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("$FreeBSD$");
   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          * Netgraph
  353          */
  354         { "ng_node", &lock_class_mtx_sleep },
  355         { "ng_worklist", &lock_class_mtx_sleep },
  356         { NULL, NULL },
  357         /*
  358          * CDEV
  359          */
  360         { "system map", &lock_class_mtx_sleep },
  361         { "vm page queue mutex", &lock_class_mtx_sleep },
  362         { "vnode interlock", &lock_class_mtx_sleep },
  363         { "cdev", &lock_class_mtx_sleep },
  364         { NULL, NULL },
  365         /*
  366          * spin locks
  367          */
  368 #ifdef SMP
  369         { "ap boot", &lock_class_mtx_spin },
  370 #endif
  371         { "rm.mutex_mtx", &lock_class_mtx_spin },
  372         { "sio", &lock_class_mtx_spin },
  373 #ifdef __i386__
  374         { "cy", &lock_class_mtx_spin },
  375 #endif
  376         { "uart_hwmtx", &lock_class_mtx_spin },
  377         { "sabtty", &lock_class_mtx_spin },
  378         { "zstty", &lock_class_mtx_spin },
  379         { "fast_taskqueue", &lock_class_mtx_spin },
  380         { "intr table", &lock_class_mtx_spin },
  381         { "sleepq chain", &lock_class_mtx_spin },
  382         { "sched lock", &lock_class_mtx_spin },
  383         { "turnstile chain", &lock_class_mtx_spin },
  384         { "td_contested", &lock_class_mtx_spin },
  385         { "callout", &lock_class_mtx_spin },
  386         { "entropy harvest mutex", &lock_class_mtx_spin },
  387         /*
  388          * leaf locks
  389          */
  390         { "allpmaps", &lock_class_mtx_spin },
  391         { "vm page queue free mutex", &lock_class_mtx_spin },
  392         { "icu", &lock_class_mtx_spin },
  393 #ifdef SMP
  394         { "smp rendezvous", &lock_class_mtx_spin },
  395 #if defined(__i386__) || defined(__amd64__)
  396         { "tlb", &lock_class_mtx_spin },
  397 #endif
  398 #ifdef __sparc64__
  399         { "ipi", &lock_class_mtx_spin },
  400         { "rtc_mtx", &lock_class_mtx_spin },
  401 #endif
  402 #endif
  403         { "clk", &lock_class_mtx_spin },
  404         { "mutex profiling lock", &lock_class_mtx_spin },
  405         { "kse zombie lock", &lock_class_mtx_spin },
  406         { "ALD Queue", &lock_class_mtx_spin },
  407 #ifdef __ia64__
  408         { "MCA spin lock", &lock_class_mtx_spin },
  409 #endif
  410 #if defined(__i386__) || defined(__amd64__)
  411         { "pcicfg", &lock_class_mtx_spin },
  412         { "NDIS thread lock", &lock_class_mtx_spin },
  413 #endif
  414         { "tw_osl_io_lock", &lock_class_mtx_spin },
  415         { "tw_osl_q_lock", &lock_class_mtx_spin },
  416         { "tw_cl_io_lock", &lock_class_mtx_spin },
  417         { "tw_cl_intr_lock", &lock_class_mtx_spin },
  418         { "tw_cl_gen_lock", &lock_class_mtx_spin },
  419         { NULL, NULL },
  420         { NULL, NULL }
  421 };
  422 
  423 #ifdef BLESSING
  424 /*
  425  * Pairs of locks which have been blessed
  426  * Don't complain about order problems with blessed locks
  427  */
  428 static struct witness_blessed blessed_list[] = {
  429 };
  430 static int blessed_count =
  431         sizeof(blessed_list) / sizeof(struct witness_blessed);
  432 #endif
  433 
  434 /*
  435  * List of all locks in the system.
  436  */
  437 TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks);
  438 
  439 static struct mtx all_mtx = {
  440         { &lock_class_mtx_sleep,        /* mtx_object.lo_class */
  441           "All locks list",             /* mtx_object.lo_name */
  442           "All locks list",             /* mtx_object.lo_type */
  443           LO_INITIALIZED,               /* mtx_object.lo_flags */
  444           { NULL, NULL },               /* mtx_object.lo_list */
  445           NULL },                       /* mtx_object.lo_witness */
  446         MTX_UNOWNED, 0                  /* mtx_lock, mtx_recurse */
  447 };
  448 
  449 /*
  450  * This global is set to 0 once it becomes safe to use the witness code.
  451  */
  452 static int witness_cold = 1;
  453 
  454 /*
  455  * This global is set to 1 once the static lock orders have been enrolled
  456  * so that a warning can be issued for any spin locks enrolled later.
  457  */
  458 static int witness_spin_warn = 0;
  459 
  460 /*
  461  * Global variables for book keeping.
  462  */
  463 static int lock_cur_cnt;
  464 static int lock_max_cnt;
  465 
  466 /*
  467  * The WITNESS-enabled diagnostic code.
  468  */
  469 static void
  470 witness_initialize(void *dummy __unused)
  471 {
  472         struct lock_object *lock;
  473         struct witness_order_list_entry *order;
  474         struct witness *w, *w1;
  475         int i;
  476 
  477         /*
  478          * We have to release Giant before initializing its witness
  479          * structure so that WITNESS doesn't get confused.
  480          */
  481         mtx_unlock(&Giant);
  482         mtx_assert(&Giant, MA_NOTOWNED);
  483 
  484         CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
  485         TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list);
  486         mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
  487             MTX_NOWITNESS);
  488         for (i = 0; i < WITNESS_COUNT; i++)
  489                 witness_free(&w_data[i]);
  490         for (i = 0; i < WITNESS_CHILDCOUNT; i++)
  491                 witness_child_free(&w_childdata[i]);
  492         for (i = 0; i < LOCK_CHILDCOUNT; i++)
  493                 witness_lock_list_free(&w_locklistdata[i]);
  494 
  495         /* First add in all the specified order lists. */
  496         for (order = order_lists; order->w_name != NULL; order++) {
  497                 w = enroll(order->w_name, order->w_class);
  498                 if (w == NULL)
  499                         continue;
  500                 w->w_file = "order list";
  501                 for (order++; order->w_name != NULL; order++) {
  502                         w1 = enroll(order->w_name, order->w_class);
  503                         if (w1 == NULL)
  504                                 continue;
  505                         w1->w_file = "order list";
  506                         if (!itismychild(w, w1))
  507                                 panic("Not enough memory for static orders!");
  508                         w = w1;
  509                 }
  510         }
  511         witness_spin_warn = 1;
  512 
  513         /* Iterate through all locks and add them to witness. */
  514         mtx_lock(&all_mtx);
  515         TAILQ_FOREACH(lock, &all_locks, lo_list) {
  516                 if (lock->lo_flags & LO_WITNESS)
  517                         lock->lo_witness = enroll(lock->lo_type,
  518                             LOCK_CLASS(lock));
  519                 else
  520                         lock->lo_witness = NULL;
  521         }
  522         mtx_unlock(&all_mtx);
  523 
  524         /* Mark the witness code as being ready for use. */
  525         witness_cold = 0;
  526 
  527         mtx_lock(&Giant);
  528 }
  529 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL)
  530 
  531 static int
  532 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
  533 {
  534         int error, value;
  535 
  536         value = witness_watch;
  537         error = sysctl_handle_int(oidp, &value, 0, req);
  538         if (error != 0 || req->newptr == NULL)
  539                 return (error);
  540         error = suser(req->td);
  541         if (error != 0)
  542                 return (error);
  543         if (value == witness_watch)
  544                 return (0);
  545         if (value != 0)
  546                 return (EINVAL);
  547         witness_watch = 0;
  548         return (0);
  549 }
  550 
  551 void
  552 witness_init(struct lock_object *lock)
  553 {
  554         struct lock_class *class;
  555 
  556         class = LOCK_CLASS(lock);
  557         if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
  558             (class->lc_flags & LC_RECURSABLE) == 0)
  559                 panic("%s: lock (%s) %s can not be recursable", __func__,
  560                     class->lc_name, lock->lo_name);
  561         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  562             (class->lc_flags & LC_SLEEPABLE) == 0)
  563                 panic("%s: lock (%s) %s can not be sleepable", __func__,
  564                     class->lc_name, lock->lo_name);
  565         if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
  566             (class->lc_flags & LC_UPGRADABLE) == 0)
  567                 panic("%s: lock (%s) %s can not be upgradable", __func__,
  568                     class->lc_name, lock->lo_name);
  569 
  570         mtx_lock(&all_mtx);
  571         TAILQ_INSERT_TAIL(&all_locks, lock, lo_list);
  572         lock_cur_cnt++;
  573         if (lock_cur_cnt > lock_max_cnt)
  574                 lock_max_cnt = lock_cur_cnt;
  575         mtx_unlock(&all_mtx);
  576         if (!witness_cold && witness_watch != 0 && panicstr == NULL &&
  577             (lock->lo_flags & LO_WITNESS) != 0)
  578                 lock->lo_witness = enroll(lock->lo_type, class);
  579         else
  580                 lock->lo_witness = NULL;
  581 }
  582 
  583 void
  584 witness_destroy(struct lock_object *lock)
  585 {
  586         struct lock_class *class;
  587         struct witness *w;
  588 
  589         class = LOCK_CLASS(lock);
  590         if (witness_cold)
  591                 panic("lock (%s) %s destroyed while witness_cold",
  592                     class->lc_name, lock->lo_name);
  593 
  594         /* XXX: need to verify that no one holds the lock */
  595         w = lock->lo_witness;
  596         if (w != NULL) {
  597                 mtx_lock_spin(&w_mtx);
  598                 MPASS(w->w_refcount > 0);
  599                 w->w_refcount--;
  600 
  601                 /*
  602                  * Lock is already released if we have an allocation failure
  603                  * and depart() fails.
  604                  */
  605                 if (w->w_refcount != 0 || depart(w))
  606                         mtx_unlock_spin(&w_mtx);
  607         }
  608 
  609         mtx_lock(&all_mtx);
  610         lock_cur_cnt--;
  611         TAILQ_REMOVE(&all_locks, lock, lo_list);
  612         mtx_unlock(&all_mtx);
  613 }
  614 
  615 #ifdef DDB
  616 static void
  617 witness_levelall (void)
  618 {
  619         struct witness_list *list;
  620         struct witness *w, *w1;
  621 
  622         /*
  623          * First clear all levels.
  624          */
  625         STAILQ_FOREACH(w, &w_all, w_list) {
  626                 w->w_level = 0;
  627         }
  628 
  629         /*
  630          * Look for locks with no parent and level all their descendants.
  631          */
  632         STAILQ_FOREACH(w, &w_all, w_list) {
  633                 /*
  634                  * This is just an optimization, technically we could get
  635                  * away just walking the all list each time.
  636                  */
  637                 if (w->w_class->lc_flags & LC_SLEEPLOCK)
  638                         list = &w_sleep;
  639                 else
  640                         list = &w_spin;
  641                 STAILQ_FOREACH(w1, list, w_typelist) {
  642                         if (isitmychild(w1, w))
  643                                 goto skip;
  644                 }
  645                 witness_leveldescendents(w, 0);
  646         skip:
  647                 ;       /* silence GCC 3.x */
  648         }
  649 }
  650 
  651 static void
  652 witness_leveldescendents(struct witness *parent, int level)
  653 {
  654         struct witness_child_list_entry *wcl;
  655         int i;
  656 
  657         if (parent->w_level < level)
  658                 parent->w_level = level;
  659         level++;
  660         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
  661                 for (i = 0; i < wcl->wcl_count; i++)
  662                         witness_leveldescendents(wcl->wcl_children[i], level);
  663 }
  664 
  665 static void
  666 witness_displaydescendants(void(*prnt)(const char *fmt, ...),
  667                            struct witness *parent, int indent)
  668 {
  669         struct witness_child_list_entry *wcl;
  670         int i, level;
  671 
  672         level = parent->w_level;
  673         prnt("%-2d", level);
  674         for (i = 0; i < indent; i++)
  675                 prnt(" ");
  676         if (parent->w_refcount > 0)
  677                 prnt("%s", parent->w_name);
  678         else
  679                 prnt("(dead)");
  680         if (parent->w_displayed) {
  681                 prnt(" -- (already displayed)\n");
  682                 return;
  683         }
  684         parent->w_displayed = 1;
  685         if (parent->w_refcount > 0) {
  686                 if (parent->w_file != NULL)
  687                         prnt(" -- last acquired @ %s:%d", parent->w_file,
  688                             parent->w_line);
  689         }
  690         prnt("\n");
  691         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next)
  692                 for (i = 0; i < wcl->wcl_count; i++)
  693                             witness_displaydescendants(prnt,
  694                                 wcl->wcl_children[i], indent + 1);
  695 }
  696 
  697 static void
  698 witness_display_list(void(*prnt)(const char *fmt, ...),
  699                      struct witness_list *list)
  700 {
  701         struct witness *w;
  702 
  703         STAILQ_FOREACH(w, list, w_typelist) {
  704                 if (w->w_file == NULL || w->w_level > 0)
  705                         continue;
  706                 /*
  707                  * This lock has no anscestors, display its descendants. 
  708                  */
  709                 witness_displaydescendants(prnt, w, 0);
  710         }
  711 }
  712         
  713 static void
  714 witness_display(void(*prnt)(const char *fmt, ...))
  715 {
  716         struct witness *w;
  717 
  718         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
  719         witness_levelall();
  720 
  721         /* Clear all the displayed flags. */
  722         STAILQ_FOREACH(w, &w_all, w_list) {
  723                 w->w_displayed = 0;
  724         }
  725 
  726         /*
  727          * First, handle sleep locks which have been acquired at least
  728          * once.
  729          */
  730         prnt("Sleep locks:\n");
  731         witness_display_list(prnt, &w_sleep);
  732         
  733         /*
  734          * Now do spin locks which have been acquired at least once.
  735          */
  736         prnt("\nSpin locks:\n");
  737         witness_display_list(prnt, &w_spin);
  738         
  739         /*
  740          * Finally, any locks which have not been acquired yet.
  741          */
  742         prnt("\nLocks which were never acquired:\n");
  743         STAILQ_FOREACH(w, &w_all, w_list) {
  744                 if (w->w_file != NULL || w->w_refcount == 0)
  745                         continue;
  746                 prnt("%s\n", w->w_name);
  747         }
  748 }
  749 #endif /* DDB */
  750 
  751 /* Trim useless garbage from filenames. */
  752 static const char *
  753 fixup_filename(const char *file)
  754 {
  755 
  756         if (file == NULL)
  757                 return (NULL);
  758         while (strncmp(file, "../", 3) == 0)
  759                 file += 3;
  760         return (file);
  761 }
  762 
  763 int
  764 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
  765 {
  766 
  767         if (witness_watch == 0 || panicstr != NULL)
  768                 return (0);
  769 
  770         /* Require locks that witness knows about. */
  771         if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
  772             lock2->lo_witness == NULL)
  773                 return (EINVAL);
  774 
  775         MPASS(!mtx_owned(&w_mtx));
  776         mtx_lock_spin(&w_mtx);
  777 
  778         /*
  779          * If we already have either an explicit or implied lock order that
  780          * is the other way around, then return an error.
  781          */
  782         if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
  783                 mtx_unlock_spin(&w_mtx);
  784                 return (EDOOFUS);
  785         }
  786         
  787         /* Try to add the new order. */
  788         CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
  789             lock2->lo_type, lock1->lo_type);
  790         if (!itismychild(lock1->lo_witness, lock2->lo_witness))
  791                 return (ENOMEM);
  792         mtx_unlock_spin(&w_mtx);
  793         return (0);
  794 }
  795 
  796 void
  797 witness_checkorder(struct lock_object *lock, int flags, const char *file,
  798     int line)
  799 {
  800         struct lock_list_entry **lock_list, *lle;
  801         struct lock_instance *lock1, *lock2;
  802         struct lock_class *class;
  803         struct witness *w, *w1;
  804         struct thread *td;
  805         int i, j;
  806 
  807         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
  808             panicstr != NULL)
  809                 return;
  810 
  811         /*
  812          * Try locks do not block if they fail to acquire the lock, thus
  813          * there is no danger of deadlocks or of switching while holding a
  814          * spin lock if we acquire a lock via a try operation.  This
  815          * function shouldn't even be called for try locks, so panic if
  816          * that happens.
  817          */
  818         if (flags & LOP_TRYLOCK)
  819                 panic("%s should not be called for try lock operations",
  820                     __func__);
  821 
  822         w = lock->lo_witness;
  823         class = LOCK_CLASS(lock);
  824         td = curthread;
  825         file = fixup_filename(file);
  826 
  827         if (class->lc_flags & LC_SLEEPLOCK) {
  828                 /*
  829                  * Since spin locks include a critical section, this check
  830                  * implicitly enforces a lock order of all sleep locks before
  831                  * all spin locks.
  832                  */
  833                 if (td->td_critnest != 0 && !kdb_active)
  834                         panic("blockable sleep lock (%s) %s @ %s:%d",
  835                             class->lc_name, lock->lo_name, file, line);
  836 
  837                 /*
  838                  * If this is the first lock acquired then just return as
  839                  * no order checking is needed.
  840                  */
  841                 if (td->td_sleeplocks == NULL)
  842                         return;
  843                 lock_list = &td->td_sleeplocks;
  844         } else {
  845                 /*
  846                  * If this is the first lock, just return as no order
  847                  * checking is needed.  We check this in both if clauses
  848                  * here as unifying the check would require us to use a
  849                  * critical section to ensure we don't migrate while doing
  850                  * the check.  Note that if this is not the first lock, we
  851                  * are already in a critical section and are safe for the
  852                  * rest of the check.
  853                  */
  854                 if (PCPU_GET(spinlocks) == NULL)
  855                         return;
  856                 lock_list = PCPU_PTR(spinlocks);
  857         }
  858 
  859         /*
  860          * Check to see if we are recursing on a lock we already own.  If
  861          * so, make sure that we don't mismatch exclusive and shared lock
  862          * acquires.
  863          */
  864         lock1 = find_instance(*lock_list, lock);
  865         if (lock1 != NULL) {
  866                 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
  867                     (flags & LOP_EXCLUSIVE) == 0) {
  868                         printf("shared lock of (%s) %s @ %s:%d\n",
  869                             class->lc_name, lock->lo_name, file, line);
  870                         printf("while exclusively locked from %s:%d\n",
  871                             lock1->li_file, lock1->li_line);
  872                         panic("share->excl");
  873                 }
  874                 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
  875                     (flags & LOP_EXCLUSIVE) != 0) {
  876                         printf("exclusive lock of (%s) %s @ %s:%d\n",
  877                             class->lc_name, lock->lo_name, file, line);
  878                         printf("while share locked from %s:%d\n",
  879                             lock1->li_file, lock1->li_line);
  880                         panic("excl->share");
  881                 }
  882                 return;
  883         }
  884 
  885         /*
  886          * Try locks do not block if they fail to acquire the lock, thus
  887          * there is no danger of deadlocks or of switching while holding a
  888          * spin lock if we acquire a lock via a try operation.
  889          */
  890         if (flags & LOP_TRYLOCK)
  891                 return;
  892 
  893         /*
  894          * Check for duplicate locks of the same type.  Note that we only
  895          * have to check for this on the last lock we just acquired.  Any
  896          * other cases will be caught as lock order violations.
  897          */
  898         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
  899         w1 = lock1->li_lock->lo_witness;
  900         if (w1 == w) {
  901                 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) ||
  902                     (flags & LOP_DUPOK))
  903                         return;
  904                 w->w_same_squawked = 1;
  905                 printf("acquiring duplicate lock of same type: \"%s\"\n", 
  906                         lock->lo_type);
  907                 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name,
  908                     lock1->li_file, lock1->li_line);
  909                 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
  910 #ifdef KDB
  911                 goto debugger;
  912 #else
  913                 return;
  914 #endif
  915         }
  916         MPASS(!mtx_owned(&w_mtx));
  917         mtx_lock_spin(&w_mtx);
  918         /*
  919          * If we know that the the lock we are acquiring comes after
  920          * the lock we most recently acquired in the lock order tree,
  921          * then there is no need for any further checks.
  922          */
  923         if (isitmychild(w1, w)) {
  924                 mtx_unlock_spin(&w_mtx);
  925                 return;
  926         }
  927         for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) {
  928                 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
  929 
  930                         MPASS(j < WITNESS_COUNT);
  931                         lock1 = &lle->ll_children[i];
  932                         w1 = lock1->li_lock->lo_witness;
  933 
  934                         /*
  935                          * If this lock doesn't undergo witness checking,
  936                          * then skip it.
  937                          */
  938                         if (w1 == NULL) {
  939                                 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
  940                                     ("lock missing witness structure"));
  941                                 continue;
  942                         }
  943                         /*
  944                          * If we are locking Giant and this is a sleepable
  945                          * lock, then skip it.
  946                          */
  947                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
  948                             lock == &Giant.mtx_object)
  949                                 continue;
  950                         /*
  951                          * If we are locking a sleepable lock and this lock
  952                          * is Giant, then skip it.
  953                          */
  954                         if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  955                             lock1->li_lock == &Giant.mtx_object)
  956                                 continue;
  957                         /*
  958                          * If we are locking a sleepable lock and this lock
  959                          * isn't sleepable, we want to treat it as a lock
  960                          * order violation to enfore a general lock order of
  961                          * sleepable locks before non-sleepable locks.
  962                          */
  963                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
  964                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
  965                                 goto reversal;
  966                         /*
  967                          * If we are locking Giant and this is a non-sleepable
  968                          * lock, then treat it as a reversal.
  969                          */
  970                         if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
  971                             lock == &Giant.mtx_object)
  972                                 goto reversal;
  973                         /*
  974                          * Check the lock order hierarchy for a reveresal.
  975                          */
  976                         if (!isitmydescendant(w, w1))
  977                                 continue;
  978                 reversal:
  979                         /*
  980                          * We have a lock order violation, check to see if it
  981                          * is allowed or has already been yelled about.
  982                          */
  983                         mtx_unlock_spin(&w_mtx);
  984 #ifdef BLESSING
  985                         /*
  986                          * If the lock order is blessed, just bail.  We don't
  987                          * look for other lock order violations though, which
  988                          * may be a bug.
  989                          */
  990                         if (blessed(w, w1))
  991                                 return;
  992 #endif
  993                         if (lock1->li_lock == &Giant.mtx_object) {
  994                                 if (w1->w_Giant_squawked)
  995                                         return;
  996                                 else
  997                                         w1->w_Giant_squawked = 1;
  998                         } else {
  999                                 if (w1->w_other_squawked)
 1000                                         return;
 1001                                 else
 1002                                         w1->w_other_squawked = 1;
 1003                         }
 1004                         /*
 1005                          * Ok, yell about it.
 1006                          */
 1007                         if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
 1008                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
 1009                                 printf(
 1010                 "lock order reversal: (sleepable after non-sleepable)\n");
 1011                         else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
 1012                             && lock == &Giant.mtx_object)
 1013                                 printf(
 1014                 "lock order reversal: (Giant after non-sleepable)\n");
 1015                         else
 1016                                 printf("lock order reversal:\n");
 1017                         /*
 1018                          * Try to locate an earlier lock with
 1019                          * witness w in our list.
 1020                          */
 1021                         do {
 1022                                 lock2 = &lle->ll_children[i];
 1023                                 MPASS(lock2->li_lock != NULL);
 1024                                 if (lock2->li_lock->lo_witness == w)
 1025                                         break;
 1026                                 if (i == 0 && lle->ll_next != NULL) {
 1027                                         lle = lle->ll_next;
 1028                                         i = lle->ll_count - 1;
 1029                                         MPASS(i >= 0 && i < LOCK_NCHILDREN);
 1030                                 } else
 1031                                         i--;
 1032                         } while (i >= 0);
 1033                         if (i < 0) {
 1034                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 1035                                     lock1->li_lock, lock1->li_lock->lo_name,
 1036                                     lock1->li_lock->lo_type, lock1->li_file,
 1037                                     lock1->li_line);
 1038                                 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
 1039                                     lock->lo_name, lock->lo_type, file, line);
 1040                         } else {
 1041                                 printf(" 1st %p %s (%s) @ %s:%d\n",
 1042                                     lock2->li_lock, lock2->li_lock->lo_name,
 1043                                     lock2->li_lock->lo_type, lock2->li_file,
 1044                                     lock2->li_line);
 1045                                 printf(" 2nd %p %s (%s) @ %s:%d\n",
 1046                                     lock1->li_lock, lock1->li_lock->lo_name,
 1047                                     lock1->li_lock->lo_type, lock1->li_file,
 1048                                     lock1->li_line);
 1049                                 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
 1050                                     lock->lo_name, lock->lo_type, file, line);
 1051                         }
 1052 #ifdef KDB
 1053                         goto debugger;
 1054 #else
 1055                         return;
 1056 #endif
 1057                 }
 1058         }
 1059         lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1];
 1060         /*
 1061          * If requested, build a new lock order.  However, don't build a new
 1062          * relationship between a sleepable lock and Giant if it is in the
 1063          * wrong direction.  The correct lock order is that sleepable locks
 1064          * always come before Giant.
 1065          */
 1066         if (flags & LOP_NEWORDER &&
 1067             !(lock1->li_lock == &Giant.mtx_object &&
 1068             (lock->lo_flags & LO_SLEEPABLE) != 0)) {
 1069                 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
 1070                     lock->lo_type, lock1->li_lock->lo_type);
 1071                 if (!itismychild(lock1->li_lock->lo_witness, w))
 1072                         /* Witness is dead. */
 1073                         return;
 1074         } 
 1075         mtx_unlock_spin(&w_mtx);
 1076         return;
 1077 
 1078 #ifdef KDB
 1079 debugger:
 1080         if (witness_trace)
 1081                 kdb_backtrace();
 1082         if (witness_kdb)
 1083                 kdb_enter(__func__);
 1084 #endif
 1085 }
 1086 
 1087 void
 1088 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
 1089 {
 1090         struct lock_list_entry **lock_list, *lle;
 1091         struct lock_instance *instance;
 1092         struct witness *w;
 1093         struct thread *td;
 1094 
 1095         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 1096             panicstr != NULL)
 1097                 return;
 1098         w = lock->lo_witness;
 1099         td = curthread;
 1100         file = fixup_filename(file);
 1101 
 1102         /* Determine lock list for this lock. */
 1103         if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
 1104                 lock_list = &td->td_sleeplocks;
 1105         else
 1106                 lock_list = PCPU_PTR(spinlocks);
 1107 
 1108         /* Check to see if we are recursing on a lock we already own. */
 1109         instance = find_instance(*lock_list, lock);
 1110         if (instance != NULL) {
 1111                 instance->li_flags++;
 1112                 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
 1113                     td->td_proc->p_pid, lock->lo_name,
 1114                     instance->li_flags & LI_RECURSEMASK);
 1115                 instance->li_file = file;
 1116                 instance->li_line = line;
 1117                 return;
 1118         }
 1119 
 1120         /* Update per-witness last file and line acquire. */
 1121         w->w_file = file;
 1122         w->w_line = line;
 1123 
 1124         /* Find the next open lock instance in the list and fill it. */
 1125         lle = *lock_list;
 1126         if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
 1127                 lle = witness_lock_list_get();
 1128                 if (lle == NULL)
 1129                         return;
 1130                 lle->ll_next = *lock_list;
 1131                 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
 1132                     td->td_proc->p_pid, lle);
 1133                 *lock_list = lle;
 1134         }
 1135         instance = &lle->ll_children[lle->ll_count++];
 1136         instance->li_lock = lock;
 1137         instance->li_line = line;
 1138         instance->li_file = file;
 1139         if ((flags & LOP_EXCLUSIVE) != 0)
 1140                 instance->li_flags = LI_EXCLUSIVE;
 1141         else
 1142                 instance->li_flags = 0;
 1143         CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
 1144             td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
 1145 }
 1146 
 1147 void
 1148 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
 1149 {
 1150         struct lock_instance *instance;
 1151         struct lock_class *class;
 1152 
 1153         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1154         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1155                 return;
 1156         class = LOCK_CLASS(lock);
 1157         file = fixup_filename(file);
 1158         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1159                 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
 1160                     class->lc_name, lock->lo_name, file, line);
 1161         if ((flags & LOP_TRYLOCK) == 0)
 1162                 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name,
 1163                     lock->lo_name, file, line);
 1164         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
 1165                 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
 1166                     class->lc_name, lock->lo_name, file, line);
 1167         instance = find_instance(curthread->td_sleeplocks, lock);
 1168         if (instance == NULL)
 1169                 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
 1170                     class->lc_name, lock->lo_name, file, line);
 1171         if ((instance->li_flags & LI_EXCLUSIVE) != 0)
 1172                 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
 1173                     class->lc_name, lock->lo_name, file, line);
 1174         if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1175                 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1176                     class->lc_name, lock->lo_name,
 1177                     instance->li_flags & LI_RECURSEMASK, file, line);
 1178         instance->li_flags |= LI_EXCLUSIVE;
 1179 }
 1180 
 1181 void
 1182 witness_downgrade(struct lock_object *lock, int flags, const char *file,
 1183     int line)
 1184 {
 1185         struct lock_instance *instance;
 1186         struct lock_class *class;
 1187 
 1188         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1189         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1190                 return;
 1191         class = LOCK_CLASS(lock);
 1192         file = fixup_filename(file);
 1193         if ((lock->lo_flags & LO_UPGRADABLE) == 0)
 1194                 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
 1195                     class->lc_name, lock->lo_name, file, line);
 1196         if ((class->lc_flags & LC_SLEEPLOCK) == 0)
 1197                 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
 1198                     class->lc_name, lock->lo_name, file, line);
 1199         instance = find_instance(curthread->td_sleeplocks, lock);
 1200         if (instance == NULL)
 1201                 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
 1202                     class->lc_name, lock->lo_name, file, line);
 1203         if ((instance->li_flags & LI_EXCLUSIVE) == 0)
 1204                 panic("downgrade of shared lock (%s) %s @ %s:%d",
 1205                     class->lc_name, lock->lo_name, file, line);
 1206         if ((instance->li_flags & LI_RECURSEMASK) != 0)
 1207                 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
 1208                     class->lc_name, lock->lo_name,
 1209                     instance->li_flags & LI_RECURSEMASK, file, line);
 1210         instance->li_flags &= ~LI_EXCLUSIVE;
 1211 }
 1212 
 1213 void
 1214 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
 1215 {
 1216         struct lock_list_entry **lock_list, *lle;
 1217         struct lock_instance *instance;
 1218         struct lock_class *class;
 1219         struct thread *td;
 1220         register_t s;
 1221         int i, j;
 1222 
 1223         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL ||
 1224             panicstr != NULL)
 1225                 return;
 1226         td = curthread;
 1227         class = LOCK_CLASS(lock);
 1228         file = fixup_filename(file);
 1229 
 1230         /* Find lock instance associated with this lock. */
 1231         if (class->lc_flags & LC_SLEEPLOCK)
 1232                 lock_list = &td->td_sleeplocks;
 1233         else
 1234                 lock_list = PCPU_PTR(spinlocks);
 1235         for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
 1236                 for (i = 0; i < (*lock_list)->ll_count; i++) {
 1237                         instance = &(*lock_list)->ll_children[i];
 1238                         if (instance->li_lock == lock)
 1239                                 goto found;
 1240                 }
 1241         panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name,
 1242             file, line);
 1243 found:
 1244 
 1245         /* First, check for shared/exclusive mismatches. */
 1246         if ((instance->li_flags & LI_EXCLUSIVE) != 0 &&
 1247             (flags & LOP_EXCLUSIVE) == 0) {
 1248                 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1249                     lock->lo_name, file, line);
 1250                 printf("while exclusively locked from %s:%d\n",
 1251                     instance->li_file, instance->li_line);
 1252                 panic("excl->ushare");
 1253         }
 1254         if ((instance->li_flags & LI_EXCLUSIVE) == 0 &&
 1255             (flags & LOP_EXCLUSIVE) != 0) {
 1256                 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
 1257                     lock->lo_name, file, line);
 1258                 printf("while share locked from %s:%d\n", instance->li_file,
 1259                     instance->li_line);
 1260                 panic("share->uexcl");
 1261         }
 1262 
 1263         /* If we are recursed, unrecurse. */
 1264         if ((instance->li_flags & LI_RECURSEMASK) > 0) {
 1265                 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
 1266                     td->td_proc->p_pid, instance->li_lock->lo_name,
 1267                     instance->li_flags);
 1268                 instance->li_flags--;
 1269                 return;
 1270         }
 1271 
 1272         /* Otherwise, remove this item from the list. */
 1273         s = intr_disable();
 1274         CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
 1275             td->td_proc->p_pid, instance->li_lock->lo_name,
 1276             (*lock_list)->ll_count - 1);
 1277         for (j = i; j < (*lock_list)->ll_count - 1; j++)
 1278                 (*lock_list)->ll_children[j] =
 1279                     (*lock_list)->ll_children[j + 1];
 1280         (*lock_list)->ll_count--;
 1281         intr_restore(s);
 1282 
 1283         /* If this lock list entry is now empty, free it. */
 1284         if ((*lock_list)->ll_count == 0) {
 1285                 lle = *lock_list;
 1286                 *lock_list = lle->ll_next;
 1287                 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
 1288                     td->td_proc->p_pid, lle);
 1289                 witness_lock_list_free(lle);
 1290         }
 1291 }
 1292 
 1293 /*
 1294  * Warn if any locks other than 'lock' are held.  Flags can be passed in to
 1295  * exempt Giant and sleepable locks from the checks as well.  If any
 1296  * non-exempt locks are held, then a supplied message is printed to the
 1297  * console along with a list of the offending locks.  If indicated in the
 1298  * flags then a failure results in a panic as well.
 1299  */
 1300 int
 1301 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
 1302 {
 1303         struct lock_list_entry *lle;
 1304         struct lock_instance *lock1;
 1305         struct thread *td;
 1306         va_list ap;
 1307         int i, n;
 1308 
 1309         if (witness_cold || witness_watch == 0 || panicstr != NULL)
 1310                 return (0);
 1311         n = 0;
 1312         td = curthread;
 1313         for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
 1314                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1315                         lock1 = &lle->ll_children[i];
 1316                         if (lock1->li_lock == lock)
 1317                                 continue;
 1318                         if (flags & WARN_GIANTOK &&
 1319                             lock1->li_lock == &Giant.mtx_object)
 1320                                 continue;
 1321                         if (flags & WARN_SLEEPOK &&
 1322                             (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
 1323                                 continue;
 1324                         if (n == 0) {
 1325                                 va_start(ap, fmt);
 1326                                 vprintf(fmt, ap);
 1327                                 va_end(ap);
 1328                                 printf(" with the following");
 1329                                 if (flags & WARN_SLEEPOK)
 1330                                         printf(" non-sleepable");
 1331                                 printf(" locks held:\n");
 1332                         }
 1333                         n++;
 1334                         witness_list_lock(lock1);
 1335                 }
 1336         if (PCPU_GET(spinlocks) != NULL) {
 1337                 /*
 1338                  * Since we already hold a spinlock preemption is
 1339                  * already blocked.
 1340                  */
 1341                 if (n == 0) {
 1342                         va_start(ap, fmt);
 1343                         vprintf(fmt, ap);
 1344                         va_end(ap);
 1345                         printf(" with the following");
 1346                         if (flags & WARN_SLEEPOK)
 1347                                 printf(" non-sleepable");
 1348                         printf(" locks held:\n");
 1349                 }
 1350                 n += witness_list_locks(PCPU_PTR(spinlocks));
 1351         }
 1352         if (flags & WARN_PANIC && n)
 1353                 panic("witness_warn");
 1354 #ifdef KDB
 1355         else if (witness_kdb && n)
 1356                 kdb_enter(__func__);
 1357         else if (witness_trace && n)
 1358                 kdb_backtrace();
 1359 #endif
 1360         return (n);
 1361 }
 1362 
 1363 const char *
 1364 witness_file(struct lock_object *lock)
 1365 {
 1366         struct witness *w;
 1367 
 1368         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
 1369                 return ("?");
 1370         w = lock->lo_witness;
 1371         return (w->w_file);
 1372 }
 1373 
 1374 int
 1375 witness_line(struct lock_object *lock)
 1376 {
 1377         struct witness *w;
 1378 
 1379         if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL)
 1380                 return (0);
 1381         w = lock->lo_witness;
 1382         return (w->w_line);
 1383 }
 1384 
 1385 static struct witness *
 1386 enroll(const char *description, struct lock_class *lock_class)
 1387 {
 1388         struct witness *w;
 1389 
 1390         if (witness_watch == 0 || panicstr != NULL)
 1391                 return (NULL);
 1392         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin)
 1393                 return (NULL);
 1394         mtx_lock_spin(&w_mtx);
 1395         STAILQ_FOREACH(w, &w_all, w_list) {
 1396                 if (w->w_name == description || (w->w_refcount > 0 &&
 1397                     strcmp(description, w->w_name) == 0)) {
 1398                         w->w_refcount++;
 1399                         mtx_unlock_spin(&w_mtx);
 1400                         if (lock_class != w->w_class)
 1401                                 panic(
 1402                                 "lock (%s) %s does not match earlier (%s) lock",
 1403                                     description, lock_class->lc_name,
 1404                                     w->w_class->lc_name);
 1405                         return (w);
 1406                 }
 1407         }
 1408         if ((w = witness_get()) == NULL)
 1409                 goto out;
 1410         w->w_name = description;
 1411         w->w_class = lock_class;
 1412         w->w_refcount = 1;
 1413         STAILQ_INSERT_HEAD(&w_all, w, w_list);
 1414         if (lock_class->lc_flags & LC_SPINLOCK) {
 1415                 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
 1416                 w_spin_cnt++;
 1417         } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
 1418                 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
 1419                 w_sleep_cnt++;
 1420         } else {
 1421                 mtx_unlock_spin(&w_mtx);
 1422                 panic("lock class %s is not sleep or spin",
 1423                     lock_class->lc_name);
 1424         }
 1425         mtx_unlock_spin(&w_mtx);
 1426 out:
 1427         /*
 1428          * We issue a warning for any spin locks not defined in the static
 1429          * order list as a way to discourage their use (folks should really
 1430          * be using non-spin mutexes most of the time).  However, several
 1431          * 3rd part device drivers use spin locks because that is all they
 1432          * have available on Windows and Linux and they think that normal
 1433          * mutexes are insufficient.
 1434          */
 1435         if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn)
 1436                 printf("WITNESS: spin lock %s not in order list\n",
 1437                     description);
 1438         return (w);
 1439 }
 1440 
 1441 /* Don't let the door bang you on the way out... */
 1442 static int
 1443 depart(struct witness *w)
 1444 {
 1445         struct witness_child_list_entry *wcl, *nwcl;
 1446         struct witness_list *list;
 1447         struct witness *parent;
 1448 
 1449         MPASS(w->w_refcount == 0);
 1450         if (w->w_class->lc_flags & LC_SLEEPLOCK) {
 1451                 list = &w_sleep;
 1452                 w_sleep_cnt--;
 1453         } else {
 1454                 list = &w_spin;
 1455                 w_spin_cnt--;
 1456         }
 1457         /*
 1458          * First, we run through the entire tree looking for any
 1459          * witnesses that the outgoing witness is a child of.  For
 1460          * each parent that we find, we reparent all the direct
 1461          * children of the outgoing witness to its parent.
 1462          */
 1463         STAILQ_FOREACH(parent, list, w_typelist) {
 1464                 if (!isitmychild(parent, w))
 1465                         continue;
 1466                 removechild(parent, w);
 1467         }
 1468 
 1469         /*
 1470          * Now we go through and free up the child list of the
 1471          * outgoing witness.
 1472          */
 1473         for (wcl = w->w_children; wcl != NULL; wcl = nwcl) {
 1474                 nwcl = wcl->wcl_next;
 1475                 w_child_cnt--;
 1476                 witness_child_free(wcl);
 1477         }
 1478 
 1479         /*
 1480          * Detach from various lists and free.
 1481          */
 1482         STAILQ_REMOVE(list, w, witness, w_typelist);
 1483         STAILQ_REMOVE(&w_all, w, witness, w_list);
 1484         witness_free(w);
 1485 
 1486         return (1);
 1487 }
 1488 
 1489 /*
 1490  * Add "child" as a direct child of "parent".  Returns false if
 1491  * we fail due to out of memory.
 1492  */
 1493 static int
 1494 insertchild(struct witness *parent, struct witness *child)
 1495 {
 1496         struct witness_child_list_entry **wcl;
 1497 
 1498         MPASS(child != NULL && parent != NULL);
 1499 
 1500         /*
 1501          * Insert "child" after "parent"
 1502          */
 1503         wcl = &parent->w_children;
 1504         while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN)
 1505                 wcl = &(*wcl)->wcl_next;
 1506         if (*wcl == NULL) {
 1507                 *wcl = witness_child_get();
 1508                 if (*wcl == NULL)
 1509                         return (0);
 1510                 w_child_cnt++;
 1511         }
 1512         (*wcl)->wcl_children[(*wcl)->wcl_count++] = child;
 1513 
 1514         return (1);
 1515 }
 1516 
 1517 
 1518 static int
 1519 itismychild(struct witness *parent, struct witness *child)
 1520 {
 1521         struct witness_list *list;
 1522 
 1523         MPASS(child != NULL && parent != NULL);
 1524         if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) !=
 1525             (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)))
 1526                 panic(
 1527                 "%s: parent (%s) and child (%s) are not the same lock type",
 1528                     __func__, parent->w_class->lc_name,
 1529                     child->w_class->lc_name);
 1530 
 1531         if (!insertchild(parent, child))
 1532                 return (0);
 1533 
 1534         if (parent->w_class->lc_flags & LC_SLEEPLOCK)
 1535                 list = &w_sleep;
 1536         else
 1537                 list = &w_spin;
 1538         return (1);
 1539 }
 1540 
 1541 static void
 1542 removechild(struct witness *parent, struct witness *child)
 1543 {
 1544         struct witness_child_list_entry **wcl, *wcl1;
 1545         int i;
 1546 
 1547         for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next)
 1548                 for (i = 0; i < (*wcl)->wcl_count; i++)
 1549                         if ((*wcl)->wcl_children[i] == child)
 1550                                 goto found;
 1551         return;
 1552 found:
 1553         (*wcl)->wcl_count--;
 1554         if ((*wcl)->wcl_count > i)
 1555                 (*wcl)->wcl_children[i] =
 1556                     (*wcl)->wcl_children[(*wcl)->wcl_count];
 1557         MPASS((*wcl)->wcl_children[i] != NULL);
 1558         if ((*wcl)->wcl_count != 0)
 1559                 return;
 1560         wcl1 = *wcl;
 1561         *wcl = wcl1->wcl_next;
 1562         w_child_cnt--;
 1563         witness_child_free(wcl1);
 1564 }
 1565 
 1566 static int
 1567 isitmychild(struct witness *parent, struct witness *child)
 1568 {
 1569         struct witness_child_list_entry *wcl;
 1570         int i;
 1571 
 1572         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
 1573                 for (i = 0; i < wcl->wcl_count; i++) {
 1574                         if (wcl->wcl_children[i] == child)
 1575                                 return (1);
 1576                 }
 1577         }
 1578         return (0);
 1579 }
 1580 
 1581 static int
 1582 isitmydescendant(struct witness *parent, struct witness *child)
 1583 {
 1584         struct witness_child_list_entry *wcl;
 1585         int i, j;
 1586 
 1587         if (isitmychild(parent, child))
 1588                 return (1);
 1589         j = 0;
 1590         for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) {
 1591                 MPASS(j < 1000);
 1592                 for (i = 0; i < wcl->wcl_count; i++) {
 1593                         if (isitmydescendant(wcl->wcl_children[i], child))
 1594                                 return (1);
 1595                 }
 1596                 j++;
 1597         }
 1598         return (0);
 1599 }
 1600 
 1601 #ifdef BLESSING
 1602 static int
 1603 blessed(struct witness *w1, struct witness *w2)
 1604 {
 1605         int i;
 1606         struct witness_blessed *b;
 1607 
 1608         for (i = 0; i < blessed_count; i++) {
 1609                 b = &blessed_list[i];
 1610                 if (strcmp(w1->w_name, b->b_lock1) == 0) {
 1611                         if (strcmp(w2->w_name, b->b_lock2) == 0)
 1612                                 return (1);
 1613                         continue;
 1614                 }
 1615                 if (strcmp(w1->w_name, b->b_lock2) == 0)
 1616                         if (strcmp(w2->w_name, b->b_lock1) == 0)
 1617                                 return (1);
 1618         }
 1619         return (0);
 1620 }
 1621 #endif
 1622 
 1623 static struct witness *
 1624 witness_get(void)
 1625 {
 1626         struct witness *w;
 1627 
 1628         if (witness_watch == 0) {
 1629                 mtx_unlock_spin(&w_mtx);
 1630                 return (NULL);
 1631         }
 1632         if (STAILQ_EMPTY(&w_free)) {
 1633                 witness_watch = 0;
 1634                 mtx_unlock_spin(&w_mtx);
 1635                 printf("%s: witness exhausted\n", __func__);
 1636                 return (NULL);
 1637         }
 1638         w = STAILQ_FIRST(&w_free);
 1639         STAILQ_REMOVE_HEAD(&w_free, w_list);
 1640         w_free_cnt--;
 1641         bzero(w, sizeof(*w));
 1642         return (w);
 1643 }
 1644 
 1645 static void
 1646 witness_free(struct witness *w)
 1647 {
 1648 
 1649         STAILQ_INSERT_HEAD(&w_free, w, w_list);
 1650         w_free_cnt++;
 1651 }
 1652 
 1653 static struct witness_child_list_entry *
 1654 witness_child_get(void)
 1655 {
 1656         struct witness_child_list_entry *wcl;
 1657 
 1658         if (witness_watch == 0) {
 1659                 mtx_unlock_spin(&w_mtx);
 1660                 return (NULL);
 1661         }
 1662         wcl = w_child_free;
 1663         if (wcl == NULL) {
 1664                 witness_watch = 0;
 1665                 mtx_unlock_spin(&w_mtx);
 1666                 printf("%s: witness exhausted\n", __func__);
 1667                 return (NULL);
 1668         }
 1669         w_child_free = wcl->wcl_next;
 1670         w_child_free_cnt--;
 1671         bzero(wcl, sizeof(*wcl));
 1672         return (wcl);
 1673 }
 1674 
 1675 static void
 1676 witness_child_free(struct witness_child_list_entry *wcl)
 1677 {
 1678 
 1679         wcl->wcl_next = w_child_free;
 1680         w_child_free = wcl;
 1681         w_child_free_cnt++;
 1682 }
 1683 
 1684 static struct lock_list_entry *
 1685 witness_lock_list_get(void)
 1686 {
 1687         struct lock_list_entry *lle;
 1688 
 1689         if (witness_watch == 0)
 1690                 return (NULL);
 1691         mtx_lock_spin(&w_mtx);
 1692         lle = w_lock_list_free;
 1693         if (lle == NULL) {
 1694                 witness_watch = 0;
 1695                 mtx_unlock_spin(&w_mtx);
 1696                 printf("%s: witness exhausted\n", __func__);
 1697                 return (NULL);
 1698         }
 1699         w_lock_list_free = lle->ll_next;
 1700         mtx_unlock_spin(&w_mtx);
 1701         bzero(lle, sizeof(*lle));
 1702         return (lle);
 1703 }
 1704                 
 1705 static void
 1706 witness_lock_list_free(struct lock_list_entry *lle)
 1707 {
 1708 
 1709         mtx_lock_spin(&w_mtx);
 1710         lle->ll_next = w_lock_list_free;
 1711         w_lock_list_free = lle;
 1712         mtx_unlock_spin(&w_mtx);
 1713 }
 1714 
 1715 static struct lock_instance *
 1716 find_instance(struct lock_list_entry *lock_list, struct lock_object *lock)
 1717 {
 1718         struct lock_list_entry *lle;
 1719         struct lock_instance *instance;
 1720         int i;
 1721 
 1722         for (lle = lock_list; lle != NULL; lle = lle->ll_next)
 1723                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1724                         instance = &lle->ll_children[i];
 1725                         if (instance->li_lock == lock)
 1726                                 return (instance);
 1727                 }
 1728         return (NULL);
 1729 }
 1730 
 1731 static void
 1732 witness_list_lock(struct lock_instance *instance)
 1733 {
 1734         struct lock_object *lock;
 1735 
 1736         lock = instance->li_lock;
 1737         printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
 1738             "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
 1739         if (lock->lo_type != lock->lo_name)
 1740                 printf(" (%s)", lock->lo_type);
 1741         printf(" r = %d (%p) locked @ %s:%d\n",
 1742             instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
 1743             instance->li_line);
 1744 }
 1745 
 1746 #ifdef DDB
 1747 static int
 1748 witness_thread_has_locks(struct thread *td)
 1749 {
 1750 
 1751         return (td->td_sleeplocks != NULL);
 1752 }
 1753 
 1754 static int
 1755 witness_proc_has_locks(struct proc *p)
 1756 {
 1757         struct thread *td;
 1758 
 1759         FOREACH_THREAD_IN_PROC(p, td) {
 1760                 if (witness_thread_has_locks(td))
 1761                         return (1);
 1762         }
 1763         return (0);
 1764 }
 1765 #endif
 1766 
 1767 int
 1768 witness_list_locks(struct lock_list_entry **lock_list)
 1769 {
 1770         struct lock_list_entry *lle;
 1771         int i, nheld;
 1772 
 1773         nheld = 0;
 1774         for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
 1775                 for (i = lle->ll_count - 1; i >= 0; i--) {
 1776                         witness_list_lock(&lle->ll_children[i]);
 1777                         nheld++;
 1778                 }
 1779         return (nheld);
 1780 }
 1781 
 1782 /*
 1783  * This is a bit risky at best.  We call this function when we have timed
 1784  * out acquiring a spin lock, and we assume that the other CPU is stuck
 1785  * with this lock held.  So, we go groveling around in the other CPU's
 1786  * per-cpu data to try to find the lock instance for this spin lock to
 1787  * see when it was last acquired.
 1788  */
 1789 void
 1790 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
 1791 {
 1792         struct lock_instance *instance;
 1793         struct pcpu *pc;
 1794 
 1795         if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
 1796                 return;
 1797         pc = pcpu_find(owner->td_oncpu);
 1798         instance = find_instance(pc->pc_spinlocks, lock);
 1799         if (instance != NULL)
 1800                 witness_list_lock(instance);
 1801 }
 1802 
 1803 void
 1804 witness_save(struct lock_object *lock, const char **filep, int *linep)
 1805 {
 1806         struct lock_list_entry *lock_list;
 1807         struct lock_instance *instance;
 1808         struct lock_class *class;
 1809 
 1810         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1811         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1812                 return;
 1813         class = LOCK_CLASS(lock);
 1814         if (class->lc_flags & LC_SLEEPLOCK)
 1815                 lock_list = curthread->td_sleeplocks;
 1816         else {
 1817                 if (witness_skipspin)
 1818                         return;
 1819                 lock_list = PCPU_GET(spinlocks);
 1820         }
 1821         instance = find_instance(lock_list, lock);
 1822         if (instance == NULL)
 1823                 panic("%s: lock (%s) %s not locked", __func__,
 1824                     class->lc_name, lock->lo_name);
 1825         *filep = instance->li_file;
 1826         *linep = instance->li_line;
 1827 }
 1828 
 1829 void
 1830 witness_restore(struct lock_object *lock, const char *file, int line)
 1831 {
 1832         struct lock_list_entry *lock_list;
 1833         struct lock_instance *instance;
 1834         struct lock_class *class;
 1835 
 1836         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1837         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1838                 return;
 1839         class = LOCK_CLASS(lock);
 1840         if (class->lc_flags & LC_SLEEPLOCK)
 1841                 lock_list = curthread->td_sleeplocks;
 1842         else {
 1843                 if (witness_skipspin)
 1844                         return;
 1845                 lock_list = PCPU_GET(spinlocks);
 1846         }
 1847         instance = find_instance(lock_list, lock);
 1848         if (instance == NULL)
 1849                 panic("%s: lock (%s) %s not locked", __func__,
 1850                     class->lc_name, lock->lo_name);
 1851         lock->lo_witness->w_file = file;
 1852         lock->lo_witness->w_line = line;
 1853         instance->li_file = file;
 1854         instance->li_line = line;
 1855 }
 1856 
 1857 void
 1858 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
 1859 {
 1860 #ifdef INVARIANT_SUPPORT
 1861         struct lock_instance *instance;
 1862         struct lock_class *class;
 1863 
 1864         if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL)
 1865                 return;
 1866         class = LOCK_CLASS(lock);
 1867         if ((class->lc_flags & LC_SLEEPLOCK) != 0)
 1868                 instance = find_instance(curthread->td_sleeplocks, lock);
 1869         else if ((class->lc_flags & LC_SPINLOCK) != 0)
 1870                 instance = find_instance(PCPU_GET(spinlocks), lock);
 1871         else {
 1872                 panic("Lock (%s) %s is not sleep or spin!",
 1873                     class->lc_name, lock->lo_name);
 1874         }
 1875         file = fixup_filename(file);
 1876         switch (flags) {
 1877         case LA_UNLOCKED:
 1878                 if (instance != NULL)
 1879                         panic("Lock (%s) %s locked @ %s:%d.",
 1880                             class->lc_name, lock->lo_name, file, line);
 1881                 break;
 1882         case LA_LOCKED:
 1883         case LA_LOCKED | LA_RECURSED:
 1884         case LA_LOCKED | LA_NOTRECURSED:
 1885         case LA_SLOCKED:
 1886         case LA_SLOCKED | LA_RECURSED:
 1887         case LA_SLOCKED | LA_NOTRECURSED:
 1888         case LA_XLOCKED:
 1889         case LA_XLOCKED | LA_RECURSED:
 1890         case LA_XLOCKED | LA_NOTRECURSED:
 1891                 if (instance == NULL) {
 1892                         panic("Lock (%s) %s not locked @ %s:%d.",
 1893                             class->lc_name, lock->lo_name, file, line);
 1894                         break;
 1895                 }
 1896                 if ((flags & LA_XLOCKED) != 0 &&
 1897                     (instance->li_flags & LI_EXCLUSIVE) == 0)
 1898                         panic("Lock (%s) %s not exclusively locked @ %s:%d.",
 1899                             class->lc_name, lock->lo_name, file, line);
 1900                 if ((flags & LA_SLOCKED) != 0 &&
 1901                     (instance->li_flags & LI_EXCLUSIVE) != 0)
 1902                         panic("Lock (%s) %s exclusively locked @ %s:%d.",
 1903                             class->lc_name, lock->lo_name, file, line);
 1904                 if ((flags & LA_RECURSED) != 0 &&
 1905                     (instance->li_flags & LI_RECURSEMASK) == 0)
 1906                         panic("Lock (%s) %s not recursed @ %s:%d.",
 1907                             class->lc_name, lock->lo_name, file, line);
 1908                 if ((flags & LA_NOTRECURSED) != 0 &&
 1909                     (instance->li_flags & LI_RECURSEMASK) != 0)
 1910                         panic("Lock (%s) %s recursed @ %s:%d.",
 1911                             class->lc_name, lock->lo_name, file, line);
 1912                 break;
 1913         default:
 1914                 panic("Invalid lock assertion at %s:%d.", file, line);
 1915 
 1916         }
 1917 #endif  /* INVARIANT_SUPPORT */
 1918 }
 1919 
 1920 #ifdef DDB
 1921 static void
 1922 witness_list(struct thread *td)
 1923 {
 1924 
 1925         KASSERT(!witness_cold, ("%s: witness_cold", __func__));
 1926         KASSERT(kdb_active, ("%s: not in the debugger", __func__));
 1927 
 1928         if (witness_watch == 0)
 1929                 return;
 1930 
 1931         witness_list_locks(&td->td_sleeplocks);
 1932 
 1933         /*
 1934          * We only handle spinlocks if td == curthread.  This is somewhat broken
 1935          * if td is currently executing on some other CPU and holds spin locks
 1936          * as we won't display those locks.  If we had a MI way of getting
 1937          * the per-cpu data for a given cpu then we could use
 1938          * td->td_oncpu to get the list of spinlocks for this thread
 1939          * and "fix" this.
 1940          *
 1941          * That still wouldn't really fix this unless we locked sched_lock
 1942          * or stopped the other CPU to make sure it wasn't changing the list
 1943          * out from under us.  It is probably best to just not try to handle
 1944          * threads on other CPU's for now.
 1945          */
 1946         if (td == curthread && PCPU_GET(spinlocks) != NULL)
 1947                 witness_list_locks(PCPU_PTR(spinlocks));
 1948 }
 1949 
 1950 DB_SHOW_COMMAND(locks, db_witness_list)
 1951 {
 1952         struct thread *td;
 1953         pid_t pid;
 1954         struct proc *p;
 1955 
 1956         if (have_addr) {
 1957                 pid = (addr % 16) + ((addr >> 4) % 16) * 10 +
 1958                     ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 +
 1959                     ((addr >> 16) % 16) * 10000;
 1960                 /* sx_slock(&allproc_lock); */
 1961                 FOREACH_PROC_IN_SYSTEM(p) {
 1962                         if (p->p_pid == pid)
 1963                                 break;
 1964                 }
 1965                 /* sx_sunlock(&allproc_lock); */
 1966                 if (p == NULL) {
 1967                         db_printf("pid %d not found\n", pid);
 1968                         return;
 1969                 }
 1970                 FOREACH_THREAD_IN_PROC(p, td) {
 1971                         witness_list(td);
 1972                 }
 1973         } else {
 1974                 td = curthread;
 1975                 witness_list(td);
 1976         }
 1977 }
 1978 
 1979 DB_SHOW_COMMAND(alllocks, db_witness_list_all)
 1980 {
 1981         struct thread *td;
 1982         struct proc *p;
 1983 
 1984         /*
 1985          * It would be nice to list only threads and processes that actually
 1986          * held sleep locks, but that information is currently not exported
 1987          * by WITNESS.
 1988          */
 1989         FOREACH_PROC_IN_SYSTEM(p) {
 1990                 if (!witness_proc_has_locks(p))
 1991                         continue;
 1992                 FOREACH_THREAD_IN_PROC(p, td) {
 1993                         if (!witness_thread_has_locks(td))
 1994                                 continue;
 1995                         db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
 1996                             p->p_comm, td, td->td_tid);
 1997                         witness_list(td);
 1998                 }
 1999         }
 2000 }
 2001 
 2002 DB_SHOW_COMMAND(witness, db_witness_display)
 2003 {
 2004 
 2005         witness_display(db_printf);
 2006 }
 2007 #endif

Cache object: 4190dc0bece75179cb5ba34091056355


[ 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.