td;
 1302         va_list ap;
 1303         int i, n;
 1304 
 1305         if (witness_cold || witness_watch == 0 || panicstr != NULL)
 1306                 return (0);
 1307         n = 0;
 1308         td = curthread;
 1309         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
 1310                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1311                         lock1 = &lle->ll_children[i];
 1312                         if (lock1->li_lock == lock)
 1313                                 continue;
 1314                         if (flags & WARN_GIANTOK &&
 1315                             lock1->li_lock == &Giant.mtx_object)
 1316                                 continue;
 1317                         if (flags & WARN_SLEEPOK &&
 1318                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
 1319                                 continue;
 1320                         if (n == 0) {
 1321                                 va_start(ap, fmt);
 1322                                 vprintf(fmt, ap);
 1323                                 va_end(ap);
 1324                                 printf(" with the following");
 1325                                 if (flags & WARN_SLEEPOK)
 1326                                         printf(" non-sleepable");
 1327                                 printf(" locks held:\n");
 1328                         }
 1329                         n++;
 1330                         witness_list_lock(lock1);
 1331                 }
 1332         if (PCPU_GET(spinlocks) != NULL) {
 1333                 /*
 1334                  * Since we already hold a spinlock preemption is
 1335                  * already blocked.
 1336                  */
 1337                 if (n == 0) {
 1338                         va_start(ap, fmt);
 1339                         vprintf(fmt, ap);
 1340                         va_end(ap);
 1341                         printf(" with the following");
 1342                         if (flags & WARN_SLEEPOK)
 1343                                 printf(" non-sleepable");
 1344                         printf(" locks held:\n");
 1345                 }
 1346                 n += witness_list_locks(PCPU_PTR(spinlocks));
 1347         }
 1348         if (flags & WARN_PANIC && n)
 1349                 panic("witness_warn");
 1350 #ifdef KDB
 1351         else if (witness_kdb && n)
 1352                 kdb_enter(__func__);
 1353         else if (witness_trace && n)
 1354                 kdb_backtrace();
 1355 #endif
 1356         return (n);
 1357 }
 1358 
 1359 const char *
 1360 witness_file(struct lock_object *lock)
 1361 {
 1362         struct witness *w;
 1363 
 1364         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
 1365                 return ("?");
 1366         w = lock->lo_witness;
 1367         return (w->w_file);
 1368 }
 1369 
 1370 int
 1371 witness_line(struct lock_object *lock)
 1372 {
 1373         struct witness *w;
 1374 
 1375         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
 1376                 return (0);
 1377         w = lock->lo_witness;
 1378         return (w->w_line);
 1379 }
 1380 
 1381 static struct witness *
 1382 enroll(const char *description, struct lock_class *lock_class)
 1383 {
 1384         struct witness *w;
 1385 
 1386         if (witness_watch == 0 || panicstr != NULL)
 1387                 return (NULL);
 1388         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
 1389                 return (NULL);
 1390         mtx_lock_spin(&w_mtx);
 1391         STAILQ_FOREACH(w, &w_all, w_list) {
 1392                 if (w->w_name == description || (w->w_refcount > 0 &&
 1393                     strcmp(description, w->w_name) == 0)) {
 1394                         w->w_refcount++;
 1395                         mtx_unlock_spin(&w_mtx);
 1396                         if (lock_class != w->w_class)
 1397                                 panic(
 1398                                 "lock (%s) %s does not match earlier (%s) lock",
 1399                                     description, lock_class->lc_name,
 1400                                     w->w_class->lc_name);
 1401                         return (w);
 1402                 }
 1403         }
 1404         if ((w = witness_get()) == NULL)
 1405                 goto out;
 1406         w->w_name = description;
 1407         w->w_class = lock_class;
 1408         w->w_refcount = 1;
 1409         STAILQ_INSERT_HEAD(&w_all, w, w_list);
 1410         if (lock_class->lc_flags & LC_SPINLOCK) {
 1411                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
 1412                 w_spin_cnt++;
 1413         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
 1414                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
 1415                 w_sleep_cnt++;
 1416         } else {
 1417                 mtx_unlock_spin(&w_mtx);
 1418                 panic("lock class %s is not sleep or spin",
 1419                     lock_class->lc_name);
 1420         }
 1421         mtx_unlock_spin(&w_mtx);
 1422 out:
 1423         /*
 1424          * We issue a warning for any spin locks not defined in the static
 1425          * order list as a way to discourage their use (folks should really
 1426          * be using non-spin mutexes most of the time).  However, several
 1427          * 3rd part device drivers use spin locks because that is all they
 1428          * have available on Windows and Linux and they think that normal
 1429          * mutexes are insufficient.
 1430          */
 1431         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
 1432                 printf("WITNESS: spin lock %s not in order list\n",
 1433                     description);
 1434         return (w);
 1435 }
 1436 
 1437 /* Don't let the door bang you on the way out... */
 1438 static int
 1439 depart(struct witness *w)
 1440 {
 1441         struct witness_child_list_entry *wcl, *nwcl;
 1442         struct witness_list *list;
 1443         struct witness *parent;
 1444 
 1445         MPASS(w->w_refcount == 0);
 1446         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
 1447                 list = &w_sleep;
 1448                 w_sleep_cnt--;
 1449         } else {
 1450                 list = &w_spin;
 1451                 w_spin_cnt--;
 1452         }
 1453         /*
 1454          * First, we run through the entire tree looking for any
 1455          * witnesses that the outgoing witness is a child of.  For
 1456          * each parent that we find, we reparent all the direct
 1457          * children of the outgoing witness to its parent.
 1458          */
 1459         STAILQ_FOREACH(parent, list, w_typelist) {
 1460                 if (!isitmychild(parent, w))
 1461                         continue;
 1462                 removechild(parent, w);
 1463         }
 1464 
 1465         /*
 1466          * Now we go through and free up the child list of the
 1467          * outgoing witness.
 1468          */
 1469         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
 1470                 nwcl = wcl->wcl_next;
 1471                 w_child_cnt--;
 1472                 witness_child_free(wcl);
 1473         }
 1474 
 1475         /*
 1476          * Detach from various lists and free.
 1477          */
 1478         STAILQ_REMOVE(list, w, witness, w_typelist);
 1479         STAILQ_REMOVE(&w_all, w, witness, w_list);
 1480         witness_free(w);
 1481 
 1482         return (1);
 1483 }
 1484 
 1485 /*
 1486  * Add "child" as a direct child of "parent".  Returns false if
 1487  * we fail due to out of memory.
 1488  */
 1489 static int
 1490 insertchild(struct witness *parent, struct witness *child)
 1491 {
 1492         struct witness_child_list_entry **wcl;
 1493 
 1494         MPASS(child != NULL && parent != NULL);
 1495 
 1496         /*
 1497          * Insert "child" after "parent"
 1498          */
 1499         wcl = &parent->w_children;
 1500         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
 1501                 wcl = &(*wcl)->wcl_next;
 1502         if (*wcl == NULL) {
 1503                 *wcl = witness_child_get();
 1504                 if (*wcl == NULL)
 1505                         return (0);
 1506                 w_child_cnt++;
 1507         }
 1508         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
 1509 
 1510         return (1);
 1511 }
 1512 
 1513 
 1514 static int
 1515 itismychild(struct witness *parent, struct witness *child)
 1516 {
 1517         struct witness_list *list;
 1518 
 1519         MPASS(child != NULL && parent != NULL);
 1520         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
 1521             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
 1522                 panic(
 1523                 "%s: parent (%s) and child (%s) are not the same lock type",
 1524                     __func__, parent->w_class->lc_name,
 1525                     child->w_class->lc_name);
 1526 
 1527         if (!insertchild(parent, child))
 1528                 return (0);
 1529 
 1530         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
 1531                 list = &w_sleep;
 1532         else
 1533                 list = &w_spin;
 1534         return (1);
 1535 }
 1536 
 1537 static void
 1538 removechild(struct witness *parent, struct witness *child)
 1539 {
 1540         struct witness_child_list_entry **wcl, *wcl1;
 1541         int i;
 1542 
 1543         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
 1544                 for (i = 0; i < (*wcl)->wcl_count; i++)
 1545                         if ((*wcl)->wcl_children[i] == child)
 1546                                 goto found;
 1547         return;
 1548 found:
 1549         (*wcl)->wcl_count--;
 1550         if ((*wcl)->wcl_count > i)
 1551                 (*wcl)->wcl_children[i] =
 1552                     (*wcl)->wcl_children[(*wcl)->wcl_count];
 1553         MPASS((*wcl)->wcl_children[i] != NULL);
 1554         if ((*wcl)->wcl_count != 0)
 1555                 return;
 1556         wcl1 = *wcl;
 1557         *wcl = wcl1->wcl_next;
 1558         w_child_cnt--;
 1559         witness_child_free(wcl1);
 1560 }
 1561 
 1562 static int
 1563 isitmychild(struct witness *parent, struct witness *child)
 1564 {
 1565         struct witness_child_list_entry *wcl;
 1566         int i;
 1567 
 1568         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
 1569                 for (i = 0; i < wcl->wcl_count; i++) {
 1570                         if (wcl->wcl_children[i] == child)
 1571                                 return (1);
 1572                 }
 1573         }
 1574         return (0);
 1575 }
 1576 
 1577 static int
 1578 isitmydescendant(struct witness *parent, struct witness *child)
 1579 {
 1580         struct witness_child_list_entry *wcl;
 1581         int i, j;
 1582 
 1583         if (isitmychild(parent, child))
 1584                 return (1);
 1585         j = 0;
 1586         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
 1587                 MPASS(j < 1000);
 1588                 for (i = 0; i < wcl->wcl_count; i++) {
 1589                         if (isitmydescendant(wcl->wcl_children[i], child))
 1590                                 return (1);
 1591                 }
 1592                 j++;
 1593         }
 1594         return (0);
 1595 }
 1596 
 1597 #ifdef BLESSING
 1598 static int
 1599 blessed(struct witness *w1, struct witness *w2)
 1600 {
 1601         int i;
 1602         struct witness_blessed *b;
 1603 
 1604         for (i = 0; i < blessed_count; i++) {
 1605                 b = &blessed_list[i];
 1606                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
 1607                         if (strcmp(w2->w_name, b->b_lock2) == 0)
 1608                                 return (1);
 1609                         continue;
 1610                 }
 1611                 if (strcmp(w1->w_name, b->b_lock2) == 0)
 1612                         if (strcmp(w2->w_name, b->b_lock1) == 0)
 1613                                 return (1);
 1614         }
 1615         return (0);
 1616 }
 1617 #endif
 1618 
 1619 static struct witness *
 1620 witness_get(void)
 1621 {
 1622         struct witness *w;
 1623 
 1624         if (witness_watch == 0) {
 1625                 mtx_unlock_spin(&w_mtx);
 1626                 return (NULL);
 1627         }
 1628         if (STAILQ_EMPTY(&w_free)) {
 1629                 witness_watch = 0;
 1630                 mtx_unlock_spin(&w_mtx);
 1631                 printf("%s: witness exhausted\n", __func__);
 1632                 return (NULL);
 1633         }
 1634         w = STAILQ_FIRST(&w_free);
 1635         STAILQ_REMOVE_HEAD(&w_free, w_list);
 1636         w_free_cnt--;
 1637         bzero(w, sizeof(*w));
 1638         return (w);
 1639 }
 1640 
 1641 static void
 1642 witness_free(struct witness *w)
 1643 {
 1644 
 1645         STAILQ_INSERT_HEAD(&w_free, w, w_list);
 1646         w_free_cnt++;
 1647 }
 1648 
 1649 static struct witness_child_list_entry *
 1650 witness_child_get(void)
 1651 {
 1652         struct witness_child_list_entry *wcl;
 1653 
 1654         if (witness_watch == 0) {
 1655                 mtx_unlock_spin(&w_mtx);
 1656                 return (NULL);
 1657         }
 1658         wcl = w_child_free;
 1659         if (wcl == NULL) {
 1660                 witness_watch = 0;
 1661                 mtx_unlock_spin(&w_mtx);
 1662                 printf("%s: witness exhausted\n", __func__);
 1663                 return (NULL);
 1664         }
 1665         w_child_free = wcl->wcl_next;
 1666         w_child_free_cnt--;
 1667         bzero(wcl, sizeof(*wcl));
 1668         return (wcl);
 1669 }
 1670 
 1671 static void
 1672 witness_child_free(struct witness_child_list_entry *wcl)
 1673 {
 1674 
 1675         wcl->wcl_next = w_child_free;
 1676         w_child_free = wcl;
 1677         w_child_free_cnt++;
 1678 }
 1679 
 1680 static struct lock_list_entry *
 1681 witness_lock_list_get(void)
 1682 {
 1683         struct lock_list_entry *lle;
 1684 
 1685         if (witness_watch == 0)
 1686                 return (NULL);
 1687         mtx_lock_spin(&w_mtx);
 1688         lle = w_lock_list_free;
 1689         if (lle == NULL) {
 1690                 witness_watch = 0;
 1691                 mtx_unlock_spin(&w_mtx);
 1692                 printf("%s: witness exhausted\n", __func__);
 1693                 return (NULL);
 1694         }
 1695         w_lock_list_free = lle->ll_next;
 1696         mtx_unlock_spin(&w_mtx);
 1697         bzero(lle, sizeof(*lle));
 1698         return (lle);
 1699 }
 1700                 
 1701 static void
 1702 witness_lock_list_free(struct lock_list_entry *lle)
 1703 {
 1704 
 1705         mtx_lock_spin(&w_mtx);
 1706         lle->ll_next = w_lock_list_free;
 1707         w_lock_list_free = lle;
 1708         mtx_unlock_spin(&w_mtx);
 1709 }
 1710 
 1711 static struct lock_instance *
 1712 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
 1713 {
 1714         struct lock_list_entry *lle;
 1715         struct lock_instance *instance;
 1716         int i;
 1717 
 1718         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
 1719                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1720                         instance = &lle->ll_children[i];
 1721                         if (instance->li_lock == lock)
 1722                                 return (instance);
 1723                 }
 1724         return (NULL);
 1725 }
 1726 
 1727 static void
 1728 witness_list_lock(struct lock_instance *instance)
 1729 {
 1730         struct lock_object *lock;
 1731 
 1732         lock = instance->li_lock;
 1733         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
 1734             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
 1735         if (lock->lo_type != lock->lo_name)
 1736                 printf(" (%s)", lock->lo_type);
 1737         printf(" r = %d (%p) locked @ %s:%d\n",
 1738             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
 1739             instance->li_line);
 1740 }
 1741 
 1742 #ifdef DDB
 1743 static int
 1744 witness_thread_has_locks(struct thread *td)
 1745 {
 1746 
 1747         return (td->td_sleeplocks != NULL);
 1748 }
 1749 
 1750 static int
 1751 witness_proc_has_locks(struct proc *p)
 1752 {
 1753         struct thread *td;
 1754 
 1755         FOREACH_THREAD_IN_PROC(p, td) {
 1756                 if (witness_thread_has_locks(td))
 1757                         return (1);
 1758         }
 1759         return (0);
 1760 }
 1761 #endif
 1762 
 1763 int
 1764 witness_list_locks(struct lock_list_entry **lock_list)
 1765 {
 1766         struct lock_list_entry *lle;
 1767         int i, nheld;
 1768 
 1769         nheld = 0;
 1770         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
 1771                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1772                         witness_list_lock(&lle->ll_children[i]);
 1773                         nheld++;
 1774                 }
 1775         return (nheld);
 1776 }
 1777 
 1778 /*
 1779  * This is a bit risky at best.  We call this function when we have timed
 1780  * out acquiring a spin lock, and we assume that the other CPU is stuck
 1781  * with this lock held.  So, we go groveling around in the other CPU's
 1782  * per-cpu data to try to find the lock instance for this spin lock to
 1783  * see when it was last acquired.
 1784  */
 1785 void
 1786 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
 1787 {
 1788         struct lock_instance *instance;
 1789         struct pcpu *pc;
 1790 
 1791         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
 1792                 return;
 1793         pc = pcpu_find(owner->td_oncpu);
 1794         instance = find_instance(pc->pc_spinlocks, lock);
 1795         if (instance != NULL)
 1796                 witness_list_lock(instance);
 1797 }
 1798 
 1799 void
 1800 witness_save(struct lock_object *lock, const char **filep, int *linep)
 1801 {
 1802         struct lock_list_entry *lock_list;
 1803         struct lock_instance *instance;
 1804         struct lock_class *class;
 1805 
 1806         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1807         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1808                 return;
 1809         class = LOCK_CLASS(lock);
 1810         if (class->lc_flags & LC_SLEEPLOCK)
 1811                 lock_list = curthread->td_sleeplocks;
 1812         else {
 1813                 if (witness_skipspin)
 1814                         return;
 1815                 lock_list = PCPU_GET(spinlocks);
 1816         }
 1817         instance = find_instance(lock_list, lock);
 1818         if (instance == NULL)
 1819                 panic("%s: lock (%s) %s not locked", __func__,
 1820                     class->lc_name, lock->lo_name);
 1821         *filep = instance->li_file;
 1822         *linep = instance->li_line;
 1823 }
 1824 
 1825 void
 1826 witness_restore(struct lock_object *lock, const char *file, int line)
 1827 {
 1828         struct lock_list_entry *lock_list;
 1829         struct lock_instance *instance;
 1830         struct lock_class *class;
 1831 
 1832         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1833         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1834                 return;
 1835         class = LOCK_CLASS(lock);
 1836         if (class->lc_flags & LC_SLEEPLOCK)
 1837                 lock_list = curthread->td_sleeplocks;
 1838         else {
 1839                 if (witness_skipspin)
 1840                         return;
 1841                 lock_list = PCPU_GET(spinlocks);
 1842         }
 1843         instance = find_instance(lock_list, lock);
 1844         if (instance == NULL)
 1845                 panic("%s: lock (%s) %s not locked", __func__,
 1846                     class->lc_name, lock->lo_name);
 1847         lock->lo_witness->w_file = file;
 1848         lock->lo_witness->w_line = line;
 1849         instance->li_file = file;
 1850         instance->li_line = line;
 1851 }
 1852 
 1853 void
 1854 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
 1855 {
 1856 #ifdef INVARIANT_SUPPORT
 1857         struct lock_instance *instance;
 1858         struct lock_class *class;
 1859 
 1860         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1861                 return;
 1862         class = LOCK_CLASS(lock);
 1863         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
 1864                 instance = find_instance(curthread->td_sleeplocks, lock);
 1865         else if ((class->lc_flags & LC_SPINLOCK) != 0)
 1866                 instance = find_instance(PCPU_GET(spinlocks), lock);
 1867         else {
 1868                 panic("Lock (%s) %s is not sleep or spin!",
 1869                     class->lc_name, lock->lo_name);
 1870         }
 1871         file = fixup_filename(file);
 1872         switch (flags) {
 1873         case LA_UNLOCKED:
 1874                 if (instance != NULL)
 1875                         panic("Lock (%s) %s locked @ %s:%d.",
 1876                             class->lc_name, lock->lo_name, file, line);
 1877                 break;
 1878         case LA_LOCKED:
 1879         case LA_LOCKED | LA_RECURSED:
 1880         case LA_LOCKED | LA_NOTRECURSED:
 1881         case LA_SLOCKED:
 1882         case LA_SLOCKED | LA_RECURSED:
 1883         case LA_SLOCKED | LA_NOTRECURSED:
 1884         case LA_XLOCKED:
 1885         case LA_XLOCKED | LA_RECURSED:
 1886         case LA_XLOCKED | LA_NOTRECURSED:
 1887                 if (instance == NULL) {
 1888                         panic("Lock (%s) %s not locked @ %s:%d.",
 1889                             class->lc_name, lock->lo_name, file, line);
 1890                         break;
 1891                 }
 1892                 if ((flags & LA_XLOCKED) != 0 &&
 1893                     (instance->li_flags & LI_EXCLUSIVE) == 0)
 1894                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
 1895                             class->lc_name, lock->lo_name, file, line);
 1896                 if ((flags & LA_SLOCKED) != 0 &&
 1897                     (instance->li_flags & LI_EXCLUSIVE) != 0)
 1898                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
 1899                             class->lc_name, lock->lo_name, file, line);
 1900                 if ((flags & LA_RECURSED) != 0 &&
 1901                     (instance->li_flags & LI_RECURSEMASK) == 0)
 1902                         panic("Lock (%s) %s not recursed @ %s:%d.",
 1903                             class->lc_name, lock->lo_name, file, line);
 1904                 if ((flags & LA_NOTRECURSED) != 0 &&
 1905                     (instance->li_flags & LI_RECURSEMASK) != 0)
 1906                         panic("Lock (%s) %s recursed @ %s:%d.",
 1907                             class->lc_name, lock->lo_name, file, line);
 1908                 break;
 1909         default:
 1910                 panic("Invalid lock assertion at %s:%d.", file, line);
 1911 
 1912         }
 1913 #endif  /* INVARIANT_SUPPORT */
 1914 }
 1915 
 1916 #ifdef DDB
 1917 static void
 1918 witness_list(struct thread *td)
 1919 {
 1920 
 1921         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1922         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
 1923 
 1924         if (witness_watch == 0)
 1925                 return;
 1926 
 1927         witness_list_locks(&td->td_sleeplocks);
 1928 
 1929         /*
 1930          * We only handle spinlocks if td == curthread.  This is somewhat broken
 1931          * if td is currently executing on some other CPU and holds spin locks
 1932          * as we won't display those locks.  If we had a MI way of getting
 1933          * the per-cpu data for a given cpu then we could use
 1934          * td->td_oncpu to get the list of spinlocks for this thread
 1935          * and "fix" this.
 1936          *
 1937          * That still wouldn't really fix this unless we locked sched_lock
 1938          * or stopped the other CPU to make sure it wasn't changing the list
 1939          * out from under us.  It is probably best to just not try to handle
 1940          * threads on other CPU's for now.
 1941          */
 1942         if (td == curthread && PCPU_GET(spinlocks) != NULL)
 1943                 witness_list_locks(PCPU_PTR(spinlocks));
 1944 }
 1945 
 1946 DB_SHOW_COMMAND(locks, db_witness_list)
 1947 {
 1948         struct thread *td;
 1949         pid_t pid;
 1950         struct proc *p;
 1951 
 1952         if (have_addr) {
 1953                 pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
 1954                     ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
 1955                     ((addr >> 16) % 16) * 10000;
 1956                 /* sx_slock(&allproc_lock); */
 1957                 FOREACH_PROC_IN_SYSTEM(p) {
 1958                         if (p->p_pid == pid)
 1959                                 break;
 1960                 }
 1961                 /* sx_sunlock(&allproc_lock); */
 1962                 if (p == NULL) {
 1963                         db_printf("pid %d not found\n", pid);
 1964                         return;
 1965                 }
 1966                 FOREACH_THREAD_IN_PROC(p, td) {
 1967                         witness_list(td);
 1968                 }
 1969         } else {
 1970                 td = curthread;
 1971                 witness_list(td);
 1972         }
 1973 }
 1974 
 1975 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
 1976 {
 1977         struct thread *td;
 1978         struct proc *p;
 1979 
 1980         /*
 1981          * It would be nice to list only threads and processes that actually
 1982          * held sleep locks, but that information is currently not exported
 1983          * by WITNESS.
 1984          */
 1985         FOREACH_PROC_IN_SYSTEM(p) {
 1986                 if (!witness_proc_has_locks(p))
 1987                         continue;
 1988                 FOREACH_THREAD_IN_PROC(p, td) {
 1989                         if (!witness_thread_has_locks(td))
 1990                                 continue;
 1991                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
 1992                             p->p_comm, td, td->td_tid);
 1993                         witness_list(td);
 1994                 }
 1995         }
 1996 }
 1997 
 1998 DB_SHOW_COMMAND(witness, db_witness_display)
 1999 {
 2000 
 2001         witness_display(db_printf);
 2002 }
 2003 #endif

Cache object: 93174edb9af687031e33067f2f0556da


[ 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.