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/sched_4bsd.c

Version: -  FREEBSD  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-2  -  FREEBSD-11-1  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-4  -  FREEBSD-10-3  -  FREEBSD-10-2  -  FREEBSD-10-1  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-3  -  FREEBSD-9-2  -  FREEBSD-9-1  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-4  -  FREEBSD-8-3  -  FREEBSD-8-2  -  FREEBSD-8-1  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-4  -  FREEBSD-7-3  -  FREEBSD-7-2  -  FREEBSD-7-1  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-4  -  FREEBSD-6-3  -  FREEBSD-6-2  -  FREEBSD-6-1  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-5  -  FREEBSD-5-4  -  FREEBSD-5-3  -  FREEBSD-5-2  -  FREEBSD-5-1  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  linux-2.6  -  linux-2.4.22  -  MK83  -  MK84  -  PLAN9  -  DFBSD  -  NETBSD  -  NETBSD5  -  NETBSD4  -  NETBSD3  -  NETBSD20  -  OPENBSD  -  xnu-517  -  xnu-792  -  xnu-792.6.70  -  xnu-1228  -  xnu-1456.1.26  -  xnu-1699.24.8  -  xnu-2050.18.24  -  OPENSOLARIS  -  minix-3-1-1 
SearchContext: -  none  -  3  -  10 

    1 /*-
    2  * Copyright (c) 1982, 1986, 1990, 1991, 1993
    3  *      The Regents of the University of California.  All rights reserved.
    4  * (c) UNIX System Laboratories, Inc.
    5  * All or some portions of this file are derived from material licensed
    6  * to the University of California by American Telephone and Telegraph
    7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
    8  * the permission of UNIX System Laboratories, Inc.
    9  *
   10  * Redistribution and use in source and binary forms, with or without
   11  * modification, are permitted provided that the following conditions
   12  * are met:
   13  * 1. Redistributions of source code must retain the above copyright
   14  *    notice, this list of conditions and the following disclaimer.
   15  * 2. Redistributions in binary form must reproduce the above copyright
   16  *    notice, this list of conditions and the following disclaimer in the
   17  *    documentation and/or other materials provided with the distribution.
   18  * 4. Neither the name of the University nor the names of its contributors
   19  *    may be used to endorse or promote products derived from this software
   20  *    without specific prior written permission.
   21  *
   22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
   23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
   26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   32  * SUCH DAMAGE.
   33  */
   34 
   35 #include <sys/cdefs.h>
   36 __FBSDID("$FreeBSD: releng/8.3/sys/kern/sched_4bsd.c 230080 2012-01-13 20:15:49Z jhb $");
   37 
   38 #include "opt_hwpmc_hooks.h"
   39 #include "opt_sched.h"
   40 #include "opt_kdtrace.h"
   41 
   42 #include <sys/param.h>
   43 #include <sys/systm.h>
   44 #include <sys/cpuset.h>
   45 #include <sys/kernel.h>
   46 #include <sys/ktr.h>
   47 #include <sys/lock.h>
   48 #include <sys/kthread.h>
   49 #include <sys/mutex.h>
   50 #include <sys/proc.h>
   51 #include <sys/resourcevar.h>
   52 #include <sys/sched.h>
   53 #include <sys/smp.h>
   54 #include <sys/sysctl.h>
   55 #include <sys/sx.h>
   56 #include <sys/turnstile.h>
   57 #include <sys/umtx.h>
   58 #include <machine/pcb.h>
   59 #include <machine/smp.h>
   60 
   61 #ifdef HWPMC_HOOKS
   62 #include <sys/pmckern.h>
   63 #endif
   64 
   65 #ifdef KDTRACE_HOOKS
   66 #include <sys/dtrace_bsd.h>
   67 int                             dtrace_vtime_active;
   68 dtrace_vtime_switch_func_t      dtrace_vtime_switch_func;
   69 #endif
   70 
   71 /*
   72  * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
   73  * the range 100-256 Hz (approximately).
   74  */
   75 #define ESTCPULIM(e) \
   76     min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
   77     RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
   78 #ifdef SMP
   79 #define INVERSE_ESTCPU_WEIGHT   (8 * smp_cpus)
   80 #else
   81 #define INVERSE_ESTCPU_WEIGHT   8       /* 1 / (priorities per estcpu level). */
   82 #endif
   83 #define NICE_WEIGHT             1       /* Priorities per nice level. */
   84 
   85 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
   86 
   87 /*
   88  * The schedulable entity that runs a context.
   89  * This is  an extension to the thread structure and is tailored to
   90  * the requirements of this scheduler
   91  */
   92 struct td_sched {
   93         fixpt_t         ts_pctcpu;      /* (j) %cpu during p_swtime. */
   94         int             ts_cpticks;     /* (j) Ticks of cpu time. */
   95         int             ts_slptime;     /* (j) Seconds !RUNNING. */
   96         int             ts_flags;
   97         struct runq     *ts_runq;       /* runq the thread is currently on */
   98 #ifdef KTR
   99         char            ts_name[TS_NAME_LEN];
  100 #endif
  101 };
  102 
  103 /* flags kept in td_flags */
  104 #define TDF_DIDRUN      TDF_SCHED0      /* thread actually ran. */
  105 #define TDF_BOUND       TDF_SCHED1      /* Bound to one CPU. */
  106 
  107 /* flags kept in ts_flags */
  108 #define TSF_AFFINITY    0x0001          /* Has a non-"full" CPU set. */
  109 
  110 #define SKE_RUNQ_PCPU(ts)                                               \
  111     ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
  112 
  113 #define THREAD_CAN_SCHED(td, cpu)       \
  114     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
  115 
  116 static struct td_sched td_sched0;
  117 struct mtx sched_lock;
  118 
  119 static int      sched_tdcnt;    /* Total runnable threads in the system. */
  120 static int      sched_quantum;  /* Roundrobin scheduling quantum in ticks. */
  121 #define SCHED_QUANTUM   (hz / 10)       /* Default sched quantum */
  122 
  123 static void     setup_runqs(void);
  124 static void     schedcpu(void);
  125 static void     schedcpu_thread(void);
  126 static void     sched_priority(struct thread *td, u_char prio);
  127 static void     sched_setup(void *dummy);
  128 static void     maybe_resched(struct thread *td);
  129 static void     updatepri(struct thread *td);
  130 static void     resetpriority(struct thread *td);
  131 static void     resetpriority_thread(struct thread *td);
  132 #ifdef SMP
  133 static int      sched_pickcpu(struct thread *td);
  134 static int      forward_wakeup(int cpunum);
  135 static void     kick_other_cpu(int pri, int cpuid);
  136 #endif
  137 
  138 static struct kproc_desc sched_kp = {
  139         "schedcpu",
  140         schedcpu_thread,
  141         NULL
  142 };
  143 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start,
  144     &sched_kp);
  145 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
  146 
  147 /*
  148  * Global run queue.
  149  */
  150 static struct runq runq;
  151 
  152 #ifdef SMP
  153 /*
  154  * Per-CPU run queues
  155  */
  156 static struct runq runq_pcpu[MAXCPU];
  157 long runq_length[MAXCPU];
  158 #endif
  159 
  160 static void
  161 setup_runqs(void)
  162 {
  163 #ifdef SMP
  164         int i;
  165 
  166         for (i = 0; i < MAXCPU; ++i)
  167                 runq_init(&runq_pcpu[i]);
  168 #endif
  169 
  170         runq_init(&runq);
  171 }
  172 
  173 static int
  174 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
  175 {
  176         int error, new_val;
  177 
  178         new_val = sched_quantum * tick;
  179         error = sysctl_handle_int(oidp, &new_val, 0, req);
  180         if (error != 0 || req->newptr == NULL)
  181                 return (error);
  182         if (new_val < tick)
  183                 return (EINVAL);
  184         sched_quantum = new_val / tick;
  185         hogticks = 2 * sched_quantum;
  186         return (0);
  187 }
  188 
  189 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
  190 
  191 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
  192     "Scheduler name");
  193 
  194 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
  195     0, sizeof sched_quantum, sysctl_kern_quantum, "I",
  196     "Roundrobin scheduling quantum in microseconds");
  197 
  198 #ifdef SMP
  199 /* Enable forwarding of wakeups to all other cpus */
  200 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
  201 
  202 static int runq_fuzz = 1;
  203 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
  204 
  205 static int forward_wakeup_enabled = 1;
  206 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
  207            &forward_wakeup_enabled, 0,
  208            "Forwarding of wakeup to idle CPUs");
  209 
  210 static int forward_wakeups_requested = 0;
  211 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
  212            &forward_wakeups_requested, 0,
  213            "Requests for Forwarding of wakeup to idle CPUs");
  214 
  215 static int forward_wakeups_delivered = 0;
  216 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
  217            &forward_wakeups_delivered, 0,
  218            "Completed Forwarding of wakeup to idle CPUs");
  219 
  220 static int forward_wakeup_use_mask = 1;
  221 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
  222            &forward_wakeup_use_mask, 0,
  223            "Use the mask of idle cpus");
  224 
  225 static int forward_wakeup_use_loop = 0;
  226 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
  227            &forward_wakeup_use_loop, 0,
  228            "Use a loop to find idle cpus");
  229 
  230 static int forward_wakeup_use_single = 0;
  231 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
  232            &forward_wakeup_use_single, 0,
  233            "Only signal one idle cpu");
  234 
  235 static int forward_wakeup_use_htt = 0;
  236 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
  237            &forward_wakeup_use_htt, 0,
  238            "account for htt");
  239 
  240 #endif
  241 #if 0
  242 static int sched_followon = 0;
  243 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
  244            &sched_followon, 0,
  245            "allow threads to share a quantum");
  246 #endif
  247 
  248 static __inline void
  249 sched_load_add(void)
  250 {
  251 
  252         sched_tdcnt++;
  253         KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
  254 }
  255 
  256 static __inline void
  257 sched_load_rem(void)
  258 {
  259 
  260         sched_tdcnt--;
  261         KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
  262 }
  263 /*
  264  * Arrange to reschedule if necessary, taking the priorities and
  265  * schedulers into account.
  266  */
  267 static void
  268 maybe_resched(struct thread *td)
  269 {
  270 
  271         THREAD_LOCK_ASSERT(td, MA_OWNED);
  272         if (td->td_priority < curthread->td_priority)
  273                 curthread->td_flags |= TDF_NEEDRESCHED;
  274 }
  275 
  276 /*
  277  * This function is called when a thread is about to be put on run queue
  278  * because it has been made runnable or its priority has been adjusted.  It
  279  * determines if the new thread should be immediately preempted to.  If so,
  280  * it switches to it and eventually returns true.  If not, it returns false
  281  * so that the caller may place the thread on an appropriate run queue.
  282  */
  283 int
  284 maybe_preempt(struct thread *td)
  285 {
  286 #ifdef PREEMPTION
  287         struct thread *ctd;
  288         int cpri, pri;
  289 
  290         /*
  291          * The new thread should not preempt the current thread if any of the
  292          * following conditions are true:
  293          *
  294          *  - The kernel is in the throes of crashing (panicstr).
  295          *  - The current thread has a higher (numerically lower) or
  296          *    equivalent priority.  Note that this prevents curthread from
  297          *    trying to preempt to itself.
  298          *  - It is too early in the boot for context switches (cold is set).
  299          *  - The current thread has an inhibitor set or is in the process of
  300          *    exiting.  In this case, the current thread is about to switch
  301          *    out anyways, so there's no point in preempting.  If we did,
  302          *    the current thread would not be properly resumed as well, so
  303          *    just avoid that whole landmine.
  304          *  - If the new thread's priority is not a realtime priority and
  305          *    the current thread's priority is not an idle priority and
  306          *    FULL_PREEMPTION is disabled.
  307          *
  308          * If all of these conditions are false, but the current thread is in
  309          * a nested critical section, then we have to defer the preemption
  310          * until we exit the critical section.  Otherwise, switch immediately
  311          * to the new thread.
  312          */
  313         ctd = curthread;
  314         THREAD_LOCK_ASSERT(td, MA_OWNED);
  315         KASSERT((td->td_inhibitors == 0),
  316                         ("maybe_preempt: trying to run inhibited thread"));
  317         pri = td->td_priority;
  318         cpri = ctd->td_priority;
  319         if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
  320             TD_IS_INHIBITED(ctd))
  321                 return (0);
  322 #ifndef FULL_PREEMPTION
  323         if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
  324                 return (0);
  325 #endif
  326 
  327         if (ctd->td_critnest > 1) {
  328                 CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
  329                     ctd->td_critnest);
  330                 ctd->td_owepreempt = 1;
  331                 return (0);
  332         }
  333         /*
  334          * Thread is runnable but not yet put on system run queue.
  335          */
  336         MPASS(ctd->td_lock == td->td_lock);
  337         MPASS(TD_ON_RUNQ(td));
  338         TD_SET_RUNNING(td);
  339         CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
  340             td->td_proc->p_pid, td->td_name);
  341         mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
  342         /*
  343          * td's lock pointer may have changed.  We have to return with it
  344          * locked.
  345          */
  346         spinlock_enter();
  347         thread_unlock(ctd);
  348         thread_lock(td);
  349         spinlock_exit();
  350         return (1);
  351 #else
  352         return (0);
  353 #endif
  354 }
  355 
  356 /*
  357  * Constants for digital decay and forget:
  358  *      90% of (td_estcpu) usage in 5 * loadav time
  359  *      95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
  360  *          Note that, as ps(1) mentions, this can let percentages
  361  *          total over 100% (I've seen 137.9% for 3 processes).
  362  *
  363  * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
  364  *
  365  * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
  366  * That is, the system wants to compute a value of decay such
  367  * that the following for loop:
  368  *      for (i = 0; i < (5 * loadavg); i++)
  369  *              td_estcpu *= decay;
  370  * will compute
  371  *      td_estcpu *= 0.1;
  372  * for all values of loadavg:
  373  *
  374  * Mathematically this loop can be expressed by saying:
  375  *      decay ** (5 * loadavg) ~= .1
  376  *
  377  * The system computes decay as:
  378  *      decay = (2 * loadavg) / (2 * loadavg + 1)
  379  *
  380  * We wish to prove that the system's computation of decay
  381  * will always fulfill the equation:
  382  *      decay ** (5 * loadavg) ~= .1
  383  *
  384  * If we compute b as:
  385  *      b = 2 * loadavg
  386  * then
  387  *      decay = b / (b + 1)
  388  *
  389  * We now need to prove two things:
  390  *      1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
  391  *      2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
  392  *
  393  * Facts:
  394  *         For x close to zero, exp(x) =~ 1 + x, since
  395  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
  396  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
  397  *         For x close to zero, ln(1+x) =~ x, since
  398  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
  399  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
  400  *         ln(.1) =~ -2.30
  401  *
  402  * Proof of (1):
  403  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
  404  *      solving for factor,
  405  *      ln(factor) =~ (-2.30/5*loadav), or
  406  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
  407  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
  408  *
  409  * Proof of (2):
  410  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
  411  *      solving for power,
  412  *      power*ln(b/(b+1)) =~ -2.30, or
  413  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
  414  *
  415  * Actual power values for the implemented algorithm are as follows:
  416  *      loadav: 1       2       3       4
  417  *      power:  5.68    10.32   14.94   19.55
  418  */
  419 
  420 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
  421 #define loadfactor(loadav)      (2 * (loadav))
  422 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
  423 
  424 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
  425 static fixpt_t  ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
  426 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
  427 
  428 /*
  429  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
  430  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
  431  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
  432  *
  433  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
  434  *      1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
  435  *
  436  * If you don't want to bother with the faster/more-accurate formula, you
  437  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
  438  * (more general) method of calculating the %age of CPU used by a process.
  439  */
  440 #define CCPU_SHIFT      11
  441 
  442 /*
  443  * Recompute process priorities, every hz ticks.
  444  * MP-safe, called without the Giant mutex.
  445  */
  446 /* ARGSUSED */
  447 static void
  448 schedcpu(void)
  449 {
  450         register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
  451         struct thread *td;
  452         struct proc *p;
  453         struct td_sched *ts;
  454         int awake, realstathz;
  455 
  456         realstathz = stathz ? stathz : hz;
  457         sx_slock(&allproc_lock);
  458         FOREACH_PROC_IN_SYSTEM(p) {
  459                 PROC_LOCK(p);
  460                 if (p->p_state == PRS_NEW) {
  461                         PROC_UNLOCK(p);
  462                         continue;
  463                 }
  464                 FOREACH_THREAD_IN_PROC(p, td) {
  465                         awake = 0;
  466                         thread_lock(td);
  467                         ts = td->td_sched;
  468                         /*
  469                          * Increment sleep time (if sleeping).  We
  470                          * ignore overflow, as above.
  471                          */
  472                         /*
  473                          * The td_sched slptimes are not touched in wakeup
  474                          * because the thread may not HAVE everything in
  475                          * memory? XXX I think this is out of date.
  476                          */
  477                         if (TD_ON_RUNQ(td)) {
  478                                 awake = 1;
  479                                 td->td_flags &= ~TDF_DIDRUN;
  480                         } else if (TD_IS_RUNNING(td)) {
  481                                 awake = 1;
  482                                 /* Do not clear TDF_DIDRUN */
  483                         } else if (td->td_flags & TDF_DIDRUN) {
  484                                 awake = 1;
  485                                 td->td_flags &= ~TDF_DIDRUN;
  486                         }
  487 
  488                         /*
  489                          * ts_pctcpu is only for ps and ttyinfo().
  490                          */
  491                         ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
  492                         /*
  493                          * If the td_sched has been idle the entire second,
  494                          * stop recalculating its priority until
  495                          * it wakes up.
  496                          */
  497                         if (ts->ts_cpticks != 0) {
  498 #if     (FSHIFT >= CCPU_SHIFT)
  499                                 ts->ts_pctcpu += (realstathz == 100)
  500                                     ? ((fixpt_t) ts->ts_cpticks) <<
  501                                     (FSHIFT - CCPU_SHIFT) :
  502                                     100 * (((fixpt_t) ts->ts_cpticks)
  503                                     << (FSHIFT - CCPU_SHIFT)) / realstathz;
  504 #else
  505                                 ts->ts_pctcpu += ((FSCALE - ccpu) *
  506                                     (ts->ts_cpticks *
  507                                     FSCALE / realstathz)) >> FSHIFT;
  508 #endif
  509                                 ts->ts_cpticks = 0;
  510                         }
  511                         /*
  512                          * If there are ANY running threads in this process,
  513                          * then don't count it as sleeping.
  514                          * XXX: this is broken.
  515                          */
  516                         if (awake) {
  517                                 if (ts->ts_slptime > 1) {
  518                                         /*
  519                                          * In an ideal world, this should not
  520                                          * happen, because whoever woke us
  521                                          * up from the long sleep should have
  522                                          * unwound the slptime and reset our
  523                                          * priority before we run at the stale
  524                                          * priority.  Should KASSERT at some
  525                                          * point when all the cases are fixed.
  526                                          */
  527                                         updatepri(td);
  528                                 }
  529                                 ts->ts_slptime = 0;
  530                         } else
  531                                 ts->ts_slptime++;
  532                         if (ts->ts_slptime > 1) {
  533                                 thread_unlock(td);
  534                                 continue;
  535                         }
  536                         td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
  537                         resetpriority(td);
  538                         resetpriority_thread(td);
  539                         thread_unlock(td);
  540                 }
  541                 PROC_UNLOCK(p);
  542         }
  543         sx_sunlock(&allproc_lock);
  544 }
  545 
  546 /*
  547  * Main loop for a kthread that executes schedcpu once a second.
  548  */
  549 static void
  550 schedcpu_thread(void)
  551 {
  552 
  553         for (;;) {
  554                 schedcpu();
  555                 pause("-", hz);
  556         }
  557 }
  558 
  559 /*
  560  * Recalculate the priority of a process after it has slept for a while.
  561  * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
  562  * least six times the loadfactor will decay td_estcpu to zero.
  563  */
  564 static void
  565 updatepri(struct thread *td)
  566 {
  567         struct td_sched *ts;
  568         fixpt_t loadfac;
  569         unsigned int newcpu;
  570 
  571         ts = td->td_sched;
  572         loadfac = loadfactor(averunnable.ldavg[0]);
  573         if (ts->ts_slptime > 5 * loadfac)
  574                 td->td_estcpu = 0;
  575         else {
  576                 newcpu = td->td_estcpu;
  577                 ts->ts_slptime--;       /* was incremented in schedcpu() */
  578                 while (newcpu && --ts->ts_slptime)
  579                         newcpu = decay_cpu(loadfac, newcpu);
  580                 td->td_estcpu = newcpu;
  581         }
  582 }
  583 
  584 /*
  585  * Compute the priority of a process when running in user mode.
  586  * Arrange to reschedule if the resulting priority is better
  587  * than that of the current process.
  588  */
  589 static void
  590 resetpriority(struct thread *td)
  591 {
  592         register unsigned int newpriority;
  593 
  594         if (td->td_pri_class == PRI_TIMESHARE) {
  595                 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
  596                     NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
  597                 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
  598                     PRI_MAX_TIMESHARE);
  599                 sched_user_prio(td, newpriority);
  600         }
  601 }
  602 
  603 /*
  604  * Update the thread's priority when the associated process's user
  605  * priority changes.
  606  */
  607 static void
  608 resetpriority_thread(struct thread *td)
  609 {
  610 
  611         /* Only change threads with a time sharing user priority. */
  612         if (td->td_priority < PRI_MIN_TIMESHARE ||
  613             td->td_priority > PRI_MAX_TIMESHARE)
  614                 return;
  615 
  616         /* XXX the whole needresched thing is broken, but not silly. */
  617         maybe_resched(td);
  618 
  619         sched_prio(td, td->td_user_pri);
  620 }
  621 
  622 /* ARGSUSED */
  623 static void
  624 sched_setup(void *dummy)
  625 {
  626         setup_runqs();
  627 
  628         if (sched_quantum == 0)
  629                 sched_quantum = SCHED_QUANTUM;
  630         hogticks = 2 * sched_quantum;
  631 
  632         /* Account for thread0. */
  633         sched_load_add();
  634 }
  635 
  636 /* External interfaces start here */
  637 
  638 /*
  639  * Very early in the boot some setup of scheduler-specific
  640  * parts of proc0 and of some scheduler resources needs to be done.
  641  * Called from:
  642  *  proc0_init()
  643  */
  644 void
  645 schedinit(void)
  646 {
  647         /*
  648          * Set up the scheduler specific parts of proc0.
  649          */
  650         proc0.p_sched = NULL; /* XXX */
  651         thread0.td_sched = &td_sched0;
  652         thread0.td_lock = &sched_lock;
  653         mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
  654 }
  655 
  656 int
  657 sched_runnable(void)
  658 {
  659 #ifdef SMP
  660         return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
  661 #else
  662         return runq_check(&runq);
  663 #endif
  664 }
  665 
  666 int
  667 sched_rr_interval(void)
  668 {
  669         if (sched_quantum == 0)
  670                 sched_quantum = SCHED_QUANTUM;
  671         return (sched_quantum);
  672 }
  673 
  674 /*
  675  * We adjust the priority of the current process.  The priority of
  676  * a process gets worse as it accumulates CPU time.  The cpu usage
  677  * estimator (td_estcpu) is increased here.  resetpriority() will
  678  * compute a different priority each time td_estcpu increases by
  679  * INVERSE_ESTCPU_WEIGHT
  680  * (until MAXPRI is reached).  The cpu usage estimator ramps up
  681  * quite quickly when the process is running (linearly), and decays
  682  * away exponentially, at a rate which is proportionally slower when
  683  * the system is busy.  The basic principle is that the system will
  684  * 90% forget that the process used a lot of CPU time in 5 * loadav
  685  * seconds.  This causes the system to favor processes which haven't
  686  * run much recently, and to round-robin among other processes.
  687  */
  688 void
  689 sched_clock(struct thread *td)
  690 {
  691         struct td_sched *ts;
  692 
  693         THREAD_LOCK_ASSERT(td, MA_OWNED);
  694         ts = td->td_sched;
  695 
  696         ts->ts_cpticks++;
  697         td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
  698         if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
  699                 resetpriority(td);
  700                 resetpriority_thread(td);
  701         }
  702 
  703         /*
  704          * Force a context switch if the current thread has used up a full
  705          * quantum (default quantum is 100ms).
  706          */
  707         if (!TD_IS_IDLETHREAD(td) &&
  708             ticks - PCPU_GET(switchticks) >= sched_quantum)
  709                 td->td_flags |= TDF_NEEDRESCHED;
  710 }
  711 
  712 /*
  713  * Charge child's scheduling CPU usage to parent.
  714  */
  715 void
  716 sched_exit(struct proc *p, struct thread *td)
  717 {
  718 
  719         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
  720             "prio:%d", td->td_priority);
  721 
  722         PROC_LOCK_ASSERT(p, MA_OWNED);
  723         sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
  724 }
  725 
  726 void
  727 sched_exit_thread(struct thread *td, struct thread *child)
  728 {
  729 
  730         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
  731             "prio:%d", child->td_priority);
  732         thread_lock(td);
  733         td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
  734         thread_unlock(td);
  735         thread_lock(child);
  736         if ((child->td_flags & TDF_NOLOAD) == 0)
  737                 sched_load_rem();
  738         thread_unlock(child);
  739 }
  740 
  741 void
  742 sched_fork(struct thread *td, struct thread *childtd)
  743 {
  744         sched_fork_thread(td, childtd);
  745 }
  746 
  747 void
  748 sched_fork_thread(struct thread *td, struct thread *childtd)
  749 {
  750         struct td_sched *ts;
  751 
  752         childtd->td_estcpu = td->td_estcpu;
  753         childtd->td_lock = &sched_lock;
  754         childtd->td_cpuset = cpuset_ref(td->td_cpuset);
  755         childtd->td_priority = childtd->td_base_pri;
  756         ts = childtd->td_sched;
  757         bzero(ts, sizeof(*ts));
  758         ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
  759 }
  760 
  761 void
  762 sched_nice(struct proc *p, int nice)
  763 {
  764         struct thread *td;
  765 
  766         PROC_LOCK_ASSERT(p, MA_OWNED);
  767         p->p_nice = nice;
  768         FOREACH_THREAD_IN_PROC(p, td) {
  769                 thread_lock(td);
  770                 resetpriority(td);
  771                 resetpriority_thread(td);
  772                 thread_unlock(td);
  773         }
  774 }
  775 
  776 void
  777 sched_class(struct thread *td, int class)
  778 {
  779         THREAD_LOCK_ASSERT(td, MA_OWNED);
  780         td->td_pri_class = class;
  781 }
  782 
  783 /*
  784  * Adjust the priority of a thread.
  785  */
  786 static void
  787 sched_priority(struct thread *td, u_char prio)
  788 {
  789 
  790 
  791         KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
  792             "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
  793             sched_tdname(curthread));
  794         if (td != curthread && prio > td->td_priority) {
  795                 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
  796                     "lend prio", "prio:%d", td->td_priority, "new prio:%d",
  797                     prio, KTR_ATTR_LINKED, sched_tdname(td));
  798         }
  799         THREAD_LOCK_ASSERT(td, MA_OWNED);
  800         if (td->td_priority == prio)
  801                 return;
  802         td->td_priority = prio;
  803         if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
  804                 sched_rem(td);
  805                 sched_add(td, SRQ_BORING);
  806         }
  807 }
  808 
  809 /*
  810  * Update a thread's priority when it is lent another thread's
  811  * priority.
  812  */
  813 void
  814 sched_lend_prio(struct thread *td, u_char prio)
  815 {
  816 
  817         td->td_flags |= TDF_BORROWING;
  818         sched_priority(td, prio);
  819 }
  820 
  821 /*
  822  * Restore a thread's priority when priority propagation is
  823  * over.  The prio argument is the minimum priority the thread
  824  * needs to have to satisfy other possible priority lending
  825  * requests.  If the thread's regulary priority is less
  826  * important than prio the thread will keep a priority boost
  827  * of prio.
  828  */
  829 void
  830 sched_unlend_prio(struct thread *td, u_char prio)
  831 {
  832         u_char base_pri;
  833 
  834         if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
  835             td->td_base_pri <= PRI_MAX_TIMESHARE)
  836                 base_pri = td->td_user_pri;
  837         else
  838                 base_pri = td->td_base_pri;
  839         if (prio >= base_pri) {
  840                 td->td_flags &= ~TDF_BORROWING;
  841                 sched_prio(td, base_pri);
  842         } else
  843                 sched_lend_prio(td, prio);
  844 }
  845 
  846 void
  847 sched_prio(struct thread *td, u_char prio)
  848 {
  849         u_char oldprio;
  850 
  851         /* First, update the base priority. */
  852         td->td_base_pri = prio;
  853 
  854         /*
  855          * If the thread is borrowing another thread's priority, don't ever
  856          * lower the priority.
  857          */
  858         if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
  859                 return;
  860 
  861         /* Change the real priority. */
  862         oldprio = td->td_priority;
  863         sched_priority(td, prio);
  864 
  865         /*
  866          * If the thread is on a turnstile, then let the turnstile update
  867          * its state.
  868          */
  869         if (TD_ON_LOCK(td) && oldprio != prio)
  870                 turnstile_adjust(td, oldprio);
  871 }
  872 
  873 void
  874 sched_user_prio(struct thread *td, u_char prio)
  875 {
  876         u_char oldprio;
  877 
  878         THREAD_LOCK_ASSERT(td, MA_OWNED);
  879         td->td_base_user_pri = prio;
  880         if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
  881                 return;
  882         oldprio = td->td_user_pri;
  883         td->td_user_pri = prio;
  884 }
  885 
  886 void
  887 sched_lend_user_prio(struct thread *td, u_char prio)
  888 {
  889         u_char oldprio;
  890 
  891         THREAD_LOCK_ASSERT(td, MA_OWNED);
  892         td->td_flags |= TDF_UBORROWING;
  893         oldprio = td->td_user_pri;
  894         td->td_user_pri = prio;
  895 }
  896 
  897 void
  898 sched_unlend_user_prio(struct thread *td, u_char prio)
  899 {
  900         u_char base_pri;
  901 
  902         THREAD_LOCK_ASSERT(td, MA_OWNED);
  903         base_pri = td->td_base_user_pri;
  904         if (prio >= base_pri) {
  905                 td->td_flags &= ~TDF_UBORROWING;
  906                 sched_user_prio(td, base_pri);
  907         } else {
  908                 sched_lend_user_prio(td, prio);
  909         }
  910 }
  911 
  912 void
  913 sched_sleep(struct thread *td, int pri)
  914 {
  915 
  916         THREAD_LOCK_ASSERT(td, MA_OWNED);
  917         td->td_slptick = ticks;
  918         td->td_sched->ts_slptime = 0;
  919         if (pri)
  920                 sched_prio(td, pri);
  921         if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
  922                 td->td_flags |= TDF_CANSWAP;
  923 }
  924 
  925 void
  926 sched_switch(struct thread *td, struct thread *newtd, int flags)
  927 {
  928         struct mtx *tmtx;
  929         struct td_sched *ts;
  930         struct proc *p;
  931 
  932         tmtx = NULL;
  933         ts = td->td_sched;
  934         p = td->td_proc;
  935 
  936         THREAD_LOCK_ASSERT(td, MA_OWNED);
  937 
  938         /* 
  939          * Switch to the sched lock to fix things up and pick
  940          * a new thread.
  941          * Block the td_lock in order to avoid breaking the critical path.
  942          */
  943         if (td->td_lock != &sched_lock) {
  944                 mtx_lock_spin(&sched_lock);
  945                 tmtx = thread_lock_block(td);
  946         }
  947 
  948         if ((td->td_flags & TDF_NOLOAD) == 0)
  949                 sched_load_rem();
  950 
  951         td->td_lastcpu = td->td_oncpu;
  952         if (!(flags & SW_PREEMPT))
  953                 td->td_flags &= ~TDF_NEEDRESCHED;
  954         td->td_owepreempt = 0;
  955         td->td_oncpu = NOCPU;
  956 
  957         /*
  958          * At the last moment, if this thread is still marked RUNNING,
  959          * then put it back on the run queue as it has not been suspended
  960          * or stopped or any thing else similar.  We never put the idle
  961          * threads on the run queue, however.
  962          */
  963         if (td->td_flags & TDF_IDLETD) {
  964                 TD_SET_CAN_RUN(td);
  965 #ifdef SMP
  966                 idle_cpus_mask &= ~PCPU_GET(cpumask);
  967 #endif
  968         } else {
  969                 if (TD_IS_RUNNING(td)) {
  970                         /* Put us back on the run queue. */
  971                         sched_add(td, (flags & SW_PREEMPT) ?
  972                             SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
  973                             SRQ_OURSELF|SRQ_YIELDING);
  974                 }
  975         }
  976         if (newtd) {
  977                 /*
  978                  * The thread we are about to run needs to be counted
  979                  * as if it had been added to the run queue and selected.
  980                  * It came from:
  981                  * * A preemption
  982                  * * An upcall
  983                  * * A followon
  984                  */
  985                 KASSERT((newtd->td_inhibitors == 0),
  986                         ("trying to run inhibited thread"));
  987                 newtd->td_flags |= TDF_DIDRUN;
  988                 TD_SET_RUNNING(newtd);
  989                 if ((newtd->td_flags & TDF_NOLOAD) == 0)
  990                         sched_load_add();
  991         } else {
  992                 newtd = choosethread();
  993                 MPASS(newtd->td_lock == &sched_lock);
  994         }
  995 
  996         if (td != newtd) {
  997 #ifdef  HWPMC_HOOKS
  998                 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
  999                         PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
 1000 #endif
 1001                 /* I feel sleepy */
 1002                 lock_profile_release_lock(&sched_lock.lock_object);
 1003 #ifdef KDTRACE_HOOKS
 1004                 /*
 1005                  * If DTrace has set the active vtime enum to anything
 1006                  * other than INACTIVE (0), then it should have set the
 1007                  * function to call.
 1008                  */
 1009                 if (dtrace_vtime_active)
 1010                         (*dtrace_vtime_switch_func)(newtd);
 1011 #endif
 1012 
 1013                 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
 1014                 lock_profile_obtain_lock_success(&sched_lock.lock_object,
 1015                     0, 0, __FILE__, __LINE__);
 1016                 /*
 1017                  * Where am I?  What year is it?
 1018                  * We are in the same thread that went to sleep above,
 1019                  * but any amount of time may have passed. All our context
 1020                  * will still be available as will local variables.
 1021                  * PCPU values however may have changed as we may have
 1022                  * changed CPU so don't trust cached values of them.
 1023                  * New threads will go to fork_exit() instead of here
 1024                  * so if you change things here you may need to change
 1025                  * things there too.
 1026                  *
 1027                  * If the thread above was exiting it will never wake
 1028                  * up again here, so either it has saved everything it
 1029                  * needed to, or the thread_wait() or wait() will
 1030                  * need to reap it.
 1031                  */
 1032 #ifdef  HWPMC_HOOKS
 1033                 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 1034                         PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
 1035 #endif
 1036         }
 1037 
 1038 #ifdef SMP
 1039         if (td->td_flags & TDF_IDLETD)
 1040                 idle_cpus_mask |= PCPU_GET(cpumask);
 1041 #endif
 1042         sched_lock.mtx_lock = (uintptr_t)td;
 1043         td->td_oncpu = PCPU_GET(cpuid);
 1044         MPASS(td->td_lock == &sched_lock);
 1045 }
 1046 
 1047 void
 1048 sched_wakeup(struct thread *td)
 1049 {
 1050         struct td_sched *ts;
 1051 
 1052         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1053         ts = td->td_sched;
 1054         td->td_flags &= ~TDF_CANSWAP;
 1055         if (ts->ts_slptime > 1) {
 1056                 updatepri(td);
 1057                 resetpriority(td);
 1058         }
 1059         td->td_slptick = 0;
 1060         ts->ts_slptime = 0;
 1061         sched_add(td, SRQ_BORING);
 1062 }
 1063 
 1064 #ifdef SMP
 1065 static int
 1066 forward_wakeup(int cpunum)
 1067 {
 1068         struct pcpu *pc;
 1069         cpumask_t dontuse, id, map, map2, map3, me;
 1070 
 1071         mtx_assert(&sched_lock, MA_OWNED);
 1072 
 1073         CTR0(KTR_RUNQ, "forward_wakeup()");
 1074 
 1075         if ((!forward_wakeup_enabled) ||
 1076              (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
 1077                 return (0);
 1078         if (!smp_started || cold || panicstr)
 1079                 return (0);
 1080 
 1081         forward_wakeups_requested++;
 1082 
 1083         /*
 1084          * Check the idle mask we received against what we calculated
 1085          * before in the old version.
 1086          */
 1087         me = PCPU_GET(cpumask);
 1088 
 1089         /* Don't bother if we should be doing it ourself. */
 1090         if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
 1091                 return (0);
 1092 
 1093         dontuse = me | stopped_cpus | hlt_cpus_mask;
 1094         map3 = 0;
 1095         if (forward_wakeup_use_loop) {
 1096                 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
 1097                         id = pc->pc_cpumask;
 1098                         if ((id & dontuse) == 0 &&
 1099                             pc->pc_curthread == pc->pc_idlethread) {
 1100                                 map3 |= id;
 1101                         }
 1102                 }
 1103         }
 1104 
 1105         if (forward_wakeup_use_mask) {
 1106                 map = 0;
 1107                 map = idle_cpus_mask & ~dontuse;
 1108 
 1109                 /* If they are both on, compare and use loop if different. */
 1110                 if (forward_wakeup_use_loop) {
 1111                         if (map != map3) {
 1112                                 printf("map (%02X) != map3 (%02X)\n", map,
 1113                                     map3);
 1114                                 map = map3;
 1115                         }
 1116                 }
 1117         } else {
 1118                 map = map3;
 1119         }
 1120 
 1121         /* If we only allow a specific CPU, then mask off all the others. */
 1122         if (cpunum != NOCPU) {
 1123                 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
 1124                 map &= (1 << cpunum);
 1125         } else {
 1126                 /* Try choose an idle die. */
 1127                 if (forward_wakeup_use_htt) {
 1128                         map2 =  (map & (map >> 1)) & 0x5555;
 1129                         if (map2) {
 1130                                 map = map2;
 1131                         }
 1132                 }
 1133 
 1134                 /* Set only one bit. */
 1135                 if (forward_wakeup_use_single) {
 1136                         map = map & ((~map) + 1);
 1137                 }
 1138         }
 1139         if (map) {
 1140                 forward_wakeups_delivered++;
 1141                 ipi_selected(map, IPI_AST);
 1142                 return (1);
 1143         }
 1144         if (cpunum == NOCPU)
 1145                 printf("forward_wakeup: Idle processor not found\n");
 1146         return (0);
 1147 }
 1148 
 1149 static void
 1150 kick_other_cpu(int pri, int cpuid)
 1151 {
 1152         struct pcpu *pcpu;
 1153         int cpri;
 1154 
 1155         pcpu = pcpu_find(cpuid);
 1156         if (idle_cpus_mask & pcpu->pc_cpumask) {
 1157                 forward_wakeups_delivered++;
 1158                 ipi_cpu(cpuid, IPI_AST);
 1159                 return;
 1160         }
 1161 
 1162         cpri = pcpu->pc_curthread->td_priority;
 1163         if (pri >= cpri)
 1164                 return;
 1165 
 1166 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
 1167 #if !defined(FULL_PREEMPTION)
 1168         if (pri <= PRI_MAX_ITHD)
 1169 #endif /* ! FULL_PREEMPTION */
 1170         {
 1171                 ipi_cpu(cpuid, IPI_PREEMPT);
 1172                 return;
 1173         }
 1174 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
 1175 
 1176         pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
 1177         ipi_cpu(cpuid, IPI_AST);
 1178         return;
 1179 }
 1180 #endif /* SMP */
 1181 
 1182 #ifdef SMP
 1183 static int
 1184 sched_pickcpu(struct thread *td)
 1185 {
 1186         int best, cpu;
 1187 
 1188         mtx_assert(&sched_lock, MA_OWNED);
 1189 
 1190         if (THREAD_CAN_SCHED(td, td->td_lastcpu))
 1191                 best = td->td_lastcpu;
 1192         else
 1193                 best = NOCPU;
 1194         CPU_FOREACH(cpu) {
 1195                 if (!THREAD_CAN_SCHED(td, cpu))
 1196                         continue;
 1197         
 1198                 if (best == NOCPU)
 1199                         best = cpu;
 1200                 else if (runq_length[cpu] < runq_length[best])
 1201                         best = cpu;
 1202         }
 1203         KASSERT(best != NOCPU, ("no valid CPUs"));
 1204 
 1205         return (best);
 1206 }
 1207 #endif
 1208 
 1209 void
 1210 sched_add(struct thread *td, int flags)
 1211 #ifdef SMP
 1212 {
 1213         struct td_sched *ts;
 1214         int forwarded = 0;
 1215         int cpu;
 1216         int single_cpu = 0;
 1217 
 1218         ts = td->td_sched;
 1219         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1220         KASSERT((td->td_inhibitors == 0),
 1221             ("sched_add: trying to run inhibited thread"));
 1222         KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
 1223             ("sched_add: bad thread state"));
 1224         KASSERT(td->td_flags & TDF_INMEM,
 1225             ("sched_add: thread swapped out"));
 1226 
 1227         KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
 1228             "prio:%d", td->td_priority, KTR_ATTR_LINKED,
 1229             sched_tdname(curthread));
 1230         KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
 1231             KTR_ATTR_LINKED, sched_tdname(td));
 1232 
 1233 
 1234         /*
 1235          * Now that the thread is moving to the run-queue, set the lock
 1236          * to the scheduler's lock.
 1237          */
 1238         if (td->td_lock != &sched_lock) {
 1239                 mtx_lock_spin(&sched_lock);
 1240                 thread_lock_set(td, &sched_lock);
 1241         }
 1242         TD_SET_RUNQ(td);
 1243 
 1244         /*
 1245          * If SMP is started and the thread is pinned or otherwise limited to
 1246          * a specific set of CPUs, queue the thread to a per-CPU run queue.
 1247          * Otherwise, queue the thread to the global run queue.
 1248          *
 1249          * If SMP has not yet been started we must use the global run queue
 1250          * as per-CPU state may not be initialized yet and we may crash if we
 1251          * try to access the per-CPU run queues.
 1252          */
 1253         if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
 1254             ts->ts_flags & TSF_AFFINITY)) {
 1255                 if (td->td_pinned != 0)
 1256                         cpu = td->td_lastcpu;
 1257                 else if (td->td_flags & TDF_BOUND) {
 1258                         /* Find CPU from bound runq. */
 1259                         KASSERT(SKE_RUNQ_PCPU(ts),
 1260                             ("sched_add: bound td_sched not on cpu runq"));
 1261                         cpu = ts->ts_runq - &runq_pcpu[0];
 1262                 } else
 1263                         /* Find a valid CPU for our cpuset */
 1264                         cpu = sched_pickcpu(td);
 1265                 ts->ts_runq = &runq_pcpu[cpu];
 1266                 single_cpu = 1;
 1267                 CTR3(KTR_RUNQ,
 1268                     "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
 1269                     cpu);
 1270         } else {
 1271                 CTR2(KTR_RUNQ,
 1272                     "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
 1273                     td);
 1274                 cpu = NOCPU;
 1275                 ts->ts_runq = &runq;
 1276         }
 1277 
 1278         if (single_cpu && (cpu != PCPU_GET(cpuid))) {
 1279                 kick_other_cpu(td->td_priority, cpu);
 1280         } else {
 1281                 if (!single_cpu) {
 1282                         cpumask_t me = PCPU_GET(cpumask);
 1283                         cpumask_t idle = idle_cpus_mask & me;
 1284 
 1285                         if (!idle && ((flags & SRQ_INTR) == 0) &&
 1286                             (idle_cpus_mask & ~(hlt_cpus_mask | me)))
 1287                                 forwarded = forward_wakeup(cpu);
 1288                 }
 1289 
 1290                 if (!forwarded) {
 1291                         if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
 1292                                 return;
 1293                         else
 1294                                 maybe_resched(td);
 1295                 }
 1296         }
 1297 
 1298         if ((td->td_flags & TDF_NOLOAD) == 0)
 1299                 sched_load_add();
 1300         runq_add(ts->ts_runq, td, flags);
 1301         if (cpu != NOCPU)
 1302                 runq_length[cpu]++;
 1303 }
 1304 #else /* SMP */
 1305 {
 1306         struct td_sched *ts;
 1307 
 1308         ts = td->td_sched;
 1309         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1310         KASSERT((td->td_inhibitors == 0),
 1311             ("sched_add: trying to run inhibited thread"));
 1312         KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
 1313             ("sched_add: bad thread state"));
 1314         KASSERT(td->td_flags & TDF_INMEM,
 1315             ("sched_add: thread swapped out"));
 1316         KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
 1317             "prio:%d", td->td_priority, KTR_ATTR_LINKED,
 1318             sched_tdname(curthread));
 1319         KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
 1320             KTR_ATTR_LINKED, sched_tdname(td));
 1321 
 1322         /*
 1323          * Now that the thread is moving to the run-queue, set the lock
 1324          * to the scheduler's lock.
 1325          */
 1326         if (td->td_lock != &sched_lock) {
 1327                 mtx_lock_spin(&sched_lock);
 1328                 thread_lock_set(td, &sched_lock);
 1329         }
 1330         TD_SET_RUNQ(td);
 1331         CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
 1332         ts->ts_runq = &runq;
 1333 
 1334         /*
 1335          * If we are yielding (on the way out anyhow) or the thread
 1336          * being saved is US, then don't try be smart about preemption
 1337          * or kicking off another CPU as it won't help and may hinder.
 1338          * In the YIEDLING case, we are about to run whoever is being
 1339          * put in the queue anyhow, and in the OURSELF case, we are
 1340          * puting ourself on the run queue which also only happens
 1341          * when we are about to yield.
 1342          */
 1343         if ((flags & SRQ_YIELDING) == 0) {
 1344                 if (maybe_preempt(td))
 1345                         return;
 1346         }
 1347         if ((td->td_flags & TDF_NOLOAD) == 0)
 1348                 sched_load_add();
 1349         runq_add(ts->ts_runq, td, flags);
 1350         maybe_resched(td);
 1351 }
 1352 #endif /* SMP */
 1353 
 1354 void
 1355 sched_rem(struct thread *td)
 1356 {
 1357         struct td_sched *ts;
 1358 
 1359         ts = td->td_sched;
 1360         KASSERT(td->td_flags & TDF_INMEM,
 1361             ("sched_rem: thread swapped out"));
 1362         KASSERT(TD_ON_RUNQ(td),
 1363             ("sched_rem: thread not on run queue"));
 1364         mtx_assert(&sched_lock, MA_OWNED);
 1365         KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
 1366             "prio:%d", td->td_priority, KTR_ATTR_LINKED,
 1367             sched_tdname(curthread));
 1368 
 1369         if ((td->td_flags & TDF_NOLOAD) == 0)
 1370                 sched_load_rem();
 1371 #ifdef SMP
 1372         if (ts->ts_runq != &runq)
 1373                 runq_length[ts->ts_runq - runq_pcpu]--;
 1374 #endif
 1375         runq_remove(ts->ts_runq, td);
 1376         TD_SET_CAN_RUN(td);
 1377 }
 1378 
 1379 /*
 1380  * Select threads to run.  Note that running threads still consume a
 1381  * slot.
 1382  */
 1383 struct thread *
 1384 sched_choose(void)
 1385 {
 1386         struct thread *td;
 1387         struct runq *rq;
 1388 
 1389         mtx_assert(&sched_lock,  MA_OWNED);
 1390 #ifdef SMP
 1391         struct thread *tdcpu;
 1392 
 1393         rq = &runq;
 1394         td = runq_choose_fuzz(&runq, runq_fuzz);
 1395         tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
 1396 
 1397         if (td == NULL ||
 1398             (tdcpu != NULL &&
 1399              tdcpu->td_priority < td->td_priority)) {
 1400                 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
 1401                      PCPU_GET(cpuid));
 1402                 td = tdcpu;
 1403                 rq = &runq_pcpu[PCPU_GET(cpuid)];
 1404         } else {
 1405                 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
 1406         }
 1407 
 1408 #else
 1409         rq = &runq;
 1410         td = runq_choose(&runq);
 1411 #endif
 1412 
 1413         if (td) {
 1414 #ifdef SMP
 1415                 if (td == tdcpu)
 1416                         runq_length[PCPU_GET(cpuid)]--;
 1417 #endif
 1418                 runq_remove(rq, td);
 1419                 td->td_flags |= TDF_DIDRUN;
 1420 
 1421                 KASSERT(td->td_flags & TDF_INMEM,
 1422                     ("sched_choose: thread swapped out"));
 1423                 return (td);
 1424         }
 1425         return (PCPU_GET(idlethread));
 1426 }
 1427 
 1428 void
 1429 sched_preempt(struct thread *td)
 1430 {
 1431         thread_lock(td);
 1432         if (td->td_critnest > 1)
 1433                 td->td_owepreempt = 1;
 1434         else
 1435                 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
 1436         thread_unlock(td);
 1437 }
 1438 
 1439 void
 1440 sched_userret(struct thread *td)
 1441 {
 1442         /*
 1443          * XXX we cheat slightly on the locking here to avoid locking in
 1444          * the usual case.  Setting td_priority here is essentially an
 1445          * incomplete workaround for not setting it properly elsewhere.
 1446          * Now that some interrupt handlers are threads, not setting it
 1447          * properly elsewhere can clobber it in the window between setting
 1448          * it here and returning to user mode, so don't waste time setting
 1449          * it perfectly here.
 1450          */
 1451         KASSERT((td->td_flags & TDF_BORROWING) == 0,
 1452             ("thread with borrowed priority returning to userland"));
 1453         if (td->td_priority != td->td_user_pri) {
 1454                 thread_lock(td);
 1455                 td->td_priority = td->td_user_pri;
 1456                 td->td_base_pri = td->td_user_pri;
 1457                 thread_unlock(td);
 1458         }
 1459 }
 1460 
 1461 void
 1462 sched_bind(struct thread *td, int cpu)
 1463 {
 1464         struct td_sched *ts;
 1465 
 1466         THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
 1467         KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
 1468 
 1469         ts = td->td_sched;
 1470 
 1471         td->td_flags |= TDF_BOUND;
 1472 #ifdef SMP
 1473         ts->ts_runq = &runq_pcpu[cpu];
 1474         if (PCPU_GET(cpuid) == cpu)
 1475                 return;
 1476 
 1477         mi_switch(SW_VOL, NULL);
 1478 #endif
 1479 }
 1480 
 1481 void
 1482 sched_unbind(struct thread* td)
 1483 {
 1484         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1485         KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
 1486         td->td_flags &= ~TDF_BOUND;
 1487 }
 1488 
 1489 int
 1490 sched_is_bound(struct thread *td)
 1491 {
 1492         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1493         return (td->td_flags & TDF_BOUND);
 1494 }
 1495 
 1496 void
 1497 sched_relinquish(struct thread *td)
 1498 {
 1499         thread_lock(td);
 1500         mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
 1501         thread_unlock(td);
 1502 }
 1503 
 1504 int
 1505 sched_load(void)
 1506 {
 1507         return (sched_tdcnt);
 1508 }
 1509 
 1510 int
 1511 sched_sizeof_proc(void)
 1512 {
 1513         return (sizeof(struct proc));
 1514 }
 1515 
 1516 int
 1517 sched_sizeof_thread(void)
 1518 {
 1519         return (sizeof(struct thread) + sizeof(struct td_sched));
 1520 }
 1521 
 1522 fixpt_t
 1523 sched_pctcpu(struct thread *td)
 1524 {
 1525         struct td_sched *ts;
 1526 
 1527         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1528         ts = td->td_sched;
 1529         return (ts->ts_pctcpu);
 1530 }
 1531 
 1532 void
 1533 sched_tick(void)
 1534 {
 1535 }
 1536 
 1537 /*
 1538  * The actual idle process.
 1539  */
 1540 void
 1541 sched_idletd(void *dummy)
 1542 {
 1543 
 1544         for (;;) {
 1545                 mtx_assert(&Giant, MA_NOTOWNED);
 1546 
 1547                 while (sched_runnable() == 0)
 1548                         cpu_idle(0);
 1549 
 1550                 mtx_lock_spin(&sched_lock);
 1551                 mi_switch(SW_VOL | SWT_IDLE, NULL);
 1552                 mtx_unlock_spin(&sched_lock);
 1553         }
 1554 }
 1555 
 1556 /*
 1557  * A CPU is entering for the first time or a thread is exiting.
 1558  */
 1559 void
 1560 sched_throw(struct thread *td)
 1561 {
 1562         /*
 1563          * Correct spinlock nesting.  The idle thread context that we are
 1564          * borrowing was created so that it would start out with a single
 1565          * spin lock (sched_lock) held in fork_trampoline().  Since we've
 1566          * explicitly acquired locks in this function, the nesting count
 1567          * is now 2 rather than 1.  Since we are nested, calling
 1568          * spinlock_exit() will simply adjust the counts without allowing
 1569          * spin lock using code to interrupt us.
 1570          */
 1571         if (td == NULL) {
 1572                 mtx_lock_spin(&sched_lock);
 1573                 spinlock_exit();
 1574                 PCPU_SET(switchtime, cpu_ticks());
 1575                 PCPU_SET(switchticks, ticks);
 1576         } else {
 1577                 lock_profile_release_lock(&sched_lock.lock_object);
 1578                 MPASS(td->td_lock == &sched_lock);
 1579         }
 1580         mtx_assert(&sched_lock, MA_OWNED);
 1581         KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
 1582         cpu_throw(td, choosethread());  /* doesn't return */
 1583 }
 1584 
 1585 void
 1586 sched_fork_exit(struct thread *td)
 1587 {
 1588 
 1589         /*
 1590          * Finish setting up thread glue so that it begins execution in a
 1591          * non-nested critical section with sched_lock held but not recursed.
 1592          */
 1593         td->td_oncpu = PCPU_GET(cpuid);
 1594         sched_lock.mtx_lock = (uintptr_t)td;
 1595         lock_profile_obtain_lock_success(&sched_lock.lock_object,
 1596             0, 0, __FILE__, __LINE__);
 1597         THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
 1598 }
 1599 
 1600 char *
 1601 sched_tdname(struct thread *td)
 1602 {
 1603 #ifdef KTR
 1604         struct td_sched *ts;
 1605 
 1606         ts = td->td_sched;
 1607         if (ts->ts_name[0] == '\0')
 1608                 snprintf(ts->ts_name, sizeof(ts->ts_name),
 1609                     "%s tid %d", td->td_name, td->td_tid);
 1610         return (ts->ts_name);
 1611 #else   
 1612         return (td->td_name);
 1613 #endif
 1614 }
 1615 
 1616 void
 1617 sched_affinity(struct thread *td)
 1618 {
 1619 #ifdef SMP
 1620         struct td_sched *ts;
 1621         int cpu;
 1622 
 1623         THREAD_LOCK_ASSERT(td, MA_OWNED);       
 1624 
 1625         /*
 1626          * Set the TSF_AFFINITY flag if there is at least one CPU this
 1627          * thread can't run on.
 1628          */
 1629         ts = td->td_sched;
 1630         ts->ts_flags &= ~TSF_AFFINITY;
 1631         CPU_FOREACH(cpu) {
 1632                 if (!THREAD_CAN_SCHED(td, cpu)) {
 1633                         ts->ts_flags |= TSF_AFFINITY;
 1634                         break;
 1635                 }
 1636         }
 1637 
 1638         /*
 1639          * If this thread can run on all CPUs, nothing else to do.
 1640          */
 1641         if (!(ts->ts_flags & TSF_AFFINITY))
 1642                 return;
 1643 
 1644         /* Pinned threads and bound threads should be left alone. */
 1645         if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
 1646                 return;
 1647 
 1648         switch (td->td_state) {
 1649         case TDS_RUNQ:
 1650                 /*
 1651                  * If we are on a per-CPU runqueue that is in the set,
 1652                  * then nothing needs to be done.
 1653                  */
 1654                 if (ts->ts_runq != &runq &&
 1655                     THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
 1656                         return;
 1657 
 1658                 /* Put this thread on a valid per-CPU runqueue. */
 1659                 sched_rem(td);
 1660                 sched_add(td, SRQ_BORING);
 1661                 break;
 1662         case TDS_RUNNING:
 1663                 /*
 1664                  * See if our current CPU is in the set.  If not, force a
 1665                  * context switch.
 1666                  */
 1667                 if (THREAD_CAN_SCHED(td, td->td_oncpu))
 1668                         return;
 1669 
 1670                 td->td_flags |= TDF_NEEDRESCHED;
 1671                 if (td != curthread)
 1672                         ipi_cpu(cpu, IPI_AST);
 1673                 break;
 1674         default:
 1675                 break;
 1676         }
 1677 #endif
 1678 }

Cache object: a9167d8d93f133e9da3fb271a569cf12


[ 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.