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/7.3/sys/kern/subr_witness.c 192494 2009-05-20 22:32:25Z kmacy $");
   86 
   87 #include "opt_ddb.h"
   88 #include "opt_hwpmc_hooks.h"
   89 #include "opt_witness.h"
   90 
   91 #include <sys/param.h>
   92 #include <sys/bus.h>
   93 #include <sys/kdb.h>
   94 #include <sys/kernel.h>
   95 #include <sys/ktr.h>
   96 #include <sys/lock.h>
   97 #include <sys/malloc.h>
   98 #include <sys/mutex.h>
   99 #include <sys/priv.h>
  100 #include <sys/proc.h>
  101 #include <sys/sysctl.h>
  102 #include <sys/systm.h>
  103 
  104 #include <ddb/ddb.h>
  105 
  106 #include <machine/stdarg.h>
  107 
  108 /* Note that these traces do not work with KTR_ALQ. */
  109 #if 0
  110 #define KTR_WITNESS     KTR_SUBSYS
  111 #else
  112 #define KTR_WITNESS     0
  113 #endif
  114 
  115 /* Easier to stay with the old names. */
  116 #define lo_list         lo_witness_data.lod_list
  117 #define lo_witness      lo_witness_data.lod_witness
  118 
  119 /* Define this to check for blessed mutexes */
  120 #undef BLESSING
  121 
  122 #define WITNESS_COUNT 1024
  123 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
  124 /*
  125  * XXX: This is somewhat bogus, as we assume here that at most 1024 threads
  126  * will hold LOCK_NCHILDREN * 2 locks.  We handle failure ok, and we should
  127  * probably be safe for the most part, but it's still a SWAG.
  128  */
  129 #define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2
  130 
  131 #define WITNESS_NCHILDREN 6
  132 
  133 struct witness_child_list_entry;
  134 
  135 struct witness {
  136         const   char *w_name;
  137         struct  lock_class *w_class;
  138         STAILQ_ENTRY(witness) w_list;           /* List of all witnesses. */
  139         STAILQ_ENTRY(witness) w_typelist;       /* Witnesses of a type. */
  140         struct  witness_child_list_entry *w_children;   /* Great evilness... */
  141         const   char *w_file;
  142         int     w_line;
  143         u_int   w_level;
  144         u_int   w_refcount;
  145         u_char  w_Giant_squawked:1;
  146         u_char  w_other_squawked:1;
  147         u_char  w_same_squawked:1;
  148         u_char  w_displayed:1;
  149 };
  150 
  151 struct witness_child_list_entry {
  152         struct  witness_child_list_entry *wcl_next;
  153         struct  witness *wcl_children[WITNESS_NCHILDREN];
  154         u_int   wcl_count;
  155 };
  156 
  157 STAILQ_HEAD(witness_list, witness);
  158 
  159 #ifdef BLESSING
  160 struct witness_blessed {
  161         const   char *b_lock1;
  162         const   char *b_lock2;
  163 };
  164 #endif
  165 
  166 struct witness_order_list_entry {
  167         const   char *w_name;
  168         struct  lock_class *w_class;
  169 };
  170 
  171 #ifdef BLESSING
  172 static int      blessed(struct witness *, struct witness *);
  173 #endif
  174 static int      depart(struct witness *w);
  175 static struct   witness *enroll(const char *description,
  176                                 struct lock_class *lock_class);
  177 static int      insertchild(struct witness *parent, struct witness *child);
  178 static int      isitmychild(struct witness *parent, struct witness *child);
  179 static int      isitmydescendant(struct witness *parent, struct witness *child);
  180 static int      itismychild(struct witness *parent, struct witness *child);
  181 static void     removechild(struct witness *parent, struct witness *child);
  182 static int      sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
  183 static const char *fixup_filename(const char *file);
  184 static struct   witness *witness_get(void);
  185 static void     witness_free(struct witness *m);
  186 static struct   witness_child_list_entry *witness_child_get(void);
  187 static void     witness_child_free(struct witness_child_list_entry *wcl);
  188 static struct   lock_list_entry *witness_lock_list_get(void);
  189 static void     witness_lock_list_free(struct lock_list_entry *lle);
  190 static struct   lock_instance *find_instance(struct lock_list_entry *lock_list,
  191                                              struct lock_object *lock);
  192 static void     witness_list_lock(struct lock_instance *instance);
  193 #ifdef DDB
  194 static void     witness_leveldescendents(struct witness *parent, int level);
  195 static void     witness_levelall(void);
  196 static void     witness_displaydescendants(void(*)(const char *fmt, ...),
  197                                            struct witness *, int indent);
  198 static void     witness_display_list(void(*prnt)(const char *fmt, ...),
  199                                      struct witness_list *list);
  200 static void     witness_display(void(*)(const char *fmt, ...));
  201 static void     witness_list(struct thread *td);
  202 #endif
  203 
  204 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking");
  205 
  206 /*
  207  * If set to 0, witness is disabled.  If set to a non-zero value, witness
  208  * performs full lock order checking for all locks.  At runtime, this
  209  * value may be set to 0 to turn off witness.  witness is not allowed be
  210  * turned on once it is turned off, however.
  211  */
  212 static int witness_watch = 1;
  213 TUNABLE_INT("debug.witness.watch", &witness_watch);
  214 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
  215     sysctl_debug_witness_watch, "I", "witness is watching lock operations");
  216 
  217 #ifdef KDB
  218 /*
  219  * When KDB is enabled and witness_kdb is set to 1, it will cause the system
  220  * to drop into kdebug() when:
  221  *      - a lock hierarchy violation occurs
  222  *      - locks are held when going to sleep.
  223  */
  224 #ifdef WITNESS_KDB
  225 int     witness_kdb = 1;
  226 #else
  227 int     witness_kdb = 0;
  228 #endif
  229 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
  230 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
  231 
  232 /*
  233  * When KDB is enabled and witness_trace is set to 1, it will cause the system
  234  * to print a stack trace:
  235  *      - a lock hierarchy violation occurs
  236  *      - locks are held when going to sleep.
  237  */
  238 int     witness_trace = 1;
  239 TUNABLE_INT("debug.witness.trace", &witness_trace);
  240 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
  241 #endif /* KDB */
  242 
  243 #ifdef WITNESS_SKIPSPIN
  244 int     witness_skipspin = 1;
  245 #else
  246 int     witness_skipspin = 0;
  247 #endif
  248 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
  249 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN,
  250     &witness_skipspin, 0, "");
  251 
  252 static struct mtx w_mtx;
  253 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
  254 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
  255 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
  256 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
  257 static struct witness_child_list_entry *w_child_free = NULL;
  258 static struct lock_list_entry *w_lock_list_free = NULL;
  259 
  260 static int w_free_cnt, w_spin_cnt, w_sleep_cnt, w_child_free_cnt, w_child_cnt;
  261 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
  262 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
  263 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
  264     "");
  265 SYSCTL_INT(_debug_witness, OID_AUTO, child_free_cnt, CTLFLAG_RD,
  266     &w_child_free_cnt, 0, "");
  267 SYSCTL_INT(_debug_witness, OID_AUTO, child_cnt, CTLFLAG_RD, &w_child_cnt, 0,
  268     "");
  269 
  270 static struct witness w_data[WITNESS_COUNT];
  271 static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT];
  272 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
  273 
  274 static struct witness_order_list_entry order_lists[] = {
  275         /*
  276          * sx locks
  277          */
  278         { "proctree", &lock_class_sx },
  279         { "allproc", &lock_class_sx },
  280         { "allprison", &lock_class_sx },
  281         { NULL, NULL },
  282         /*
  283          * Various mutexes
  284          */
  285         { "Giant", &lock_class_mtx_sleep },
  286         { "pipe mutex", &lock_class_mtx_sleep },
  287         { "sigio lock", &lock_class_mtx_sleep },
  288         { "process group", &lock_class_mtx_sleep },
  289         { "process lock", &lock_class_mtx_sleep },
  290         { "session", &lock_class_mtx_sleep },
  291         { "uidinfo hash", &lock_class_mtx_sleep },
  292         { "uidinfo struct", &lock_class_mtx_sleep },
  293 #ifdef  HWPMC_HOOKS
  294         { "pmc-sleep", &lock_class_mtx_sleep },
  295 #endif
  296         { NULL, NULL },
  297         /*
  298          * Sockets
  299          */
  300         { "accept", &lock_class_mtx_sleep },
  301         { "so_snd", &lock_class_mtx_sleep },
  302         { "so_rcv", &lock_class_mtx_sleep },
  303         { "sellck", &lock_class_mtx_sleep },
  304         { NULL, NULL },
  305         /*
  306          * Routing
  307          */
  308         { "so_rcv", &lock_class_mtx_sleep },
  309         { "radix node head", &lock_class_mtx_sleep },
  310         { "rtentry", &lock_class_mtx_sleep },
  311         { "ifaddr", &lock_class_mtx_sleep },
  312         { NULL, NULL },
  313         /*
  314          * Multicast - protocol locks before interface locks, after UDP locks.
  315          */
  316         { "udpinp", &lock_class_rw },
  317         { "in_multi_mtx", &lock_class_mtx_sleep },
  318         { "igmp_mtx", &lock_class_mtx_sleep },
  319         { "if_addr_mtx", &lock_class_mtx_sleep },
  320         { NULL, NULL },
  321         /*
  322          * UNIX Domain Sockets
  323          */
  324         { "unp", &lock_class_mtx_sleep },
  325         { "so_snd", &lock_class_mtx_sleep },
  326         { NULL, NULL },
  327         /*
  328          * UDP/IP
  329          */
  330         { "udp", &lock_class_rw },
  331         { "udpinp", &lock_class_rw },
  332         { "so_snd", &lock_class_mtx_sleep },
  333         { NULL, NULL },
  334         /*
  335          * TCP/IP
  336          */
  337         { "tcp", &lock_class_rw },
  338         { "tcpinp", &lock_class_rw },
  339         { "so_snd", &lock_class_mtx_sleep },
  340         { NULL, NULL },
  341         /*
  342          * SLIP
  343          */
  344         { "slip_mtx", &lock_class_mtx_sleep },
  345         { "slip sc_mtx", &lock_class_mtx_sleep },
  346         { NULL, NULL },
  347         /*
  348          * netatalk
  349          */
  350         { "ddp_list_mtx", &lock_class_mtx_sleep },
  351         { "ddp_mtx", &lock_class_mtx_sleep },
  352         { NULL, NULL },
  353         /*
  354          * BPF
  355          */
  356         { "bpf global lock", &lock_class_mtx_sleep },
  357         { "bpf interface lock", &lock_class_mtx_sleep },
  358         { "bpf cdev lock", &lock_class_mtx_sleep },
  359         { NULL, NULL },
  360         /*
  361          * NFS server
  362          */
  363         { "nfsd_mtx", &lock_class_mtx_sleep },
  364         { "so_snd", &lock_class_mtx_sleep },
  365         { NULL, NULL },
  366 
  367         /*
  368          * IEEE 802.11
  369          */
  370         { "802.11 com lock", &lock_class_mtx_sleep},
  371         { NULL, NULL },
  372         /*
  373          * Network drivers
  374          */
  375         { "network driver", &lock_class_mtx_sleep},
  376         { NULL, NULL },
  377 
  378         /*
  379          * Netgraph
  380          */
  381         { "ng_node", &lock_class_mtx_sleep },
  382         { "ng_worklist", &lock_class_mtx_sleep },
  383         { NULL, NULL },
  384         /*
  385          * CDEV
  386          */
  387         { "system map", &lock_class_mtx_sleep },
  388         { "vm page queue mutex", &lock_class_mtx_sleep },
  389         { "vnode interlock", &lock_class_mtx_sleep },
  390         { "cdev", &lock_class_mtx_sleep },
  391         { NULL, NULL },
  392         /*
  393          * kqueue/VFS interaction
  394          */
  395         { "kqueue", &lock_class_mtx_sleep },
  396         { "struct mount mtx", &lock_class_mtx_sleep },
  397         { "vnode interlock", &lock_class_mtx_sleep },
  398         { NULL, NULL },
  399         /*
  400          * ZFS locking
  401          */
  402         { "dn->dn_mtx", &lock_class_sx },
  403         { "dr->dt.di.dr_mtx", &lock_class_sx },
  404         { "db->db_mtx", &lock_class_sx },
  405         { NULL, NULL },
  406         /*
  407          * spin locks
  408          */
  409 #ifdef SMP
  410         { "ap boot", &lock_class_mtx_spin },
  411 #endif
  412         { "rm.mutex_mtx", &lock_class_mtx_spin },
  413         { "sio", &lock_class_mtx_spin },
  414         { "scrlock", &lock_class_mtx_spin },
  415 #ifdef __i386__
  416         { "cy", &lock_class_mtx_spin },
  417 #endif
  418 #ifdef __sparc64__
  419         { "pcib_mtx", &lock_class_mtx_spin },
  420         { "rtc_mtx", &lock_class_mtx_spin },
  421 #endif
  422         { "scc_hwmtx", &lock_class_mtx_spin },
  423         { "uart_hwmtx", &lock_class_mtx_spin },
  424         { "fast_taskqueue", &lock_class_mtx_spin },
  425         { "intr table", &lock_class_mtx_spin },
  426 #ifdef  HWPMC_HOOKS
  427         { "pmc-per-proc", &lock_class_mtx_spin },
  428 #endif
  429         { "process slock", &lock_class_mtx_spin },
  430         { "sleepq chain", &lock_class_mtx_spin },
  431         { "umtx lock", &lock_class_mtx_spin },
  432         { "turnstile chain", &lock_class_mtx_spin },
  433         { "turnstile lock", &lock_class_mtx_spin },
  434         { "sched lock", &lock_class_mtx_spin },
  435         { "td_contested", &lock_class_mtx_spin },
  436         { "callout", &lock_class_mtx_spin },
  437         { "entropy harvest mutex", &lock_class_mtx_spin },
  438         { "syscons video lock", &lock_class_mtx_spin },
  439         { "time lock", &lock_class_mtx_spin },
  440 #ifdef SMP
  441         { "smp rendezvous", &lock_class_mtx_spin },
  442 #endif
  443         /*
  444          * leaf locks
  445          */
  446         { "icu", &lock_class_mtx_spin },
  447 #if defined(SMP) && defined(__sparc64__)
  448         { "ipi", &lock_class_mtx_spin },
  449 #endif
  450 #ifdef __i386__
  451         { "allpmaps", &lock_class_mtx_spin },
  452         { "descriptor tables", &lock_class_mtx_spin },
  453 #endif
  454         { "clk", &lock_class_mtx_spin },
  455         { "mprof lock", &lock_class_mtx_spin },
  456         { "kse lock", &lock_class_mtx_spin },
  457         { "zombie lock", &lock_class_mtx_spin },
  458         { "ALD Queue", &lock_class_mtx_spin },
  459 #ifdef __ia64__
  460         { "MCA spin lock", &lock_class_mtx_spin },
  461 #endif
  462 #if defined(__i386__) || defined(__amd64__)
  463         { "pcicfg", &lock_class_mtx_spin },
  464         { "NDIS thread lock", &lock_class_mtx_spin },
  465 #endif
  466         { "tw_osl_io_lock", &lock_class_mtx_spin },
  467         { "tw_osl_q_lock", &lock_class_mtx_spin },
  468         { "tw_cl_io_lock", &lock_class_mtx_spin },
  469         { "tw_cl_intr_lock", &lock_class_mtx_spin },
  470         { "tw_cl_gen_lock", &lock_class_mtx_spin },
  471 #ifdef  HWPMC_HOOKS
  472         { "pmc-leaf", &lock_class_mtx_spin },
  473 #endif
  474         { "blocked lock", &lock_class_mtx_spin },
  475         { NULL, NULL },
  476         { NULL, NULL }
  477 };
  478 
  479 #ifdef BLESSING
  480 /*
  481  * Pairs of locks which have been blessed
  482  * Don't complain about order problems with blessed locks
  483  */
  484 static struct witness_blessed blessed_list[] = {
  485 };
  486 static int blessed_count =
  487         sizeof(blessed_list) / sizeof(struct witness_blessed);
  488 #endif
  489 
  490 /*
  491  * List of locks initialized prior to witness being initialized whose
  492  * enrollment is currently deferred.
  493  */
  494 STAILQ_HEAD(, lock_object) pending_locks =
  495     STAILQ_HEAD_INITIALIZER(pending_locks);
  496 
  497 /*
  498  * This global is set to 0 once it becomes safe to use the witness code.
  499  */
  500 static int witness_cold = 1;
  501 
  502 /*
  503  * This global is set to 1 once the static lock orders have been enrolled
  504  * so that a warning can be issued for any spin locks enrolled later.
  505  */
  506 static int witness_spin_warn = 0;
  507 
  508 /*
  509  * The WITNESS-enabled diagnostic code.  Note that the witness code does
  510  * assume that the early boot is single-threaded at least until after this
  511  * routine is completed.
  512  */
  513 static void
  514 witness_initialize(void *dummy __unused)
  515 {
  516         struct lock_object *lock;
  517         struct witness_order_list_entry *order;
  518         struct witness *w, *w1;
  519         int i;
  520 
  521         /*
  522          * We have to release Giant before initializing its witness
  523          * structure so that WITNESS doesn't get confused.
  524          */
  525         mtx_unlock(&Giant);
  526         mtx_assert(&Giant, MA_NOTOWNED);
  527 
  528         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
  529         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
  530             MTX_NOWITNESS | MTX_NOPROFILE);
  531         for (i = 0; i < WITNESS_COUNT; i++)
  532                 witness_free(&w_data[i]);
  533         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
  534                 witness_child_free(&w_childdata[i]);
  535         for (i = 0; i < LOCK_CHILDCOUNT; i++)
  536                 witness_lock_list_free(&w_locklistdata[i]);
  537 
  538         /* First add in all the specified order lists. */
  539         for (order = order_lists; order->w_name != NULL; order++) {
  540                 w = enroll(order->w_name, order->w_class);
  541                 if (w == NULL)
  542                         continue;
  543                 w->w_file = "order list";
  544                 for (order++; order->w_name != NULL; order++) {
  545                         w1 = enroll(order->w_name, order->w_class);
  546                         if (w1 == NULL)
  547                                 continue;
  548                         w1->w_file = "order list";
  549                         if (!itismychild(w, w1))
  550                                 panic("Not enough memory for static orders!");
  551                         w = w1;
  552                 }
  553         }
  554         witness_spin_warn = 1;
  555 
  556         /* Iterate through all locks and add them to witness. */
  557         while (!STAILQ_EMPTY(&pending_locks)) {
  558                 lock = STAILQ_FIRST(&pending_locks);
  559                 STAILQ_REMOVE_HEAD(&pending_locks, lo_list);
  560                 KASSERT(lock->lo_flags & LO_WITNESS,
  561                     ("%s: lock %s is on pending list but not LO_WITNESS",
  562                     __func__, lock->lo_name));
  563                 lock->lo_witness = enroll(lock->lo_type, LOCK_CLASS(lock));
  564         }
  565 
  566         /* Mark the witness code as being ready for use. */
  567         witness_cold = 0;
  568 
  569         mtx_lock(&Giant);
  570 }
  571 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
  572     NULL);
  573 
  574 static int
  575 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
  576 {
  577         int error, value;
  578 
  579         value = witness_watch;
  580         error = sysctl_handle_int(oidp, &value, 0, req);
  581         if (error != 0 || req->newptr == NULL)
  582                 return (error);
  583         if (value == witness_watch)
  584                 return (0);
  585         if (value != 0)
  586                 return (EINVAL);
  587         witness_watch = 0;
  588         return (0);
  589 }
  590 
  591 void
  592 witness_init(struct lock_object *lock)
  593 {
  594         struct lock_class *class;
  595 
  596         /* Various sanity checks. */
  597         class = LOCK_CLASS(lock);
  598         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
  599             (class->lc_flags & LC_RECURSABLE) == 0)
  600                 panic("%s: lock (%s) %s can not be recursable", __func__,
  601                     class->lc_name, lock->lo_name);
  602         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  603             (class->lc_flags & LC_SLEEPABLE) == 0)
  604                 panic("%s: lock (%s) %s can not be sleepable", __func__,
  605                     class->lc_name, lock->lo_name);
  606         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
  607             (class->lc_flags & LC_UPGRADABLE) == 0)
  608                 panic("%s: lock (%s) %s can not be upgradable", __func__,
  609                     class->lc_name, lock->lo_name);
  610 
  611         /*
  612          * If we shouldn't watch this lock, then just clear lo_witness.
  613          * Otherwise, if witness_cold is set, then it is too early to
  614          * enroll this lock, so defer it to witness_initialize() by adding
  615          * it to the pending_locks list.  If it is not too early, then enroll
  616          * the lock now.
  617          */
  618         if (witness_watch == 0 || panicstr != NULL ||
  619             (lock->lo_flags & LO_WITNESS) == 0)
  620                 lock->lo_witness = NULL;
  621         else if (witness_cold) {
  622                 STAILQ_INSERT_TAIL(&pending_locks, lock, lo_list);
  623                 lock->lo_flags |= LO_ENROLLPEND;
  624         } else
  625                 lock->lo_witness = enroll(lock->lo_type, class);
  626 }
  627 
  628 void
  629 witness_destroy(struct lock_object *lock)
  630 {
  631         struct lock_class *class;
  632         struct witness *w;
  633 
  634         class = LOCK_CLASS(lock);
  635         if (witness_cold)
  636                 panic("lock (%s) %s destroyed while witness_cold",
  637                     class->lc_name, lock->lo_name);
  638 
  639         /* XXX: need to verify that no one holds the lock */
  640         if ((lock->lo_flags & (LO_WITNESS | LO_ENROLLPEND)) == LO_WITNESS &&
  641             lock->lo_witness != NULL) {
  642                 w = lock->lo_witness;
  643                 mtx_lock_spin(&w_mtx);
  644                 MPASS(w->w_refcount > 0);
  645                 w->w_refcount--;
  646 
  647                 /*
  648                  * Lock is already released if we have an allocation failure
  649                  * and depart() fails.
  650                  */
  651                 if (w->w_refcount != 0 || depart(w))
  652                         mtx_unlock_spin(&w_mtx);
  653         }
  654 
  655         /*
  656          * If this lock is destroyed before witness is up and running,
  657          * remove it from the pending list.
  658          */
  659         if (lock->lo_flags & LO_ENROLLPEND) {
  660                 STAILQ_REMOVE(&pending_locks, lock, lock_object, lo_list);
  661                 lock->lo_flags &= ~LO_ENROLLPEND;
  662         }
  663 }
  664 
  665 #ifdef DDB
  666 static void
  667 witness_levelall (void)
  668 {
  669         struct witness_list *list;
  670         struct witness *w, *w1;
  671 
  672         /*
  673          * First clear all levels.
  674          */
  675         STAILQ_FOREACH(w, &w_all, w_list) {
  676                 w->w_level = 0;
  677         }
  678 
  679         /*
  680          * Look for locks with no parent and level all their descendants.
  681          */
  682         STAILQ_FOREACH(w, &w_all, w_list) {
  683                 /*
  684                  * This is just an optimization, technically we could get
  685                  * away just walking the all list each time.
  686                  */
  687                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
  688                         list = &w_sleep;
  689                 else
  690                         list = &w_spin;
  691                 STAILQ_FOREACH(w1, list, w_typelist) {
  692                         if (isitmychild(w1, w))
  693                                 goto skip;
  694                 }
  695                 witness_leveldescendents(w, 0);
  696         skip:
  697                 ;       /* silence GCC 3.x */
  698         }
  699 }
  700 
  701 static void
  702 witness_leveldescendents(struct witness *parent, int level)
  703 {
  704         struct witness_child_list_entry *wcl;
  705         int i;
  706 
  707         if (parent->w_level < level)
  708                 parent->w_level = level;
  709         level++;
  710         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
  711                 for (i = 0; i < wcl->wcl_count; i++)
  712                         witness_leveldescendents(wcl->wcl_children[i], level);
  713 }
  714 
  715 static void
  716 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
  717                            struct witness *parent, int indent)
  718 {
  719         struct witness_child_list_entry *wcl;
  720         int i, level;
  721 
  722         level = parent->w_level;
  723         prnt("%-2d", level);
  724         for (i = 0; i < indent; i++)
  725                 prnt(" ");
  726         if (parent->w_refcount > 0)
  727                 prnt("%s", parent->w_name);
  728         else
  729                 prnt("(dead)");
  730         if (parent->w_displayed) {
  731                 prnt(" -- (already displayed)\n");
  732                 return;
  733         }
  734         parent->w_displayed = 1;
  735         if (parent->w_refcount > 0) {
  736                 if (parent->w_file != NULL)
  737                         prnt(" -- last acquired @ %s:%d", parent->w_file,
  738                             parent->w_line);
  739         }
  740         prnt("\n");
  741         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
  742                 for (i = 0; i < wcl->wcl_count; i++)
  743                             witness_displaydescendants(prnt,
  744                                 wcl->wcl_children[i], indent + 1);
  745 }
  746 
  747 static void
  748 witness_display_list(void(*prnt)(const char *fmt, ...),
  749                      struct witness_list *list)
  750 {
  751         struct witness *w;
  752 
  753         STAILQ_FOREACH(w, list, w_typelist) {
  754                 if (w->w_file == NULL || w->w_level > 0)
  755                         continue;
  756                 /*
  757                  * This lock has no anscestors, display its descendants. 
  758                  */
  759                 witness_displaydescendants(prnt, w, 0);
  760         }
  761 }
  762         
  763 static void
  764 witness_display(void(*prnt)(const char *fmt, ...))
  765 {
  766         struct witness *w;
  767 
  768         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
  769         witness_levelall();
  770 
  771         /* Clear all the displayed flags. */
  772         STAILQ_FOREACH(w, &w_all, w_list) {
  773                 w->w_displayed = 0;
  774         }
  775 
  776         /*
  777          * First, handle sleep locks which have been acquired at least
  778          * once.
  779          */
  780         prnt("Sleep locks:\n");
  781         witness_display_list(prnt, &w_sleep);
  782         
  783         /*
  784          * Now do spin locks which have been acquired at least once.
  785          */
  786         prnt("\nSpin locks:\n");
  787         witness_display_list(prnt, &w_spin);
  788         
  789         /*
  790          * Finally, any locks which have not been acquired yet.
  791          */
  792         prnt("\nLocks which were never acquired:\n");
  793         STAILQ_FOREACH(w, &w_all, w_list) {
  794                 if (w->w_file != NULL || w->w_refcount == 0)
  795                         continue;
  796                 prnt("%s\n", w->w_name);
  797         }
  798 }
  799 #endif /* DDB */
  800 
  801 /* Trim useless garbage from filenames. */
  802 static const char *
  803 fixup_filename(const char *file)
  804 {
  805 
  806         if (file == NULL)
  807                 return (NULL);
  808         while (strncmp(file, "../", 3) == 0)
  809                 file += 3;
  810         return (file);
  811 }
  812 
  813 int
  814 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
  815 {
  816 
  817         if (witness_watch == 0 || panicstr != NULL)
  818                 return (0);
  819 
  820         /* Require locks that witness knows about. */
  821         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
  822             lock2->lo_witness == NULL)
  823                 return (EINVAL);
  824 
  825         MPASS(!mtx_owned(&w_mtx));
  826         mtx_lock_spin(&w_mtx);
  827 
  828         /*
  829          * If we already have either an explicit or implied lock order that
  830          * is the other way around, then return an error.
  831          */
  832         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
  833                 mtx_unlock_spin(&w_mtx);
  834                 return (EDOOFUS);
  835         }
  836         
  837         /* Try to add the new order. */
  838         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
  839             lock2->lo_type, lock1->lo_type);
  840         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
  841                 return (ENOMEM);
  842         mtx_unlock_spin(&w_mtx);
  843         return (0);
  844 }
  845 
  846 void
  847 witness_checkorder(struct lock_object *lock, int flags, const char *file,
  848     int line)
  849 {
  850         struct lock_list_entry **lock_list, *lle;
  851         struct lock_instance *lock1, *lock2;
  852         struct lock_class *class;
  853         struct witness *w, *w1;
  854         struct thread *td;
  855         int i, j;
  856 
  857         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
  858             panicstr != NULL)
  859                 return;
  860 
  861         /*
  862          * Try locks do not block if they fail to acquire the lock, thus
  863          * there is no danger of deadlocks or of switching while holding a
  864          * spin lock if we acquire a lock via a try operation.  This
  865          * function shouldn't even be called for try locks, so panic if
  866          * that happens.
  867          */
  868         if (flags & LOP_TRYLOCK)
  869                 panic("%s should not be called for try lock operations",
  870                     __func__);
  871 
  872         w = lock->lo_witness;
  873         class = LOCK_CLASS(lock);
  874         td = curthread;
  875         file = fixup_filename(file);
  876 
  877         if (class->lc_flags & LC_SLEEPLOCK) {
  878                 /*
  879                  * Since spin locks include a critical section, this check
  880                  * implicitly enforces a lock order of all sleep locks before
  881                  * all spin locks.
  882                  */
  883                 if (td->td_critnest != 0 && !kdb_active)
  884                         panic("blockable sleep lock (%s) %s @ %s:%d",
  885                             class->lc_name, lock->lo_name, file, line);
  886 
  887                 /*
  888                  * If this is the first lock acquired then just return as
  889                  * no order checking is needed.
  890                  */
  891                 if (td->td_sleeplocks == NULL)
  892                         return;
  893                 lock_list = &td->td_sleeplocks;
  894         } else {
  895                 /*
  896                  * If this is the first lock, just return as no order
  897                  * checking is needed.  We check this in both if clauses
  898                  * here as unifying the check would require us to use a
  899                  * critical section to ensure we don't migrate while doing
  900                  * the check.  Note that if this is not the first lock, we
  901                  * are already in a critical section and are safe for the
  902                  * rest of the check.
  903                  */
  904                 if (PCPU_GET(spinlocks) == NULL)
  905                         return;
  906                 lock_list = PCPU_PTR(spinlocks);
  907         }
  908 
  909         /*
  910          * Check to see if we are recursing on a lock we already own.  If
  911          * so, make sure that we don't mismatch exclusive and shared lock
  912          * acquires.
  913          */
  914         lock1 = find_instance(*lock_list, lock);
  915         if (lock1 != NULL) {
  916                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
  917                     (flags & LOP_EXCLUSIVE) == 0) {
  918                         printf("shared lock of (%s) %s @ %s:%d\n",
  919                             class->lc_name, lock->lo_name, file, line);
  920                         printf("while exclusively locked from %s:%d\n",
  921                             lock1->li_file, lock1->li_line);
  922                         panic("share->excl");
  923                 }
  924                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
  925                     (flags & LOP_EXCLUSIVE) != 0) {
  926                         printf("exclusive lock of (%s) %s @ %s:%d\n",
  927                             class->lc_name, lock->lo_name, file, line);
  928                         printf("while share locked from %s:%d\n",
  929                             lock1->li_file, lock1->li_line);
  930                         panic("excl->share");
  931                 }
  932                 return;
  933         }
  934 
  935         /*
  936          * Try locks do not block if they fail to acquire the lock, thus
  937          * there is no danger of deadlocks or of switching while holding a
  938          * spin lock if we acquire a lock via a try operation.
  939          */
  940         if (flags & LOP_TRYLOCK)
  941                 return;
  942 
  943         /*
  944          * Check for duplicate locks of the same type.  Note that we only
  945          * have to check for this on the last lock we just acquired.  Any
  946          * other cases will be caught as lock order violations.
  947          */
  948         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
  949         w1 = lock1->li_lock->lo_witness;
  950         if (w1 == w) {
  951                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
  952                     (flags & LOP_DUPOK))
  953                         return;
  954                 w->w_same_squawked = 1;
  955                 printf("acquiring duplicate lock of same type: \"%s\"\n", 
  956                         lock->lo_type);
  957                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
  958                     lock1->li_file, lock1->li_line);
  959                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
  960 #ifdef KDB
  961                 goto debugger;
  962 #else
  963                 return;
  964 #endif
  965         }
  966         MPASS(!mtx_owned(&w_mtx));
  967         mtx_lock_spin(&w_mtx);
  968         /*
  969          * If we know that the the lock we are acquiring comes after
  970          * the lock we most recently acquired in the lock order tree,
  971          * then there is no need for any further checks.
  972          */
  973         if (isitmychild(w1, w)) {
  974                 mtx_unlock_spin(&w_mtx);
  975                 return;
  976         }
  977         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
  978                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
  979 
  980                         MPASS(j < WITNESS_COUNT);
  981                         lock1 = &lle->ll_children[i];
  982                         w1 = lock1->li_lock->lo_witness;
  983 
  984                         /*
  985                          * If this lock doesn't undergo witness checking,
  986                          * then skip it.
  987                          */
  988                         if (w1 == NULL) {
  989                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
  990                                     ("lock missing witness structure"));
  991                                 continue;
  992                         }
  993                         /*
  994                          * If we are locking Giant and this is a sleepable
  995                          * lock, then skip it.
  996                          */
  997                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
  998                             lock == &Giant.lock_object)
  999                                 continue;
 1000                         /*
 1001                          * If we are locking a sleepable lock and this lock
 1002                          * is Giant, then skip it.
 1003                          */
 1004                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1005                             lock1->li_lock == &Giant.lock_object)
 1006                                 continue;
 1007                         /*
 1008                          * If we are locking a sleepable lock and this lock
 1009                          * isn't sleepable, we want to treat it as a lock
 1010                          * order violation to enfore a general lock order of
 1011                          * sleepable locks before non-sleepable locks.
 1012                          */
 1013                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1014                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
 1015                                 goto reversal;
 1016                         /*
 1017                          * If we are locking Giant and this is a non-sleepable
 1018                          * lock, then treat it as a reversal.
 1019                          */
 1020                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
 1021                             lock == &Giant.lock_object)
 1022                                 goto reversal;
 1023                         /*
 1024                          * Check the lock order hierarchy for a reveresal.
 1025                          */
 1026                         if (!isitmydescendant(w, w1))
 1027                                 continue;
 1028                 reversal:
 1029                         /*
 1030                          * We have a lock order violation, check to see if it
 1031                          * is allowed or has already been yelled about.
 1032                          */
 1033                         mtx_unlock_spin(&w_mtx);
 1034 #ifdef BLESSING
 1035                         /*
 1036                          * If the lock order is blessed, just bail.  We don't
 1037                          * look for other lock order violations though, which
 1038                          * may be a bug.
 1039                          */
 1040                         if (blessed(w, w1))
 1041                                 return;
 1042 #endif
 1043                         if (lock1->li_lock == &Giant.lock_object) {
 1044                                 if (w1->w_Giant_squawked)
 1045                                         return;
 1046                                 else
 1047                                         w1->w_Giant_squawked = 1;
 1048                         } else {
 1049                                 if (w1->w_other_squawked)
 1050                                         return;
 1051                                 else
 1052                                         w1->w_other_squawked = 1;
 1053                         }
 1054                         /*
 1055                          * Ok, yell about it.
 1056                          */
 1057                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1058                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
 1059                                 printf(
 1060                 "lock order reversal: (sleepable after non-sleepable)\n");
 1061                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
 1062                             && lock == &Giant.lock_object)
 1063                                 printf(
 1064                 "lock order reversal: (Giant after non-sleepable)\n");
 1065                         else
 1066                                 printf("lock order reversal:\n");
 1067                         /*
 1068                          * Try to locate an earlier lock with
 1069                          * witness w in our list.
 1070                          */
 1071                         do {
 1072                                 lock2 = &lle->ll_children[i];
 1073                                 MPASS(lock2->li_lock != NULL);
 1074                                 if (lock2->li_lock->lo_witness == w)
 1075                                         break;
 1076                                 if (i == 0 && lle->ll_next != NULL) {
 1077                                         lle = lle->ll_next;
 1078                                         i = lle->ll_count - 1;
 1079                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
 1080                                 } else
 1081                                         i--;
 1082                         } while (i >= 0);
 1083                         if (i < 0) {
 1084                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 1085                                     lock1->li_lock, lock1->li_lock->lo_name,
 1086                                     lock1->li_lock->lo_type, lock1->li_file,
 1087                                     lock1->li_line);
 1088                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
 1089                                     lock->lo_name, lock->lo_type, file, line);
 1090                         } else {
 1091                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 1092                                     lock2->li_lock, lock2->li_lock->lo_name,
 1093                                     lock2->li_lock->lo_type, lock2->li_file,
 1094                                     lock2->li_line);
 1095                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
 1096                                     lock1->li_lock, lock1->li_lock->lo_name,
 1097                                     lock1->li_lock->lo_type, lock1->li_file,
 1098                                     lock1->li_line);
 1099                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
 1100                                     lock->lo_name, lock->lo_type, file, line);
 1101                         }
 1102 #ifdef KDB
 1103                         goto debugger;
 1104 #else
 1105                         return;
 1106 #endif
 1107                 }
 1108         }
 1109         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 1110         /*
 1111          * If requested, build a new lock order.  However, don't build a new
 1112          * relationship between a sleepable lock and Giant if it is in the
 1113          * wrong direction.  The correct lock order is that sleepable locks
 1114          * always come before Giant.
 1115          */
 1116         if (flags & LOP_NEWORDER &&
 1117             !(lock1->li_lock == &Giant.lock_object &&
 1118             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
 1119                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 1120                     lock->lo_type, lock1->li_lock->lo_type);
 1121                 if (!itismychild(lock1->li_lock->lo_witness, w))
 1122                         /* Witness is dead. */
 1123                         return;
 1124         } 
 1125         mtx_unlock_spin(&w_mtx);
 1126         return;
 1127 
 1128 #ifdef KDB
 1129 debugger:
 1130         if (witness_trace)
 1131                 kdb_backtrace();
 1132         if (witness_kdb)
 1133                 kdb_enter_why(KDB_WHY_WITNESS, __func__);
 1134 #endif
 1135 }
 1136 
 1137 void
 1138 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
 1139 {
 1140         struct lock_list_entry **lock_list, *lle;
 1141         struct lock_instance *instance;
 1142         struct witness *w;
 1143         struct thread *td;
 1144 
 1145         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 1146             panicstr != NULL)
 1147                 return;
 1148         w = lock->lo_witness;
 1149         td = curthread;
 1150         file = fixup_filename(file);
 1151 
 1152         /* Determine lock list for this lock. */
 1153         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
 1154                 lock_list = &td->td_sleeplocks;
 1155         else
 1156                 lock_list = PCPU_PTR(spinlocks);
 1157 
 1158         /* Check to see if we are recursing on a lock we already own. */
 1159         instance = find_instance(*lock_list, lock);
 1160         if (instance != NULL) {
 1161                 instance->li_flags++;
 1162                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
 1163                     td->td_proc->p_pid, lock->lo_name,
 1164                     instance->li_flags & LI_RECURSEMASK);
 1165                 instance->li_file = file;
 1166                 instance->li_line = line;
 1167                 return;
 1168         }
 1169 
 1170         /* Update per-witness last file and line acquire. */
 1171         w->w_file = file;
 1172         w->w_line = line;
 1173 
 1174         /* Find the next open lock instance in the list and fill it. */
 1175         lle = *lock_list;
 1176         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
 1177                 lle = witness_lock_list_get();
 1178                 if (lle == NULL)
 1179                         return;
 1180                 lle->ll_next = *lock_list;
 1181                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
 1182                     td->td_proc->p_pid, lle);
 1183                 *lock_list = lle;
 1184         }
 1185         instance = &lle->ll_children[lle->ll_count++];
 1186         instance->li_lock = lock;
 1187         instance->li_line = line;
 1188         instance->li_file = file;
 1189         if ((flags & LOP_EXCLUSIVE) != 0)
 1190                 instance->li_flags = LI_EXCLUSIVE;
 1191         else
 1192                 instance->li_flags = 0;
 1193         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
 1194             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
 1195 }
 1196 
 1197 void
 1198 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
 1199 {
 1200         struct lock_instance *instance;
 1201         struct lock_class *class;
 1202 
 1203         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1204         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1205                 return;
 1206         class = LOCK_CLASS(lock);
 1207         file = fixup_filename(file);
 1208         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1209                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
 1210                     class->lc_name, lock->lo_name, file, line);
 1211         if ((flags & LOP_TRYLOCK) == 0)
 1212                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
 1213                     lock->lo_name, file, line);
 1214         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
 1215                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
 1216                     class->lc_name, lock->lo_name, file, line);
 1217         instance = find_instance(curthread->td_sleeplocks, lock);
 1218         if (instance == NULL)
 1219                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
 1220                     class->lc_name, lock->lo_name, file, line);
 1221         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
 1222                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
 1223                     class->lc_name, lock->lo_name, file, line);
 1224         if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1225                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1226                     class->lc_name, lock->lo_name,
 1227                     instance->li_flags & LI_RECURSEMASK, file, line);
 1228         instance->li_flags |= LI_EXCLUSIVE;
 1229 }
 1230 
 1231 void
 1232 witness_downgrade(struct lock_object *lock, int flags, const char *file,
 1233     int line)
 1234 {
 1235         struct lock_instance *instance;
 1236         struct lock_class *class;
 1237 
 1238         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1239         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1240                 return;
 1241         class = LOCK_CLASS(lock);
 1242         file = fixup_filename(file);
 1243         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1244                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
 1245                     class->lc_name, lock->lo_name, file, line);
 1246         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
 1247                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
 1248                     class->lc_name, lock->lo_name, file, line);
 1249         instance = find_instance(curthread->td_sleeplocks, lock);
 1250         if (instance == NULL)
 1251                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
 1252                     class->lc_name, lock->lo_name, file, line);
 1253         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
 1254                 panic("downgrade of shared lock (%s) %s @ %s:%d",
 1255                     class->lc_name, lock->lo_name, file, line);
 1256         if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1257                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1258                     class->lc_name, lock->lo_name,
 1259                     instance->li_flags & LI_RECURSEMASK, file, line);
 1260         instance->li_flags &= ~LI_EXCLUSIVE;
 1261 }
 1262 
 1263 void
 1264 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
 1265 {
 1266         struct lock_list_entry **lock_list, *lle;
 1267         struct lock_instance *instance;
 1268         struct lock_class *class;
 1269         struct thread *td;
 1270         register_t s;
 1271         int i, j;
 1272 
 1273         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 1274             panicstr != NULL)
 1275                 return;
 1276         td = curthread;
 1277         class = LOCK_CLASS(lock);
 1278         file = fixup_filename(file);
 1279 
 1280         /* Find lock instance associated with this lock. */
 1281         if (class->lc_flags & LC_SLEEPLOCK)
 1282                 lock_list = &td->td_sleeplocks;
 1283         else
 1284                 lock_list = PCPU_PTR(spinlocks);
 1285         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
 1286                 for (i = 0; i < (*lock_list)->ll_count; i++) {
 1287                         instance = &(*lock_list)->ll_children[i];
 1288                         if (instance->li_lock == lock)
 1289                                 goto found;
 1290                 }
 1291         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
 1292             file, line);
 1293 found:
 1294 
 1295         /* First, check for shared/exclusive mismatches. */
 1296         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
 1297             (flags & LOP_EXCLUSIVE) == 0) {
 1298                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1299                     lock->lo_name, file, line);
 1300                 printf("while exclusively locked from %s:%d\n",
 1301                     instance->li_file, instance->li_line);
 1302                 panic("excl->ushare");
 1303         }
 1304         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
 1305             (flags & LOP_EXCLUSIVE) != 0) {
 1306                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1307                     lock->lo_name, file, line);
 1308                 printf("while share locked from %s:%d\n", instance->li_file,
 1309                     instance->li_line);
 1310                 panic("share->uexcl");
 1311         }
 1312 
 1313         /* If we are recursed, unrecurse. */
 1314         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
 1315                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
 1316                     td->td_proc->p_pid, instance->li_lock->lo_name,
 1317                     instance->li_flags);
 1318                 instance->li_flags--;
 1319                 return;
 1320         }
 1321 
 1322         /* Otherwise, remove this item from the list. */
 1323         s = intr_disable();
 1324         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
 1325             td->td_proc->p_pid, instance->li_lock->lo_name,
 1326             (*lock_list)->ll_count - 1);
 1327         for (j = i; j < (*lock_list)->ll_count - 1; j++)
 1328                 (*lock_list)->ll_children[j] =
 1329                     (*lock_list)->ll_children[j + 1];
 1330         (*lock_list)->ll_count--;
 1331         intr_restore(s);
 1332 
 1333         /* If this lock list entry is now empty, free it. */
 1334         if ((*lock_list)->ll_count == 0) {
 1335                 lle = *lock_list;
 1336                 *lock_list = lle->ll_next;
 1337                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
 1338                     td->td_proc->p_pid, lle);
 1339                 witness_lock_list_free(lle);
 1340         }
 1341 }
 1342 
 1343 /*
 1344  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
 1345  * exempt Giant and sleepable locks from the checks as well.  If any
 1346  * non-exempt locks are held, then a supplied message is printed to the
 1347  * console along with a list of the offending locks.  If indicated in the
 1348  * flags then a failure results in a panic as well.
 1349  */
 1350 int
 1351 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
 1352 {
 1353         struct lock_list_entry *lle;
 1354         struct lock_instance *lock1;
 1355         struct thread *td;
 1356         va_list ap;
 1357         int i, n;
 1358 
 1359         if (witness_cold || witness_watch == 0 || panicstr != NULL)
 1360                 return (0);
 1361         n = 0;
 1362         td = curthread;
 1363         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
 1364                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1365                         lock1 = &lle->ll_children[i];
 1366                         if (lock1->li_lock == lock)
 1367                                 continue;
 1368                         if (flags & WARN_GIANTOK &&
 1369                             lock1->li_lock == &Giant.lock_object)
 1370                                 continue;
 1371                         if (flags & WARN_SLEEPOK &&
 1372                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
 1373                                 continue;
 1374                         if (n == 0) {
 1375                                 va_start(ap, fmt);
 1376                                 vprintf(fmt, ap);
 1377                                 va_end(ap);
 1378                                 printf(" with the following");
 1379                                 if (flags & WARN_SLEEPOK)
 1380                                         printf(" non-sleepable");
 1381                                 printf(" locks held:\n");
 1382                         }
 1383                         n++;
 1384                         witness_list_lock(lock1);
 1385                 }
 1386         if (PCPU_GET(spinlocks) != NULL) {
 1387                 /*
 1388                  * Since we already hold a spinlock preemption is
 1389                  * already blocked.
 1390                  */
 1391                 if (n == 0) {
 1392                         va_start(ap, fmt);
 1393                         vprintf(fmt, ap);
 1394                         va_end(ap);
 1395                         printf(" with the following");
 1396                         if (flags & WARN_SLEEPOK)
 1397                                 printf(" non-sleepable");
 1398                         printf(" locks held:\n");
 1399                 }
 1400                 n += witness_list_locks(PCPU_PTR(spinlocks));
 1401         }
 1402         if (flags & WARN_PANIC && n)
 1403                 panic("witness_warn");
 1404 #ifdef KDB
 1405         else if (witness_kdb && n)
 1406                 kdb_enter_why(KDB_WHY_WITNESS, __func__);
 1407         else if (witness_trace && n)
 1408                 kdb_backtrace();
 1409 #endif
 1410         return (n);
 1411 }
 1412 
 1413 const char *
 1414 witness_file(struct lock_object *lock)
 1415 {
 1416         struct witness *w;
 1417 
 1418         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
 1419                 return ("?");
 1420         w = lock->lo_witness;
 1421         return (w->w_file);
 1422 }
 1423 
 1424 int
 1425 witness_line(struct lock_object *lock)
 1426 {
 1427         struct witness *w;
 1428 
 1429         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
 1430                 return (0);
 1431         w = lock->lo_witness;
 1432         return (w->w_line);
 1433 }
 1434 
 1435 static struct witness *
 1436 enroll(const char *description, struct lock_class *lock_class)
 1437 {
 1438         struct witness *w;
 1439 
 1440         if (witness_watch == 0 || panicstr != NULL)
 1441                 return (NULL);
 1442         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
 1443                 return (NULL);
 1444         mtx_lock_spin(&w_mtx);
 1445         STAILQ_FOREACH(w, &w_all, w_list) {
 1446                 if (w->w_name == description || (w->w_refcount > 0 &&
 1447                     strcmp(description, w->w_name) == 0)) {
 1448                         w->w_refcount++;
 1449                         mtx_unlock_spin(&w_mtx);
 1450                         if (lock_class != w->w_class)
 1451                                 panic(
 1452                                 "lock (%s) %s does not match earlier (%s) lock",
 1453                                     description, lock_class->lc_name,
 1454                                     w->w_class->lc_name);
 1455                         return (w);
 1456                 }
 1457         }
 1458         if ((w = witness_get()) == NULL)
 1459                 goto out;
 1460         w->w_name = description;
 1461         w->w_class = lock_class;
 1462         w->w_refcount = 1;
 1463         STAILQ_INSERT_HEAD(&w_all, w, w_list);
 1464         if (lock_class->lc_flags & LC_SPINLOCK) {
 1465                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
 1466                 w_spin_cnt++;
 1467         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
 1468                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
 1469                 w_sleep_cnt++;
 1470         } else {
 1471                 mtx_unlock_spin(&w_mtx);
 1472                 panic("lock class %s is not sleep or spin",
 1473                     lock_class->lc_name);
 1474         }
 1475         mtx_unlock_spin(&w_mtx);
 1476 out:
 1477         /*
 1478          * We issue a warning for any spin locks not defined in the static
 1479          * order list as a way to discourage their use (folks should really
 1480          * be using non-spin mutexes most of the time).  However, several
 1481          * 3rd part device drivers use spin locks because that is all they
 1482          * have available on Windows and Linux and they think that normal
 1483          * mutexes are insufficient.
 1484          */
 1485         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
 1486                 printf("WITNESS: spin lock %s not in order list\n",
 1487                     description);
 1488         return (w);
 1489 }
 1490 
 1491 /* Don't let the door bang you on the way out... */
 1492 static int
 1493 depart(struct witness *w)
 1494 {
 1495         struct witness_child_list_entry *wcl, *nwcl;
 1496         struct witness_list *list;
 1497         struct witness *parent;
 1498 
 1499         MPASS(w->w_refcount == 0);
 1500         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
 1501                 list = &w_sleep;
 1502                 w_sleep_cnt--;
 1503         } else {
 1504                 list = &w_spin;
 1505                 w_spin_cnt--;
 1506         }
 1507         /*
 1508          * First, we run through the entire tree looking for any
 1509          * witnesses that the outgoing witness is a child of.  For
 1510          * each parent that we find, we reparent all the direct
 1511          * children of the outgoing witness to its parent.
 1512          */
 1513         STAILQ_FOREACH(parent, list, w_typelist) {
 1514                 if (!isitmychild(parent, w))
 1515                         continue;
 1516                 removechild(parent, w);
 1517         }
 1518 
 1519         /*
 1520          * Now we go through and free up the child list of the
 1521          * outgoing witness.
 1522          */
 1523         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
 1524                 nwcl = wcl->wcl_next;
 1525                 w_child_cnt--;
 1526                 witness_child_free(wcl);
 1527         }
 1528 
 1529         /*
 1530          * Detach from various lists and free.
 1531          */
 1532         STAILQ_REMOVE(list, w, witness, w_typelist);
 1533         STAILQ_REMOVE(&w_all, w, witness, w_list);
 1534         witness_free(w);
 1535 
 1536         return (1);
 1537 }
 1538 
 1539 /*
 1540  * Add "child" as a direct child of "parent".  Returns false if
 1541  * we fail due to out of memory.
 1542  */
 1543 static int
 1544 insertchild(struct witness *parent, struct witness *child)
 1545 {
 1546         struct witness_child_list_entry **wcl;
 1547 
 1548         MPASS(child != NULL && parent != NULL);
 1549 
 1550         /*
 1551          * Insert "child" after "parent"
 1552          */
 1553         wcl = &parent->w_children;
 1554         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
 1555                 wcl = &(*wcl)->wcl_next;
 1556         if (*wcl == NULL) {
 1557                 *wcl = witness_child_get();
 1558                 if (*wcl == NULL)
 1559                         return (0);
 1560                 w_child_cnt++;
 1561         }
 1562         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
 1563 
 1564         return (1);
 1565 }
 1566 
 1567 
 1568 static int
 1569 itismychild(struct witness *parent, struct witness *child)
 1570 {
 1571         struct witness_list *list;
 1572 
 1573         MPASS(child != NULL && parent != NULL);
 1574         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
 1575             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
 1576                 panic(
 1577                 "%s: parent (%s) and child (%s) are not the same lock type",
 1578                     __func__, parent->w_class->lc_name,
 1579                     child->w_class->lc_name);
 1580 
 1581         if (!insertchild(parent, child))
 1582                 return (0);
 1583 
 1584         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
 1585                 list = &w_sleep;
 1586         else
 1587                 list = &w_spin;
 1588         return (1);
 1589 }
 1590 
 1591 static void
 1592 removechild(struct witness *parent, struct witness *child)
 1593 {
 1594         struct witness_child_list_entry **wcl, *wcl1;
 1595         int i;
 1596 
 1597         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
 1598                 for (i = 0; i < (*wcl)->wcl_count; i++)
 1599                         if ((*wcl)->wcl_children[i] == child)
 1600                                 goto found;
 1601         return;
 1602 found:
 1603         (*wcl)->wcl_count--;
 1604         if ((*wcl)->wcl_count > i)
 1605                 (*wcl)->wcl_children[i] =
 1606                     (*wcl)->wcl_children[(*wcl)->wcl_count];
 1607         MPASS((*wcl)->wcl_children[i] != NULL);
 1608         if ((*wcl)->wcl_count != 0)
 1609                 return;
 1610         wcl1 = *wcl;
 1611         *wcl = wcl1->wcl_next;
 1612         w_child_cnt--;
 1613         witness_child_free(wcl1);
 1614 }
 1615 
 1616 static int
 1617 isitmychild(struct witness *parent, struct witness *child)
 1618 {
 1619         struct witness_child_list_entry *wcl;
 1620         int i;
 1621 
 1622         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
 1623                 for (i = 0; i < wcl->wcl_count; i++) {
 1624                         if (wcl->wcl_children[i] == child)
 1625                                 return (1);
 1626                 }
 1627         }
 1628         return (0);
 1629 }
 1630 
 1631 static int
 1632 isitmydescendant(struct witness *parent, struct witness *child)
 1633 {
 1634         struct witness_child_list_entry *wcl;
 1635         int i, j;
 1636 
 1637         if (isitmychild(parent, child))
 1638                 return (1);
 1639         j = 0;
 1640         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
 1641                 MPASS(j < 1000);
 1642                 for (i = 0; i < wcl->wcl_count; i++) {
 1643                         if (isitmydescendant(wcl->wcl_children[i], child))
 1644                                 return (1);
 1645                 }
 1646                 j++;
 1647         }
 1648         return (0);
 1649 }
 1650 
 1651 #ifdef BLESSING
 1652 static int
 1653 blessed(struct witness *w1, struct witness *w2)
 1654 {
 1655         int i;
 1656         struct witness_blessed *b;
 1657 
 1658         for (i = 0; i < blessed_count; i++) {
 1659                 b = &blessed_list[i];
 1660                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
 1661                         if (strcmp(w2->w_name, b->b_lock2) == 0)
 1662                                 return (1);
 1663                         continue;
 1664                 }
 1665                 if (strcmp(w1->w_name, b->b_lock2) == 0)
 1666                         if (strcmp(w2->w_name, b->b_lock1) == 0)
 1667                                 return (1);
 1668         }
 1669         return (0);
 1670 }
 1671 #endif
 1672 
 1673 static struct witness *
 1674 witness_get(void)
 1675 {
 1676         struct witness *w;
 1677 
 1678         if (witness_watch == 0) {
 1679                 mtx_unlock_spin(&w_mtx);
 1680                 return (NULL);
 1681         }
 1682         if (STAILQ_EMPTY(&w_free)) {
 1683                 witness_watch = 0;
 1684                 mtx_unlock_spin(&w_mtx);
 1685                 printf("%s: witness exhausted\n", __func__);
 1686                 return (NULL);
 1687         }
 1688         w = STAILQ_FIRST(&w_free);
 1689         STAILQ_REMOVE_HEAD(&w_free, w_list);
 1690         w_free_cnt--;
 1691         bzero(w, sizeof(*w));
 1692         return (w);
 1693 }
 1694 
 1695 static void
 1696 witness_free(struct witness *w)
 1697 {
 1698 
 1699         STAILQ_INSERT_HEAD(&w_free, w, w_list);
 1700         w_free_cnt++;
 1701 }
 1702 
 1703 static struct witness_child_list_entry *
 1704 witness_child_get(void)
 1705 {
 1706         struct witness_child_list_entry *wcl;
 1707 
 1708         if (witness_watch == 0) {
 1709                 mtx_unlock_spin(&w_mtx);
 1710                 return (NULL);
 1711         }
 1712         wcl = w_child_free;
 1713         if (wcl == NULL) {
 1714                 witness_watch = 0;
 1715                 mtx_unlock_spin(&w_mtx);
 1716                 printf("%s: witness exhausted\n", __func__);
 1717                 return (NULL);
 1718         }
 1719         w_child_free = wcl->wcl_next;
 1720         w_child_free_cnt--;
 1721         bzero(wcl, sizeof(*wcl));
 1722         return (wcl);
 1723 }
 1724 
 1725 static void
 1726 witness_child_free(struct witness_child_list_entry *wcl)
 1727 {
 1728 
 1729         wcl->wcl_next = w_child_free;
 1730         w_child_free = wcl;
 1731         w_child_free_cnt++;
 1732 }
 1733 
 1734 static struct lock_list_entry *
 1735 witness_lock_list_get(void)
 1736 {
 1737         struct lock_list_entry *lle;
 1738 
 1739         if (witness_watch == 0)
 1740                 return (NULL);
 1741         mtx_lock_spin(&w_mtx);
 1742         lle = w_lock_list_free;
 1743         if (lle == NULL) {
 1744                 witness_watch = 0;
 1745                 mtx_unlock_spin(&w_mtx);
 1746                 printf("%s: witness exhausted\n", __func__);
 1747                 return (NULL);
 1748         }
 1749         w_lock_list_free = lle->ll_next;
 1750         mtx_unlock_spin(&w_mtx);
 1751         bzero(lle, sizeof(*lle));
 1752         return (lle);
 1753 }
 1754                 
 1755 static void
 1756 witness_lock_list_free(struct lock_list_entry *lle)
 1757 {
 1758 
 1759         mtx_lock_spin(&w_mtx);
 1760         lle->ll_next = w_lock_list_free;
 1761         w_lock_list_free = lle;
 1762         mtx_unlock_spin(&w_mtx);
 1763 }
 1764 
 1765 static struct lock_instance *
 1766 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
 1767 {
 1768         struct lock_list_entry *lle;
 1769         struct lock_instance *instance;
 1770         int i;
 1771 
 1772         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
 1773                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1774                         instance = &lle->ll_children[i];
 1775                         if (instance->li_lock == lock)
 1776                                 return (instance);
 1777                 }
 1778         return (NULL);
 1779 }
 1780 
 1781 static void
 1782 witness_list_lock(struct lock_instance *instance)
 1783 {
 1784         struct lock_object *lock;
 1785 
 1786         lock = instance->li_lock;
 1787         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
 1788             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
 1789         if (lock->lo_type != lock->lo_name)
 1790                 printf(" (%s)", lock->lo_type);
 1791         printf(" r = %d (%p) locked @ %s:%d\n",
 1792             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
 1793             instance->li_line);
 1794 }
 1795 
 1796 #ifdef DDB
 1797 static int
 1798 witness_thread_has_locks(struct thread *td)
 1799 {
 1800 
 1801         return (td->td_sleeplocks != NULL);
 1802 }
 1803 
 1804 static int
 1805 witness_proc_has_locks(struct proc *p)
 1806 {
 1807         struct thread *td;
 1808 
 1809         FOREACH_THREAD_IN_PROC(p, td) {
 1810                 if (witness_thread_has_locks(td))
 1811                         return (1);
 1812         }
 1813         return (0);
 1814 }
 1815 #endif
 1816 
 1817 int
 1818 witness_list_locks(struct lock_list_entry **lock_list)
 1819 {
 1820         struct lock_list_entry *lle;
 1821         int i, nheld;
 1822 
 1823         nheld = 0;
 1824         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
 1825                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1826                         witness_list_lock(&lle->ll_children[i]);
 1827                         nheld++;
 1828                 }
 1829         return (nheld);
 1830 }
 1831 
 1832 /*
 1833  * This is a bit risky at best.  We call this function when we have timed
 1834  * out acquiring a spin lock, and we assume that the other CPU is stuck
 1835  * with this lock held.  So, we go groveling around in the other CPU's
 1836  * per-cpu data to try to find the lock instance for this spin lock to
 1837  * see when it was last acquired.
 1838  */
 1839 void
 1840 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
 1841 {
 1842         struct lock_instance *instance;
 1843         struct pcpu *pc;
 1844 
 1845         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
 1846                 return;
 1847         pc = pcpu_find(owner->td_oncpu);
 1848         instance = find_instance(pc->pc_spinlocks, lock);
 1849         if (instance != NULL)
 1850                 witness_list_lock(instance);
 1851 }
 1852 
 1853 void
 1854 witness_save(struct lock_object *lock, const char **filep, int *linep)
 1855 {
 1856         struct lock_list_entry *lock_list;
 1857         struct lock_instance *instance;
 1858         struct lock_class *class;
 1859 
 1860         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1861         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1862                 return;
 1863         class = LOCK_CLASS(lock);
 1864         if (class->lc_flags & LC_SLEEPLOCK)
 1865                 lock_list = curthread->td_sleeplocks;
 1866         else {
 1867                 if (witness_skipspin)
 1868                         return;
 1869                 lock_list = PCPU_GET(spinlocks);
 1870         }
 1871         instance = find_instance(lock_list, lock);
 1872         if (instance == NULL)
 1873                 panic("%s: lock (%s) %s not locked", __func__,
 1874                     class->lc_name, lock->lo_name);
 1875         *filep = instance->li_file;
 1876         *linep = instance->li_line;
 1877 }
 1878 
 1879 void
 1880 witness_restore(struct lock_object *lock, const char *file, int line)
 1881 {
 1882         struct lock_list_entry *lock_list;
 1883         struct lock_instance *instance;
 1884         struct lock_class *class;
 1885 
 1886         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1887         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1888                 return;
 1889         class = LOCK_CLASS(lock);
 1890         if (class->lc_flags & LC_SLEEPLOCK)
 1891                 lock_list = curthread->td_sleeplocks;
 1892         else {
 1893                 if (witness_skipspin)
 1894                         return;
 1895                 lock_list = PCPU_GET(spinlocks);
 1896         }
 1897         instance = find_instance(lock_list, lock);
 1898         if (instance == NULL)
 1899                 panic("%s: lock (%s) %s not locked", __func__,
 1900                     class->lc_name, lock->lo_name);
 1901         lock->lo_witness->w_file = file;
 1902         lock->lo_witness->w_line = line;
 1903         instance->li_file = file;
 1904         instance->li_line = line;
 1905 }
 1906 
 1907 void
 1908 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
 1909 {
 1910 #ifdef INVARIANT_SUPPORT
 1911         struct lock_instance *instance;
 1912         struct lock_class *class;
 1913 
 1914         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1915                 return;
 1916         class = LOCK_CLASS(lock);
 1917         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
 1918                 instance = find_instance(curthread->td_sleeplocks, lock);
 1919         else if ((class->lc_flags & LC_SPINLOCK) != 0)
 1920                 instance = find_instance(PCPU_GET(spinlocks), lock);
 1921         else {
 1922                 panic("Lock (%s) %s is not sleep or spin!",
 1923                     class->lc_name, lock->lo_name);
 1924         }
 1925         file = fixup_filename(file);
 1926         switch (flags) {
 1927         case LA_UNLOCKED:
 1928                 if (instance != NULL)
 1929                         panic("Lock (%s) %s locked @ %s:%d.",
 1930                             class->lc_name, lock->lo_name, file, line);
 1931                 break;
 1932         case LA_LOCKED:
 1933         case LA_LOCKED | LA_RECURSED:
 1934         case LA_LOCKED | LA_NOTRECURSED:
 1935         case LA_SLOCKED:
 1936         case LA_SLOCKED | LA_RECURSED:
 1937         case LA_SLOCKED | LA_NOTRECURSED:
 1938         case LA_XLOCKED:
 1939         case LA_XLOCKED | LA_RECURSED:
 1940         case LA_XLOCKED | LA_NOTRECURSED:
 1941                 if (instance == NULL) {
 1942                         panic("Lock (%s) %s not locked @ %s:%d.",
 1943                             class->lc_name, lock->lo_name, file, line);
 1944                         break;
 1945                 }
 1946                 if ((flags & LA_XLOCKED) != 0 &&
 1947                     (instance->li_flags & LI_EXCLUSIVE) == 0)
 1948                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
 1949                             class->lc_name, lock->lo_name, file, line);
 1950                 if ((flags & LA_SLOCKED) != 0 &&
 1951                     (instance->li_flags & LI_EXCLUSIVE) != 0)
 1952                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
 1953                             class->lc_name, lock->lo_name, file, line);
 1954                 if ((flags & LA_RECURSED) != 0 &&
 1955                     (instance->li_flags & LI_RECURSEMASK) == 0)
 1956                         panic("Lock (%s) %s not recursed @ %s:%d.",
 1957                             class->lc_name, lock->lo_name, file, line);
 1958                 if ((flags & LA_NOTRECURSED) != 0 &&
 1959                     (instance->li_flags & LI_RECURSEMASK) != 0)
 1960                         panic("Lock (%s) %s recursed @ %s:%d.",
 1961                             class->lc_name, lock->lo_name, file, line);
 1962                 break;
 1963         default:
 1964                 panic("Invalid lock assertion at %s:%d.", file, line);
 1965 
 1966         }
 1967 #endif  /* INVARIANT_SUPPORT */
 1968 }
 1969 
 1970 #ifdef DDB
 1971 static void
 1972 witness_list(struct thread *td)
 1973 {
 1974 
 1975         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1976         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
 1977 
 1978         if (witness_watch == 0)
 1979                 return;
 1980 
 1981         witness_list_locks(&td->td_sleeplocks);
 1982 
 1983         /*
 1984          * We only handle spinlocks if td == curthread.  This is somewhat broken
 1985          * if td is currently executing on some other CPU and holds spin locks
 1986          * as we won't display those locks.  If we had a MI way of getting
 1987          * the per-cpu data for a given cpu then we could use
 1988          * td->td_oncpu to get the list of spinlocks for this thread
 1989          * and "fix" this.
 1990          *
 1991          * That still wouldn't really fix this unless we locked the scheduler
 1992          * lock or stopped the other CPU to make sure it wasn't changing the
 1993          * list out from under us.  It is probably best to just not try to
 1994          * handle threads on other CPU's for now.
 1995          */
 1996         if (td == curthread && PCPU_GET(spinlocks) != NULL)
 1997                 witness_list_locks(PCPU_PTR(spinlocks));
 1998 }
 1999 
 2000 DB_SHOW_COMMAND(locks, db_witness_list)
 2001 {
 2002         struct thread *td;
 2003 
 2004         if (have_addr)
 2005                 td = db_lookup_thread(addr, TRUE);
 2006         else
 2007                 td = kdb_thread;
 2008         witness_list(td);
 2009 }
 2010 
 2011 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
 2012 {
 2013         struct thread *td;
 2014         struct proc *p;
 2015 
 2016         /*
 2017          * It would be nice to list only threads and processes that actually
 2018          * held sleep locks, but that information is currently not exported
 2019          * by WITNESS.
 2020          */
 2021         FOREACH_PROC_IN_SYSTEM(p) {
 2022                 if (!witness_proc_has_locks(p))
 2023                         continue;
 2024                 FOREACH_THREAD_IN_PROC(p, td) {
 2025                         if (!witness_thread_has_locks(td))
 2026                                 continue;
 2027                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
 2028                             p->p_comm, td, td->td_tid);
 2029                         witness_list(td);
 2030                 }
 2031         }
 2032 }
 2033 
 2034 DB_SHOW_COMMAND(witness, db_witness_display)
 2035 {
 2036 
 2037         witness_display(db_printf);
 2038 }
 2039 #endif

Cache object: cb0c7c9f887344933c919b25b2609dd5


[ 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.