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-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) 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("$FreeBSD$");
   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/sdt.h>
   54 #include <sys/smp.h>
   55 #include <sys/sysctl.h>
   56 #include <sys/sx.h>
   57 #include <sys/turnstile.h>
   58 #include <sys/umtx.h>
   59 #include <machine/pcb.h>
   60 #include <machine/smp.h>
   61 
   62 #ifdef HWPMC_HOOKS
   63 #include <sys/pmckern.h>
   64 #endif
   65 
   66 #ifdef KDTRACE_HOOKS
   67 #include <sys/dtrace_bsd.h>
   68 int                             dtrace_vtime_active;
   69 dtrace_vtime_switch_func_t      dtrace_vtime_switch_func;
   70 #endif
   71 
   72 /*
   73  * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in
   74  * the range 100-256 Hz (approximately).
   75  */
   76 #define ESTCPULIM(e) \
   77     min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \
   78     RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1)
   79 #ifdef SMP
   80 #define INVERSE_ESTCPU_WEIGHT   (8 * smp_cpus)
   81 #else
   82 #define INVERSE_ESTCPU_WEIGHT   8       /* 1 / (priorities per estcpu level). */
   83 #endif
   84 #define NICE_WEIGHT             1       /* Priorities per nice level. */
   85 
   86 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
   87 
   88 /*
   89  * The schedulable entity that runs a context.
   90  * This is  an extension to the thread structure and is tailored to
   91  * the requirements of this scheduler
   92  */
   93 struct td_sched {
   94         fixpt_t         ts_pctcpu;      /* (j) %cpu during p_swtime. */
   95         int             ts_cpticks;     /* (j) Ticks of cpu time. */
   96         int             ts_slptime;     /* (j) Seconds !RUNNING. */
   97         int             ts_slice;       /* Remaining part of time slice. */
   98         int             ts_flags;
   99         struct runq     *ts_runq;       /* runq the thread is currently on */
  100 #ifdef KTR
  101         char            ts_name[TS_NAME_LEN];
  102 #endif
  103 };
  104 
  105 /* flags kept in td_flags */
  106 #define TDF_DIDRUN      TDF_SCHED0      /* thread actually ran. */
  107 #define TDF_BOUND       TDF_SCHED1      /* Bound to one CPU. */
  108 #define TDF_SLICEEND    TDF_SCHED2      /* Thread time slice is over. */
  109 
  110 /* flags kept in ts_flags */
  111 #define TSF_AFFINITY    0x0001          /* Has a non-"full" CPU set. */
  112 
  113 #define SKE_RUNQ_PCPU(ts)                                               \
  114     ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq)
  115 
  116 #define THREAD_CAN_SCHED(td, cpu)       \
  117     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
  118 
  119 static struct td_sched td_sched0;
  120 struct mtx sched_lock;
  121 
  122 static int      realstathz = 127; /* stathz is sometimes 0 and run off of hz. */
  123 static int      sched_tdcnt;    /* Total runnable threads in the system. */
  124 static int      sched_slice = 12; /* Thread run time before rescheduling. */
  125 
  126 static void     setup_runqs(void);
  127 static void     schedcpu(void);
  128 static void     schedcpu_thread(void);
  129 static void     sched_priority(struct thread *td, u_char prio);
  130 static void     sched_setup(void *dummy);
  131 static void     maybe_resched(struct thread *td);
  132 static void     updatepri(struct thread *td);
  133 static void     resetpriority(struct thread *td);
  134 static void     resetpriority_thread(struct thread *td);
  135 #ifdef SMP
  136 static int      sched_pickcpu(struct thread *td);
  137 static int      forward_wakeup(int cpunum);
  138 static void     kick_other_cpu(int pri, int cpuid);
  139 #endif
  140 
  141 static struct kproc_desc sched_kp = {
  142         "schedcpu",
  143         schedcpu_thread,
  144         NULL
  145 };
  146 SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start,
  147     &sched_kp);
  148 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
  149 
  150 static void sched_initticks(void *dummy);
  151 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
  152     NULL);
  153 
  154 /*
  155  * Global run queue.
  156  */
  157 static struct runq runq;
  158 
  159 #ifdef SMP
  160 /*
  161  * Per-CPU run queues
  162  */
  163 static struct runq runq_pcpu[MAXCPU];
  164 long runq_length[MAXCPU];
  165 #endif
  166 
  167 struct pcpuidlestat {
  168         u_int idlecalls;
  169         u_int oldidlecalls;
  170 };
  171 static DPCPU_DEFINE(struct pcpuidlestat, idlestat);
  172 
  173 static void
  174 setup_runqs(void)
  175 {
  176 #ifdef SMP
  177         int i;
  178 
  179         for (i = 0; i < MAXCPU; ++i)
  180                 runq_init(&runq_pcpu[i]);
  181 #endif
  182 
  183         runq_init(&runq);
  184 }
  185 
  186 static int
  187 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
  188 {
  189         int error, new_val, period;
  190 
  191         period = 1000000 / realstathz;
  192         new_val = period * sched_slice;
  193         error = sysctl_handle_int(oidp, &new_val, 0, req);
  194         if (error != 0 || req->newptr == NULL)
  195                 return (error);
  196         if (new_val <= 0)
  197                 return (EINVAL);
  198         sched_slice = imax(1, (new_val + period / 2) / period);
  199         hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
  200             realstathz);
  201         return (0);
  202 }
  203 
  204 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler");
  205 
  206 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0,
  207     "Scheduler name");
  208 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
  209     NULL, 0, sysctl_kern_quantum, "I",
  210     "Quantum for timeshare threads in microseconds");
  211 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
  212     "Quantum for timeshare threads in stathz ticks");
  213 #ifdef SMP
  214 /* Enable forwarding of wakeups to all other cpus */
  215 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP");
  216 
  217 static int runq_fuzz = 1;
  218 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, "");
  219 
  220 static int forward_wakeup_enabled = 1;
  221 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW,
  222            &forward_wakeup_enabled, 0,
  223            "Forwarding of wakeup to idle CPUs");
  224 
  225 static int forward_wakeups_requested = 0;
  226 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD,
  227            &forward_wakeups_requested, 0,
  228            "Requests for Forwarding of wakeup to idle CPUs");
  229 
  230 static int forward_wakeups_delivered = 0;
  231 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD,
  232            &forward_wakeups_delivered, 0,
  233            "Completed Forwarding of wakeup to idle CPUs");
  234 
  235 static int forward_wakeup_use_mask = 1;
  236 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW,
  237            &forward_wakeup_use_mask, 0,
  238            "Use the mask of idle cpus");
  239 
  240 static int forward_wakeup_use_loop = 0;
  241 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW,
  242            &forward_wakeup_use_loop, 0,
  243            "Use a loop to find idle cpus");
  244 
  245 static int forward_wakeup_use_single = 0;
  246 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW,
  247            &forward_wakeup_use_single, 0,
  248            "Only signal one idle cpu");
  249 
  250 static int forward_wakeup_use_htt = 0;
  251 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW,
  252            &forward_wakeup_use_htt, 0,
  253            "account for htt");
  254 
  255 #endif
  256 #if 0
  257 static int sched_followon = 0;
  258 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW,
  259            &sched_followon, 0,
  260            "allow threads to share a quantum");
  261 #endif
  262 
  263 SDT_PROVIDER_DEFINE(sched);
  264 
  265 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *", 
  266     "struct proc *", "uint8_t");
  267 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *", 
  268     "struct proc *", "void *");
  269 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *", 
  270     "struct proc *", "void *", "int");
  271 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *", 
  272     "struct proc *", "uint8_t", "struct thread *");
  273 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
  274 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
  275     "struct proc *");
  276 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
  277 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
  278 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
  279     "struct proc *");
  280 
  281 static __inline void
  282 sched_load_add(void)
  283 {
  284 
  285         sched_tdcnt++;
  286         KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
  287         SDT_PROBE2(sched, , , load_change, NOCPU, sched_tdcnt);
  288 }
  289 
  290 static __inline void
  291 sched_load_rem(void)
  292 {
  293 
  294         sched_tdcnt--;
  295         KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
  296         SDT_PROBE2(sched, , , load_change, NOCPU, sched_tdcnt);
  297 }
  298 /*
  299  * Arrange to reschedule if necessary, taking the priorities and
  300  * schedulers into account.
  301  */
  302 static void
  303 maybe_resched(struct thread *td)
  304 {
  305 
  306         THREAD_LOCK_ASSERT(td, MA_OWNED);
  307         if (td->td_priority < curthread->td_priority)
  308                 curthread->td_flags |= TDF_NEEDRESCHED;
  309 }
  310 
  311 /*
  312  * This function is called when a thread is about to be put on run queue
  313  * because it has been made runnable or its priority has been adjusted.  It
  314  * determines if the new thread should be immediately preempted to.  If so,
  315  * it switches to it and eventually returns true.  If not, it returns false
  316  * so that the caller may place the thread on an appropriate run queue.
  317  */
  318 int
  319 maybe_preempt(struct thread *td)
  320 {
  321 #ifdef PREEMPTION
  322         struct thread *ctd;
  323         int cpri, pri;
  324 
  325         /*
  326          * The new thread should not preempt the current thread if any of the
  327          * following conditions are true:
  328          *
  329          *  - The kernel is in the throes of crashing (panicstr).
  330          *  - The current thread has a higher (numerically lower) or
  331          *    equivalent priority.  Note that this prevents curthread from
  332          *    trying to preempt to itself.
  333          *  - It is too early in the boot for context switches (cold is set).
  334          *  - The current thread has an inhibitor set or is in the process of
  335          *    exiting.  In this case, the current thread is about to switch
  336          *    out anyways, so there's no point in preempting.  If we did,
  337          *    the current thread would not be properly resumed as well, so
  338          *    just avoid that whole landmine.
  339          *  - If the new thread's priority is not a realtime priority and
  340          *    the current thread's priority is not an idle priority and
  341          *    FULL_PREEMPTION is disabled.
  342          *
  343          * If all of these conditions are false, but the current thread is in
  344          * a nested critical section, then we have to defer the preemption
  345          * until we exit the critical section.  Otherwise, switch immediately
  346          * to the new thread.
  347          */
  348         ctd = curthread;
  349         THREAD_LOCK_ASSERT(td, MA_OWNED);
  350         KASSERT((td->td_inhibitors == 0),
  351                         ("maybe_preempt: trying to run inhibited thread"));
  352         pri = td->td_priority;
  353         cpri = ctd->td_priority;
  354         if (panicstr != NULL || pri >= cpri || cold /* || dumping */ ||
  355             TD_IS_INHIBITED(ctd))
  356                 return (0);
  357 #ifndef FULL_PREEMPTION
  358         if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
  359                 return (0);
  360 #endif
  361 
  362         if (ctd->td_critnest > 1) {
  363                 CTR1(KTR_PROC, "maybe_preempt: in critical section %d",
  364                     ctd->td_critnest);
  365                 ctd->td_owepreempt = 1;
  366                 return (0);
  367         }
  368         /*
  369          * Thread is runnable but not yet put on system run queue.
  370          */
  371         MPASS(ctd->td_lock == td->td_lock);
  372         MPASS(TD_ON_RUNQ(td));
  373         TD_SET_RUNNING(td);
  374         CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
  375             td->td_proc->p_pid, td->td_name);
  376         mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td);
  377         /*
  378          * td's lock pointer may have changed.  We have to return with it
  379          * locked.
  380          */
  381         spinlock_enter();
  382         thread_unlock(ctd);
  383         thread_lock(td);
  384         spinlock_exit();
  385         return (1);
  386 #else
  387         return (0);
  388 #endif
  389 }
  390 
  391 /*
  392  * Constants for digital decay and forget:
  393  *      90% of (td_estcpu) usage in 5 * loadav time
  394  *      95% of (ts_pctcpu) usage in 60 seconds (load insensitive)
  395  *          Note that, as ps(1) mentions, this can let percentages
  396  *          total over 100% (I've seen 137.9% for 3 processes).
  397  *
  398  * Note that schedclock() updates td_estcpu and p_cpticks asynchronously.
  399  *
  400  * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds.
  401  * That is, the system wants to compute a value of decay such
  402  * that the following for loop:
  403  *      for (i = 0; i < (5 * loadavg); i++)
  404  *              td_estcpu *= decay;
  405  * will compute
  406  *      td_estcpu *= 0.1;
  407  * for all values of loadavg:
  408  *
  409  * Mathematically this loop can be expressed by saying:
  410  *      decay ** (5 * loadavg) ~= .1
  411  *
  412  * The system computes decay as:
  413  *      decay = (2 * loadavg) / (2 * loadavg + 1)
  414  *
  415  * We wish to prove that the system's computation of decay
  416  * will always fulfill the equation:
  417  *      decay ** (5 * loadavg) ~= .1
  418  *
  419  * If we compute b as:
  420  *      b = 2 * loadavg
  421  * then
  422  *      decay = b / (b + 1)
  423  *
  424  * We now need to prove two things:
  425  *      1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
  426  *      2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
  427  *
  428  * Facts:
  429  *         For x close to zero, exp(x) =~ 1 + x, since
  430  *              exp(x) = 0! + x**1/1! + x**2/2! + ... .
  431  *              therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
  432  *         For x close to zero, ln(1+x) =~ x, since
  433  *              ln(1+x) = x - x**2/2 + x**3/3 - ...     -1 < x < 1
  434  *              therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
  435  *         ln(.1) =~ -2.30
  436  *
  437  * Proof of (1):
  438  *    Solve (factor)**(power) =~ .1 given power (5*loadav):
  439  *      solving for factor,
  440  *      ln(factor) =~ (-2.30/5*loadav), or
  441  *      factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
  442  *          exp(-1/b) =~ (b-1)/b =~ b/(b+1).                    QED
  443  *
  444  * Proof of (2):
  445  *    Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
  446  *      solving for power,
  447  *      power*ln(b/(b+1)) =~ -2.30, or
  448  *      power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav.  QED
  449  *
  450  * Actual power values for the implemented algorithm are as follows:
  451  *      loadav: 1       2       3       4
  452  *      power:  5.68    10.32   14.94   19.55
  453  */
  454 
  455 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */
  456 #define loadfactor(loadav)      (2 * (loadav))
  457 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
  458 
  459 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
  460 static fixpt_t  ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
  461 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
  462 
  463 /*
  464  * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
  465  * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
  466  * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
  467  *
  468  * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
  469  *      1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
  470  *
  471  * If you don't want to bother with the faster/more-accurate formula, you
  472  * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
  473  * (more general) method of calculating the %age of CPU used by a process.
  474  */
  475 #define CCPU_SHIFT      11
  476 
  477 /*
  478  * Recompute process priorities, every hz ticks.
  479  * MP-safe, called without the Giant mutex.
  480  */
  481 /* ARGSUSED */
  482 static void
  483 schedcpu(void)
  484 {
  485         register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
  486         struct thread *td;
  487         struct proc *p;
  488         struct td_sched *ts;
  489         int awake;
  490 
  491         sx_slock(&allproc_lock);
  492         FOREACH_PROC_IN_SYSTEM(p) {
  493                 PROC_LOCK(p);
  494                 if (p->p_state == PRS_NEW) {
  495                         PROC_UNLOCK(p);
  496                         continue;
  497                 }
  498                 FOREACH_THREAD_IN_PROC(p, td) {
  499                         awake = 0;
  500                         thread_lock(td);
  501                         ts = td->td_sched;
  502                         /*
  503                          * Increment sleep time (if sleeping).  We
  504                          * ignore overflow, as above.
  505                          */
  506                         /*
  507                          * The td_sched slptimes are not touched in wakeup
  508                          * because the thread may not HAVE everything in
  509                          * memory? XXX I think this is out of date.
  510                          */
  511                         if (TD_ON_RUNQ(td)) {
  512                                 awake = 1;
  513                                 td->td_flags &= ~TDF_DIDRUN;
  514                         } else if (TD_IS_RUNNING(td)) {
  515                                 awake = 1;
  516                                 /* Do not clear TDF_DIDRUN */
  517                         } else if (td->td_flags & TDF_DIDRUN) {
  518                                 awake = 1;
  519                                 td->td_flags &= ~TDF_DIDRUN;
  520                         }
  521 
  522                         /*
  523                          * ts_pctcpu is only for ps and ttyinfo().
  524                          */
  525                         ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT;
  526                         /*
  527                          * If the td_sched has been idle the entire second,
  528                          * stop recalculating its priority until
  529                          * it wakes up.
  530                          */
  531                         if (ts->ts_cpticks != 0) {
  532 #if     (FSHIFT >= CCPU_SHIFT)
  533                                 ts->ts_pctcpu += (realstathz == 100)
  534                                     ? ((fixpt_t) ts->ts_cpticks) <<
  535                                     (FSHIFT - CCPU_SHIFT) :
  536                                     100 * (((fixpt_t) ts->ts_cpticks)
  537                                     << (FSHIFT - CCPU_SHIFT)) / realstathz;
  538 #else
  539                                 ts->ts_pctcpu += ((FSCALE - ccpu) *
  540                                     (ts->ts_cpticks *
  541                                     FSCALE / realstathz)) >> FSHIFT;
  542 #endif
  543                                 ts->ts_cpticks = 0;
  544                         }
  545                         /*
  546                          * If there are ANY running threads in this process,
  547                          * then don't count it as sleeping.
  548                          * XXX: this is broken.
  549                          */
  550                         if (awake) {
  551                                 if (ts->ts_slptime > 1) {
  552                                         /*
  553                                          * In an ideal world, this should not
  554                                          * happen, because whoever woke us
  555                                          * up from the long sleep should have
  556                                          * unwound the slptime and reset our
  557                                          * priority before we run at the stale
  558                                          * priority.  Should KASSERT at some
  559                                          * point when all the cases are fixed.
  560                                          */
  561                                         updatepri(td);
  562                                 }
  563                                 ts->ts_slptime = 0;
  564                         } else
  565                                 ts->ts_slptime++;
  566                         if (ts->ts_slptime > 1) {
  567                                 thread_unlock(td);
  568                                 continue;
  569                         }
  570                         td->td_estcpu = decay_cpu(loadfac, td->td_estcpu);
  571                         resetpriority(td);
  572                         resetpriority_thread(td);
  573                         thread_unlock(td);
  574                 }
  575                 PROC_UNLOCK(p);
  576         }
  577         sx_sunlock(&allproc_lock);
  578 }
  579 
  580 /*
  581  * Main loop for a kthread that executes schedcpu once a second.
  582  */
  583 static void
  584 schedcpu_thread(void)
  585 {
  586 
  587         for (;;) {
  588                 schedcpu();
  589                 pause("-", hz);
  590         }
  591 }
  592 
  593 /*
  594  * Recalculate the priority of a process after it has slept for a while.
  595  * For all load averages >= 1 and max td_estcpu of 255, sleeping for at
  596  * least six times the loadfactor will decay td_estcpu to zero.
  597  */
  598 static void
  599 updatepri(struct thread *td)
  600 {
  601         struct td_sched *ts;
  602         fixpt_t loadfac;
  603         unsigned int newcpu;
  604 
  605         ts = td->td_sched;
  606         loadfac = loadfactor(averunnable.ldavg[0]);
  607         if (ts->ts_slptime > 5 * loadfac)
  608                 td->td_estcpu = 0;
  609         else {
  610                 newcpu = td->td_estcpu;
  611                 ts->ts_slptime--;       /* was incremented in schedcpu() */
  612                 while (newcpu && --ts->ts_slptime)
  613                         newcpu = decay_cpu(loadfac, newcpu);
  614                 td->td_estcpu = newcpu;
  615         }
  616 }
  617 
  618 /*
  619  * Compute the priority of a process when running in user mode.
  620  * Arrange to reschedule if the resulting priority is better
  621  * than that of the current process.
  622  */
  623 static void
  624 resetpriority(struct thread *td)
  625 {
  626         register unsigned int newpriority;
  627 
  628         if (td->td_pri_class == PRI_TIMESHARE) {
  629                 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT +
  630                     NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN);
  631                 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE),
  632                     PRI_MAX_TIMESHARE);
  633                 sched_user_prio(td, newpriority);
  634         }
  635 }
  636 
  637 /*
  638  * Update the thread's priority when the associated process's user
  639  * priority changes.
  640  */
  641 static void
  642 resetpriority_thread(struct thread *td)
  643 {
  644 
  645         /* Only change threads with a time sharing user priority. */
  646         if (td->td_priority < PRI_MIN_TIMESHARE ||
  647             td->td_priority > PRI_MAX_TIMESHARE)
  648                 return;
  649 
  650         /* XXX the whole needresched thing is broken, but not silly. */
  651         maybe_resched(td);
  652 
  653         sched_prio(td, td->td_user_pri);
  654 }
  655 
  656 /* ARGSUSED */
  657 static void
  658 sched_setup(void *dummy)
  659 {
  660 
  661         setup_runqs();
  662 
  663         /* Account for thread0. */
  664         sched_load_add();
  665 }
  666 
  667 /*
  668  * This routine determines time constants after stathz and hz are setup.
  669  */
  670 static void
  671 sched_initticks(void *dummy)
  672 {
  673 
  674         realstathz = stathz ? stathz : hz;
  675         sched_slice = realstathz / 10;  /* ~100ms */
  676         hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
  677             realstathz);
  678 }
  679 
  680 /* External interfaces start here */
  681 
  682 /*
  683  * Very early in the boot some setup of scheduler-specific
  684  * parts of proc0 and of some scheduler resources needs to be done.
  685  * Called from:
  686  *  proc0_init()
  687  */
  688 void
  689 schedinit(void)
  690 {
  691         /*
  692          * Set up the scheduler specific parts of proc0.
  693          */
  694         proc0.p_sched = NULL; /* XXX */
  695         thread0.td_sched = &td_sched0;
  696         thread0.td_lock = &sched_lock;
  697         td_sched0.ts_slice = sched_slice;
  698         mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE);
  699 }
  700 
  701 int
  702 sched_runnable(void)
  703 {
  704 #ifdef SMP
  705         return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]);
  706 #else
  707         return runq_check(&runq);
  708 #endif
  709 }
  710 
  711 int
  712 sched_rr_interval(void)
  713 {
  714 
  715         /* Convert sched_slice from stathz to hz. */
  716         return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
  717 }
  718 
  719 /*
  720  * We adjust the priority of the current process.  The priority of
  721  * a process gets worse as it accumulates CPU time.  The cpu usage
  722  * estimator (td_estcpu) is increased here.  resetpriority() will
  723  * compute a different priority each time td_estcpu increases by
  724  * INVERSE_ESTCPU_WEIGHT
  725  * (until MAXPRI is reached).  The cpu usage estimator ramps up
  726  * quite quickly when the process is running (linearly), and decays
  727  * away exponentially, at a rate which is proportionally slower when
  728  * the system is busy.  The basic principle is that the system will
  729  * 90% forget that the process used a lot of CPU time in 5 * loadav
  730  * seconds.  This causes the system to favor processes which haven't
  731  * run much recently, and to round-robin among other processes.
  732  */
  733 void
  734 sched_clock(struct thread *td)
  735 {
  736         struct pcpuidlestat *stat;
  737         struct td_sched *ts;
  738 
  739         THREAD_LOCK_ASSERT(td, MA_OWNED);
  740         ts = td->td_sched;
  741 
  742         ts->ts_cpticks++;
  743         td->td_estcpu = ESTCPULIM(td->td_estcpu + 1);
  744         if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) {
  745                 resetpriority(td);
  746                 resetpriority_thread(td);
  747         }
  748 
  749         /*
  750          * Force a context switch if the current thread has used up a full
  751          * time slice (default is 100ms).
  752          */
  753         if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
  754                 ts->ts_slice = sched_slice;
  755                 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
  756         }
  757 
  758         stat = DPCPU_PTR(idlestat);
  759         stat->oldidlecalls = stat->idlecalls;
  760         stat->idlecalls = 0;
  761 }
  762 
  763 /*
  764  * Charge child's scheduling CPU usage to parent.
  765  */
  766 void
  767 sched_exit(struct proc *p, struct thread *td)
  768 {
  769 
  770         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit",
  771             "prio:%d", td->td_priority);
  772 
  773         PROC_LOCK_ASSERT(p, MA_OWNED);
  774         sched_exit_thread(FIRST_THREAD_IN_PROC(p), td);
  775 }
  776 
  777 void
  778 sched_exit_thread(struct thread *td, struct thread *child)
  779 {
  780 
  781         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit",
  782             "prio:%d", child->td_priority);
  783         thread_lock(td);
  784         td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu);
  785         thread_unlock(td);
  786         thread_lock(child);
  787         if ((child->td_flags & TDF_NOLOAD) == 0)
  788                 sched_load_rem();
  789         thread_unlock(child);
  790 }
  791 
  792 void
  793 sched_fork(struct thread *td, struct thread *childtd)
  794 {
  795         sched_fork_thread(td, childtd);
  796 }
  797 
  798 void
  799 sched_fork_thread(struct thread *td, struct thread *childtd)
  800 {
  801         struct td_sched *ts;
  802 
  803         childtd->td_estcpu = td->td_estcpu;
  804         childtd->td_lock = &sched_lock;
  805         childtd->td_cpuset = cpuset_ref(td->td_cpuset);
  806         childtd->td_priority = childtd->td_base_pri;
  807         ts = childtd->td_sched;
  808         bzero(ts, sizeof(*ts));
  809         ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY);
  810         ts->ts_slice = 1;
  811 }
  812 
  813 void
  814 sched_nice(struct proc *p, int nice)
  815 {
  816         struct thread *td;
  817 
  818         PROC_LOCK_ASSERT(p, MA_OWNED);
  819         p->p_nice = nice;
  820         FOREACH_THREAD_IN_PROC(p, td) {
  821                 thread_lock(td);
  822                 resetpriority(td);
  823                 resetpriority_thread(td);
  824                 thread_unlock(td);
  825         }
  826 }
  827 
  828 void
  829 sched_class(struct thread *td, int class)
  830 {
  831         THREAD_LOCK_ASSERT(td, MA_OWNED);
  832         td->td_pri_class = class;
  833 }
  834 
  835 /*
  836  * Adjust the priority of a thread.
  837  */
  838 static void
  839 sched_priority(struct thread *td, u_char prio)
  840 {
  841 
  842 
  843         KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change",
  844             "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED,
  845             sched_tdname(curthread));
  846         SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
  847         if (td != curthread && prio > td->td_priority) {
  848                 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
  849                     "lend prio", "prio:%d", td->td_priority, "new prio:%d",
  850                     prio, KTR_ATTR_LINKED, sched_tdname(td));
  851                 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio, 
  852                     curthread);
  853         }
  854         THREAD_LOCK_ASSERT(td, MA_OWNED);
  855         if (td->td_priority == prio)
  856                 return;
  857         td->td_priority = prio;
  858         if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) {
  859                 sched_rem(td);
  860                 sched_add(td, SRQ_BORING);
  861         }
  862 }
  863 
  864 /*
  865  * Update a thread's priority when it is lent another thread's
  866  * priority.
  867  */
  868 void
  869 sched_lend_prio(struct thread *td, u_char prio)
  870 {
  871 
  872         td->td_flags |= TDF_BORROWING;
  873         sched_priority(td, prio);
  874 }
  875 
  876 /*
  877  * Restore a thread's priority when priority propagation is
  878  * over.  The prio argument is the minimum priority the thread
  879  * needs to have to satisfy other possible priority lending
  880  * requests.  If the thread's regulary priority is less
  881  * important than prio the thread will keep a priority boost
  882  * of prio.
  883  */
  884 void
  885 sched_unlend_prio(struct thread *td, u_char prio)
  886 {
  887         u_char base_pri;
  888 
  889         if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
  890             td->td_base_pri <= PRI_MAX_TIMESHARE)
  891                 base_pri = td->td_user_pri;
  892         else
  893                 base_pri = td->td_base_pri;
  894         if (prio >= base_pri) {
  895                 td->td_flags &= ~TDF_BORROWING;
  896                 sched_prio(td, base_pri);
  897         } else
  898                 sched_lend_prio(td, prio);
  899 }
  900 
  901 void
  902 sched_prio(struct thread *td, u_char prio)
  903 {
  904         u_char oldprio;
  905 
  906         /* First, update the base priority. */
  907         td->td_base_pri = prio;
  908 
  909         /*
  910          * If the thread is borrowing another thread's priority, don't ever
  911          * lower the priority.
  912          */
  913         if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
  914                 return;
  915 
  916         /* Change the real priority. */
  917         oldprio = td->td_priority;
  918         sched_priority(td, prio);
  919 
  920         /*
  921          * If the thread is on a turnstile, then let the turnstile update
  922          * its state.
  923          */
  924         if (TD_ON_LOCK(td) && oldprio != prio)
  925                 turnstile_adjust(td, oldprio);
  926 }
  927 
  928 void
  929 sched_user_prio(struct thread *td, u_char prio)
  930 {
  931         u_char oldprio;
  932 
  933         THREAD_LOCK_ASSERT(td, MA_OWNED);
  934         td->td_base_user_pri = prio;
  935         if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
  936                 return;
  937         oldprio = td->td_user_pri;
  938         td->td_user_pri = prio;
  939 }
  940 
  941 void
  942 sched_lend_user_prio(struct thread *td, u_char prio)
  943 {
  944         u_char oldprio;
  945 
  946         THREAD_LOCK_ASSERT(td, MA_OWNED);
  947         td->td_flags |= TDF_UBORROWING;
  948         oldprio = td->td_user_pri;
  949         td->td_user_pri = prio;
  950 }
  951 
  952 void
  953 sched_unlend_user_prio(struct thread *td, u_char prio)
  954 {
  955         u_char base_pri;
  956 
  957         THREAD_LOCK_ASSERT(td, MA_OWNED);
  958         base_pri = td->td_base_user_pri;
  959         if (prio >= base_pri) {
  960                 td->td_flags &= ~TDF_UBORROWING;
  961                 sched_user_prio(td, base_pri);
  962         } else {
  963                 sched_lend_user_prio(td, prio);
  964         }
  965 }
  966 
  967 void
  968 sched_sleep(struct thread *td, int pri)
  969 {
  970 
  971         THREAD_LOCK_ASSERT(td, MA_OWNED);
  972         td->td_slptick = ticks;
  973         td->td_sched->ts_slptime = 0;
  974         if (pri)
  975                 sched_prio(td, pri);
  976         if (TD_IS_SUSPENDED(td) || pri >= PSOCK)
  977                 td->td_flags |= TDF_CANSWAP;
  978 }
  979 
  980 void
  981 sched_switch(struct thread *td, struct thread *newtd, int flags)
  982 {
  983         struct mtx *tmtx;
  984         struct td_sched *ts;
  985         struct proc *p;
  986         int preempted;
  987 
  988         tmtx = NULL;
  989         ts = td->td_sched;
  990         p = td->td_proc;
  991 
  992         THREAD_LOCK_ASSERT(td, MA_OWNED);
  993 
  994         /* 
  995          * Switch to the sched lock to fix things up and pick
  996          * a new thread.
  997          * Block the td_lock in order to avoid breaking the critical path.
  998          */
  999         if (td->td_lock != &sched_lock) {
 1000                 mtx_lock_spin(&sched_lock);
 1001                 tmtx = thread_lock_block(td);
 1002         }
 1003 
 1004         if ((td->td_flags & TDF_NOLOAD) == 0)
 1005                 sched_load_rem();
 1006 
 1007         td->td_lastcpu = td->td_oncpu;
 1008         preempted = !(td->td_flags & TDF_SLICEEND);
 1009         td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
 1010         td->td_owepreempt = 0;
 1011         td->td_oncpu = NOCPU;
 1012 
 1013         /*
 1014          * At the last moment, if this thread is still marked RUNNING,
 1015          * then put it back on the run queue as it has not been suspended
 1016          * or stopped or any thing else similar.  We never put the idle
 1017          * threads on the run queue, however.
 1018          */
 1019         if (td->td_flags & TDF_IDLETD) {
 1020                 TD_SET_CAN_RUN(td);
 1021 #ifdef SMP
 1022                 idle_cpus_mask &= ~PCPU_GET(cpumask);
 1023 #endif
 1024         } else {
 1025                 if (TD_IS_RUNNING(td)) {
 1026                         /* Put us back on the run queue. */
 1027                         sched_add(td, preempted ?
 1028                             SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
 1029                             SRQ_OURSELF|SRQ_YIELDING);
 1030                 }
 1031         }
 1032         if (newtd) {
 1033                 /*
 1034                  * The thread we are about to run needs to be counted
 1035                  * as if it had been added to the run queue and selected.
 1036                  * It came from:
 1037                  * * A preemption
 1038                  * * An upcall
 1039                  * * A followon
 1040                  */
 1041                 KASSERT((newtd->td_inhibitors == 0),
 1042                         ("trying to run inhibited thread"));
 1043                 newtd->td_flags |= TDF_DIDRUN;
 1044                 TD_SET_RUNNING(newtd);
 1045                 if ((newtd->td_flags & TDF_NOLOAD) == 0)
 1046                         sched_load_add();
 1047         } else {
 1048                 newtd = choosethread();
 1049                 MPASS(newtd->td_lock == &sched_lock);
 1050         }
 1051 
 1052         if (td != newtd) {
 1053 #ifdef  HWPMC_HOOKS
 1054                 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 1055                         PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
 1056 #endif
 1057 
 1058                 SDT_PROBE2(sched, , , off_cpu, newtd, newtd->td_proc);
 1059 
 1060                 /* I feel sleepy */
 1061                 lock_profile_release_lock(&sched_lock.lock_object);
 1062 #ifdef KDTRACE_HOOKS
 1063                 /*
 1064                  * If DTrace has set the active vtime enum to anything
 1065                  * other than INACTIVE (0), then it should have set the
 1066                  * function to call.
 1067                  */
 1068                 if (dtrace_vtime_active)
 1069                         (*dtrace_vtime_switch_func)(newtd);
 1070 #endif
 1071 
 1072                 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
 1073                 lock_profile_obtain_lock_success(&sched_lock.lock_object,
 1074                     0, 0, __FILE__, __LINE__);
 1075                 /*
 1076                  * Where am I?  What year is it?
 1077                  * We are in the same thread that went to sleep above,
 1078                  * but any amount of time may have passed. All our context
 1079                  * will still be available as will local variables.
 1080                  * PCPU values however may have changed as we may have
 1081                  * changed CPU so don't trust cached values of them.
 1082                  * New threads will go to fork_exit() instead of here
 1083                  * so if you change things here you may need to change
 1084                  * things there too.
 1085                  *
 1086                  * If the thread above was exiting it will never wake
 1087                  * up again here, so either it has saved everything it
 1088                  * needed to, or the thread_wait() or wait() will
 1089                  * need to reap it.
 1090                  */
 1091 
 1092                 SDT_PROBE0(sched, , , on_cpu);
 1093 #ifdef  HWPMC_HOOKS
 1094                 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 1095                         PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
 1096 #endif
 1097         } else
 1098                 SDT_PROBE0(sched, , , remain_cpu);
 1099 
 1100 #ifdef SMP
 1101         if (td->td_flags & TDF_IDLETD)
 1102                 idle_cpus_mask |= PCPU_GET(cpumask);
 1103 #endif
 1104         sched_lock.mtx_lock = (uintptr_t)td;
 1105         td->td_oncpu = PCPU_GET(cpuid);
 1106         MPASS(td->td_lock == &sched_lock);
 1107 }
 1108 
 1109 void
 1110 sched_wakeup(struct thread *td)
 1111 {
 1112         struct td_sched *ts;
 1113 
 1114         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1115         ts = td->td_sched;
 1116         td->td_flags &= ~TDF_CANSWAP;
 1117         if (ts->ts_slptime > 1) {
 1118                 updatepri(td);
 1119                 resetpriority(td);
 1120         }
 1121         td->td_slptick = 0;
 1122         ts->ts_slptime = 0;
 1123         ts->ts_slice = sched_slice;
 1124         sched_add(td, SRQ_BORING);
 1125 }
 1126 
 1127 #ifdef SMP
 1128 static int
 1129 forward_wakeup(int cpunum)
 1130 {
 1131         struct pcpu *pc;
 1132         cpumask_t dontuse, id, map, map2, map3, me;
 1133 
 1134         mtx_assert(&sched_lock, MA_OWNED);
 1135 
 1136         CTR0(KTR_RUNQ, "forward_wakeup()");
 1137 
 1138         if ((!forward_wakeup_enabled) ||
 1139              (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0))
 1140                 return (0);
 1141         if (!smp_started || cold || panicstr)
 1142                 return (0);
 1143 
 1144         forward_wakeups_requested++;
 1145 
 1146         /*
 1147          * Check the idle mask we received against what we calculated
 1148          * before in the old version.
 1149          */
 1150         me = PCPU_GET(cpumask);
 1151 
 1152         /* Don't bother if we should be doing it ourself. */
 1153         if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum)))
 1154                 return (0);
 1155 
 1156         dontuse = me | stopped_cpus | hlt_cpus_mask;
 1157         map3 = 0;
 1158         if (forward_wakeup_use_loop) {
 1159                 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
 1160                         id = pc->pc_cpumask;
 1161                         if ((id & dontuse) == 0 &&
 1162                             pc->pc_curthread == pc->pc_idlethread) {
 1163                                 map3 |= id;
 1164                         }
 1165                 }
 1166         }
 1167 
 1168         if (forward_wakeup_use_mask) {
 1169                 map = 0;
 1170                 map = idle_cpus_mask & ~dontuse;
 1171 
 1172                 /* If they are both on, compare and use loop if different. */
 1173                 if (forward_wakeup_use_loop) {
 1174                         if (map != map3) {
 1175                                 printf("map (%02X) != map3 (%02X)\n", map,
 1176                                     map3);
 1177                                 map = map3;
 1178                         }
 1179                 }
 1180         } else {
 1181                 map = map3;
 1182         }
 1183 
 1184         /* If we only allow a specific CPU, then mask off all the others. */
 1185         if (cpunum != NOCPU) {
 1186                 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum."));
 1187                 map &= (1 << cpunum);
 1188         } else {
 1189                 /* Try choose an idle die. */
 1190                 if (forward_wakeup_use_htt) {
 1191                         map2 =  (map & (map >> 1)) & 0x5555;
 1192                         if (map2) {
 1193                                 map = map2;
 1194                         }
 1195                 }
 1196 
 1197                 /* Set only one bit. */
 1198                 if (forward_wakeup_use_single) {
 1199                         map = map & ((~map) + 1);
 1200                 }
 1201         }
 1202         if (map) {
 1203                 forward_wakeups_delivered++;
 1204                 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) {
 1205                         id = pc->pc_cpumask;
 1206                         if ((map & id) == 0)
 1207                                 continue;
 1208                         if (cpu_idle_wakeup(pc->pc_cpuid))
 1209                                 map &= ~id;
 1210                 }
 1211                 if (map)
 1212                         ipi_selected(map, IPI_AST);
 1213                 return (1);
 1214         }
 1215         if (cpunum == NOCPU)
 1216                 printf("forward_wakeup: Idle processor not found\n");
 1217         return (0);
 1218 }
 1219 
 1220 static void
 1221 kick_other_cpu(int pri, int cpuid)
 1222 {
 1223         struct pcpu *pcpu;
 1224         int cpri;
 1225 
 1226         pcpu = pcpu_find(cpuid);
 1227         if (idle_cpus_mask & pcpu->pc_cpumask) {
 1228                 forward_wakeups_delivered++;
 1229                 if (!cpu_idle_wakeup(cpuid))
 1230                         ipi_cpu(cpuid, IPI_AST);
 1231                 return;
 1232         }
 1233 
 1234         cpri = pcpu->pc_curthread->td_priority;
 1235         if (pri >= cpri)
 1236                 return;
 1237 
 1238 #if defined(IPI_PREEMPTION) && defined(PREEMPTION)
 1239 #if !defined(FULL_PREEMPTION)
 1240         if (pri <= PRI_MAX_ITHD)
 1241 #endif /* ! FULL_PREEMPTION */
 1242         {
 1243                 ipi_cpu(cpuid, IPI_PREEMPT);
 1244                 return;
 1245         }
 1246 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */
 1247 
 1248         pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED;
 1249         ipi_cpu(cpuid, IPI_AST);
 1250         return;
 1251 }
 1252 #endif /* SMP */
 1253 
 1254 #ifdef SMP
 1255 static int
 1256 sched_pickcpu(struct thread *td)
 1257 {
 1258         int best, cpu;
 1259 
 1260         mtx_assert(&sched_lock, MA_OWNED);
 1261 
 1262         if (THREAD_CAN_SCHED(td, td->td_lastcpu))
 1263                 best = td->td_lastcpu;
 1264         else
 1265                 best = NOCPU;
 1266         CPU_FOREACH(cpu) {
 1267                 if (!THREAD_CAN_SCHED(td, cpu))
 1268                         continue;
 1269         
 1270                 if (best == NOCPU)
 1271                         best = cpu;
 1272                 else if (runq_length[cpu] < runq_length[best])
 1273                         best = cpu;
 1274         }
 1275         KASSERT(best != NOCPU, ("no valid CPUs"));
 1276 
 1277         return (best);
 1278 }
 1279 #endif
 1280 
 1281 void
 1282 sched_add(struct thread *td, int flags)
 1283 #ifdef SMP
 1284 {
 1285         struct td_sched *ts;
 1286         int forwarded = 0;
 1287         int cpu;
 1288         int single_cpu = 0;
 1289 
 1290         ts = td->td_sched;
 1291         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1292         KASSERT((td->td_inhibitors == 0),
 1293             ("sched_add: trying to run inhibited thread"));
 1294         KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
 1295             ("sched_add: bad thread state"));
 1296         KASSERT(td->td_flags & TDF_INMEM,
 1297             ("sched_add: thread swapped out"));
 1298 
 1299         KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
 1300             "prio:%d", td->td_priority, KTR_ATTR_LINKED,
 1301             sched_tdname(curthread));
 1302         KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
 1303             KTR_ATTR_LINKED, sched_tdname(td));
 1304         SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 
 1305             flags & SRQ_PREEMPTED);
 1306 
 1307 
 1308         /*
 1309          * Now that the thread is moving to the run-queue, set the lock
 1310          * to the scheduler's lock.
 1311          */
 1312         if (td->td_lock != &sched_lock) {
 1313                 mtx_lock_spin(&sched_lock);
 1314                 thread_lock_set(td, &sched_lock);
 1315         }
 1316         TD_SET_RUNQ(td);
 1317 
 1318         /*
 1319          * If SMP is started and the thread is pinned or otherwise limited to
 1320          * a specific set of CPUs, queue the thread to a per-CPU run queue.
 1321          * Otherwise, queue the thread to the global run queue.
 1322          *
 1323          * If SMP has not yet been started we must use the global run queue
 1324          * as per-CPU state may not be initialized yet and we may crash if we
 1325          * try to access the per-CPU run queues.
 1326          */
 1327         if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND ||
 1328             ts->ts_flags & TSF_AFFINITY)) {
 1329                 if (td->td_pinned != 0)
 1330                         cpu = td->td_lastcpu;
 1331                 else if (td->td_flags & TDF_BOUND) {
 1332                         /* Find CPU from bound runq. */
 1333                         KASSERT(SKE_RUNQ_PCPU(ts),
 1334                             ("sched_add: bound td_sched not on cpu runq"));
 1335                         cpu = ts->ts_runq - &runq_pcpu[0];
 1336                 } else
 1337                         /* Find a valid CPU for our cpuset */
 1338                         cpu = sched_pickcpu(td);
 1339                 ts->ts_runq = &runq_pcpu[cpu];
 1340                 single_cpu = 1;
 1341                 CTR3(KTR_RUNQ,
 1342                     "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td,
 1343                     cpu);
 1344         } else {
 1345                 CTR2(KTR_RUNQ,
 1346                     "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts,
 1347                     td);
 1348                 cpu = NOCPU;
 1349                 ts->ts_runq = &runq;
 1350         }
 1351 
 1352         if (single_cpu && (cpu != PCPU_GET(cpuid))) {
 1353                 kick_other_cpu(td->td_priority, cpu);
 1354         } else {
 1355                 if (!single_cpu) {
 1356                         cpumask_t me = PCPU_GET(cpumask);
 1357                         cpumask_t idle = idle_cpus_mask & me;
 1358 
 1359                         if (!idle && ((flags & SRQ_INTR) == 0) &&
 1360                             (idle_cpus_mask & ~(hlt_cpus_mask | me)))
 1361                                 forwarded = forward_wakeup(cpu);
 1362                 }
 1363 
 1364                 if (!forwarded) {
 1365                         if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td))
 1366                                 return;
 1367                         else
 1368                                 maybe_resched(td);
 1369                 }
 1370         }
 1371 
 1372         if ((td->td_flags & TDF_NOLOAD) == 0)
 1373                 sched_load_add();
 1374         runq_add(ts->ts_runq, td, flags);
 1375         if (cpu != NOCPU)
 1376                 runq_length[cpu]++;
 1377 }
 1378 #else /* SMP */
 1379 {
 1380         struct td_sched *ts;
 1381 
 1382         ts = td->td_sched;
 1383         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1384         KASSERT((td->td_inhibitors == 0),
 1385             ("sched_add: trying to run inhibited thread"));
 1386         KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
 1387             ("sched_add: bad thread state"));
 1388         KASSERT(td->td_flags & TDF_INMEM,
 1389             ("sched_add: thread swapped out"));
 1390         KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
 1391             "prio:%d", td->td_priority, KTR_ATTR_LINKED,
 1392             sched_tdname(curthread));
 1393         KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
 1394             KTR_ATTR_LINKED, sched_tdname(td));
 1395         SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 
 1396             flags & SRQ_PREEMPTED);
 1397 
 1398         /*
 1399          * Now that the thread is moving to the run-queue, set the lock
 1400          * to the scheduler's lock.
 1401          */
 1402         if (td->td_lock != &sched_lock) {
 1403                 mtx_lock_spin(&sched_lock);
 1404                 thread_lock_set(td, &sched_lock);
 1405         }
 1406         TD_SET_RUNQ(td);
 1407         CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td);
 1408         ts->ts_runq = &runq;
 1409 
 1410         /*
 1411          * If we are yielding (on the way out anyhow) or the thread
 1412          * being saved is US, then don't try be smart about preemption
 1413          * or kicking off another CPU as it won't help and may hinder.
 1414          * In the YIEDLING case, we are about to run whoever is being
 1415          * put in the queue anyhow, and in the OURSELF case, we are
 1416          * puting ourself on the run queue which also only happens
 1417          * when we are about to yield.
 1418          */
 1419         if ((flags & SRQ_YIELDING) == 0) {
 1420                 if (maybe_preempt(td))
 1421                         return;
 1422         }
 1423         if ((td->td_flags & TDF_NOLOAD) == 0)
 1424                 sched_load_add();
 1425         runq_add(ts->ts_runq, td, flags);
 1426         maybe_resched(td);
 1427 }
 1428 #endif /* SMP */
 1429 
 1430 void
 1431 sched_rem(struct thread *td)
 1432 {
 1433         struct td_sched *ts;
 1434 
 1435         ts = td->td_sched;
 1436         KASSERT(td->td_flags & TDF_INMEM,
 1437             ("sched_rem: thread swapped out"));
 1438         KASSERT(TD_ON_RUNQ(td),
 1439             ("sched_rem: thread not on run queue"));
 1440         mtx_assert(&sched_lock, MA_OWNED);
 1441         KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
 1442             "prio:%d", td->td_priority, KTR_ATTR_LINKED,
 1443             sched_tdname(curthread));
 1444         SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
 1445 
 1446         if ((td->td_flags & TDF_NOLOAD) == 0)
 1447                 sched_load_rem();
 1448 #ifdef SMP
 1449         if (ts->ts_runq != &runq)
 1450                 runq_length[ts->ts_runq - runq_pcpu]--;
 1451 #endif
 1452         runq_remove(ts->ts_runq, td);
 1453         TD_SET_CAN_RUN(td);
 1454 }
 1455 
 1456 /*
 1457  * Select threads to run.  Note that running threads still consume a
 1458  * slot.
 1459  */
 1460 struct thread *
 1461 sched_choose(void)
 1462 {
 1463         struct thread *td;
 1464         struct runq *rq;
 1465 
 1466         mtx_assert(&sched_lock,  MA_OWNED);
 1467 #ifdef SMP
 1468         struct thread *tdcpu;
 1469 
 1470         rq = &runq;
 1471         td = runq_choose_fuzz(&runq, runq_fuzz);
 1472         tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]);
 1473 
 1474         if (td == NULL ||
 1475             (tdcpu != NULL &&
 1476              tdcpu->td_priority < td->td_priority)) {
 1477                 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu,
 1478                      PCPU_GET(cpuid));
 1479                 td = tdcpu;
 1480                 rq = &runq_pcpu[PCPU_GET(cpuid)];
 1481         } else {
 1482                 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td);
 1483         }
 1484 
 1485 #else
 1486         rq = &runq;
 1487         td = runq_choose(&runq);
 1488 #endif
 1489 
 1490         if (td) {
 1491 #ifdef SMP
 1492                 if (td == tdcpu)
 1493                         runq_length[PCPU_GET(cpuid)]--;
 1494 #endif
 1495                 runq_remove(rq, td);
 1496                 td->td_flags |= TDF_DIDRUN;
 1497 
 1498                 KASSERT(td->td_flags & TDF_INMEM,
 1499                     ("sched_choose: thread swapped out"));
 1500                 return (td);
 1501         }
 1502         return (PCPU_GET(idlethread));
 1503 }
 1504 
 1505 void
 1506 sched_preempt(struct thread *td)
 1507 {
 1508 
 1509         SDT_PROBE2(sched, , , surrender, td, td->td_proc);
 1510         thread_lock(td);
 1511         if (td->td_critnest > 1)
 1512                 td->td_owepreempt = 1;
 1513         else
 1514                 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL);
 1515         thread_unlock(td);
 1516 }
 1517 
 1518 void
 1519 sched_userret(struct thread *td)
 1520 {
 1521         /*
 1522          * XXX we cheat slightly on the locking here to avoid locking in
 1523          * the usual case.  Setting td_priority here is essentially an
 1524          * incomplete workaround for not setting it properly elsewhere.
 1525          * Now that some interrupt handlers are threads, not setting it
 1526          * properly elsewhere can clobber it in the window between setting
 1527          * it here and returning to user mode, so don't waste time setting
 1528          * it perfectly here.
 1529          */
 1530         KASSERT((td->td_flags & TDF_BORROWING) == 0,
 1531             ("thread with borrowed priority returning to userland"));
 1532         if (td->td_priority != td->td_user_pri) {
 1533                 thread_lock(td);
 1534                 td->td_priority = td->td_user_pri;
 1535                 td->td_base_pri = td->td_user_pri;
 1536                 thread_unlock(td);
 1537         }
 1538 }
 1539 
 1540 void
 1541 sched_bind(struct thread *td, int cpu)
 1542 {
 1543         struct td_sched *ts;
 1544 
 1545         THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
 1546         KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
 1547 
 1548         ts = td->td_sched;
 1549 
 1550         td->td_flags |= TDF_BOUND;
 1551 #ifdef SMP
 1552         ts->ts_runq = &runq_pcpu[cpu];
 1553         if (PCPU_GET(cpuid) == cpu)
 1554                 return;
 1555 
 1556         mi_switch(SW_VOL, NULL);
 1557 #endif
 1558 }
 1559 
 1560 void
 1561 sched_unbind(struct thread* td)
 1562 {
 1563         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1564         KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
 1565         td->td_flags &= ~TDF_BOUND;
 1566 }
 1567 
 1568 int
 1569 sched_is_bound(struct thread *td)
 1570 {
 1571         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1572         return (td->td_flags & TDF_BOUND);
 1573 }
 1574 
 1575 void
 1576 sched_relinquish(struct thread *td)
 1577 {
 1578         thread_lock(td);
 1579         mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
 1580         thread_unlock(td);
 1581 }
 1582 
 1583 int
 1584 sched_load(void)
 1585 {
 1586         return (sched_tdcnt);
 1587 }
 1588 
 1589 int
 1590 sched_sizeof_proc(void)
 1591 {
 1592         return (sizeof(struct proc));
 1593 }
 1594 
 1595 int
 1596 sched_sizeof_thread(void)
 1597 {
 1598         return (sizeof(struct thread) + sizeof(struct td_sched));
 1599 }
 1600 
 1601 fixpt_t
 1602 sched_pctcpu(struct thread *td)
 1603 {
 1604         struct td_sched *ts;
 1605 
 1606         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1607         ts = td->td_sched;
 1608         return (ts->ts_pctcpu);
 1609 }
 1610 
 1611 void
 1612 sched_tick(void)
 1613 {
 1614 }
 1615 
 1616 /*
 1617  * The actual idle process.
 1618  */
 1619 void
 1620 sched_idletd(void *dummy)
 1621 {
 1622         struct pcpuidlestat *stat;
 1623 
 1624         stat = DPCPU_PTR(idlestat);
 1625         for (;;) {
 1626                 mtx_assert(&Giant, MA_NOTOWNED);
 1627 
 1628                 while (sched_runnable() == 0) {
 1629                         cpu_idle(stat->idlecalls + stat->oldidlecalls > 64);
 1630                         stat->idlecalls++;
 1631                 }
 1632 
 1633                 mtx_lock_spin(&sched_lock);
 1634                 mi_switch(SW_VOL | SWT_IDLE, NULL);
 1635                 mtx_unlock_spin(&sched_lock);
 1636         }
 1637 }
 1638 
 1639 /*
 1640  * A CPU is entering for the first time or a thread is exiting.
 1641  */
 1642 void
 1643 sched_throw(struct thread *td)
 1644 {
 1645         /*
 1646          * Correct spinlock nesting.  The idle thread context that we are
 1647          * borrowing was created so that it would start out with a single
 1648          * spin lock (sched_lock) held in fork_trampoline().  Since we've
 1649          * explicitly acquired locks in this function, the nesting count
 1650          * is now 2 rather than 1.  Since we are nested, calling
 1651          * spinlock_exit() will simply adjust the counts without allowing
 1652          * spin lock using code to interrupt us.
 1653          */
 1654         if (td == NULL) {
 1655                 mtx_lock_spin(&sched_lock);
 1656                 spinlock_exit();
 1657                 PCPU_SET(switchtime, cpu_ticks());
 1658                 PCPU_SET(switchticks, ticks);
 1659         } else {
 1660                 lock_profile_release_lock(&sched_lock.lock_object);
 1661                 MPASS(td->td_lock == &sched_lock);
 1662         }
 1663         mtx_assert(&sched_lock, MA_OWNED);
 1664         KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
 1665         cpu_throw(td, choosethread());  /* doesn't return */
 1666 }
 1667 
 1668 void
 1669 sched_fork_exit(struct thread *td)
 1670 {
 1671 
 1672         /*
 1673          * Finish setting up thread glue so that it begins execution in a
 1674          * non-nested critical section with sched_lock held but not recursed.
 1675          */
 1676         td->td_oncpu = PCPU_GET(cpuid);
 1677         sched_lock.mtx_lock = (uintptr_t)td;
 1678         lock_profile_obtain_lock_success(&sched_lock.lock_object,
 1679             0, 0, __FILE__, __LINE__);
 1680         THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
 1681 }
 1682 
 1683 char *
 1684 sched_tdname(struct thread *td)
 1685 {
 1686 #ifdef KTR
 1687         struct td_sched *ts;
 1688 
 1689         ts = td->td_sched;
 1690         if (ts->ts_name[0] == '\0')
 1691                 snprintf(ts->ts_name, sizeof(ts->ts_name),
 1692                     "%s tid %d", td->td_name, td->td_tid);
 1693         return (ts->ts_name);
 1694 #else   
 1695         return (td->td_name);
 1696 #endif
 1697 }
 1698 
 1699 #ifdef KTR
 1700 void
 1701 sched_clear_tdname(struct thread *td)
 1702 {
 1703         struct td_sched *ts;
 1704 
 1705         ts = td->td_sched;
 1706         ts->ts_name[0] = '\0';
 1707 }
 1708 #endif
 1709 
 1710 void
 1711 sched_affinity(struct thread *td)
 1712 {
 1713 #ifdef SMP
 1714         struct td_sched *ts;
 1715         int cpu;
 1716 
 1717         THREAD_LOCK_ASSERT(td, MA_OWNED);       
 1718 
 1719         /*
 1720          * Set the TSF_AFFINITY flag if there is at least one CPU this
 1721          * thread can't run on.
 1722          */
 1723         ts = td->td_sched;
 1724         ts->ts_flags &= ~TSF_AFFINITY;
 1725         CPU_FOREACH(cpu) {
 1726                 if (!THREAD_CAN_SCHED(td, cpu)) {
 1727                         ts->ts_flags |= TSF_AFFINITY;
 1728                         break;
 1729                 }
 1730         }
 1731 
 1732         /*
 1733          * If this thread can run on all CPUs, nothing else to do.
 1734          */
 1735         if (!(ts->ts_flags & TSF_AFFINITY))
 1736                 return;
 1737 
 1738         /* Pinned threads and bound threads should be left alone. */
 1739         if (td->td_pinned != 0 || td->td_flags & TDF_BOUND)
 1740                 return;
 1741 
 1742         switch (td->td_state) {
 1743         case TDS_RUNQ:
 1744                 /*
 1745                  * If we are on a per-CPU runqueue that is in the set,
 1746                  * then nothing needs to be done.
 1747                  */
 1748                 if (ts->ts_runq != &runq &&
 1749                     THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu))
 1750                         return;
 1751 
 1752                 /* Put this thread on a valid per-CPU runqueue. */
 1753                 sched_rem(td);
 1754                 sched_add(td, SRQ_BORING);
 1755                 break;
 1756         case TDS_RUNNING:
 1757                 /*
 1758                  * See if our current CPU is in the set.  If not, force a
 1759                  * context switch.
 1760                  */
 1761                 if (THREAD_CAN_SCHED(td, td->td_oncpu))
 1762                         return;
 1763 
 1764                 td->td_flags |= TDF_NEEDRESCHED;
 1765                 if (td != curthread)
 1766                         ipi_cpu(cpu, IPI_AST);
 1767                 break;
 1768         default:
 1769                 break;
 1770         }
 1771 #endif
 1772 }

Cache object: 1d85e7c6f0433a4cc9aeec3713cf6dac


[ 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.