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_sleepqueue.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) 2004 John Baldwin <jhb@FreeBSD.org>
    3  * All rights reserved.
    4  *
    5  * Redistribution and use in source and binary forms, with or without
    6  * modification, are permitted provided that the following conditions
    7  * are met:
    8  * 1. Redistributions of source code must retain the above copyright
    9  *    notice, this list of conditions and the following disclaimer.
   10  * 2. Redistributions in binary form must reproduce the above copyright
   11  *    notice, this list of conditions and the following disclaimer in the
   12  *    documentation and/or other materials provided with the distribution.
   13  *
   14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
   15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
   18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   24  * SUCH DAMAGE.
   25  */
   26 
   27 /*
   28  * Implementation of sleep queues used to hold queue of threads blocked on
   29  * a wait channel.  Sleep queues different from turnstiles in that wait
   30  * channels are not owned by anyone, so there is no priority propagation.
   31  * Sleep queues can also provide a timeout and can also be interrupted by
   32  * signals.  That said, there are several similarities between the turnstile
   33  * and sleep queue implementations.  (Note: turnstiles were implemented
   34  * first.)  For example, both use a hash table of the same size where each
   35  * bucket is referred to as a "chain" that contains both a spin lock and
   36  * a linked list of queues.  An individual queue is located by using a hash
   37  * to pick a chain, locking the chain, and then walking the chain searching
   38  * for the queue.  This means that a wait channel object does not need to
   39  * embed it's queue head just as locks do not embed their turnstile queue
   40  * head.  Threads also carry around a sleep queue that they lend to the
   41  * wait channel when blocking.  Just as in turnstiles, the queue includes
   42  * a free list of the sleep queues of other threads blocked on the same
   43  * wait channel in the case of multiple waiters.
   44  *
   45  * Some additional functionality provided by sleep queues include the
   46  * ability to set a timeout.  The timeout is managed using a per-thread
   47  * callout that resumes a thread if it is asleep.  A thread may also
   48  * catch signals while it is asleep (aka an interruptible sleep).  The
   49  * signal code uses sleepq_abort() to interrupt a sleeping thread.  Finally,
   50  * sleep queues also provide some extra assertions.  One is not allowed to
   51  * mix the sleep/wakeup and cv APIs for a given wait channel.  Also, one
   52  * must consistently use the same lock to synchronize with a wait channel,
   53  * though this check is currently only a warning for sleep/wakeup due to
   54  * pre-existing abuse of that API.  The same lock must also be held when
   55  * awakening threads, though that is currently only enforced for condition
   56  * variables.
   57  */
   58 
   59 #include <sys/cdefs.h>
   60 __FBSDID("$FreeBSD: releng/11.0/sys/kern/subr_sleepqueue.c 302350 2016-07-05 18:47:17Z glebius $");
   61 
   62 #include "opt_sleepqueue_profiling.h"
   63 #include "opt_ddb.h"
   64 #include "opt_sched.h"
   65 #include "opt_stack.h"
   66 
   67 #include <sys/param.h>
   68 #include <sys/systm.h>
   69 #include <sys/lock.h>
   70 #include <sys/kernel.h>
   71 #include <sys/ktr.h>
   72 #include <sys/mutex.h>
   73 #include <sys/proc.h>
   74 #include <sys/sbuf.h>
   75 #include <sys/sched.h>
   76 #include <sys/sdt.h>
   77 #include <sys/signalvar.h>
   78 #include <sys/sleepqueue.h>
   79 #include <sys/stack.h>
   80 #include <sys/sysctl.h>
   81 
   82 #include <vm/uma.h>
   83 
   84 #ifdef DDB
   85 #include <ddb/ddb.h>
   86 #endif
   87 
   88 
   89 /*
   90  * Constants for the hash table of sleep queue chains.
   91  * SC_TABLESIZE must be a power of two for SC_MASK to work properly.
   92  */
   93 #define SC_TABLESIZE    256                     /* Must be power of 2. */
   94 #define SC_MASK         (SC_TABLESIZE - 1)
   95 #define SC_SHIFT        8
   96 #define SC_HASH(wc)     ((((uintptr_t)(wc) >> SC_SHIFT) ^ (uintptr_t)(wc)) & \
   97                             SC_MASK)
   98 #define SC_LOOKUP(wc)   &sleepq_chains[SC_HASH(wc)]
   99 #define NR_SLEEPQS      2
  100 /*
  101  * There two different lists of sleep queues.  Both lists are connected
  102  * via the sq_hash entries.  The first list is the sleep queue chain list
  103  * that a sleep queue is on when it is attached to a wait channel.  The
  104  * second list is the free list hung off of a sleep queue that is attached
  105  * to a wait channel.
  106  *
  107  * Each sleep queue also contains the wait channel it is attached to, the
  108  * list of threads blocked on that wait channel, flags specific to the
  109  * wait channel, and the lock used to synchronize with a wait channel.
  110  * The flags are used to catch mismatches between the various consumers
  111  * of the sleep queue API (e.g. sleep/wakeup and condition variables).
  112  * The lock pointer is only used when invariants are enabled for various
  113  * debugging checks.
  114  *
  115  * Locking key:
  116  *  c - sleep queue chain lock
  117  */
  118 struct sleepqueue {
  119         TAILQ_HEAD(, thread) sq_blocked[NR_SLEEPQS];    /* (c) Blocked threads. */
  120         u_int sq_blockedcnt[NR_SLEEPQS];        /* (c) N. of blocked threads. */
  121         LIST_ENTRY(sleepqueue) sq_hash;         /* (c) Chain and free list. */
  122         LIST_HEAD(, sleepqueue) sq_free;        /* (c) Free queues. */
  123         void    *sq_wchan;                      /* (c) Wait channel. */
  124         int     sq_type;                        /* (c) Queue type. */
  125 #ifdef INVARIANTS
  126         struct lock_object *sq_lock;            /* (c) Associated lock. */
  127 #endif
  128 };
  129 
  130 struct sleepqueue_chain {
  131         LIST_HEAD(, sleepqueue) sc_queues;      /* List of sleep queues. */
  132         struct mtx sc_lock;                     /* Spin lock for this chain. */
  133 #ifdef SLEEPQUEUE_PROFILING
  134         u_int   sc_depth;                       /* Length of sc_queues. */
  135         u_int   sc_max_depth;                   /* Max length of sc_queues. */
  136 #endif
  137 };
  138 
  139 #ifdef SLEEPQUEUE_PROFILING
  140 u_int sleepq_max_depth;
  141 static SYSCTL_NODE(_debug, OID_AUTO, sleepq, CTLFLAG_RD, 0, "sleepq profiling");
  142 static SYSCTL_NODE(_debug_sleepq, OID_AUTO, chains, CTLFLAG_RD, 0,
  143     "sleepq chain stats");
  144 SYSCTL_UINT(_debug_sleepq, OID_AUTO, max_depth, CTLFLAG_RD, &sleepq_max_depth,
  145     0, "maxmimum depth achieved of a single chain");
  146 
  147 static void     sleepq_profile(const char *wmesg);
  148 static int      prof_enabled;
  149 #endif
  150 static struct sleepqueue_chain sleepq_chains[SC_TABLESIZE];
  151 static uma_zone_t sleepq_zone;
  152 
  153 /*
  154  * Prototypes for non-exported routines.
  155  */
  156 static int      sleepq_catch_signals(void *wchan, int pri);
  157 static int      sleepq_check_signals(void);
  158 static int      sleepq_check_timeout(void);
  159 #ifdef INVARIANTS
  160 static void     sleepq_dtor(void *mem, int size, void *arg);
  161 #endif
  162 static int      sleepq_init(void *mem, int size, int flags);
  163 static int      sleepq_resume_thread(struct sleepqueue *sq, struct thread *td,
  164                     int pri);
  165 static void     sleepq_switch(void *wchan, int pri);
  166 static void     sleepq_timeout(void *arg);
  167 
  168 SDT_PROBE_DECLARE(sched, , , sleep);
  169 SDT_PROBE_DECLARE(sched, , , wakeup);
  170 
  171 /*
  172  * Initialize SLEEPQUEUE_PROFILING specific sysctl nodes.
  173  * Note that it must happen after sleepinit() has been fully executed, so
  174  * it must happen after SI_SUB_KMEM SYSINIT() subsystem setup.
  175  */
  176 #ifdef SLEEPQUEUE_PROFILING
  177 static void
  178 init_sleepqueue_profiling(void)
  179 {
  180         char chain_name[10];
  181         struct sysctl_oid *chain_oid;
  182         u_int i;
  183 
  184         for (i = 0; i < SC_TABLESIZE; i++) {
  185                 snprintf(chain_name, sizeof(chain_name), "%u", i);
  186                 chain_oid = SYSCTL_ADD_NODE(NULL, 
  187                     SYSCTL_STATIC_CHILDREN(_debug_sleepq_chains), OID_AUTO,
  188                     chain_name, CTLFLAG_RD, NULL, "sleepq chain stats");
  189                 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
  190                     "depth", CTLFLAG_RD, &sleepq_chains[i].sc_depth, 0, NULL);
  191                 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
  192                     "max_depth", CTLFLAG_RD, &sleepq_chains[i].sc_max_depth, 0,
  193                     NULL);
  194         }
  195 }
  196 
  197 SYSINIT(sleepqueue_profiling, SI_SUB_LOCK, SI_ORDER_ANY,
  198     init_sleepqueue_profiling, NULL);
  199 #endif
  200 
  201 /*
  202  * Early initialization of sleep queues that is called from the sleepinit()
  203  * SYSINIT.
  204  */
  205 void
  206 init_sleepqueues(void)
  207 {
  208         int i;
  209 
  210         for (i = 0; i < SC_TABLESIZE; i++) {
  211                 LIST_INIT(&sleepq_chains[i].sc_queues);
  212                 mtx_init(&sleepq_chains[i].sc_lock, "sleepq chain", NULL,
  213                     MTX_SPIN | MTX_RECURSE);
  214         }
  215         sleepq_zone = uma_zcreate("SLEEPQUEUE", sizeof(struct sleepqueue),
  216 #ifdef INVARIANTS
  217             NULL, sleepq_dtor, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
  218 #else
  219             NULL, NULL, sleepq_init, NULL, UMA_ALIGN_CACHE, 0);
  220 #endif
  221         
  222         thread0.td_sleepqueue = sleepq_alloc();
  223 }
  224 
  225 /*
  226  * Get a sleep queue for a new thread.
  227  */
  228 struct sleepqueue *
  229 sleepq_alloc(void)
  230 {
  231 
  232         return (uma_zalloc(sleepq_zone, M_WAITOK));
  233 }
  234 
  235 /*
  236  * Free a sleep queue when a thread is destroyed.
  237  */
  238 void
  239 sleepq_free(struct sleepqueue *sq)
  240 {
  241 
  242         uma_zfree(sleepq_zone, sq);
  243 }
  244 
  245 /*
  246  * Lock the sleep queue chain associated with the specified wait channel.
  247  */
  248 void
  249 sleepq_lock(void *wchan)
  250 {
  251         struct sleepqueue_chain *sc;
  252 
  253         sc = SC_LOOKUP(wchan);
  254         mtx_lock_spin(&sc->sc_lock);
  255 }
  256 
  257 /*
  258  * Look up the sleep queue associated with a given wait channel in the hash
  259  * table locking the associated sleep queue chain.  If no queue is found in
  260  * the table, NULL is returned.
  261  */
  262 struct sleepqueue *
  263 sleepq_lookup(void *wchan)
  264 {
  265         struct sleepqueue_chain *sc;
  266         struct sleepqueue *sq;
  267 
  268         KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
  269         sc = SC_LOOKUP(wchan);
  270         mtx_assert(&sc->sc_lock, MA_OWNED);
  271         LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
  272                 if (sq->sq_wchan == wchan)
  273                         return (sq);
  274         return (NULL);
  275 }
  276 
  277 /*
  278  * Unlock the sleep queue chain associated with a given wait channel.
  279  */
  280 void
  281 sleepq_release(void *wchan)
  282 {
  283         struct sleepqueue_chain *sc;
  284 
  285         sc = SC_LOOKUP(wchan);
  286         mtx_unlock_spin(&sc->sc_lock);
  287 }
  288 
  289 /*
  290  * Places the current thread on the sleep queue for the specified wait
  291  * channel.  If INVARIANTS is enabled, then it associates the passed in
  292  * lock with the sleepq to make sure it is held when that sleep queue is
  293  * woken up.
  294  */
  295 void
  296 sleepq_add(void *wchan, struct lock_object *lock, const char *wmesg, int flags,
  297     int queue)
  298 {
  299         struct sleepqueue_chain *sc;
  300         struct sleepqueue *sq;
  301         struct thread *td;
  302 
  303         td = curthread;
  304         sc = SC_LOOKUP(wchan);
  305         mtx_assert(&sc->sc_lock, MA_OWNED);
  306         MPASS(td->td_sleepqueue != NULL);
  307         MPASS(wchan != NULL);
  308         MPASS((queue >= 0) && (queue < NR_SLEEPQS));
  309 
  310         /* If this thread is not allowed to sleep, die a horrible death. */
  311         KASSERT(td->td_no_sleeping == 0,
  312             ("%s: td %p to sleep on wchan %p with sleeping prohibited",
  313             __func__, td, wchan));
  314 
  315         /* Look up the sleep queue associated with the wait channel 'wchan'. */
  316         sq = sleepq_lookup(wchan);
  317 
  318         /*
  319          * If the wait channel does not already have a sleep queue, use
  320          * this thread's sleep queue.  Otherwise, insert the current thread
  321          * into the sleep queue already in use by this wait channel.
  322          */
  323         if (sq == NULL) {
  324 #ifdef INVARIANTS
  325                 int i;
  326 
  327                 sq = td->td_sleepqueue;
  328                 for (i = 0; i < NR_SLEEPQS; i++) {
  329                         KASSERT(TAILQ_EMPTY(&sq->sq_blocked[i]),
  330                             ("thread's sleep queue %d is not empty", i));
  331                         KASSERT(sq->sq_blockedcnt[i] == 0,
  332                             ("thread's sleep queue %d count mismatches", i));
  333                 }
  334                 KASSERT(LIST_EMPTY(&sq->sq_free),
  335                     ("thread's sleep queue has a non-empty free list"));
  336                 KASSERT(sq->sq_wchan == NULL, ("stale sq_wchan pointer"));
  337                 sq->sq_lock = lock;
  338 #endif
  339 #ifdef SLEEPQUEUE_PROFILING
  340                 sc->sc_depth++;
  341                 if (sc->sc_depth > sc->sc_max_depth) {
  342                         sc->sc_max_depth = sc->sc_depth;
  343                         if (sc->sc_max_depth > sleepq_max_depth)
  344                                 sleepq_max_depth = sc->sc_max_depth;
  345                 }
  346 #endif
  347                 sq = td->td_sleepqueue;
  348                 LIST_INSERT_HEAD(&sc->sc_queues, sq, sq_hash);
  349                 sq->sq_wchan = wchan;
  350                 sq->sq_type = flags & SLEEPQ_TYPE;
  351         } else {
  352                 MPASS(wchan == sq->sq_wchan);
  353                 MPASS(lock == sq->sq_lock);
  354                 MPASS((flags & SLEEPQ_TYPE) == sq->sq_type);
  355                 LIST_INSERT_HEAD(&sq->sq_free, td->td_sleepqueue, sq_hash);
  356         }
  357         thread_lock(td);
  358         TAILQ_INSERT_TAIL(&sq->sq_blocked[queue], td, td_slpq);
  359         sq->sq_blockedcnt[queue]++;
  360         td->td_sleepqueue = NULL;
  361         td->td_sqqueue = queue;
  362         td->td_wchan = wchan;
  363         td->td_wmesg = wmesg;
  364         if (flags & SLEEPQ_INTERRUPTIBLE) {
  365                 td->td_flags |= TDF_SINTR;
  366                 td->td_flags &= ~TDF_SLEEPABORT;
  367         }
  368         thread_unlock(td);
  369 }
  370 
  371 /*
  372  * Sets a timeout that will remove the current thread from the specified
  373  * sleep queue after timo ticks if the thread has not already been awakened.
  374  */
  375 void
  376 sleepq_set_timeout_sbt(void *wchan, sbintime_t sbt, sbintime_t pr,
  377     int flags)
  378 {
  379         struct sleepqueue_chain *sc;
  380         struct thread *td;
  381 
  382         td = curthread;
  383         sc = SC_LOOKUP(wchan);
  384         mtx_assert(&sc->sc_lock, MA_OWNED);
  385         MPASS(TD_ON_SLEEPQ(td));
  386         MPASS(td->td_sleepqueue == NULL);
  387         MPASS(wchan != NULL);
  388         if (cold)
  389                 panic("timed sleep before timers are working");
  390         callout_reset_sbt_on(&td->td_slpcallout, sbt, pr,
  391             sleepq_timeout, td, PCPU_GET(cpuid), flags | C_DIRECT_EXEC);
  392 }
  393 
  394 /*
  395  * Return the number of actual sleepers for the specified queue.
  396  */
  397 u_int
  398 sleepq_sleepcnt(void *wchan, int queue)
  399 {
  400         struct sleepqueue *sq;
  401 
  402         KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
  403         MPASS((queue >= 0) && (queue < NR_SLEEPQS));
  404         sq = sleepq_lookup(wchan);
  405         if (sq == NULL)
  406                 return (0);
  407         return (sq->sq_blockedcnt[queue]);
  408 }
  409 
  410 /*
  411  * Marks the pending sleep of the current thread as interruptible and
  412  * makes an initial check for pending signals before putting a thread
  413  * to sleep. Enters and exits with the thread lock held.  Thread lock
  414  * may have transitioned from the sleepq lock to a run lock.
  415  */
  416 static int
  417 sleepq_catch_signals(void *wchan, int pri)
  418 {
  419         struct sleepqueue_chain *sc;
  420         struct sleepqueue *sq;
  421         struct thread *td;
  422         struct proc *p;
  423         struct sigacts *ps;
  424         int sig, ret;
  425 
  426         td = curthread;
  427         p = curproc;
  428         sc = SC_LOOKUP(wchan);
  429         mtx_assert(&sc->sc_lock, MA_OWNED);
  430         MPASS(wchan != NULL);
  431         if ((td->td_pflags & TDP_WAKEUP) != 0) {
  432                 td->td_pflags &= ~TDP_WAKEUP;
  433                 ret = EINTR;
  434                 thread_lock(td);
  435                 goto out;
  436         }
  437 
  438         /*
  439          * See if there are any pending signals for this thread.  If not
  440          * we can switch immediately.  Otherwise do the signal processing
  441          * directly.
  442          */
  443         thread_lock(td);
  444         if ((td->td_flags & (TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK)) == 0) {
  445                 sleepq_switch(wchan, pri);
  446                 return (0);
  447         }
  448         thread_unlock(td);
  449         mtx_unlock_spin(&sc->sc_lock);
  450         CTR3(KTR_PROC, "sleepq catching signals: thread %p (pid %ld, %s)",
  451                 (void *)td, (long)p->p_pid, td->td_name);
  452         PROC_LOCK(p);
  453         ps = p->p_sigacts;
  454         mtx_lock(&ps->ps_mtx);
  455         sig = cursig(td);
  456         if (sig == -1) {
  457                 mtx_unlock(&ps->ps_mtx);
  458                 KASSERT((td->td_flags & TDF_SBDRY) != 0, ("lost TDF_SBDRY"));
  459                 KASSERT(TD_SBDRY_INTR(td),
  460                     ("lost TDF_SERESTART of TDF_SEINTR"));
  461                 KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) !=
  462                     (TDF_SEINTR | TDF_SERESTART),
  463                     ("both TDF_SEINTR and TDF_SERESTART"));
  464                 ret = TD_SBDRY_ERRNO(td);
  465         } else if (sig == 0) {
  466                 mtx_unlock(&ps->ps_mtx);
  467                 ret = thread_suspend_check(1);
  468                 MPASS(ret == 0 || ret == EINTR || ret == ERESTART);
  469         } else {
  470                 if (SIGISMEMBER(ps->ps_sigintr, sig))
  471                         ret = EINTR;
  472                 else
  473                         ret = ERESTART;
  474                 mtx_unlock(&ps->ps_mtx);
  475         }
  476         /*
  477          * Lock the per-process spinlock prior to dropping the PROC_LOCK
  478          * to avoid a signal delivery race.  PROC_LOCK, PROC_SLOCK, and
  479          * thread_lock() are currently held in tdsendsignal().
  480          */
  481         PROC_SLOCK(p);
  482         mtx_lock_spin(&sc->sc_lock);
  483         PROC_UNLOCK(p);
  484         thread_lock(td);
  485         PROC_SUNLOCK(p);
  486         if (ret == 0) {
  487                 sleepq_switch(wchan, pri);
  488                 return (0);
  489         }
  490 out:
  491         /*
  492          * There were pending signals and this thread is still
  493          * on the sleep queue, remove it from the sleep queue.
  494          */
  495         if (TD_ON_SLEEPQ(td)) {
  496                 sq = sleepq_lookup(wchan);
  497                 if (sleepq_resume_thread(sq, td, 0)) {
  498 #ifdef INVARIANTS
  499                         /*
  500                          * This thread hasn't gone to sleep yet, so it
  501                          * should not be swapped out.
  502                          */
  503                         panic("not waking up swapper");
  504 #endif
  505                 }
  506         }
  507         mtx_unlock_spin(&sc->sc_lock);
  508         MPASS(td->td_lock != &sc->sc_lock);
  509         return (ret);
  510 }
  511 
  512 /*
  513  * Switches to another thread if we are still asleep on a sleep queue.
  514  * Returns with thread lock.
  515  */
  516 static void
  517 sleepq_switch(void *wchan, int pri)
  518 {
  519         struct sleepqueue_chain *sc;
  520         struct sleepqueue *sq;
  521         struct thread *td;
  522 
  523         td = curthread;
  524         sc = SC_LOOKUP(wchan);
  525         mtx_assert(&sc->sc_lock, MA_OWNED);
  526         THREAD_LOCK_ASSERT(td, MA_OWNED);
  527 
  528         /* 
  529          * If we have a sleep queue, then we've already been woken up, so
  530          * just return.
  531          */
  532         if (td->td_sleepqueue != NULL) {
  533                 mtx_unlock_spin(&sc->sc_lock);
  534                 return;
  535         }
  536 
  537         /*
  538          * If TDF_TIMEOUT is set, then our sleep has been timed out
  539          * already but we are still on the sleep queue, so dequeue the
  540          * thread and return.
  541          */
  542         if (td->td_flags & TDF_TIMEOUT) {
  543                 MPASS(TD_ON_SLEEPQ(td));
  544                 sq = sleepq_lookup(wchan);
  545                 if (sleepq_resume_thread(sq, td, 0)) {
  546 #ifdef INVARIANTS
  547                         /*
  548                          * This thread hasn't gone to sleep yet, so it
  549                          * should not be swapped out.
  550                          */
  551                         panic("not waking up swapper");
  552 #endif
  553                 }
  554                 mtx_unlock_spin(&sc->sc_lock);
  555                 return;         
  556         }
  557 #ifdef SLEEPQUEUE_PROFILING
  558         if (prof_enabled)
  559                 sleepq_profile(td->td_wmesg);
  560 #endif
  561         MPASS(td->td_sleepqueue == NULL);
  562         sched_sleep(td, pri);
  563         thread_lock_set(td, &sc->sc_lock);
  564         SDT_PROBE0(sched, , , sleep);
  565         TD_SET_SLEEPING(td);
  566         mi_switch(SW_VOL | SWT_SLEEPQ, NULL);
  567         KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING"));
  568         CTR3(KTR_PROC, "sleepq resume: thread %p (pid %ld, %s)",
  569             (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
  570 }
  571 
  572 /*
  573  * Check to see if we timed out.
  574  */
  575 static int
  576 sleepq_check_timeout(void)
  577 {
  578         struct thread *td;
  579 
  580         td = curthread;
  581         THREAD_LOCK_ASSERT(td, MA_OWNED);
  582 
  583         /*
  584          * If TDF_TIMEOUT is set, we timed out.
  585          */
  586         if (td->td_flags & TDF_TIMEOUT) {
  587                 td->td_flags &= ~TDF_TIMEOUT;
  588                 return (EWOULDBLOCK);
  589         }
  590 
  591         /*
  592          * If TDF_TIMOFAIL is set, the timeout ran after we had
  593          * already been woken up.
  594          */
  595         if (td->td_flags & TDF_TIMOFAIL)
  596                 td->td_flags &= ~TDF_TIMOFAIL;
  597 
  598         /*
  599          * If callout_stop() fails, then the timeout is running on
  600          * another CPU, so synchronize with it to avoid having it
  601          * accidentally wake up a subsequent sleep.
  602          */
  603         else if (_callout_stop_safe(&td->td_slpcallout, CS_EXECUTING, NULL)
  604             == 0) {
  605                 td->td_flags |= TDF_TIMEOUT;
  606                 TD_SET_SLEEPING(td);
  607                 mi_switch(SW_INVOL | SWT_SLEEPQTIMO, NULL);
  608         }
  609         return (0);
  610 }
  611 
  612 /*
  613  * Check to see if we were awoken by a signal.
  614  */
  615 static int
  616 sleepq_check_signals(void)
  617 {
  618         struct thread *td;
  619 
  620         td = curthread;
  621         THREAD_LOCK_ASSERT(td, MA_OWNED);
  622 
  623         /* We are no longer in an interruptible sleep. */
  624         if (td->td_flags & TDF_SINTR)
  625                 td->td_flags &= ~TDF_SINTR;
  626 
  627         if (td->td_flags & TDF_SLEEPABORT) {
  628                 td->td_flags &= ~TDF_SLEEPABORT;
  629                 return (td->td_intrval);
  630         }
  631 
  632         return (0);
  633 }
  634 
  635 /*
  636  * Block the current thread until it is awakened from its sleep queue.
  637  */
  638 void
  639 sleepq_wait(void *wchan, int pri)
  640 {
  641         struct thread *td;
  642 
  643         td = curthread;
  644         MPASS(!(td->td_flags & TDF_SINTR));
  645         thread_lock(td);
  646         sleepq_switch(wchan, pri);
  647         thread_unlock(td);
  648 }
  649 
  650 /*
  651  * Block the current thread until it is awakened from its sleep queue
  652  * or it is interrupted by a signal.
  653  */
  654 int
  655 sleepq_wait_sig(void *wchan, int pri)
  656 {
  657         int rcatch;
  658         int rval;
  659 
  660         rcatch = sleepq_catch_signals(wchan, pri);
  661         rval = sleepq_check_signals();
  662         thread_unlock(curthread);
  663         if (rcatch)
  664                 return (rcatch);
  665         return (rval);
  666 }
  667 
  668 /*
  669  * Block the current thread until it is awakened from its sleep queue
  670  * or it times out while waiting.
  671  */
  672 int
  673 sleepq_timedwait(void *wchan, int pri)
  674 {
  675         struct thread *td;
  676         int rval;
  677 
  678         td = curthread;
  679         MPASS(!(td->td_flags & TDF_SINTR));
  680         thread_lock(td);
  681         sleepq_switch(wchan, pri);
  682         rval = sleepq_check_timeout();
  683         thread_unlock(td);
  684 
  685         return (rval);
  686 }
  687 
  688 /*
  689  * Block the current thread until it is awakened from its sleep queue,
  690  * it is interrupted by a signal, or it times out waiting to be awakened.
  691  */
  692 int
  693 sleepq_timedwait_sig(void *wchan, int pri)
  694 {
  695         int rcatch, rvalt, rvals;
  696 
  697         rcatch = sleepq_catch_signals(wchan, pri);
  698         rvalt = sleepq_check_timeout();
  699         rvals = sleepq_check_signals();
  700         thread_unlock(curthread);
  701         if (rcatch)
  702                 return (rcatch);
  703         if (rvals)
  704                 return (rvals);
  705         return (rvalt);
  706 }
  707 
  708 /*
  709  * Returns the type of sleepqueue given a waitchannel.
  710  */
  711 int
  712 sleepq_type(void *wchan)
  713 {
  714         struct sleepqueue *sq;
  715         int type;
  716 
  717         MPASS(wchan != NULL);
  718 
  719         sleepq_lock(wchan);
  720         sq = sleepq_lookup(wchan);
  721         if (sq == NULL) {
  722                 sleepq_release(wchan);
  723                 return (-1);
  724         }
  725         type = sq->sq_type;
  726         sleepq_release(wchan);
  727         return (type);
  728 }
  729 
  730 /*
  731  * Removes a thread from a sleep queue and makes it
  732  * runnable.
  733  */
  734 static int
  735 sleepq_resume_thread(struct sleepqueue *sq, struct thread *td, int pri)
  736 {
  737         struct sleepqueue_chain *sc;
  738 
  739         MPASS(td != NULL);
  740         MPASS(sq->sq_wchan != NULL);
  741         MPASS(td->td_wchan == sq->sq_wchan);
  742         MPASS(td->td_sqqueue < NR_SLEEPQS && td->td_sqqueue >= 0);
  743         THREAD_LOCK_ASSERT(td, MA_OWNED);
  744         sc = SC_LOOKUP(sq->sq_wchan);
  745         mtx_assert(&sc->sc_lock, MA_OWNED);
  746 
  747         SDT_PROBE2(sched, , , wakeup, td, td->td_proc);
  748 
  749         /* Remove the thread from the queue. */
  750         sq->sq_blockedcnt[td->td_sqqueue]--;
  751         TAILQ_REMOVE(&sq->sq_blocked[td->td_sqqueue], td, td_slpq);
  752 
  753         /*
  754          * Get a sleep queue for this thread.  If this is the last waiter,
  755          * use the queue itself and take it out of the chain, otherwise,
  756          * remove a queue from the free list.
  757          */
  758         if (LIST_EMPTY(&sq->sq_free)) {
  759                 td->td_sleepqueue = sq;
  760 #ifdef INVARIANTS
  761                 sq->sq_wchan = NULL;
  762 #endif
  763 #ifdef SLEEPQUEUE_PROFILING
  764                 sc->sc_depth--;
  765 #endif
  766         } else
  767                 td->td_sleepqueue = LIST_FIRST(&sq->sq_free);
  768         LIST_REMOVE(td->td_sleepqueue, sq_hash);
  769 
  770         td->td_wmesg = NULL;
  771         td->td_wchan = NULL;
  772         td->td_flags &= ~TDF_SINTR;
  773 
  774         CTR3(KTR_PROC, "sleepq_wakeup: thread %p (pid %ld, %s)",
  775             (void *)td, (long)td->td_proc->p_pid, td->td_name);
  776 
  777         /* Adjust priority if requested. */
  778         MPASS(pri == 0 || (pri >= PRI_MIN && pri <= PRI_MAX));
  779         if (pri != 0 && td->td_priority > pri &&
  780             PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
  781                 sched_prio(td, pri);
  782 
  783         /*
  784          * Note that thread td might not be sleeping if it is running
  785          * sleepq_catch_signals() on another CPU or is blocked on its
  786          * proc lock to check signals.  There's no need to mark the
  787          * thread runnable in that case.
  788          */
  789         if (TD_IS_SLEEPING(td)) {
  790                 TD_CLR_SLEEPING(td);
  791                 return (setrunnable(td));
  792         }
  793         return (0);
  794 }
  795 
  796 #ifdef INVARIANTS
  797 /*
  798  * UMA zone item deallocator.
  799  */
  800 static void
  801 sleepq_dtor(void *mem, int size, void *arg)
  802 {
  803         struct sleepqueue *sq;
  804         int i;
  805 
  806         sq = mem;
  807         for (i = 0; i < NR_SLEEPQS; i++) {
  808                 MPASS(TAILQ_EMPTY(&sq->sq_blocked[i]));
  809                 MPASS(sq->sq_blockedcnt[i] == 0);
  810         }
  811 }
  812 #endif
  813 
  814 /*
  815  * UMA zone item initializer.
  816  */
  817 static int
  818 sleepq_init(void *mem, int size, int flags)
  819 {
  820         struct sleepqueue *sq;
  821         int i;
  822 
  823         bzero(mem, size);
  824         sq = mem;
  825         for (i = 0; i < NR_SLEEPQS; i++) {
  826                 TAILQ_INIT(&sq->sq_blocked[i]);
  827                 sq->sq_blockedcnt[i] = 0;
  828         }
  829         LIST_INIT(&sq->sq_free);
  830         return (0);
  831 }
  832 
  833 /*
  834  * Find the highest priority thread sleeping on a wait channel and resume it.
  835  */
  836 int
  837 sleepq_signal(void *wchan, int flags, int pri, int queue)
  838 {
  839         struct sleepqueue *sq;
  840         struct thread *td, *besttd;
  841         int wakeup_swapper;
  842 
  843         CTR2(KTR_PROC, "sleepq_signal(%p, %d)", wchan, flags);
  844         KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
  845         MPASS((queue >= 0) && (queue < NR_SLEEPQS));
  846         sq = sleepq_lookup(wchan);
  847         if (sq == NULL)
  848                 return (0);
  849         KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
  850             ("%s: mismatch between sleep/wakeup and cv_*", __func__));
  851 
  852         /*
  853          * Find the highest priority thread on the queue.  If there is a
  854          * tie, use the thread that first appears in the queue as it has
  855          * been sleeping the longest since threads are always added to
  856          * the tail of sleep queues.
  857          */
  858         besttd = NULL;
  859         TAILQ_FOREACH(td, &sq->sq_blocked[queue], td_slpq) {
  860                 if (besttd == NULL || td->td_priority < besttd->td_priority)
  861                         besttd = td;
  862         }
  863         MPASS(besttd != NULL);
  864         thread_lock(besttd);
  865         wakeup_swapper = sleepq_resume_thread(sq, besttd, pri);
  866         thread_unlock(besttd);
  867         return (wakeup_swapper);
  868 }
  869 
  870 /*
  871  * Resume all threads sleeping on a specified wait channel.
  872  */
  873 int
  874 sleepq_broadcast(void *wchan, int flags, int pri, int queue)
  875 {
  876         struct sleepqueue *sq;
  877         struct thread *td;
  878         int wakeup_swapper;
  879 
  880         CTR2(KTR_PROC, "sleepq_broadcast(%p, %d)", wchan, flags);
  881         KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
  882         MPASS((queue >= 0) && (queue < NR_SLEEPQS));
  883         sq = sleepq_lookup(wchan);
  884         if (sq == NULL)
  885                 return (0);
  886         KASSERT(sq->sq_type == (flags & SLEEPQ_TYPE),
  887             ("%s: mismatch between sleep/wakeup and cv_*", __func__));
  888 
  889         /* Resume all blocked threads on the sleep queue. */
  890         wakeup_swapper = 0;
  891         while ((td = TAILQ_FIRST(&sq->sq_blocked[queue])) != NULL) {
  892                 thread_lock(td);
  893                 wakeup_swapper |= sleepq_resume_thread(sq, td, pri);
  894                 thread_unlock(td);
  895         }
  896         return (wakeup_swapper);
  897 }
  898 
  899 /*
  900  * Time sleeping threads out.  When the timeout expires, the thread is
  901  * removed from the sleep queue and made runnable if it is still asleep.
  902  */
  903 static void
  904 sleepq_timeout(void *arg)
  905 {
  906         struct sleepqueue_chain *sc;
  907         struct sleepqueue *sq;
  908         struct thread *td;
  909         void *wchan;
  910         int wakeup_swapper;
  911 
  912         td = arg;
  913         wakeup_swapper = 0;
  914         CTR3(KTR_PROC, "sleepq_timeout: thread %p (pid %ld, %s)",
  915             (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
  916 
  917         /*
  918          * First, see if the thread is asleep and get the wait channel if
  919          * it is.
  920          */
  921         thread_lock(td);
  922         if (TD_IS_SLEEPING(td) && TD_ON_SLEEPQ(td)) {
  923                 wchan = td->td_wchan;
  924                 sc = SC_LOOKUP(wchan);
  925                 THREAD_LOCKPTR_ASSERT(td, &sc->sc_lock);
  926                 sq = sleepq_lookup(wchan);
  927                 MPASS(sq != NULL);
  928                 td->td_flags |= TDF_TIMEOUT;
  929                 wakeup_swapper = sleepq_resume_thread(sq, td, 0);
  930                 thread_unlock(td);
  931                 if (wakeup_swapper)
  932                         kick_proc0();
  933                 return;
  934         }
  935 
  936         /*
  937          * If the thread is on the SLEEPQ but isn't sleeping yet, it
  938          * can either be on another CPU in between sleepq_add() and
  939          * one of the sleepq_*wait*() routines or it can be in
  940          * sleepq_catch_signals().
  941          */
  942         if (TD_ON_SLEEPQ(td)) {
  943                 td->td_flags |= TDF_TIMEOUT;
  944                 thread_unlock(td);
  945                 return;
  946         }
  947 
  948         /*
  949          * Now check for the edge cases.  First, if TDF_TIMEOUT is set,
  950          * then the other thread has already yielded to us, so clear
  951          * the flag and resume it.  If TDF_TIMEOUT is not set, then the
  952          * we know that the other thread is not on a sleep queue, but it
  953          * hasn't resumed execution yet.  In that case, set TDF_TIMOFAIL
  954          * to let it know that the timeout has already run and doesn't
  955          * need to be canceled.
  956          */
  957         if (td->td_flags & TDF_TIMEOUT) {
  958                 MPASS(TD_IS_SLEEPING(td));
  959                 td->td_flags &= ~TDF_TIMEOUT;
  960                 TD_CLR_SLEEPING(td);
  961                 wakeup_swapper = setrunnable(td);
  962         } else
  963                 td->td_flags |= TDF_TIMOFAIL;
  964         thread_unlock(td);
  965         if (wakeup_swapper)
  966                 kick_proc0();
  967 }
  968 
  969 /*
  970  * Resumes a specific thread from the sleep queue associated with a specific
  971  * wait channel if it is on that queue.
  972  */
  973 void
  974 sleepq_remove(struct thread *td, void *wchan)
  975 {
  976         struct sleepqueue *sq;
  977         int wakeup_swapper;
  978 
  979         /*
  980          * Look up the sleep queue for this wait channel, then re-check
  981          * that the thread is asleep on that channel, if it is not, then
  982          * bail.
  983          */
  984         MPASS(wchan != NULL);
  985         sleepq_lock(wchan);
  986         sq = sleepq_lookup(wchan);
  987         /*
  988          * We can not lock the thread here as it may be sleeping on a
  989          * different sleepq.  However, holding the sleepq lock for this
  990          * wchan can guarantee that we do not miss a wakeup for this
  991          * channel.  The asserts below will catch any false positives.
  992          */
  993         if (!TD_ON_SLEEPQ(td) || td->td_wchan != wchan) {
  994                 sleepq_release(wchan);
  995                 return;
  996         }
  997         /* Thread is asleep on sleep queue sq, so wake it up. */
  998         thread_lock(td);
  999         MPASS(sq != NULL);
 1000         MPASS(td->td_wchan == wchan);
 1001         wakeup_swapper = sleepq_resume_thread(sq, td, 0);
 1002         thread_unlock(td);
 1003         sleepq_release(wchan);
 1004         if (wakeup_swapper)
 1005                 kick_proc0();
 1006 }
 1007 
 1008 /*
 1009  * Abort a thread as if an interrupt had occurred.  Only abort
 1010  * interruptible waits (unfortunately it isn't safe to abort others).
 1011  */
 1012 int
 1013 sleepq_abort(struct thread *td, int intrval)
 1014 {
 1015         struct sleepqueue *sq;
 1016         void *wchan;
 1017 
 1018         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1019         MPASS(TD_ON_SLEEPQ(td));
 1020         MPASS(td->td_flags & TDF_SINTR);
 1021         MPASS(intrval == EINTR || intrval == ERESTART);
 1022 
 1023         /*
 1024          * If the TDF_TIMEOUT flag is set, just leave. A
 1025          * timeout is scheduled anyhow.
 1026          */
 1027         if (td->td_flags & TDF_TIMEOUT)
 1028                 return (0);
 1029 
 1030         CTR3(KTR_PROC, "sleepq_abort: thread %p (pid %ld, %s)",
 1031             (void *)td, (long)td->td_proc->p_pid, (void *)td->td_name);
 1032         td->td_intrval = intrval;
 1033         td->td_flags |= TDF_SLEEPABORT;
 1034         /*
 1035          * If the thread has not slept yet it will find the signal in
 1036          * sleepq_catch_signals() and call sleepq_resume_thread.  Otherwise
 1037          * we have to do it here.
 1038          */
 1039         if (!TD_IS_SLEEPING(td))
 1040                 return (0);
 1041         wchan = td->td_wchan;
 1042         MPASS(wchan != NULL);
 1043         sq = sleepq_lookup(wchan);
 1044         MPASS(sq != NULL);
 1045 
 1046         /* Thread is asleep on sleep queue sq, so wake it up. */
 1047         return (sleepq_resume_thread(sq, td, 0));
 1048 }
 1049 
 1050 /*
 1051  * Prints the stacks of all threads presently sleeping on wchan/queue to
 1052  * the sbuf sb.  Sets count_stacks_printed to the number of stacks actually
 1053  * printed.  Typically, this will equal the number of threads sleeping on the
 1054  * queue, but may be less if sb overflowed before all stacks were printed.
 1055  */
 1056 #ifdef STACK
 1057 int
 1058 sleepq_sbuf_print_stacks(struct sbuf *sb, void *wchan, int queue,
 1059     int *count_stacks_printed)
 1060 {
 1061         struct thread *td, *td_next;
 1062         struct sleepqueue *sq;
 1063         struct stack **st;
 1064         struct sbuf **td_infos;
 1065         int i, stack_idx, error, stacks_to_allocate;
 1066         bool finished, partial_print;
 1067 
 1068         error = 0;
 1069         finished = false;
 1070         partial_print = false;
 1071 
 1072         KASSERT(wchan != NULL, ("%s: invalid NULL wait channel", __func__));
 1073         MPASS((queue >= 0) && (queue < NR_SLEEPQS));
 1074 
 1075         stacks_to_allocate = 10;
 1076         for (i = 0; i < 3 && !finished ; i++) {
 1077                 /* We cannot malloc while holding the queue's spinlock, so
 1078                  * we do our mallocs now, and hope it is enough.  If it
 1079                  * isn't, we will free these, drop the lock, malloc more,
 1080                  * and try again, up to a point.  After that point we will
 1081                  * give up and report ENOMEM. We also cannot write to sb
 1082                  * during this time since the client may have set the
 1083                  * SBUF_AUTOEXTEND flag on their sbuf, which could cause a
 1084                  * malloc as we print to it.  So we defer actually printing
 1085                  * to sb until after we drop the spinlock.
 1086                  */
 1087 
 1088                 /* Where we will store the stacks. */
 1089                 st = malloc(sizeof(struct stack *) * stacks_to_allocate,
 1090                     M_TEMP, M_WAITOK);
 1091                 for (stack_idx = 0; stack_idx < stacks_to_allocate;
 1092                     stack_idx++)
 1093                         st[stack_idx] = stack_create();
 1094 
 1095                 /* Where we will store the td name, tid, etc. */
 1096                 td_infos = malloc(sizeof(struct sbuf *) * stacks_to_allocate,
 1097                     M_TEMP, M_WAITOK);
 1098                 for (stack_idx = 0; stack_idx < stacks_to_allocate;
 1099                     stack_idx++)
 1100                         td_infos[stack_idx] = sbuf_new(NULL, NULL,
 1101                             MAXCOMLEN + sizeof(struct thread *) * 2 + 40,
 1102                             SBUF_FIXEDLEN);
 1103 
 1104                 sleepq_lock(wchan);
 1105                 sq = sleepq_lookup(wchan);
 1106                 if (sq == NULL) {
 1107                         /* This sleepq does not exist; exit and return ENOENT. */
 1108                         error = ENOENT;
 1109                         finished = true;
 1110                         sleepq_release(wchan);
 1111                         goto loop_end;
 1112                 }
 1113 
 1114                 stack_idx = 0;
 1115                 /* Save thread info */
 1116                 TAILQ_FOREACH_SAFE(td, &sq->sq_blocked[queue], td_slpq,
 1117                     td_next) {
 1118                         if (stack_idx >= stacks_to_allocate)
 1119                                 goto loop_end;
 1120 
 1121                         /* Note the td_lock is equal to the sleepq_lock here. */
 1122                         stack_save_td(st[stack_idx], td);
 1123 
 1124                         sbuf_printf(td_infos[stack_idx], "%d: %s %p",
 1125                             td->td_tid, td->td_name, td);
 1126 
 1127                         ++stack_idx;
 1128                 }
 1129 
 1130                 finished = true;
 1131                 sleepq_release(wchan);
 1132 
 1133                 /* Print the stacks */
 1134                 for (i = 0; i < stack_idx; i++) {
 1135                         sbuf_finish(td_infos[i]);
 1136                         sbuf_printf(sb, "--- thread %s: ---\n", sbuf_data(td_infos[i]));
 1137                         stack_sbuf_print(sb, st[i]);
 1138                         sbuf_printf(sb, "\n");
 1139 
 1140                         error = sbuf_error(sb);
 1141                         if (error == 0)
 1142                                 *count_stacks_printed = stack_idx;
 1143                 }
 1144 
 1145 loop_end:
 1146                 if (!finished)
 1147                         sleepq_release(wchan);
 1148                 for (stack_idx = 0; stack_idx < stacks_to_allocate;
 1149                     stack_idx++)
 1150                         stack_destroy(st[stack_idx]);
 1151                 for (stack_idx = 0; stack_idx < stacks_to_allocate;
 1152                     stack_idx++)
 1153                         sbuf_delete(td_infos[stack_idx]);
 1154                 free(st, M_TEMP);
 1155                 free(td_infos, M_TEMP);
 1156                 stacks_to_allocate *= 10;
 1157         }
 1158 
 1159         if (!finished && error == 0)
 1160                 error = ENOMEM;
 1161 
 1162         return (error);
 1163 }
 1164 #endif
 1165 
 1166 #ifdef SLEEPQUEUE_PROFILING
 1167 #define SLEEPQ_PROF_LOCATIONS   1024
 1168 #define SLEEPQ_SBUFSIZE         512
 1169 struct sleepq_prof {
 1170         LIST_ENTRY(sleepq_prof) sp_link;
 1171         const char      *sp_wmesg;
 1172         long            sp_count;
 1173 };
 1174 
 1175 LIST_HEAD(sqphead, sleepq_prof);
 1176 
 1177 struct sqphead sleepq_prof_free;
 1178 struct sqphead sleepq_hash[SC_TABLESIZE];
 1179 static struct sleepq_prof sleepq_profent[SLEEPQ_PROF_LOCATIONS];
 1180 static struct mtx sleepq_prof_lock;
 1181 MTX_SYSINIT(sleepq_prof_lock, &sleepq_prof_lock, "sleepq_prof", MTX_SPIN);
 1182 
 1183 static void
 1184 sleepq_profile(const char *wmesg)
 1185 {
 1186         struct sleepq_prof *sp;
 1187 
 1188         mtx_lock_spin(&sleepq_prof_lock);
 1189         if (prof_enabled == 0)
 1190                 goto unlock;
 1191         LIST_FOREACH(sp, &sleepq_hash[SC_HASH(wmesg)], sp_link)
 1192                 if (sp->sp_wmesg == wmesg)
 1193                         goto done;
 1194         sp = LIST_FIRST(&sleepq_prof_free);
 1195         if (sp == NULL)
 1196                 goto unlock;
 1197         sp->sp_wmesg = wmesg;
 1198         LIST_REMOVE(sp, sp_link);
 1199         LIST_INSERT_HEAD(&sleepq_hash[SC_HASH(wmesg)], sp, sp_link);
 1200 done:
 1201         sp->sp_count++;
 1202 unlock:
 1203         mtx_unlock_spin(&sleepq_prof_lock);
 1204         return;
 1205 }
 1206 
 1207 static void
 1208 sleepq_prof_reset(void)
 1209 {
 1210         struct sleepq_prof *sp;
 1211         int enabled;
 1212         int i;
 1213 
 1214         mtx_lock_spin(&sleepq_prof_lock);
 1215         enabled = prof_enabled;
 1216         prof_enabled = 0;
 1217         for (i = 0; i < SC_TABLESIZE; i++)
 1218                 LIST_INIT(&sleepq_hash[i]);
 1219         LIST_INIT(&sleepq_prof_free);
 1220         for (i = 0; i < SLEEPQ_PROF_LOCATIONS; i++) {
 1221                 sp = &sleepq_profent[i];
 1222                 sp->sp_wmesg = NULL;
 1223                 sp->sp_count = 0;
 1224                 LIST_INSERT_HEAD(&sleepq_prof_free, sp, sp_link);
 1225         }
 1226         prof_enabled = enabled;
 1227         mtx_unlock_spin(&sleepq_prof_lock);
 1228 }
 1229 
 1230 static int
 1231 enable_sleepq_prof(SYSCTL_HANDLER_ARGS)
 1232 {
 1233         int error, v;
 1234 
 1235         v = prof_enabled;
 1236         error = sysctl_handle_int(oidp, &v, v, req);
 1237         if (error)
 1238                 return (error);
 1239         if (req->newptr == NULL)
 1240                 return (error);
 1241         if (v == prof_enabled)
 1242                 return (0);
 1243         if (v == 1)
 1244                 sleepq_prof_reset();
 1245         mtx_lock_spin(&sleepq_prof_lock);
 1246         prof_enabled = !!v;
 1247         mtx_unlock_spin(&sleepq_prof_lock);
 1248 
 1249         return (0);
 1250 }
 1251 
 1252 static int
 1253 reset_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
 1254 {
 1255         int error, v;
 1256 
 1257         v = 0;
 1258         error = sysctl_handle_int(oidp, &v, 0, req);
 1259         if (error)
 1260                 return (error);
 1261         if (req->newptr == NULL)
 1262                 return (error);
 1263         if (v == 0)
 1264                 return (0);
 1265         sleepq_prof_reset();
 1266 
 1267         return (0);
 1268 }
 1269 
 1270 static int
 1271 dump_sleepq_prof_stats(SYSCTL_HANDLER_ARGS)
 1272 {
 1273         struct sleepq_prof *sp;
 1274         struct sbuf *sb;
 1275         int enabled;
 1276         int error;
 1277         int i;
 1278 
 1279         error = sysctl_wire_old_buffer(req, 0);
 1280         if (error != 0)
 1281                 return (error);
 1282         sb = sbuf_new_for_sysctl(NULL, NULL, SLEEPQ_SBUFSIZE, req);
 1283         sbuf_printf(sb, "\nwmesg\tcount\n");
 1284         enabled = prof_enabled;
 1285         mtx_lock_spin(&sleepq_prof_lock);
 1286         prof_enabled = 0;
 1287         mtx_unlock_spin(&sleepq_prof_lock);
 1288         for (i = 0; i < SC_TABLESIZE; i++) {
 1289                 LIST_FOREACH(sp, &sleepq_hash[i], sp_link) {
 1290                         sbuf_printf(sb, "%s\t%ld\n",
 1291                             sp->sp_wmesg, sp->sp_count);
 1292                 }
 1293         }
 1294         mtx_lock_spin(&sleepq_prof_lock);
 1295         prof_enabled = enabled;
 1296         mtx_unlock_spin(&sleepq_prof_lock);
 1297 
 1298         error = sbuf_finish(sb);
 1299         sbuf_delete(sb);
 1300         return (error);
 1301 }
 1302 
 1303 SYSCTL_PROC(_debug_sleepq, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
 1304     NULL, 0, dump_sleepq_prof_stats, "A", "Sleepqueue profiling statistics");
 1305 SYSCTL_PROC(_debug_sleepq, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
 1306     NULL, 0, reset_sleepq_prof_stats, "I",
 1307     "Reset sleepqueue profiling statistics");
 1308 SYSCTL_PROC(_debug_sleepq, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
 1309     NULL, 0, enable_sleepq_prof, "I", "Enable sleepqueue profiling");
 1310 #endif
 1311 
 1312 #ifdef DDB
 1313 DB_SHOW_COMMAND(sleepq, db_show_sleepqueue)
 1314 {
 1315         struct sleepqueue_chain *sc;
 1316         struct sleepqueue *sq;
 1317 #ifdef INVARIANTS
 1318         struct lock_object *lock;
 1319 #endif
 1320         struct thread *td;
 1321         void *wchan;
 1322         int i;
 1323 
 1324         if (!have_addr)
 1325                 return;
 1326 
 1327         /*
 1328          * First, see if there is an active sleep queue for the wait channel
 1329          * indicated by the address.
 1330          */
 1331         wchan = (void *)addr;
 1332         sc = SC_LOOKUP(wchan);
 1333         LIST_FOREACH(sq, &sc->sc_queues, sq_hash)
 1334                 if (sq->sq_wchan == wchan)
 1335                         goto found;
 1336 
 1337         /*
 1338          * Second, see if there is an active sleep queue at the address
 1339          * indicated.
 1340          */
 1341         for (i = 0; i < SC_TABLESIZE; i++)
 1342                 LIST_FOREACH(sq, &sleepq_chains[i].sc_queues, sq_hash) {
 1343                         if (sq == (struct sleepqueue *)addr)
 1344                                 goto found;
 1345                 }
 1346 
 1347         db_printf("Unable to locate a sleep queue via %p\n", (void *)addr);
 1348         return;
 1349 found:
 1350         db_printf("Wait channel: %p\n", sq->sq_wchan);
 1351         db_printf("Queue type: %d\n", sq->sq_type);
 1352 #ifdef INVARIANTS
 1353         if (sq->sq_lock) {
 1354                 lock = sq->sq_lock;
 1355                 db_printf("Associated Interlock: %p - (%s) %s\n", lock,
 1356                     LOCK_CLASS(lock)->lc_name, lock->lo_name);
 1357         }
 1358 #endif
 1359         db_printf("Blocked threads:\n");
 1360         for (i = 0; i < NR_SLEEPQS; i++) {
 1361                 db_printf("\nQueue[%d]:\n", i);
 1362                 if (TAILQ_EMPTY(&sq->sq_blocked[i]))
 1363                         db_printf("\tempty\n");
 1364                 else
 1365                         TAILQ_FOREACH(td, &sq->sq_blocked[0],
 1366                                       td_slpq) {
 1367                                 db_printf("\t%p (tid %d, pid %d, \"%s\")\n", td,
 1368                                           td->td_tid, td->td_proc->p_pid,
 1369                                           td->td_name);
 1370                         }
 1371                 db_printf("(expected: %u)\n", sq->sq_blockedcnt[i]);
 1372         }
 1373 }
 1374 
 1375 /* Alias 'show sleepqueue' to 'show sleepq'. */
 1376 DB_SHOW_ALIAS(sleepqueue, db_show_sleepqueue);
 1377 #endif

Cache object: 7aea6c4d71eafcf6aa959d8d6b2480f1


[ 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.