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_ule.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) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
    3  * All rights reserved.
    4  *
    5  * Redistribution and use in source and binary forms, with or without
    6  * modification, are permitted provided that the following conditions
    7  * are met:
    8  * 1. Redistributions of source code must retain the above copyright
    9  *    notice unmodified, this list of conditions, and the following
   10  *    disclaimer.
   11  * 2. Redistributions in binary form must reproduce the above copyright
   12  *    notice, this list of conditions and the following disclaimer in the
   13  *    documentation and/or other materials provided with the distribution.
   14  *
   15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
   16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
   17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
   18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
   19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
   20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
   21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
   22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
   23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
   24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
   25  */
   26 
   27 /*
   28  * This file implements the ULE scheduler.  ULE supports independent CPU
   29  * run queues and fine grain locking.  It has superior interactive
   30  * performance under load even on uni-processor systems.
   31  *
   32  * etymology:
   33  *   ULE is the last three letters in schedule.  It owes its name to a
   34  * generic user created for a scheduling system by Paul Mikesell at
   35  * Isilon Systems and a general lack of creativity on the part of the author.
   36  */
   37 
   38 #include <sys/cdefs.h>
   39 __FBSDID("$FreeBSD$");
   40 
   41 #include "opt_hwpmc_hooks.h"
   42 #include "opt_kdtrace.h"
   43 #include "opt_sched.h"
   44 
   45 #include <sys/param.h>
   46 #include <sys/systm.h>
   47 #include <sys/kdb.h>
   48 #include <sys/kernel.h>
   49 #include <sys/ktr.h>
   50 #include <sys/lock.h>
   51 #include <sys/mutex.h>
   52 #include <sys/proc.h>
   53 #include <sys/resource.h>
   54 #include <sys/resourcevar.h>
   55 #include <sys/sched.h>
   56 #include <sys/sdt.h>
   57 #include <sys/smp.h>
   58 #include <sys/sx.h>
   59 #include <sys/sysctl.h>
   60 #include <sys/sysproto.h>
   61 #include <sys/turnstile.h>
   62 #include <sys/umtx.h>
   63 #include <sys/vmmeter.h>
   64 #include <sys/cpuset.h>
   65 #include <sys/sbuf.h>
   66 
   67 #ifdef HWPMC_HOOKS
   68 #include <sys/pmckern.h>
   69 #endif
   70 
   71 #ifdef KDTRACE_HOOKS
   72 #include <sys/dtrace_bsd.h>
   73 int                             dtrace_vtime_active;
   74 dtrace_vtime_switch_func_t      dtrace_vtime_switch_func;
   75 #endif
   76 
   77 #include <machine/cpu.h>
   78 #include <machine/smp.h>
   79 
   80 #if defined(__powerpc__) && defined(E500)
   81 #error "This architecture is not currently compatible with ULE"
   82 #endif
   83 
   84 #define KTR_ULE 0
   85 
   86 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
   87 #define TDQ_NAME_LEN    (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
   88 #define TDQ_LOADNAME_LEN        (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
   89 
   90 /*
   91  * Thread scheduler specific section.  All fields are protected
   92  * by the thread lock.
   93  */
   94 struct td_sched {       
   95         struct runq     *ts_runq;       /* Run-queue we're queued on. */
   96         short           ts_flags;       /* TSF_* flags. */
   97         u_char          ts_cpu;         /* CPU that we have affinity for. */
   98         int             ts_rltick;      /* Real last tick, for affinity. */
   99         int             ts_slice;       /* Ticks of slice remaining. */
  100         u_int           ts_slptime;     /* Number of ticks we vol. slept */
  101         u_int           ts_runtime;     /* Number of ticks we were running */
  102         int             ts_ltick;       /* Last tick that we were running on */
  103         int             ts_ftick;       /* First tick that we were running on */
  104         int             ts_ticks;       /* Tick count */
  105 #ifdef KTR
  106         char            ts_name[TS_NAME_LEN];
  107 #endif
  108 };
  109 /* flags kept in ts_flags */
  110 #define TSF_BOUND       0x0001          /* Thread can not migrate. */
  111 #define TSF_XFERABLE    0x0002          /* Thread was added as transferable. */
  112 
  113 static struct td_sched td_sched0;
  114 
  115 #define THREAD_CAN_MIGRATE(td)  ((td)->td_pinned == 0)
  116 #define THREAD_CAN_SCHED(td, cpu)       \
  117     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
  118 
  119 /*
  120  * Priority ranges used for interactive and non-interactive timeshare
  121  * threads.  The timeshare priorities are split up into four ranges.
  122  * The first range handles interactive threads.  The last three ranges
  123  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
  124  * ranges supporting nice values.
  125  */
  126 #define PRI_TIMESHARE_RANGE     (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
  127 #define PRI_INTERACT_RANGE      ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
  128 #define PRI_BATCH_RANGE         (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
  129 
  130 #define PRI_MIN_INTERACT        PRI_MIN_TIMESHARE
  131 #define PRI_MAX_INTERACT        (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
  132 #define PRI_MIN_BATCH           (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
  133 #define PRI_MAX_BATCH           PRI_MAX_TIMESHARE
  134 
  135 /*
  136  * Cpu percentage computation macros and defines.
  137  *
  138  * SCHED_TICK_SECS:     Number of seconds to average the cpu usage across.
  139  * SCHED_TICK_TARG:     Number of hz ticks to average the cpu usage across.
  140  * SCHED_TICK_MAX:      Maximum number of ticks before scaling back.
  141  * SCHED_TICK_SHIFT:    Shift factor to avoid rounding away results.
  142  * SCHED_TICK_HZ:       Compute the number of hz ticks for a given ticks count.
  143  * SCHED_TICK_TOTAL:    Gives the amount of time we've been recording ticks.
  144  */
  145 #define SCHED_TICK_SECS         10
  146 #define SCHED_TICK_TARG         (hz * SCHED_TICK_SECS)
  147 #define SCHED_TICK_MAX          (SCHED_TICK_TARG + hz)
  148 #define SCHED_TICK_SHIFT        10
  149 #define SCHED_TICK_HZ(ts)       ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
  150 #define SCHED_TICK_TOTAL(ts)    (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
  151 
  152 /*
  153  * These macros determine priorities for non-interactive threads.  They are
  154  * assigned a priority based on their recent cpu utilization as expressed
  155  * by the ratio of ticks to the tick total.  NHALF priorities at the start
  156  * and end of the MIN to MAX timeshare range are only reachable with negative
  157  * or positive nice respectively.
  158  *
  159  * PRI_RANGE:   Priority range for utilization dependent priorities.
  160  * PRI_NRESV:   Number of nice values.
  161  * PRI_TICKS:   Compute a priority in PRI_RANGE from the ticks count and total.
  162  * PRI_NICE:    Determines the part of the priority inherited from nice.
  163  */
  164 #define SCHED_PRI_NRESV         (PRIO_MAX - PRIO_MIN)
  165 #define SCHED_PRI_NHALF         (SCHED_PRI_NRESV / 2)
  166 #define SCHED_PRI_MIN           (PRI_MIN_BATCH + SCHED_PRI_NHALF)
  167 #define SCHED_PRI_MAX           (PRI_MAX_BATCH - SCHED_PRI_NHALF)
  168 #define SCHED_PRI_RANGE         (SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
  169 #define SCHED_PRI_TICKS(ts)                                             \
  170     (SCHED_TICK_HZ((ts)) /                                              \
  171     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
  172 #define SCHED_PRI_NICE(nice)    (nice)
  173 
  174 /*
  175  * These determine the interactivity of a process.  Interactivity differs from
  176  * cpu utilization in that it expresses the voluntary time slept vs time ran
  177  * while cpu utilization includes all time not running.  This more accurately
  178  * models the intent of the thread.
  179  *
  180  * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
  181  *              before throttling back.
  182  * SLP_RUN_FORK:        Maximum slp+run time to inherit at fork time.
  183  * INTERACT_MAX:        Maximum interactivity value.  Smaller is better.
  184  * INTERACT_THRESH:     Threshold for placement on the current runq.
  185  */
  186 #define SCHED_SLP_RUN_MAX       ((hz * 5) << SCHED_TICK_SHIFT)
  187 #define SCHED_SLP_RUN_FORK      ((hz / 2) << SCHED_TICK_SHIFT)
  188 #define SCHED_INTERACT_MAX      (100)
  189 #define SCHED_INTERACT_HALF     (SCHED_INTERACT_MAX / 2)
  190 #define SCHED_INTERACT_THRESH   (30)
  191 
  192 /* Flags kept in td_flags. */
  193 #define TDF_SLICEEND    TDF_SCHED2      /* Thread time slice is over. */
  194 
  195 /*
  196  * tickincr:            Converts a stathz tick into a hz domain scaled by
  197  *                      the shift factor.  Without the shift the error rate
  198  *                      due to rounding would be unacceptably high.
  199  * realstathz:          stathz is sometimes 0 and run off of hz.
  200  * sched_slice:         Runtime of each thread before rescheduling.
  201  * preempt_thresh:      Priority threshold for preemption and remote IPIs.
  202  */
  203 static int sched_interact = SCHED_INTERACT_THRESH;
  204 static int realstathz = 127;
  205 static int tickincr = 8 << SCHED_TICK_SHIFT;
  206 static int sched_slice = 12;
  207 #ifdef PREEMPTION
  208 #ifdef FULL_PREEMPTION
  209 static int preempt_thresh = PRI_MAX_IDLE;
  210 #else
  211 static int preempt_thresh = PRI_MIN_KERN;
  212 #endif
  213 #else 
  214 static int preempt_thresh = 0;
  215 #endif
  216 static int static_boost = PRI_MIN_BATCH;
  217 static int sched_idlespins = 10000;
  218 static int sched_idlespinthresh = -1;
  219 
  220 /*
  221  * tdq - per processor runqs and statistics.  All fields are protected by the
  222  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
  223  * locking in sched_pickcpu();
  224  */
  225 struct tdq {
  226         /* Ordered to improve efficiency of cpu_search() and switch(). */
  227         struct mtx      tdq_lock;               /* run queue lock. */
  228         struct cpu_group *tdq_cg;               /* Pointer to cpu topology. */
  229         volatile int    tdq_load;               /* Aggregate load. */
  230         volatile int    tdq_cpu_idle;           /* cpu_idle() is active. */
  231         int             tdq_sysload;            /* For loadavg, !ITHD load. */
  232         int             tdq_transferable;       /* Transferable thread count. */
  233         short           tdq_switchcnt;          /* Switches this tick. */
  234         short           tdq_oldswitchcnt;       /* Switches last tick. */
  235         u_char          tdq_lowpri;             /* Lowest priority thread. */
  236         u_char          tdq_ipipending;         /* IPI pending. */
  237         u_char          tdq_idx;                /* Current insert index. */
  238         u_char          tdq_ridx;               /* Current removal index. */
  239         struct runq     tdq_realtime;           /* real-time run queue. */
  240         struct runq     tdq_timeshare;          /* timeshare run queue. */
  241         struct runq     tdq_idle;               /* Queue of IDLE threads. */
  242         char            tdq_name[TDQ_NAME_LEN];
  243 #ifdef KTR
  244         char            tdq_loadname[TDQ_LOADNAME_LEN];
  245 #endif
  246 } __aligned(64);
  247 
  248 /* Idle thread states and config. */
  249 #define TDQ_RUNNING     1
  250 #define TDQ_IDLE        2
  251 
  252 #ifdef SMP
  253 struct cpu_group *cpu_top;              /* CPU topology */
  254 
  255 #define SCHED_AFFINITY_DEFAULT  (max(1, hz / 1000))
  256 #define SCHED_AFFINITY(ts, t)   ((ts)->ts_rltick > ticks - ((t) * affinity))
  257 
  258 /*
  259  * Run-time tunables.
  260  */
  261 static int rebalance = 1;
  262 static int balance_interval = 128;      /* Default set in sched_initticks(). */
  263 static int affinity;
  264 static int steal_idle = 1;
  265 static int steal_thresh = 2;
  266 
  267 /*
  268  * One thread queue per processor.
  269  */
  270 static struct tdq       tdq_cpu[MAXCPU];
  271 static struct tdq       *balance_tdq;
  272 static int balance_ticks;
  273 static DPCPU_DEFINE(uint32_t, randomval);
  274 
  275 #define TDQ_SELF()      (&tdq_cpu[PCPU_GET(cpuid)])
  276 #define TDQ_CPU(x)      (&tdq_cpu[(x)])
  277 #define TDQ_ID(x)       ((int)((x) - tdq_cpu))
  278 #else   /* !SMP */
  279 static struct tdq       tdq_cpu;
  280 
  281 #define TDQ_ID(x)       (0)
  282 #define TDQ_SELF()      (&tdq_cpu)
  283 #define TDQ_CPU(x)      (&tdq_cpu)
  284 #endif
  285 
  286 #define TDQ_LOCK_ASSERT(t, type)        mtx_assert(TDQ_LOCKPTR((t)), (type))
  287 #define TDQ_LOCK(t)             mtx_lock_spin(TDQ_LOCKPTR((t)))
  288 #define TDQ_LOCK_FLAGS(t, f)    mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
  289 #define TDQ_UNLOCK(t)           mtx_unlock_spin(TDQ_LOCKPTR((t)))
  290 #define TDQ_LOCKPTR(t)          (&(t)->tdq_lock)
  291 
  292 static void sched_priority(struct thread *);
  293 static void sched_thread_priority(struct thread *, u_char);
  294 static int sched_interact_score(struct thread *);
  295 static void sched_interact_update(struct thread *);
  296 static void sched_interact_fork(struct thread *);
  297 static void sched_pctcpu_update(struct td_sched *, int);
  298 
  299 /* Operations on per processor queues */
  300 static struct thread *tdq_choose(struct tdq *);
  301 static void tdq_setup(struct tdq *);
  302 static void tdq_load_add(struct tdq *, struct thread *);
  303 static void tdq_load_rem(struct tdq *, struct thread *);
  304 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
  305 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
  306 static inline int sched_shouldpreempt(int, int, int);
  307 void tdq_print(int cpu);
  308 static void runq_print(struct runq *rq);
  309 static void tdq_add(struct tdq *, struct thread *, int);
  310 #ifdef SMP
  311 static int tdq_move(struct tdq *, struct tdq *);
  312 static int tdq_idled(struct tdq *);
  313 static void tdq_notify(struct tdq *, struct thread *);
  314 static struct thread *tdq_steal(struct tdq *, int);
  315 static struct thread *runq_steal(struct runq *, int);
  316 static int sched_pickcpu(struct thread *, int);
  317 static void sched_balance(void);
  318 static int sched_balance_pair(struct tdq *, struct tdq *);
  319 static inline struct tdq *sched_setcpu(struct thread *, int, int);
  320 static inline void thread_unblock_switch(struct thread *, struct mtx *);
  321 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
  322 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
  323 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 
  324     struct cpu_group *cg, int indent);
  325 #endif
  326 
  327 static void sched_setup(void *dummy);
  328 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
  329 
  330 static void sched_initticks(void *dummy);
  331 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
  332     NULL);
  333 
  334 SDT_PROVIDER_DEFINE(sched);
  335 
  336 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 
  337     "struct proc *", "uint8_t");
  338 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 
  339     "struct proc *", "void *");
  340 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 
  341     "struct proc *", "void *", "int");
  342 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 
  343     "struct proc *", "uint8_t", "struct thread *");
  344 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
  345 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 
  346     "struct proc *");
  347 SDT_PROBE_DEFINE(sched, , , on__cpu);
  348 SDT_PROBE_DEFINE(sched, , , remain__cpu);
  349 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 
  350     "struct proc *");
  351 
  352 /*
  353  * Print the threads waiting on a run-queue.
  354  */
  355 static void
  356 runq_print(struct runq *rq)
  357 {
  358         struct rqhead *rqh;
  359         struct thread *td;
  360         int pri;
  361         int j;
  362         int i;
  363 
  364         for (i = 0; i < RQB_LEN; i++) {
  365                 printf("\t\trunq bits %d 0x%zx\n",
  366                     i, rq->rq_status.rqb_bits[i]);
  367                 for (j = 0; j < RQB_BPW; j++)
  368                         if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
  369                                 pri = j + (i << RQB_L2BPW);
  370                                 rqh = &rq->rq_queues[pri];
  371                                 TAILQ_FOREACH(td, rqh, td_runq) {
  372                                         printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
  373                                             td, td->td_name, td->td_priority,
  374                                             td->td_rqindex, pri);
  375                                 }
  376                         }
  377         }
  378 }
  379 
  380 /*
  381  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
  382  */
  383 void
  384 tdq_print(int cpu)
  385 {
  386         struct tdq *tdq;
  387 
  388         tdq = TDQ_CPU(cpu);
  389 
  390         printf("tdq %d:\n", TDQ_ID(tdq));
  391         printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
  392         printf("\tLock name:      %s\n", tdq->tdq_name);
  393         printf("\tload:           %d\n", tdq->tdq_load);
  394         printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
  395         printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
  396         printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
  397         printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
  398         printf("\tload transferable: %d\n", tdq->tdq_transferable);
  399         printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
  400         printf("\trealtime runq:\n");
  401         runq_print(&tdq->tdq_realtime);
  402         printf("\ttimeshare runq:\n");
  403         runq_print(&tdq->tdq_timeshare);
  404         printf("\tidle runq:\n");
  405         runq_print(&tdq->tdq_idle);
  406 }
  407 
  408 static inline int
  409 sched_shouldpreempt(int pri, int cpri, int remote)
  410 {
  411         /*
  412          * If the new priority is not better than the current priority there is
  413          * nothing to do.
  414          */
  415         if (pri >= cpri)
  416                 return (0);
  417         /*
  418          * Always preempt idle.
  419          */
  420         if (cpri >= PRI_MIN_IDLE)
  421                 return (1);
  422         /*
  423          * If preemption is disabled don't preempt others.
  424          */
  425         if (preempt_thresh == 0)
  426                 return (0);
  427         /*
  428          * Preempt if we exceed the threshold.
  429          */
  430         if (pri <= preempt_thresh)
  431                 return (1);
  432         /*
  433          * If we're interactive or better and there is non-interactive
  434          * or worse running preempt only remote processors.
  435          */
  436         if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
  437                 return (1);
  438         return (0);
  439 }
  440 
  441 /*
  442  * Add a thread to the actual run-queue.  Keeps transferable counts up to
  443  * date with what is actually on the run-queue.  Selects the correct
  444  * queue position for timeshare threads.
  445  */
  446 static __inline void
  447 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
  448 {
  449         struct td_sched *ts;
  450         u_char pri;
  451 
  452         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  453         THREAD_LOCK_ASSERT(td, MA_OWNED);
  454 
  455         pri = td->td_priority;
  456         ts = td->td_sched;
  457         TD_SET_RUNQ(td);
  458         if (THREAD_CAN_MIGRATE(td)) {
  459                 tdq->tdq_transferable++;
  460                 ts->ts_flags |= TSF_XFERABLE;
  461         }
  462         if (pri < PRI_MIN_BATCH) {
  463                 ts->ts_runq = &tdq->tdq_realtime;
  464         } else if (pri <= PRI_MAX_BATCH) {
  465                 ts->ts_runq = &tdq->tdq_timeshare;
  466                 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
  467                         ("Invalid priority %d on timeshare runq", pri));
  468                 /*
  469                  * This queue contains only priorities between MIN and MAX
  470                  * realtime.  Use the whole queue to represent these values.
  471                  */
  472                 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
  473                         pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
  474                         pri = (pri + tdq->tdq_idx) % RQ_NQS;
  475                         /*
  476                          * This effectively shortens the queue by one so we
  477                          * can have a one slot difference between idx and
  478                          * ridx while we wait for threads to drain.
  479                          */
  480                         if (tdq->tdq_ridx != tdq->tdq_idx &&
  481                             pri == tdq->tdq_ridx)
  482                                 pri = (unsigned char)(pri - 1) % RQ_NQS;
  483                 } else
  484                         pri = tdq->tdq_ridx;
  485                 runq_add_pri(ts->ts_runq, td, pri, flags);
  486                 return;
  487         } else
  488                 ts->ts_runq = &tdq->tdq_idle;
  489         runq_add(ts->ts_runq, td, flags);
  490 }
  491 
  492 /* 
  493  * Remove a thread from a run-queue.  This typically happens when a thread
  494  * is selected to run.  Running threads are not on the queue and the
  495  * transferable count does not reflect them.
  496  */
  497 static __inline void
  498 tdq_runq_rem(struct tdq *tdq, struct thread *td)
  499 {
  500         struct td_sched *ts;
  501 
  502         ts = td->td_sched;
  503         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  504         KASSERT(ts->ts_runq != NULL,
  505             ("tdq_runq_remove: thread %p null ts_runq", td));
  506         if (ts->ts_flags & TSF_XFERABLE) {
  507                 tdq->tdq_transferable--;
  508                 ts->ts_flags &= ~TSF_XFERABLE;
  509         }
  510         if (ts->ts_runq == &tdq->tdq_timeshare) {
  511                 if (tdq->tdq_idx != tdq->tdq_ridx)
  512                         runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
  513                 else
  514                         runq_remove_idx(ts->ts_runq, td, NULL);
  515         } else
  516                 runq_remove(ts->ts_runq, td);
  517 }
  518 
  519 /*
  520  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
  521  * for this thread to the referenced thread queue.
  522  */
  523 static void
  524 tdq_load_add(struct tdq *tdq, struct thread *td)
  525 {
  526 
  527         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  528         THREAD_LOCK_ASSERT(td, MA_OWNED);
  529 
  530         tdq->tdq_load++;
  531         if ((td->td_flags & TDF_NOLOAD) == 0)
  532                 tdq->tdq_sysload++;
  533         KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
  534         SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
  535 }
  536 
  537 /*
  538  * Remove the load from a thread that is transitioning to a sleep state or
  539  * exiting.
  540  */
  541 static void
  542 tdq_load_rem(struct tdq *tdq, struct thread *td)
  543 {
  544 
  545         THREAD_LOCK_ASSERT(td, MA_OWNED);
  546         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  547         KASSERT(tdq->tdq_load != 0,
  548             ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
  549 
  550         tdq->tdq_load--;
  551         if ((td->td_flags & TDF_NOLOAD) == 0)
  552                 tdq->tdq_sysload--;
  553         KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
  554         SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
  555 }
  556 
  557 /*
  558  * Set lowpri to its exact value by searching the run-queue and
  559  * evaluating curthread.  curthread may be passed as an optimization.
  560  */
  561 static void
  562 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
  563 {
  564         struct thread *td;
  565 
  566         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  567         if (ctd == NULL)
  568                 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
  569         td = tdq_choose(tdq);
  570         if (td == NULL || td->td_priority > ctd->td_priority)
  571                 tdq->tdq_lowpri = ctd->td_priority;
  572         else
  573                 tdq->tdq_lowpri = td->td_priority;
  574 }
  575 
  576 #ifdef SMP
  577 struct cpu_search {
  578         cpuset_t cs_mask;
  579         u_int   cs_prefer;
  580         int     cs_pri;         /* Min priority for low. */
  581         int     cs_limit;       /* Max load for low, min load for high. */
  582         int     cs_cpu;
  583         int     cs_load;
  584 };
  585 
  586 #define CPU_SEARCH_LOWEST       0x1
  587 #define CPU_SEARCH_HIGHEST      0x2
  588 #define CPU_SEARCH_BOTH         (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
  589 
  590 #define CPUSET_FOREACH(cpu, mask)                               \
  591         for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)             \
  592                 if (CPU_ISSET(cpu, &mask))
  593 
  594 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
  595     struct cpu_search *high, const int match);
  596 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
  597 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
  598 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
  599     struct cpu_search *high);
  600 
  601 /*
  602  * Search the tree of cpu_groups for the lowest or highest loaded cpu
  603  * according to the match argument.  This routine actually compares the
  604  * load on all paths through the tree and finds the least loaded cpu on
  605  * the least loaded path, which may differ from the least loaded cpu in
  606  * the system.  This balances work among caches and busses.
  607  *
  608  * This inline is instantiated in three forms below using constants for the
  609  * match argument.  It is reduced to the minimum set for each case.  It is
  610  * also recursive to the depth of the tree.
  611  */
  612 static __inline int
  613 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
  614     struct cpu_search *high, const int match)
  615 {
  616         struct cpu_search lgroup;
  617         struct cpu_search hgroup;
  618         cpuset_t cpumask;
  619         struct cpu_group *child;
  620         struct tdq *tdq;
  621         int cpu, i, hload, lload, load, total, rnd, *rndptr;
  622 
  623         total = 0;
  624         cpumask = cg->cg_mask;
  625         if (match & CPU_SEARCH_LOWEST) {
  626                 lload = INT_MAX;
  627                 lgroup = *low;
  628         }
  629         if (match & CPU_SEARCH_HIGHEST) {
  630                 hload = INT_MIN;
  631                 hgroup = *high;
  632         }
  633 
  634         /* Iterate through the child CPU groups and then remaining CPUs. */
  635         for (i = cg->cg_children, cpu = mp_maxid; ; ) {
  636                 if (i == 0) {
  637 #ifdef HAVE_INLINE_FFSL
  638                         cpu = CPU_FFS(&cpumask) - 1;
  639 #else
  640                         while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
  641                                 cpu--;
  642 #endif
  643                         if (cpu < 0)
  644                                 break;
  645                         child = NULL;
  646                 } else
  647                         child = &cg->cg_child[i - 1];
  648 
  649                 if (match & CPU_SEARCH_LOWEST)
  650                         lgroup.cs_cpu = -1;
  651                 if (match & CPU_SEARCH_HIGHEST)
  652                         hgroup.cs_cpu = -1;
  653                 if (child) {                    /* Handle child CPU group. */
  654                         CPU_NAND(&cpumask, &child->cg_mask);
  655                         switch (match) {
  656                         case CPU_SEARCH_LOWEST:
  657                                 load = cpu_search_lowest(child, &lgroup);
  658                                 break;
  659                         case CPU_SEARCH_HIGHEST:
  660                                 load = cpu_search_highest(child, &hgroup);
  661                                 break;
  662                         case CPU_SEARCH_BOTH:
  663                                 load = cpu_search_both(child, &lgroup, &hgroup);
  664                                 break;
  665                         }
  666                 } else {                        /* Handle child CPU. */
  667                         CPU_CLR(cpu, &cpumask);
  668                         tdq = TDQ_CPU(cpu);
  669                         load = tdq->tdq_load * 256;
  670                         rndptr = DPCPU_PTR(randomval);
  671                         rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
  672                         if (match & CPU_SEARCH_LOWEST) {
  673                                 if (cpu == low->cs_prefer)
  674                                         load -= 64;
  675                                 /* If that CPU is allowed and get data. */
  676                                 if (tdq->tdq_lowpri > lgroup.cs_pri &&
  677                                     tdq->tdq_load <= lgroup.cs_limit &&
  678                                     CPU_ISSET(cpu, &lgroup.cs_mask)) {
  679                                         lgroup.cs_cpu = cpu;
  680                                         lgroup.cs_load = load - rnd;
  681                                 }
  682                         }
  683                         if (match & CPU_SEARCH_HIGHEST)
  684                                 if (tdq->tdq_load >= hgroup.cs_limit &&
  685                                     tdq->tdq_transferable &&
  686                                     CPU_ISSET(cpu, &hgroup.cs_mask)) {
  687                                         hgroup.cs_cpu = cpu;
  688                                         hgroup.cs_load = load - rnd;
  689                                 }
  690                 }
  691                 total += load;
  692 
  693                 /* We have info about child item. Compare it. */
  694                 if (match & CPU_SEARCH_LOWEST) {
  695                         if (lgroup.cs_cpu >= 0 &&
  696                             (load < lload ||
  697                              (load == lload && lgroup.cs_load < low->cs_load))) {
  698                                 lload = load;
  699                                 low->cs_cpu = lgroup.cs_cpu;
  700                                 low->cs_load = lgroup.cs_load;
  701                         }
  702                 }
  703                 if (match & CPU_SEARCH_HIGHEST)
  704                         if (hgroup.cs_cpu >= 0 &&
  705                             (load > hload ||
  706                              (load == hload && hgroup.cs_load > high->cs_load))) {
  707                                 hload = load;
  708                                 high->cs_cpu = hgroup.cs_cpu;
  709                                 high->cs_load = hgroup.cs_load;
  710                         }
  711                 if (child) {
  712                         i--;
  713                         if (i == 0 && CPU_EMPTY(&cpumask))
  714                                 break;
  715                 }
  716 #ifndef HAVE_INLINE_FFSL
  717                 else
  718                         cpu--;
  719 #endif
  720         }
  721         return (total);
  722 }
  723 
  724 /*
  725  * cpu_search instantiations must pass constants to maintain the inline
  726  * optimization.
  727  */
  728 int
  729 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
  730 {
  731         return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
  732 }
  733 
  734 int
  735 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
  736 {
  737         return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
  738 }
  739 
  740 int
  741 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
  742     struct cpu_search *high)
  743 {
  744         return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
  745 }
  746 
  747 /*
  748  * Find the cpu with the least load via the least loaded path that has a
  749  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
  750  * acceptable.
  751  */
  752 static inline int
  753 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
  754     int prefer)
  755 {
  756         struct cpu_search low;
  757 
  758         low.cs_cpu = -1;
  759         low.cs_prefer = prefer;
  760         low.cs_mask = mask;
  761         low.cs_pri = pri;
  762         low.cs_limit = maxload;
  763         cpu_search_lowest(cg, &low);
  764         return low.cs_cpu;
  765 }
  766 
  767 /*
  768  * Find the cpu with the highest load via the highest loaded path.
  769  */
  770 static inline int
  771 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
  772 {
  773         struct cpu_search high;
  774 
  775         high.cs_cpu = -1;
  776         high.cs_mask = mask;
  777         high.cs_limit = minload;
  778         cpu_search_highest(cg, &high);
  779         return high.cs_cpu;
  780 }
  781 
  782 static void
  783 sched_balance_group(struct cpu_group *cg)
  784 {
  785         cpuset_t hmask, lmask;
  786         int high, low, anylow;
  787 
  788         CPU_FILL(&hmask);
  789         for (;;) {
  790                 high = sched_highest(cg, hmask, 1);
  791                 /* Stop if there is no more CPU with transferrable threads. */
  792                 if (high == -1)
  793                         break;
  794                 CPU_CLR(high, &hmask);
  795                 CPU_COPY(&hmask, &lmask);
  796                 /* Stop if there is no more CPU left for low. */
  797                 if (CPU_EMPTY(&lmask))
  798                         break;
  799                 anylow = 1;
  800 nextlow:
  801                 low = sched_lowest(cg, lmask, -1,
  802                     TDQ_CPU(high)->tdq_load - 1, high);
  803                 /* Stop if we looked well and found no less loaded CPU. */
  804                 if (anylow && low == -1)
  805                         break;
  806                 /* Go to next high if we found no less loaded CPU. */
  807                 if (low == -1)
  808                         continue;
  809                 /* Transfer thread from high to low. */
  810                 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
  811                         /* CPU that got thread can no longer be a donor. */
  812                         CPU_CLR(low, &hmask);
  813                 } else {
  814                         /*
  815                          * If failed, then there is no threads on high
  816                          * that can run on this low. Drop low from low
  817                          * mask and look for different one.
  818                          */
  819                         CPU_CLR(low, &lmask);
  820                         anylow = 0;
  821                         goto nextlow;
  822                 }
  823         }
  824 }
  825 
  826 static void
  827 sched_balance(void)
  828 {
  829         struct tdq *tdq;
  830 
  831         /*
  832          * Select a random time between .5 * balance_interval and
  833          * 1.5 * balance_interval.
  834          */
  835         balance_ticks = max(balance_interval / 2, 1);
  836         balance_ticks += random() % balance_interval;
  837         if (smp_started == 0 || rebalance == 0)
  838                 return;
  839         tdq = TDQ_SELF();
  840         TDQ_UNLOCK(tdq);
  841         sched_balance_group(cpu_top);
  842         TDQ_LOCK(tdq);
  843 }
  844 
  845 /*
  846  * Lock two thread queues using their address to maintain lock order.
  847  */
  848 static void
  849 tdq_lock_pair(struct tdq *one, struct tdq *two)
  850 {
  851         if (one < two) {
  852                 TDQ_LOCK(one);
  853                 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
  854         } else {
  855                 TDQ_LOCK(two);
  856                 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
  857         }
  858 }
  859 
  860 /*
  861  * Unlock two thread queues.  Order is not important here.
  862  */
  863 static void
  864 tdq_unlock_pair(struct tdq *one, struct tdq *two)
  865 {
  866         TDQ_UNLOCK(one);
  867         TDQ_UNLOCK(two);
  868 }
  869 
  870 /*
  871  * Transfer load between two imbalanced thread queues.
  872  */
  873 static int
  874 sched_balance_pair(struct tdq *high, struct tdq *low)
  875 {
  876         int moved;
  877         int cpu;
  878 
  879         tdq_lock_pair(high, low);
  880         moved = 0;
  881         /*
  882          * Determine what the imbalance is and then adjust that to how many
  883          * threads we actually have to give up (transferable).
  884          */
  885         if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
  886             (moved = tdq_move(high, low)) > 0) {
  887                 /*
  888                  * In case the target isn't the current cpu IPI it to force a
  889                  * reschedule with the new workload.
  890                  */
  891                 cpu = TDQ_ID(low);
  892                 sched_pin();
  893                 if (cpu != PCPU_GET(cpuid))
  894                         ipi_cpu(cpu, IPI_PREEMPT);
  895                 sched_unpin();
  896         }
  897         tdq_unlock_pair(high, low);
  898         return (moved);
  899 }
  900 
  901 /*
  902  * Move a thread from one thread queue to another.
  903  */
  904 static int
  905 tdq_move(struct tdq *from, struct tdq *to)
  906 {
  907         struct td_sched *ts;
  908         struct thread *td;
  909         struct tdq *tdq;
  910         int cpu;
  911 
  912         TDQ_LOCK_ASSERT(from, MA_OWNED);
  913         TDQ_LOCK_ASSERT(to, MA_OWNED);
  914 
  915         tdq = from;
  916         cpu = TDQ_ID(to);
  917         td = tdq_steal(tdq, cpu);
  918         if (td == NULL)
  919                 return (0);
  920         ts = td->td_sched;
  921         /*
  922          * Although the run queue is locked the thread may be blocked.  Lock
  923          * it to clear this and acquire the run-queue lock.
  924          */
  925         thread_lock(td);
  926         /* Drop recursive lock on from acquired via thread_lock(). */
  927         TDQ_UNLOCK(from);
  928         sched_rem(td);
  929         ts->ts_cpu = cpu;
  930         td->td_lock = TDQ_LOCKPTR(to);
  931         tdq_add(to, td, SRQ_YIELDING);
  932         return (1);
  933 }
  934 
  935 /*
  936  * This tdq has idled.  Try to steal a thread from another cpu and switch
  937  * to it.
  938  */
  939 static int
  940 tdq_idled(struct tdq *tdq)
  941 {
  942         struct cpu_group *cg;
  943         struct tdq *steal;
  944         cpuset_t mask;
  945         int thresh;
  946         int cpu;
  947 
  948         if (smp_started == 0 || steal_idle == 0)
  949                 return (1);
  950         CPU_FILL(&mask);
  951         CPU_CLR(PCPU_GET(cpuid), &mask);
  952         /* We don't want to be preempted while we're iterating. */
  953         spinlock_enter();
  954         for (cg = tdq->tdq_cg; cg != NULL; ) {
  955                 if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
  956                         thresh = steal_thresh;
  957                 else
  958                         thresh = 1;
  959                 cpu = sched_highest(cg, mask, thresh);
  960                 if (cpu == -1) {
  961                         cg = cg->cg_parent;
  962                         continue;
  963                 }
  964                 steal = TDQ_CPU(cpu);
  965                 CPU_CLR(cpu, &mask);
  966                 tdq_lock_pair(tdq, steal);
  967                 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
  968                         tdq_unlock_pair(tdq, steal);
  969                         continue;
  970                 }
  971                 /*
  972                  * If a thread was added while interrupts were disabled don't
  973                  * steal one here.  If we fail to acquire one due to affinity
  974                  * restrictions loop again with this cpu removed from the
  975                  * set.
  976                  */
  977                 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
  978                         tdq_unlock_pair(tdq, steal);
  979                         continue;
  980                 }
  981                 spinlock_exit();
  982                 TDQ_UNLOCK(steal);
  983                 mi_switch(SW_VOL | SWT_IDLE, NULL);
  984                 thread_unlock(curthread);
  985 
  986                 return (0);
  987         }
  988         spinlock_exit();
  989         return (1);
  990 }
  991 
  992 /*
  993  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
  994  */
  995 static void
  996 tdq_notify(struct tdq *tdq, struct thread *td)
  997 {
  998         struct thread *ctd;
  999         int pri;
 1000         int cpu;
 1001 
 1002         if (tdq->tdq_ipipending)
 1003                 return;
 1004         cpu = td->td_sched->ts_cpu;
 1005         pri = td->td_priority;
 1006         ctd = pcpu_find(cpu)->pc_curthread;
 1007         if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
 1008                 return;
 1009 
 1010         /*
 1011          * Make sure that tdq_load updated before calling this function
 1012          * is globally visible before we read tdq_cpu_idle.  Idle thread
 1013          * accesses both of them without locks, and the order is important.
 1014          */
 1015         mb();
 1016 
 1017         if (TD_IS_IDLETHREAD(ctd)) {
 1018                 /*
 1019                  * If the MD code has an idle wakeup routine try that before
 1020                  * falling back to IPI.
 1021                  */
 1022                 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
 1023                         return;
 1024         }
 1025         tdq->tdq_ipipending = 1;
 1026         ipi_cpu(cpu, IPI_PREEMPT);
 1027 }
 1028 
 1029 /*
 1030  * Steals load from a timeshare queue.  Honors the rotating queue head
 1031  * index.
 1032  */
 1033 static struct thread *
 1034 runq_steal_from(struct runq *rq, int cpu, u_char start)
 1035 {
 1036         struct rqbits *rqb;
 1037         struct rqhead *rqh;
 1038         struct thread *td, *first;
 1039         int bit;
 1040         int pri;
 1041         int i;
 1042 
 1043         rqb = &rq->rq_status;
 1044         bit = start & (RQB_BPW -1);
 1045         pri = 0;
 1046         first = NULL;
 1047 again:
 1048         for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
 1049                 if (rqb->rqb_bits[i] == 0)
 1050                         continue;
 1051                 if (bit != 0) {
 1052                         for (pri = bit; pri < RQB_BPW; pri++)
 1053                                 if (rqb->rqb_bits[i] & (1ul << pri))
 1054                                         break;
 1055                         if (pri >= RQB_BPW)
 1056                                 continue;
 1057                 } else
 1058                         pri = RQB_FFS(rqb->rqb_bits[i]);
 1059                 pri += (i << RQB_L2BPW);
 1060                 rqh = &rq->rq_queues[pri];
 1061                 TAILQ_FOREACH(td, rqh, td_runq) {
 1062                         if (first && THREAD_CAN_MIGRATE(td) &&
 1063                             THREAD_CAN_SCHED(td, cpu))
 1064                                 return (td);
 1065                         first = td;
 1066                 }
 1067         }
 1068         if (start != 0) {
 1069                 start = 0;
 1070                 goto again;
 1071         }
 1072 
 1073         if (first && THREAD_CAN_MIGRATE(first) &&
 1074             THREAD_CAN_SCHED(first, cpu))
 1075                 return (first);
 1076         return (NULL);
 1077 }
 1078 
 1079 /*
 1080  * Steals load from a standard linear queue.
 1081  */
 1082 static struct thread *
 1083 runq_steal(struct runq *rq, int cpu)
 1084 {
 1085         struct rqhead *rqh;
 1086         struct rqbits *rqb;
 1087         struct thread *td;
 1088         int word;
 1089         int bit;
 1090 
 1091         rqb = &rq->rq_status;
 1092         for (word = 0; word < RQB_LEN; word++) {
 1093                 if (rqb->rqb_bits[word] == 0)
 1094                         continue;
 1095                 for (bit = 0; bit < RQB_BPW; bit++) {
 1096                         if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
 1097                                 continue;
 1098                         rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
 1099                         TAILQ_FOREACH(td, rqh, td_runq)
 1100                                 if (THREAD_CAN_MIGRATE(td) &&
 1101                                     THREAD_CAN_SCHED(td, cpu))
 1102                                         return (td);
 1103                 }
 1104         }
 1105         return (NULL);
 1106 }
 1107 
 1108 /*
 1109  * Attempt to steal a thread in priority order from a thread queue.
 1110  */
 1111 static struct thread *
 1112 tdq_steal(struct tdq *tdq, int cpu)
 1113 {
 1114         struct thread *td;
 1115 
 1116         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 1117         if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
 1118                 return (td);
 1119         if ((td = runq_steal_from(&tdq->tdq_timeshare,
 1120             cpu, tdq->tdq_ridx)) != NULL)
 1121                 return (td);
 1122         return (runq_steal(&tdq->tdq_idle, cpu));
 1123 }
 1124 
 1125 /*
 1126  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
 1127  * current lock and returns with the assigned queue locked.
 1128  */
 1129 static inline struct tdq *
 1130 sched_setcpu(struct thread *td, int cpu, int flags)
 1131 {
 1132 
 1133         struct tdq *tdq;
 1134 
 1135         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1136         tdq = TDQ_CPU(cpu);
 1137         td->td_sched->ts_cpu = cpu;
 1138         /*
 1139          * If the lock matches just return the queue.
 1140          */
 1141         if (td->td_lock == TDQ_LOCKPTR(tdq))
 1142                 return (tdq);
 1143 #ifdef notyet
 1144         /*
 1145          * If the thread isn't running its lockptr is a
 1146          * turnstile or a sleepqueue.  We can just lock_set without
 1147          * blocking.
 1148          */
 1149         if (TD_CAN_RUN(td)) {
 1150                 TDQ_LOCK(tdq);
 1151                 thread_lock_set(td, TDQ_LOCKPTR(tdq));
 1152                 return (tdq);
 1153         }
 1154 #endif
 1155         /*
 1156          * The hard case, migration, we need to block the thread first to
 1157          * prevent order reversals with other cpus locks.
 1158          */
 1159         spinlock_enter();
 1160         thread_lock_block(td);
 1161         TDQ_LOCK(tdq);
 1162         thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
 1163         spinlock_exit();
 1164         return (tdq);
 1165 }
 1166 
 1167 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
 1168 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
 1169 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
 1170 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
 1171 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
 1172 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
 1173 
 1174 static int
 1175 sched_pickcpu(struct thread *td, int flags)
 1176 {
 1177         struct cpu_group *cg, *ccg;
 1178         struct td_sched *ts;
 1179         struct tdq *tdq;
 1180         cpuset_t mask;
 1181         int cpu, pri, self;
 1182 
 1183         self = PCPU_GET(cpuid);
 1184         ts = td->td_sched;
 1185         if (smp_started == 0)
 1186                 return (self);
 1187         /*
 1188          * Don't migrate a running thread from sched_switch().
 1189          */
 1190         if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
 1191                 return (ts->ts_cpu);
 1192         /*
 1193          * Prefer to run interrupt threads on the processors that generate
 1194          * the interrupt.
 1195          */
 1196         pri = td->td_priority;
 1197         if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
 1198             curthread->td_intr_nesting_level && ts->ts_cpu != self) {
 1199                 SCHED_STAT_INC(pickcpu_intrbind);
 1200                 ts->ts_cpu = self;
 1201                 if (TDQ_CPU(self)->tdq_lowpri > pri) {
 1202                         SCHED_STAT_INC(pickcpu_affinity);
 1203                         return (ts->ts_cpu);
 1204                 }
 1205         }
 1206         /*
 1207          * If the thread can run on the last cpu and the affinity has not
 1208          * expired or it is idle run it there.
 1209          */
 1210         tdq = TDQ_CPU(ts->ts_cpu);
 1211         cg = tdq->tdq_cg;
 1212         if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
 1213             tdq->tdq_lowpri >= PRI_MIN_IDLE &&
 1214             SCHED_AFFINITY(ts, CG_SHARE_L2)) {
 1215                 if (cg->cg_flags & CG_FLAG_THREAD) {
 1216                         CPUSET_FOREACH(cpu, cg->cg_mask) {
 1217                                 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
 1218                                         break;
 1219                         }
 1220                 } else
 1221                         cpu = INT_MAX;
 1222                 if (cpu > mp_maxid) {
 1223                         SCHED_STAT_INC(pickcpu_idle_affinity);
 1224                         return (ts->ts_cpu);
 1225                 }
 1226         }
 1227         /*
 1228          * Search for the last level cache CPU group in the tree.
 1229          * Skip caches with expired affinity time and SMT groups.
 1230          * Affinity to higher level caches will be handled less aggressively.
 1231          */
 1232         for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
 1233                 if (cg->cg_flags & CG_FLAG_THREAD)
 1234                         continue;
 1235                 if (!SCHED_AFFINITY(ts, cg->cg_level))
 1236                         continue;
 1237                 ccg = cg;
 1238         }
 1239         if (ccg != NULL)
 1240                 cg = ccg;
 1241         cpu = -1;
 1242         /* Search the group for the less loaded idle CPU we can run now. */
 1243         mask = td->td_cpuset->cs_mask;
 1244         if (cg != NULL && cg != cpu_top &&
 1245             CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
 1246                 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
 1247                     INT_MAX, ts->ts_cpu);
 1248         /* Search globally for the less loaded CPU we can run now. */
 1249         if (cpu == -1)
 1250                 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
 1251         /* Search globally for the less loaded CPU. */
 1252         if (cpu == -1)
 1253                 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
 1254         KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
 1255         /*
 1256          * Compare the lowest loaded cpu to current cpu.
 1257          */
 1258         if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
 1259             TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
 1260             TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
 1261                 SCHED_STAT_INC(pickcpu_local);
 1262                 cpu = self;
 1263         } else
 1264                 SCHED_STAT_INC(pickcpu_lowest);
 1265         if (cpu != ts->ts_cpu)
 1266                 SCHED_STAT_INC(pickcpu_migration);
 1267         return (cpu);
 1268 }
 1269 #endif
 1270 
 1271 /*
 1272  * Pick the highest priority task we have and return it.
 1273  */
 1274 static struct thread *
 1275 tdq_choose(struct tdq *tdq)
 1276 {
 1277         struct thread *td;
 1278 
 1279         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 1280         td = runq_choose(&tdq->tdq_realtime);
 1281         if (td != NULL)
 1282                 return (td);
 1283         td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
 1284         if (td != NULL) {
 1285                 KASSERT(td->td_priority >= PRI_MIN_BATCH,
 1286                     ("tdq_choose: Invalid priority on timeshare queue %d",
 1287                     td->td_priority));
 1288                 return (td);
 1289         }
 1290         td = runq_choose(&tdq->tdq_idle);
 1291         if (td != NULL) {
 1292                 KASSERT(td->td_priority >= PRI_MIN_IDLE,
 1293                     ("tdq_choose: Invalid priority on idle queue %d",
 1294                     td->td_priority));
 1295                 return (td);
 1296         }
 1297 
 1298         return (NULL);
 1299 }
 1300 
 1301 /*
 1302  * Initialize a thread queue.
 1303  */
 1304 static void
 1305 tdq_setup(struct tdq *tdq)
 1306 {
 1307 
 1308         if (bootverbose)
 1309                 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
 1310         runq_init(&tdq->tdq_realtime);
 1311         runq_init(&tdq->tdq_timeshare);
 1312         runq_init(&tdq->tdq_idle);
 1313         snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
 1314             "sched lock %d", (int)TDQ_ID(tdq));
 1315         mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
 1316             MTX_SPIN | MTX_RECURSE);
 1317 #ifdef KTR
 1318         snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
 1319             "CPU %d load", (int)TDQ_ID(tdq));
 1320 #endif
 1321 }
 1322 
 1323 #ifdef SMP
 1324 static void
 1325 sched_setup_smp(void)
 1326 {
 1327         struct tdq *tdq;
 1328         int i;
 1329 
 1330         cpu_top = smp_topo();
 1331         CPU_FOREACH(i) {
 1332                 tdq = TDQ_CPU(i);
 1333                 tdq_setup(tdq);
 1334                 tdq->tdq_cg = smp_topo_find(cpu_top, i);
 1335                 if (tdq->tdq_cg == NULL)
 1336                         panic("Can't find cpu group for %d\n", i);
 1337         }
 1338         balance_tdq = TDQ_SELF();
 1339         sched_balance();
 1340 }
 1341 #endif
 1342 
 1343 /*
 1344  * Setup the thread queues and initialize the topology based on MD
 1345  * information.
 1346  */
 1347 static void
 1348 sched_setup(void *dummy)
 1349 {
 1350         struct tdq *tdq;
 1351 
 1352         tdq = TDQ_SELF();
 1353 #ifdef SMP
 1354         sched_setup_smp();
 1355 #else
 1356         tdq_setup(tdq);
 1357 #endif
 1358 
 1359         /* Add thread0's load since it's running. */
 1360         TDQ_LOCK(tdq);
 1361         thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
 1362         tdq_load_add(tdq, &thread0);
 1363         tdq->tdq_lowpri = thread0.td_priority;
 1364         TDQ_UNLOCK(tdq);
 1365 }
 1366 
 1367 /*
 1368  * This routine determines time constants after stathz and hz are setup.
 1369  */
 1370 /* ARGSUSED */
 1371 static void
 1372 sched_initticks(void *dummy)
 1373 {
 1374         int incr;
 1375 
 1376         realstathz = stathz ? stathz : hz;
 1377         sched_slice = realstathz / 10;  /* ~100ms */
 1378         hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
 1379             realstathz);
 1380 
 1381         /*
 1382          * tickincr is shifted out by 10 to avoid rounding errors due to
 1383          * hz not being evenly divisible by stathz on all platforms.
 1384          */
 1385         incr = (hz << SCHED_TICK_SHIFT) / realstathz;
 1386         /*
 1387          * This does not work for values of stathz that are more than
 1388          * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
 1389          */
 1390         if (incr == 0)
 1391                 incr = 1;
 1392         tickincr = incr;
 1393 #ifdef SMP
 1394         /*
 1395          * Set the default balance interval now that we know
 1396          * what realstathz is.
 1397          */
 1398         balance_interval = realstathz;
 1399         affinity = SCHED_AFFINITY_DEFAULT;
 1400 #endif
 1401         if (sched_idlespinthresh < 0)
 1402                 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
 1403 }
 1404 
 1405 
 1406 /*
 1407  * This is the core of the interactivity algorithm.  Determines a score based
 1408  * on past behavior.  It is the ratio of sleep time to run time scaled to
 1409  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
 1410  * differs from the cpu usage because it does not account for time spent
 1411  * waiting on a run-queue.  Would be prettier if we had floating point.
 1412  */
 1413 static int
 1414 sched_interact_score(struct thread *td)
 1415 {
 1416         struct td_sched *ts;
 1417         int div;
 1418 
 1419         ts = td->td_sched;
 1420         /*
 1421          * The score is only needed if this is likely to be an interactive
 1422          * task.  Don't go through the expense of computing it if there's
 1423          * no chance.
 1424          */
 1425         if (sched_interact <= SCHED_INTERACT_HALF &&
 1426                 ts->ts_runtime >= ts->ts_slptime)
 1427                         return (SCHED_INTERACT_HALF);
 1428 
 1429         if (ts->ts_runtime > ts->ts_slptime) {
 1430                 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
 1431                 return (SCHED_INTERACT_HALF +
 1432                     (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
 1433         }
 1434         if (ts->ts_slptime > ts->ts_runtime) {
 1435                 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
 1436                 return (ts->ts_runtime / div);
 1437         }
 1438         /* runtime == slptime */
 1439         if (ts->ts_runtime)
 1440                 return (SCHED_INTERACT_HALF);
 1441 
 1442         /*
 1443          * This can happen if slptime and runtime are 0.
 1444          */
 1445         return (0);
 1446 
 1447 }
 1448 
 1449 /*
 1450  * Scale the scheduling priority according to the "interactivity" of this
 1451  * process.
 1452  */
 1453 static void
 1454 sched_priority(struct thread *td)
 1455 {
 1456         int score;
 1457         int pri;
 1458 
 1459         if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
 1460                 return;
 1461         /*
 1462          * If the score is interactive we place the thread in the realtime
 1463          * queue with a priority that is less than kernel and interrupt
 1464          * priorities.  These threads are not subject to nice restrictions.
 1465          *
 1466          * Scores greater than this are placed on the normal timeshare queue
 1467          * where the priority is partially decided by the most recent cpu
 1468          * utilization and the rest is decided by nice value.
 1469          *
 1470          * The nice value of the process has a linear effect on the calculated
 1471          * score.  Negative nice values make it easier for a thread to be
 1472          * considered interactive.
 1473          */
 1474         score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
 1475         if (score < sched_interact) {
 1476                 pri = PRI_MIN_INTERACT;
 1477                 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
 1478                     sched_interact) * score;
 1479                 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
 1480                     ("sched_priority: invalid interactive priority %d score %d",
 1481                     pri, score));
 1482         } else {
 1483                 pri = SCHED_PRI_MIN;
 1484                 if (td->td_sched->ts_ticks)
 1485                         pri += min(SCHED_PRI_TICKS(td->td_sched),
 1486                             SCHED_PRI_RANGE - 1);
 1487                 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
 1488                 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
 1489                     ("sched_priority: invalid priority %d: nice %d, " 
 1490                     "ticks %d ftick %d ltick %d tick pri %d",
 1491                     pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
 1492                     td->td_sched->ts_ftick, td->td_sched->ts_ltick,
 1493                     SCHED_PRI_TICKS(td->td_sched)));
 1494         }
 1495         sched_user_prio(td, pri);
 1496 
 1497         return;
 1498 }
 1499 
 1500 /*
 1501  * This routine enforces a maximum limit on the amount of scheduling history
 1502  * kept.  It is called after either the slptime or runtime is adjusted.  This
 1503  * function is ugly due to integer math.
 1504  */
 1505 static void
 1506 sched_interact_update(struct thread *td)
 1507 {
 1508         struct td_sched *ts;
 1509         u_int sum;
 1510 
 1511         ts = td->td_sched;
 1512         sum = ts->ts_runtime + ts->ts_slptime;
 1513         if (sum < SCHED_SLP_RUN_MAX)
 1514                 return;
 1515         /*
 1516          * This only happens from two places:
 1517          * 1) We have added an unusual amount of run time from fork_exit.
 1518          * 2) We have added an unusual amount of sleep time from sched_sleep().
 1519          */
 1520         if (sum > SCHED_SLP_RUN_MAX * 2) {
 1521                 if (ts->ts_runtime > ts->ts_slptime) {
 1522                         ts->ts_runtime = SCHED_SLP_RUN_MAX;
 1523                         ts->ts_slptime = 1;
 1524                 } else {
 1525                         ts->ts_slptime = SCHED_SLP_RUN_MAX;
 1526                         ts->ts_runtime = 1;
 1527                 }
 1528                 return;
 1529         }
 1530         /*
 1531          * If we have exceeded by more than 1/5th then the algorithm below
 1532          * will not bring us back into range.  Dividing by two here forces
 1533          * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
 1534          */
 1535         if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
 1536                 ts->ts_runtime /= 2;
 1537                 ts->ts_slptime /= 2;
 1538                 return;
 1539         }
 1540         ts->ts_runtime = (ts->ts_runtime / 5) * 4;
 1541         ts->ts_slptime = (ts->ts_slptime / 5) * 4;
 1542 }
 1543 
 1544 /*
 1545  * Scale back the interactivity history when a child thread is created.  The
 1546  * history is inherited from the parent but the thread may behave totally
 1547  * differently.  For example, a shell spawning a compiler process.  We want
 1548  * to learn that the compiler is behaving badly very quickly.
 1549  */
 1550 static void
 1551 sched_interact_fork(struct thread *td)
 1552 {
 1553         int ratio;
 1554         int sum;
 1555 
 1556         sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
 1557         if (sum > SCHED_SLP_RUN_FORK) {
 1558                 ratio = sum / SCHED_SLP_RUN_FORK;
 1559                 td->td_sched->ts_runtime /= ratio;
 1560                 td->td_sched->ts_slptime /= ratio;
 1561         }
 1562 }
 1563 
 1564 /*
 1565  * Called from proc0_init() to setup the scheduler fields.
 1566  */
 1567 void
 1568 schedinit(void)
 1569 {
 1570 
 1571         /*
 1572          * Set up the scheduler specific parts of proc0.
 1573          */
 1574         proc0.p_sched = NULL; /* XXX */
 1575         thread0.td_sched = &td_sched0;
 1576         td_sched0.ts_ltick = ticks;
 1577         td_sched0.ts_ftick = ticks;
 1578         td_sched0.ts_slice = sched_slice;
 1579 }
 1580 
 1581 /*
 1582  * This is only somewhat accurate since given many processes of the same
 1583  * priority they will switch when their slices run out, which will be
 1584  * at most sched_slice stathz ticks.
 1585  */
 1586 int
 1587 sched_rr_interval(void)
 1588 {
 1589 
 1590         /* Convert sched_slice from stathz to hz. */
 1591         return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
 1592 }
 1593 
 1594 /*
 1595  * Update the percent cpu tracking information when it is requested or
 1596  * the total history exceeds the maximum.  We keep a sliding history of
 1597  * tick counts that slowly decays.  This is less precise than the 4BSD
 1598  * mechanism since it happens with less regular and frequent events.
 1599  */
 1600 static void
 1601 sched_pctcpu_update(struct td_sched *ts, int run)
 1602 {
 1603         int t = ticks;
 1604 
 1605         if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
 1606                 ts->ts_ticks = 0;
 1607                 ts->ts_ftick = t - SCHED_TICK_TARG;
 1608         } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
 1609                 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
 1610                     (ts->ts_ltick - (t - SCHED_TICK_TARG));
 1611                 ts->ts_ftick = t - SCHED_TICK_TARG;
 1612         }
 1613         if (run)
 1614                 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
 1615         ts->ts_ltick = t;
 1616 }
 1617 
 1618 /*
 1619  * Adjust the priority of a thread.  Move it to the appropriate run-queue
 1620  * if necessary.  This is the back-end for several priority related
 1621  * functions.
 1622  */
 1623 static void
 1624 sched_thread_priority(struct thread *td, u_char prio)
 1625 {
 1626         struct td_sched *ts;
 1627         struct tdq *tdq;
 1628         int oldpri;
 1629 
 1630         KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
 1631             "prio:%d", td->td_priority, "new prio:%d", prio,
 1632             KTR_ATTR_LINKED, sched_tdname(curthread));
 1633         SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
 1634         if (td != curthread && prio < td->td_priority) {
 1635                 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
 1636                     "lend prio", "prio:%d", td->td_priority, "new prio:%d",
 1637                     prio, KTR_ATTR_LINKED, sched_tdname(td));
 1638                 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 
 1639                     curthread);
 1640         } 
 1641         ts = td->td_sched;
 1642         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1643         if (td->td_priority == prio)
 1644                 return;
 1645         /*
 1646          * If the priority has been elevated due to priority
 1647          * propagation, we may have to move ourselves to a new
 1648          * queue.  This could be optimized to not re-add in some
 1649          * cases.
 1650          */
 1651         if (TD_ON_RUNQ(td) && prio < td->td_priority) {
 1652                 sched_rem(td);
 1653                 td->td_priority = prio;
 1654                 sched_add(td, SRQ_BORROWING);
 1655                 return;
 1656         }
 1657         /*
 1658          * If the thread is currently running we may have to adjust the lowpri
 1659          * information so other cpus are aware of our current priority.
 1660          */
 1661         if (TD_IS_RUNNING(td)) {
 1662                 tdq = TDQ_CPU(ts->ts_cpu);
 1663                 oldpri = td->td_priority;
 1664                 td->td_priority = prio;
 1665                 if (prio < tdq->tdq_lowpri)
 1666                         tdq->tdq_lowpri = prio;
 1667                 else if (tdq->tdq_lowpri == oldpri)
 1668                         tdq_setlowpri(tdq, td);
 1669                 return;
 1670         }
 1671         td->td_priority = prio;
 1672 }
 1673 
 1674 /*
 1675  * Update a thread's priority when it is lent another thread's
 1676  * priority.
 1677  */
 1678 void
 1679 sched_lend_prio(struct thread *td, u_char prio)
 1680 {
 1681 
 1682         td->td_flags |= TDF_BORROWING;
 1683         sched_thread_priority(td, prio);
 1684 }
 1685 
 1686 /*
 1687  * Restore a thread's priority when priority propagation is
 1688  * over.  The prio argument is the minimum priority the thread
 1689  * needs to have to satisfy other possible priority lending
 1690  * requests.  If the thread's regular priority is less
 1691  * important than prio, the thread will keep a priority boost
 1692  * of prio.
 1693  */
 1694 void
 1695 sched_unlend_prio(struct thread *td, u_char prio)
 1696 {
 1697         u_char base_pri;
 1698 
 1699         if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
 1700             td->td_base_pri <= PRI_MAX_TIMESHARE)
 1701                 base_pri = td->td_user_pri;
 1702         else
 1703                 base_pri = td->td_base_pri;
 1704         if (prio >= base_pri) {
 1705                 td->td_flags &= ~TDF_BORROWING;
 1706                 sched_thread_priority(td, base_pri);
 1707         } else
 1708                 sched_lend_prio(td, prio);
 1709 }
 1710 
 1711 /*
 1712  * Standard entry for setting the priority to an absolute value.
 1713  */
 1714 void
 1715 sched_prio(struct thread *td, u_char prio)
 1716 {
 1717         u_char oldprio;
 1718 
 1719         /* First, update the base priority. */
 1720         td->td_base_pri = prio;
 1721 
 1722         /*
 1723          * If the thread is borrowing another thread's priority, don't
 1724          * ever lower the priority.
 1725          */
 1726         if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
 1727                 return;
 1728 
 1729         /* Change the real priority. */
 1730         oldprio = td->td_priority;
 1731         sched_thread_priority(td, prio);
 1732 
 1733         /*
 1734          * If the thread is on a turnstile, then let the turnstile update
 1735          * its state.
 1736          */
 1737         if (TD_ON_LOCK(td) && oldprio != prio)
 1738                 turnstile_adjust(td, oldprio);
 1739 }
 1740 
 1741 /*
 1742  * Set the base user priority, does not effect current running priority.
 1743  */
 1744 void
 1745 sched_user_prio(struct thread *td, u_char prio)
 1746 {
 1747 
 1748         td->td_base_user_pri = prio;
 1749         if (td->td_lend_user_pri <= prio)
 1750                 return;
 1751         td->td_user_pri = prio;
 1752 }
 1753 
 1754 void
 1755 sched_lend_user_prio(struct thread *td, u_char prio)
 1756 {
 1757 
 1758         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1759         td->td_lend_user_pri = prio;
 1760         td->td_user_pri = min(prio, td->td_base_user_pri);
 1761         if (td->td_priority > td->td_user_pri)
 1762                 sched_prio(td, td->td_user_pri);
 1763         else if (td->td_priority != td->td_user_pri)
 1764                 td->td_flags |= TDF_NEEDRESCHED;
 1765 }
 1766 
 1767 /*
 1768  * Handle migration from sched_switch().  This happens only for
 1769  * cpu binding.
 1770  */
 1771 static struct mtx *
 1772 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
 1773 {
 1774         struct tdq *tdn;
 1775 
 1776         tdn = TDQ_CPU(td->td_sched->ts_cpu);
 1777 #ifdef SMP
 1778         tdq_load_rem(tdq, td);
 1779         /*
 1780          * Do the lock dance required to avoid LOR.  We grab an extra
 1781          * spinlock nesting to prevent preemption while we're
 1782          * not holding either run-queue lock.
 1783          */
 1784         spinlock_enter();
 1785         thread_lock_block(td);  /* This releases the lock on tdq. */
 1786 
 1787         /*
 1788          * Acquire both run-queue locks before placing the thread on the new
 1789          * run-queue to avoid deadlocks created by placing a thread with a
 1790          * blocked lock on the run-queue of a remote processor.  The deadlock
 1791          * occurs when a third processor attempts to lock the two queues in
 1792          * question while the target processor is spinning with its own
 1793          * run-queue lock held while waiting for the blocked lock to clear.
 1794          */
 1795         tdq_lock_pair(tdn, tdq);
 1796         tdq_add(tdn, td, flags);
 1797         tdq_notify(tdn, td);
 1798         TDQ_UNLOCK(tdn);
 1799         spinlock_exit();
 1800 #endif
 1801         return (TDQ_LOCKPTR(tdn));
 1802 }
 1803 
 1804 /*
 1805  * Variadic version of thread_lock_unblock() that does not assume td_lock
 1806  * is blocked.
 1807  */
 1808 static inline void
 1809 thread_unblock_switch(struct thread *td, struct mtx *mtx)
 1810 {
 1811         atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
 1812             (uintptr_t)mtx);
 1813 }
 1814 
 1815 /*
 1816  * Switch threads.  This function has to handle threads coming in while
 1817  * blocked for some reason, running, or idle.  It also must deal with
 1818  * migrating a thread from one queue to another as running threads may
 1819  * be assigned elsewhere via binding.
 1820  */
 1821 void
 1822 sched_switch(struct thread *td, struct thread *newtd, int flags)
 1823 {
 1824         struct tdq *tdq;
 1825         struct td_sched *ts;
 1826         struct mtx *mtx;
 1827         int srqflag;
 1828         int cpuid, preempted;
 1829 
 1830         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1831         KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
 1832 
 1833         cpuid = PCPU_GET(cpuid);
 1834         tdq = TDQ_CPU(cpuid);
 1835         ts = td->td_sched;
 1836         mtx = td->td_lock;
 1837         sched_pctcpu_update(ts, 1);
 1838         ts->ts_rltick = ticks;
 1839         td->td_lastcpu = td->td_oncpu;
 1840         td->td_oncpu = NOCPU;
 1841         preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
 1842             (flags & SW_PREEMPT) != 0;
 1843         td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
 1844         td->td_owepreempt = 0;
 1845         if (!TD_IS_IDLETHREAD(td))
 1846                 tdq->tdq_switchcnt++;
 1847         /*
 1848          * The lock pointer in an idle thread should never change.  Reset it
 1849          * to CAN_RUN as well.
 1850          */
 1851         if (TD_IS_IDLETHREAD(td)) {
 1852                 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 1853                 TD_SET_CAN_RUN(td);
 1854         } else if (TD_IS_RUNNING(td)) {
 1855                 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 1856                 srqflag = preempted ?
 1857                     SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
 1858                     SRQ_OURSELF|SRQ_YIELDING;
 1859 #ifdef SMP
 1860                 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
 1861                         ts->ts_cpu = sched_pickcpu(td, 0);
 1862 #endif
 1863                 if (ts->ts_cpu == cpuid)
 1864                         tdq_runq_add(tdq, td, srqflag);
 1865                 else {
 1866                         KASSERT(THREAD_CAN_MIGRATE(td) ||
 1867                             (ts->ts_flags & TSF_BOUND) != 0,
 1868                             ("Thread %p shouldn't migrate", td));
 1869                         mtx = sched_switch_migrate(tdq, td, srqflag);
 1870                 }
 1871         } else {
 1872                 /* This thread must be going to sleep. */
 1873                 TDQ_LOCK(tdq);
 1874                 mtx = thread_lock_block(td);
 1875                 tdq_load_rem(tdq, td);
 1876         }
 1877         /*
 1878          * We enter here with the thread blocked and assigned to the
 1879          * appropriate cpu run-queue or sleep-queue and with the current
 1880          * thread-queue locked.
 1881          */
 1882         TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
 1883         newtd = choosethread();
 1884         /*
 1885          * Call the MD code to switch contexts if necessary.
 1886          */
 1887         if (td != newtd) {
 1888 #ifdef  HWPMC_HOOKS
 1889                 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 1890                         PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
 1891 #endif
 1892                 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
 1893                 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
 1894                 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
 1895                 sched_pctcpu_update(newtd->td_sched, 0);
 1896 
 1897 #ifdef KDTRACE_HOOKS
 1898                 /*
 1899                  * If DTrace has set the active vtime enum to anything
 1900                  * other than INACTIVE (0), then it should have set the
 1901                  * function to call.
 1902                  */
 1903                 if (dtrace_vtime_active)
 1904                         (*dtrace_vtime_switch_func)(newtd);
 1905 #endif
 1906 
 1907                 cpu_switch(td, newtd, mtx);
 1908                 /*
 1909                  * We may return from cpu_switch on a different cpu.  However,
 1910                  * we always return with td_lock pointing to the current cpu's
 1911                  * run queue lock.
 1912                  */
 1913                 cpuid = PCPU_GET(cpuid);
 1914                 tdq = TDQ_CPU(cpuid);
 1915                 lock_profile_obtain_lock_success(
 1916                     &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
 1917 
 1918                 SDT_PROBE0(sched, , , on__cpu);
 1919 #ifdef  HWPMC_HOOKS
 1920                 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 1921                         PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
 1922 #endif
 1923         } else {
 1924                 thread_unblock_switch(td, mtx);
 1925                 SDT_PROBE0(sched, , , remain__cpu);
 1926         }
 1927         /*
 1928          * Assert that all went well and return.
 1929          */
 1930         TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
 1931         MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 1932         td->td_oncpu = cpuid;
 1933 }
 1934 
 1935 /*
 1936  * Adjust thread priorities as a result of a nice request.
 1937  */
 1938 void
 1939 sched_nice(struct proc *p, int nice)
 1940 {
 1941         struct thread *td;
 1942 
 1943         PROC_LOCK_ASSERT(p, MA_OWNED);
 1944 
 1945         p->p_nice = nice;
 1946         FOREACH_THREAD_IN_PROC(p, td) {
 1947                 thread_lock(td);
 1948                 sched_priority(td);
 1949                 sched_prio(td, td->td_base_user_pri);
 1950                 thread_unlock(td);
 1951         }
 1952 }
 1953 
 1954 /*
 1955  * Record the sleep time for the interactivity scorer.
 1956  */
 1957 void
 1958 sched_sleep(struct thread *td, int prio)
 1959 {
 1960 
 1961         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1962 
 1963         td->td_slptick = ticks;
 1964         if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
 1965                 td->td_flags |= TDF_CANSWAP;
 1966         if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
 1967                 return;
 1968         if (static_boost == 1 && prio)
 1969                 sched_prio(td, prio);
 1970         else if (static_boost && td->td_priority > static_boost)
 1971                 sched_prio(td, static_boost);
 1972 }
 1973 
 1974 /*
 1975  * Schedule a thread to resume execution and record how long it voluntarily
 1976  * slept.  We also update the pctcpu, interactivity, and priority.
 1977  */
 1978 void
 1979 sched_wakeup(struct thread *td)
 1980 {
 1981         struct td_sched *ts;
 1982         int slptick;
 1983 
 1984         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1985         ts = td->td_sched;
 1986         td->td_flags &= ~TDF_CANSWAP;
 1987         /*
 1988          * If we slept for more than a tick update our interactivity and
 1989          * priority.
 1990          */
 1991         slptick = td->td_slptick;
 1992         td->td_slptick = 0;
 1993         if (slptick && slptick != ticks) {
 1994                 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
 1995                 sched_interact_update(td);
 1996                 sched_pctcpu_update(ts, 0);
 1997         }
 1998         /* Reset the slice value after we sleep. */
 1999         ts->ts_slice = sched_slice;
 2000         sched_add(td, SRQ_BORING);
 2001 }
 2002 
 2003 /*
 2004  * Penalize the parent for creating a new child and initialize the child's
 2005  * priority.
 2006  */
 2007 void
 2008 sched_fork(struct thread *td, struct thread *child)
 2009 {
 2010         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2011         sched_pctcpu_update(td->td_sched, 1);
 2012         sched_fork_thread(td, child);
 2013         /*
 2014          * Penalize the parent and child for forking.
 2015          */
 2016         sched_interact_fork(child);
 2017         sched_priority(child);
 2018         td->td_sched->ts_runtime += tickincr;
 2019         sched_interact_update(td);
 2020         sched_priority(td);
 2021 }
 2022 
 2023 /*
 2024  * Fork a new thread, may be within the same process.
 2025  */
 2026 void
 2027 sched_fork_thread(struct thread *td, struct thread *child)
 2028 {
 2029         struct td_sched *ts;
 2030         struct td_sched *ts2;
 2031 
 2032         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2033         /*
 2034          * Initialize child.
 2035          */
 2036         ts = td->td_sched;
 2037         ts2 = child->td_sched;
 2038         child->td_oncpu = NOCPU;
 2039         child->td_lastcpu = NOCPU;
 2040         child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
 2041         child->td_cpuset = cpuset_ref(td->td_cpuset);
 2042         ts2->ts_cpu = ts->ts_cpu;
 2043         ts2->ts_flags = 0;
 2044         /*
 2045          * Grab our parents cpu estimation information.
 2046          */
 2047         ts2->ts_ticks = ts->ts_ticks;
 2048         ts2->ts_ltick = ts->ts_ltick;
 2049         ts2->ts_ftick = ts->ts_ftick;
 2050         /*
 2051          * Do not inherit any borrowed priority from the parent.
 2052          */
 2053         child->td_priority = child->td_base_pri;
 2054         /*
 2055          * And update interactivity score.
 2056          */
 2057         ts2->ts_slptime = ts->ts_slptime;
 2058         ts2->ts_runtime = ts->ts_runtime;
 2059         ts2->ts_slice = 1;      /* Attempt to quickly learn interactivity. */
 2060 #ifdef KTR
 2061         bzero(ts2->ts_name, sizeof(ts2->ts_name));
 2062 #endif
 2063 }
 2064 
 2065 /*
 2066  * Adjust the priority class of a thread.
 2067  */
 2068 void
 2069 sched_class(struct thread *td, int class)
 2070 {
 2071 
 2072         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2073         if (td->td_pri_class == class)
 2074                 return;
 2075         td->td_pri_class = class;
 2076 }
 2077 
 2078 /*
 2079  * Return some of the child's priority and interactivity to the parent.
 2080  */
 2081 void
 2082 sched_exit(struct proc *p, struct thread *child)
 2083 {
 2084         struct thread *td;
 2085 
 2086         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
 2087             "prio:%d", child->td_priority);
 2088         PROC_LOCK_ASSERT(p, MA_OWNED);
 2089         td = FIRST_THREAD_IN_PROC(p);
 2090         sched_exit_thread(td, child);
 2091 }
 2092 
 2093 /*
 2094  * Penalize another thread for the time spent on this one.  This helps to
 2095  * worsen the priority and interactivity of processes which schedule batch
 2096  * jobs such as make.  This has little effect on the make process itself but
 2097  * causes new processes spawned by it to receive worse scores immediately.
 2098  */
 2099 void
 2100 sched_exit_thread(struct thread *td, struct thread *child)
 2101 {
 2102 
 2103         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
 2104             "prio:%d", child->td_priority);
 2105         /*
 2106          * Give the child's runtime to the parent without returning the
 2107          * sleep time as a penalty to the parent.  This causes shells that
 2108          * launch expensive things to mark their children as expensive.
 2109          */
 2110         thread_lock(td);
 2111         td->td_sched->ts_runtime += child->td_sched->ts_runtime;
 2112         sched_interact_update(td);
 2113         sched_priority(td);
 2114         thread_unlock(td);
 2115 }
 2116 
 2117 void
 2118 sched_preempt(struct thread *td)
 2119 {
 2120         struct tdq *tdq;
 2121 
 2122         SDT_PROBE2(sched, , , surrender, td, td->td_proc);
 2123 
 2124         thread_lock(td);
 2125         tdq = TDQ_SELF();
 2126         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 2127         tdq->tdq_ipipending = 0;
 2128         if (td->td_priority > tdq->tdq_lowpri) {
 2129                 int flags;
 2130 
 2131                 flags = SW_INVOL | SW_PREEMPT;
 2132                 if (td->td_critnest > 1)
 2133                         td->td_owepreempt = 1;
 2134                 else if (TD_IS_IDLETHREAD(td))
 2135                         mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
 2136                 else
 2137                         mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
 2138         }
 2139         thread_unlock(td);
 2140 }
 2141 
 2142 /*
 2143  * Fix priorities on return to user-space.  Priorities may be elevated due
 2144  * to static priorities in msleep() or similar.
 2145  */
 2146 void
 2147 sched_userret(struct thread *td)
 2148 {
 2149         /*
 2150          * XXX we cheat slightly on the locking here to avoid locking in  
 2151          * the usual case.  Setting td_priority here is essentially an
 2152          * incomplete workaround for not setting it properly elsewhere.
 2153          * Now that some interrupt handlers are threads, not setting it
 2154          * properly elsewhere can clobber it in the window between setting
 2155          * it here and returning to user mode, so don't waste time setting
 2156          * it perfectly here.
 2157          */
 2158         KASSERT((td->td_flags & TDF_BORROWING) == 0,
 2159             ("thread with borrowed priority returning to userland"));
 2160         if (td->td_priority != td->td_user_pri) {
 2161                 thread_lock(td);
 2162                 td->td_priority = td->td_user_pri;
 2163                 td->td_base_pri = td->td_user_pri;
 2164                 tdq_setlowpri(TDQ_SELF(), td);
 2165                 thread_unlock(td);
 2166         }
 2167 }
 2168 
 2169 /*
 2170  * Handle a stathz tick.  This is really only relevant for timeshare
 2171  * threads.
 2172  */
 2173 void
 2174 sched_clock(struct thread *td)
 2175 {
 2176         struct tdq *tdq;
 2177         struct td_sched *ts;
 2178 
 2179         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2180         tdq = TDQ_SELF();
 2181 #ifdef SMP
 2182         /*
 2183          * We run the long term load balancer infrequently on the first cpu.
 2184          */
 2185         if (balance_tdq == tdq) {
 2186                 if (balance_ticks && --balance_ticks == 0)
 2187                         sched_balance();
 2188         }
 2189 #endif
 2190         /*
 2191          * Save the old switch count so we have a record of the last ticks
 2192          * activity.   Initialize the new switch count based on our load.
 2193          * If there is some activity seed it to reflect that.
 2194          */
 2195         tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
 2196         tdq->tdq_switchcnt = tdq->tdq_load;
 2197         /*
 2198          * Advance the insert index once for each tick to ensure that all
 2199          * threads get a chance to run.
 2200          */
 2201         if (tdq->tdq_idx == tdq->tdq_ridx) {
 2202                 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
 2203                 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
 2204                         tdq->tdq_ridx = tdq->tdq_idx;
 2205         }
 2206         ts = td->td_sched;
 2207         sched_pctcpu_update(ts, 1);
 2208         if (td->td_pri_class & PRI_FIFO_BIT)
 2209                 return;
 2210         if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
 2211                 /*
 2212                  * We used a tick; charge it to the thread so
 2213                  * that we can compute our interactivity.
 2214                  */
 2215                 td->td_sched->ts_runtime += tickincr;
 2216                 sched_interact_update(td);
 2217                 sched_priority(td);
 2218         }
 2219 
 2220         /*
 2221          * Force a context switch if the current thread has used up a full
 2222          * time slice (default is 100ms).
 2223          */
 2224         if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) {
 2225                 ts->ts_slice = sched_slice;
 2226                 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
 2227         }
 2228 }
 2229 
 2230 /*
 2231  * Called once per hz tick.
 2232  */
 2233 void
 2234 sched_tick(int cnt)
 2235 {
 2236 
 2237 }
 2238 
 2239 /*
 2240  * Return whether the current CPU has runnable tasks.  Used for in-kernel
 2241  * cooperative idle threads.
 2242  */
 2243 int
 2244 sched_runnable(void)
 2245 {
 2246         struct tdq *tdq;
 2247         int load;
 2248 
 2249         load = 1;
 2250 
 2251         tdq = TDQ_SELF();
 2252         if ((curthread->td_flags & TDF_IDLETD) != 0) {
 2253                 if (tdq->tdq_load > 0)
 2254                         goto out;
 2255         } else
 2256                 if (tdq->tdq_load - 1 > 0)
 2257                         goto out;
 2258         load = 0;
 2259 out:
 2260         return (load);
 2261 }
 2262 
 2263 /*
 2264  * Choose the highest priority thread to run.  The thread is removed from
 2265  * the run-queue while running however the load remains.  For SMP we set
 2266  * the tdq in the global idle bitmask if it idles here.
 2267  */
 2268 struct thread *
 2269 sched_choose(void)
 2270 {
 2271         struct thread *td;
 2272         struct tdq *tdq;
 2273 
 2274         tdq = TDQ_SELF();
 2275         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 2276         td = tdq_choose(tdq);
 2277         if (td) {
 2278                 tdq_runq_rem(tdq, td);
 2279                 tdq->tdq_lowpri = td->td_priority;
 2280                 return (td);
 2281         }
 2282         tdq->tdq_lowpri = PRI_MAX_IDLE;
 2283         return (PCPU_GET(idlethread));
 2284 }
 2285 
 2286 /*
 2287  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
 2288  * we always request it once we exit a critical section.
 2289  */
 2290 static inline void
 2291 sched_setpreempt(struct thread *td)
 2292 {
 2293         struct thread *ctd;
 2294         int cpri;
 2295         int pri;
 2296 
 2297         THREAD_LOCK_ASSERT(curthread, MA_OWNED);
 2298 
 2299         ctd = curthread;
 2300         pri = td->td_priority;
 2301         cpri = ctd->td_priority;
 2302         if (pri < cpri)
 2303                 ctd->td_flags |= TDF_NEEDRESCHED;
 2304         if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
 2305                 return;
 2306         if (!sched_shouldpreempt(pri, cpri, 0))
 2307                 return;
 2308         ctd->td_owepreempt = 1;
 2309 }
 2310 
 2311 /*
 2312  * Add a thread to a thread queue.  Select the appropriate runq and add the
 2313  * thread to it.  This is the internal function called when the tdq is
 2314  * predetermined.
 2315  */
 2316 void
 2317 tdq_add(struct tdq *tdq, struct thread *td, int flags)
 2318 {
 2319 
 2320         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 2321         KASSERT((td->td_inhibitors == 0),
 2322             ("sched_add: trying to run inhibited thread"));
 2323         KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
 2324             ("sched_add: bad thread state"));
 2325         KASSERT(td->td_flags & TDF_INMEM,
 2326             ("sched_add: thread swapped out"));
 2327 
 2328         if (td->td_priority < tdq->tdq_lowpri)
 2329                 tdq->tdq_lowpri = td->td_priority;
 2330         tdq_runq_add(tdq, td, flags);
 2331         tdq_load_add(tdq, td);
 2332 }
 2333 
 2334 /*
 2335  * Select the target thread queue and add a thread to it.  Request
 2336  * preemption or IPI a remote processor if required.
 2337  */
 2338 void
 2339 sched_add(struct thread *td, int flags)
 2340 {
 2341         struct tdq *tdq;
 2342 #ifdef SMP
 2343         int cpu;
 2344 #endif
 2345 
 2346         KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
 2347             "prio:%d", td->td_priority, KTR_ATTR_LINKED,
 2348             sched_tdname(curthread));
 2349         KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
 2350             KTR_ATTR_LINKED, sched_tdname(td));
 2351         SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 
 2352             flags & SRQ_PREEMPTED);
 2353         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2354         /*
 2355          * Recalculate the priority before we select the target cpu or
 2356          * run-queue.
 2357          */
 2358         if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
 2359                 sched_priority(td);
 2360 #ifdef SMP
 2361         /*
 2362          * Pick the destination cpu and if it isn't ours transfer to the
 2363          * target cpu.
 2364          */
 2365         cpu = sched_pickcpu(td, flags);
 2366         tdq = sched_setcpu(td, cpu, flags);
 2367         tdq_add(tdq, td, flags);
 2368         if (cpu != PCPU_GET(cpuid)) {
 2369                 tdq_notify(tdq, td);
 2370                 return;
 2371         }
 2372 #else
 2373         tdq = TDQ_SELF();
 2374         TDQ_LOCK(tdq);
 2375         /*
 2376          * Now that the thread is moving to the run-queue, set the lock
 2377          * to the scheduler's lock.
 2378          */
 2379         thread_lock_set(td, TDQ_LOCKPTR(tdq));
 2380         tdq_add(tdq, td, flags);
 2381 #endif
 2382         if (!(flags & SRQ_YIELDING))
 2383                 sched_setpreempt(td);
 2384 }
 2385 
 2386 /*
 2387  * Remove a thread from a run-queue without running it.  This is used
 2388  * when we're stealing a thread from a remote queue.  Otherwise all threads
 2389  * exit by calling sched_exit_thread() and sched_throw() themselves.
 2390  */
 2391 void
 2392 sched_rem(struct thread *td)
 2393 {
 2394         struct tdq *tdq;
 2395 
 2396         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
 2397             "prio:%d", td->td_priority);
 2398         SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
 2399         tdq = TDQ_CPU(td->td_sched->ts_cpu);
 2400         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 2401         MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 2402         KASSERT(TD_ON_RUNQ(td),
 2403             ("sched_rem: thread not on run queue"));
 2404         tdq_runq_rem(tdq, td);
 2405         tdq_load_rem(tdq, td);
 2406         TD_SET_CAN_RUN(td);
 2407         if (td->td_priority == tdq->tdq_lowpri)
 2408                 tdq_setlowpri(tdq, NULL);
 2409 }
 2410 
 2411 /*
 2412  * Fetch cpu utilization information.  Updates on demand.
 2413  */
 2414 fixpt_t
 2415 sched_pctcpu(struct thread *td)
 2416 {
 2417         fixpt_t pctcpu;
 2418         struct td_sched *ts;
 2419 
 2420         pctcpu = 0;
 2421         ts = td->td_sched;
 2422         if (ts == NULL)
 2423                 return (0);
 2424 
 2425         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2426         sched_pctcpu_update(ts, TD_IS_RUNNING(td));
 2427         if (ts->ts_ticks) {
 2428                 int rtick;
 2429 
 2430                 /* How many rtick per second ? */
 2431                 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
 2432                 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
 2433         }
 2434 
 2435         return (pctcpu);
 2436 }
 2437 
 2438 /*
 2439  * Enforce affinity settings for a thread.  Called after adjustments to
 2440  * cpumask.
 2441  */
 2442 void
 2443 sched_affinity(struct thread *td)
 2444 {
 2445 #ifdef SMP
 2446         struct td_sched *ts;
 2447 
 2448         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2449         ts = td->td_sched;
 2450         if (THREAD_CAN_SCHED(td, ts->ts_cpu))
 2451                 return;
 2452         if (TD_ON_RUNQ(td)) {
 2453                 sched_rem(td);
 2454                 sched_add(td, SRQ_BORING);
 2455                 return;
 2456         }
 2457         if (!TD_IS_RUNNING(td))
 2458                 return;
 2459         /*
 2460          * Force a switch before returning to userspace.  If the
 2461          * target thread is not running locally send an ipi to force
 2462          * the issue.
 2463          */
 2464         td->td_flags |= TDF_NEEDRESCHED;
 2465         if (td != curthread)
 2466                 ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
 2467 #endif
 2468 }
 2469 
 2470 /*
 2471  * Bind a thread to a target cpu.
 2472  */
 2473 void
 2474 sched_bind(struct thread *td, int cpu)
 2475 {
 2476         struct td_sched *ts;
 2477 
 2478         THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
 2479         KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
 2480         ts = td->td_sched;
 2481         if (ts->ts_flags & TSF_BOUND)
 2482                 sched_unbind(td);
 2483         KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
 2484         ts->ts_flags |= TSF_BOUND;
 2485         sched_pin();
 2486         if (PCPU_GET(cpuid) == cpu)
 2487                 return;
 2488         ts->ts_cpu = cpu;
 2489         /* When we return from mi_switch we'll be on the correct cpu. */
 2490         mi_switch(SW_VOL, NULL);
 2491 }
 2492 
 2493 /*
 2494  * Release a bound thread.
 2495  */
 2496 void
 2497 sched_unbind(struct thread *td)
 2498 {
 2499         struct td_sched *ts;
 2500 
 2501         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2502         KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
 2503         ts = td->td_sched;
 2504         if ((ts->ts_flags & TSF_BOUND) == 0)
 2505                 return;
 2506         ts->ts_flags &= ~TSF_BOUND;
 2507         sched_unpin();
 2508 }
 2509 
 2510 int
 2511 sched_is_bound(struct thread *td)
 2512 {
 2513         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2514         return (td->td_sched->ts_flags & TSF_BOUND);
 2515 }
 2516 
 2517 /*
 2518  * Basic yield call.
 2519  */
 2520 void
 2521 sched_relinquish(struct thread *td)
 2522 {
 2523         thread_lock(td);
 2524         mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
 2525         thread_unlock(td);
 2526 }
 2527 
 2528 /*
 2529  * Return the total system load.
 2530  */
 2531 int
 2532 sched_load(void)
 2533 {
 2534 #ifdef SMP
 2535         int total;
 2536         int i;
 2537 
 2538         total = 0;
 2539         CPU_FOREACH(i)
 2540                 total += TDQ_CPU(i)->tdq_sysload;
 2541         return (total);
 2542 #else
 2543         return (TDQ_SELF()->tdq_sysload);
 2544 #endif
 2545 }
 2546 
 2547 int
 2548 sched_sizeof_proc(void)
 2549 {
 2550         return (sizeof(struct proc));
 2551 }
 2552 
 2553 int
 2554 sched_sizeof_thread(void)
 2555 {
 2556         return (sizeof(struct thread) + sizeof(struct td_sched));
 2557 }
 2558 
 2559 #ifdef SMP
 2560 #define TDQ_IDLESPIN(tdq)                                               \
 2561     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
 2562 #else
 2563 #define TDQ_IDLESPIN(tdq)       1
 2564 #endif
 2565 
 2566 /*
 2567  * The actual idle process.
 2568  */
 2569 void
 2570 sched_idletd(void *dummy)
 2571 {
 2572         struct thread *td;
 2573         struct tdq *tdq;
 2574         int oldswitchcnt, switchcnt;
 2575         int i;
 2576 
 2577         mtx_assert(&Giant, MA_NOTOWNED);
 2578         td = curthread;
 2579         tdq = TDQ_SELF();
 2580         oldswitchcnt = -1;
 2581         for (;;) {
 2582                 if (tdq->tdq_load) {
 2583                         thread_lock(td);
 2584                         mi_switch(SW_VOL | SWT_IDLE, NULL);
 2585                         thread_unlock(td);
 2586                 }
 2587                 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
 2588 #ifdef SMP
 2589                 if (switchcnt != oldswitchcnt) {
 2590                         oldswitchcnt = switchcnt;
 2591                         if (tdq_idled(tdq) == 0)
 2592                                 continue;
 2593                 }
 2594                 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
 2595 #else
 2596                 oldswitchcnt = switchcnt;
 2597 #endif
 2598                 /*
 2599                  * If we're switching very frequently, spin while checking
 2600                  * for load rather than entering a low power state that 
 2601                  * may require an IPI.  However, don't do any busy
 2602                  * loops while on SMT machines as this simply steals
 2603                  * cycles from cores doing useful work.
 2604                  */
 2605                 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
 2606                         for (i = 0; i < sched_idlespins; i++) {
 2607                                 if (tdq->tdq_load)
 2608                                         break;
 2609                                 cpu_spinwait();
 2610                         }
 2611                 }
 2612 
 2613                 /* If there was context switch during spin, restart it. */
 2614                 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
 2615                 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
 2616                         continue;
 2617 
 2618                 /* Run main MD idle handler. */
 2619                 tdq->tdq_cpu_idle = 1;
 2620                 /*
 2621                  * Make sure that tdq_cpu_idle update is globally visible
 2622                  * before cpu_idle() read tdq_load.  The order is important
 2623                  * to avoid race with tdq_notify.
 2624                  */
 2625                 mb();
 2626                 cpu_idle(switchcnt * 4 > sched_idlespinthresh);
 2627                 tdq->tdq_cpu_idle = 0;
 2628 
 2629                 /*
 2630                  * Account thread-less hardware interrupts and
 2631                  * other wakeup reasons equal to context switches.
 2632                  */
 2633                 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
 2634                 if (switchcnt != oldswitchcnt)
 2635                         continue;
 2636                 tdq->tdq_switchcnt++;
 2637                 oldswitchcnt++;
 2638         }
 2639 }
 2640 
 2641 /*
 2642  * A CPU is entering for the first time or a thread is exiting.
 2643  */
 2644 void
 2645 sched_throw(struct thread *td)
 2646 {
 2647         struct thread *newtd;
 2648         struct tdq *tdq;
 2649 
 2650         tdq = TDQ_SELF();
 2651         if (td == NULL) {
 2652                 /* Correct spinlock nesting and acquire the correct lock. */
 2653                 TDQ_LOCK(tdq);
 2654                 spinlock_exit();
 2655                 PCPU_SET(switchtime, cpu_ticks());
 2656                 PCPU_SET(switchticks, ticks);
 2657         } else {
 2658                 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 2659                 tdq_load_rem(tdq, td);
 2660                 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
 2661                 td->td_lastcpu = td->td_oncpu;
 2662                 td->td_oncpu = NOCPU;
 2663         }
 2664         KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
 2665         newtd = choosethread();
 2666         TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
 2667         cpu_throw(td, newtd);           /* doesn't return */
 2668 }
 2669 
 2670 /*
 2671  * This is called from fork_exit().  Just acquire the correct locks and
 2672  * let fork do the rest of the work.
 2673  */
 2674 void
 2675 sched_fork_exit(struct thread *td)
 2676 {
 2677         struct td_sched *ts;
 2678         struct tdq *tdq;
 2679         int cpuid;
 2680 
 2681         /*
 2682          * Finish setting up thread glue so that it begins execution in a
 2683          * non-nested critical section with the scheduler lock held.
 2684          */
 2685         cpuid = PCPU_GET(cpuid);
 2686         tdq = TDQ_CPU(cpuid);
 2687         ts = td->td_sched;
 2688         if (TD_IS_IDLETHREAD(td))
 2689                 td->td_lock = TDQ_LOCKPTR(tdq);
 2690         MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 2691         td->td_oncpu = cpuid;
 2692         TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
 2693         lock_profile_obtain_lock_success(
 2694             &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
 2695 }
 2696 
 2697 /*
 2698  * Create on first use to catch odd startup conditons.
 2699  */
 2700 char *
 2701 sched_tdname(struct thread *td)
 2702 {
 2703 #ifdef KTR
 2704         struct td_sched *ts;
 2705 
 2706         ts = td->td_sched;
 2707         if (ts->ts_name[0] == '\0')
 2708                 snprintf(ts->ts_name, sizeof(ts->ts_name),
 2709                     "%s tid %d", td->td_name, td->td_tid);
 2710         return (ts->ts_name);
 2711 #else
 2712         return (td->td_name);
 2713 #endif
 2714 }
 2715 
 2716 #ifdef KTR
 2717 void
 2718 sched_clear_tdname(struct thread *td)
 2719 {
 2720         struct td_sched *ts;
 2721 
 2722         ts = td->td_sched;
 2723         ts->ts_name[0] = '\0';
 2724 }
 2725 #endif
 2726 
 2727 #ifdef SMP
 2728 
 2729 /*
 2730  * Build the CPU topology dump string. Is recursively called to collect
 2731  * the topology tree.
 2732  */
 2733 static int
 2734 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
 2735     int indent)
 2736 {
 2737         char cpusetbuf[CPUSETBUFSIZ];
 2738         int i, first;
 2739 
 2740         sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
 2741             "", 1 + indent / 2, cg->cg_level);
 2742         sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
 2743             cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
 2744         first = TRUE;
 2745         for (i = 0; i < MAXCPU; i++) {
 2746                 if (CPU_ISSET(i, &cg->cg_mask)) {
 2747                         if (!first)
 2748                                 sbuf_printf(sb, ", ");
 2749                         else
 2750                                 first = FALSE;
 2751                         sbuf_printf(sb, "%d", i);
 2752                 }
 2753         }
 2754         sbuf_printf(sb, "</cpu>\n");
 2755 
 2756         if (cg->cg_flags != 0) {
 2757                 sbuf_printf(sb, "%*s <flags>", indent, "");
 2758                 if ((cg->cg_flags & CG_FLAG_HTT) != 0)
 2759                         sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
 2760                 if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
 2761                         sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
 2762                 if ((cg->cg_flags & CG_FLAG_SMT) != 0)
 2763                         sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
 2764                 sbuf_printf(sb, "</flags>\n");
 2765         }
 2766 
 2767         if (cg->cg_children > 0) {
 2768                 sbuf_printf(sb, "%*s <children>\n", indent, "");
 2769                 for (i = 0; i < cg->cg_children; i++)
 2770                         sysctl_kern_sched_topology_spec_internal(sb, 
 2771                             &cg->cg_child[i], indent+2);
 2772                 sbuf_printf(sb, "%*s </children>\n", indent, "");
 2773         }
 2774         sbuf_printf(sb, "%*s</group>\n", indent, "");
 2775         return (0);
 2776 }
 2777 
 2778 /*
 2779  * Sysctl handler for retrieving topology dump. It's a wrapper for
 2780  * the recursive sysctl_kern_smp_topology_spec_internal().
 2781  */
 2782 static int
 2783 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
 2784 {
 2785         struct sbuf *topo;
 2786         int err;
 2787 
 2788         KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
 2789 
 2790         topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
 2791         if (topo == NULL)
 2792                 return (ENOMEM);
 2793 
 2794         sbuf_printf(topo, "<groups>\n");
 2795         err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
 2796         sbuf_printf(topo, "</groups>\n");
 2797 
 2798         if (err == 0) {
 2799                 sbuf_finish(topo);
 2800                 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
 2801         }
 2802         sbuf_delete(topo);
 2803         return (err);
 2804 }
 2805 
 2806 #endif
 2807 
 2808 static int
 2809 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
 2810 {
 2811         int error, new_val, period;
 2812 
 2813         period = 1000000 / realstathz;
 2814         new_val = period * sched_slice;
 2815         error = sysctl_handle_int(oidp, &new_val, 0, req);
 2816         if (error != 0 || req->newptr == NULL)
 2817                 return (error);
 2818         if (new_val <= 0)
 2819                 return (EINVAL);
 2820         sched_slice = imax(1, (new_val + period / 2) / period);
 2821         hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
 2822             realstathz);
 2823         return (0);
 2824 }
 2825 
 2826 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
 2827 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
 2828     "Scheduler name");
 2829 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
 2830     NULL, 0, sysctl_kern_quantum, "I",
 2831     "Quantum for timeshare threads in microseconds");
 2832 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
 2833     "Quantum for timeshare threads in stathz ticks");
 2834 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
 2835     "Interactivity score threshold");
 2836 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
 2837     &preempt_thresh, 0,
 2838     "Maximal (lowest) priority for preemption");
 2839 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
 2840     "Assign static kernel priorities to sleeping threads");
 2841 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
 2842     "Number of times idle thread will spin waiting for new work");
 2843 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
 2844     &sched_idlespinthresh, 0,
 2845     "Threshold before we will permit idle thread spinning");
 2846 #ifdef SMP
 2847 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
 2848     "Number of hz ticks to keep thread affinity for");
 2849 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
 2850     "Enables the long-term load balancer");
 2851 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
 2852     &balance_interval, 0,
 2853     "Average period in stathz ticks to run the long-term balancer");
 2854 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
 2855     "Attempts to steal work from other cores before idling");
 2856 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
 2857     "Minimum load on remote CPU before we'll steal");
 2858 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
 2859     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
 2860     "XML dump of detected CPU topology");
 2861 #endif
 2862 
 2863 /* ps compat.  All cpu percentages from ULE are weighted. */
 2864 static int ccpu = 0;
 2865 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");

Cache object: 290cef32ee365c95359b335f4b39167a


[ 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.