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: releng/7.2/sys/kern/sched_ule.c 204409 2010-02-27 10:55:43Z cperciva $");
   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/smp.h>
   57 #include <sys/sx.h>
   58 #include <sys/sysctl.h>
   59 #include <sys/sysproto.h>
   60 #include <sys/turnstile.h>
   61 #include <sys/umtx.h>
   62 #include <sys/vmmeter.h>
   63 #include <sys/cpuset.h>
   64 #ifdef KTRACE
   65 #include <sys/uio.h>
   66 #include <sys/ktrace.h>
   67 #endif
   68 
   69 #ifdef HWPMC_HOOKS
   70 #include <sys/pmckern.h>
   71 #endif
   72 
   73 #ifdef KDTRACE_HOOKS
   74 #include <sys/dtrace_bsd.h>
   75 int                             dtrace_vtime_active;
   76 dtrace_vtime_switch_func_t      dtrace_vtime_switch_func;
   77 #endif
   78 
   79 #include <machine/cpu.h>
   80 #include <machine/smp.h>
   81 
   82 #if !defined(__i386__) && !defined(__amd64__) && !defined(__arm__)
   83 #error "This architecture is not currently compatible with ULE"
   84 #endif
   85 
   86 #define KTR_ULE 0
   87 
   88 /*
   89  * Thread scheduler specific section.  All fields are protected
   90  * by the thread lock.
   91  */
   92 struct td_sched {       
   93         TAILQ_ENTRY(td_sched) ts_procq; /* Run queue. */
   94         struct thread   *ts_thread;     /* Active associated thread. */
   95         struct runq     *ts_runq;       /* Run-queue we're queued on. */
   96         short           ts_flags;       /* TSF_* flags. */
   97         u_char          ts_rqindex;     /* Run queue index. */
   98         u_char          ts_cpu;         /* CPU that we have affinity for. */
   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         /* The following variables are only used for pctcpu calculation */
  103         int             ts_ltick;       /* Last tick that we were running on */
  104         int             ts_ftick;       /* First tick that we were running on */
  105         int             ts_ticks;       /* Tick count */
  106 #ifdef SMP
  107         int             ts_rltick;      /* Real last tick, for affinity. */
  108 #endif
  109 };
  110 /* flags kept in ts_flags */
  111 #define TSF_BOUND       0x0001          /* Thread can not migrate. */
  112 #define TSF_XFERABLE    0x0002          /* Thread was added as transferable. */
  113 
  114 static struct td_sched td_sched0;
  115 
  116 #define THREAD_CAN_MIGRATE(td)  ((td)->td_pinned == 0)
  117 #define THREAD_CAN_SCHED(td, cpu)       \
  118     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
  119 
  120 /*
  121  * Cpu percentage computation macros and defines.
  122  *
  123  * SCHED_TICK_SECS:     Number of seconds to average the cpu usage across.
  124  * SCHED_TICK_TARG:     Number of hz ticks to average the cpu usage across.
  125  * SCHED_TICK_MAX:      Maximum number of ticks before scaling back.
  126  * SCHED_TICK_SHIFT:    Shift factor to avoid rounding away results.
  127  * SCHED_TICK_HZ:       Compute the number of hz ticks for a given ticks count.
  128  * SCHED_TICK_TOTAL:    Gives the amount of time we've been recording ticks.
  129  */
  130 #define SCHED_TICK_SECS         10
  131 #define SCHED_TICK_TARG         (hz * SCHED_TICK_SECS)
  132 #define SCHED_TICK_MAX          (SCHED_TICK_TARG + hz)
  133 #define SCHED_TICK_SHIFT        10
  134 #define SCHED_TICK_HZ(ts)       ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
  135 #define SCHED_TICK_TOTAL(ts)    (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
  136 
  137 /*
  138  * These macros determine priorities for non-interactive threads.  They are
  139  * assigned a priority based on their recent cpu utilization as expressed
  140  * by the ratio of ticks to the tick total.  NHALF priorities at the start
  141  * and end of the MIN to MAX timeshare range are only reachable with negative
  142  * or positive nice respectively.
  143  *
  144  * PRI_RANGE:   Priority range for utilization dependent priorities.
  145  * PRI_NRESV:   Number of nice values.
  146  * PRI_TICKS:   Compute a priority in PRI_RANGE from the ticks count and total.
  147  * PRI_NICE:    Determines the part of the priority inherited from nice.
  148  */
  149 #define SCHED_PRI_NRESV         (PRIO_MAX - PRIO_MIN)
  150 #define SCHED_PRI_NHALF         (SCHED_PRI_NRESV / 2)
  151 #define SCHED_PRI_MIN           (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
  152 #define SCHED_PRI_MAX           (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
  153 #define SCHED_PRI_RANGE         (SCHED_PRI_MAX - SCHED_PRI_MIN)
  154 #define SCHED_PRI_TICKS(ts)                                             \
  155     (SCHED_TICK_HZ((ts)) /                                              \
  156     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
  157 #define SCHED_PRI_NICE(nice)    (nice)
  158 
  159 /*
  160  * These determine the interactivity of a process.  Interactivity differs from
  161  * cpu utilization in that it expresses the voluntary time slept vs time ran
  162  * while cpu utilization includes all time not running.  This more accurately
  163  * models the intent of the thread.
  164  *
  165  * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
  166  *              before throttling back.
  167  * SLP_RUN_FORK:        Maximum slp+run time to inherit at fork time.
  168  * INTERACT_MAX:        Maximum interactivity value.  Smaller is better.
  169  * INTERACT_THRESH:     Threshhold for placement on the current runq.
  170  */
  171 #define SCHED_SLP_RUN_MAX       ((hz * 5) << SCHED_TICK_SHIFT)
  172 #define SCHED_SLP_RUN_FORK      ((hz / 2) << SCHED_TICK_SHIFT)
  173 #define SCHED_INTERACT_MAX      (100)
  174 #define SCHED_INTERACT_HALF     (SCHED_INTERACT_MAX / 2)
  175 #define SCHED_INTERACT_THRESH   (30)
  176 
  177 /*
  178  * tickincr:            Converts a stathz tick into a hz domain scaled by
  179  *                      the shift factor.  Without the shift the error rate
  180  *                      due to rounding would be unacceptably high.
  181  * realstathz:          stathz is sometimes 0 and run off of hz.
  182  * sched_slice:         Runtime of each thread before rescheduling.
  183  * preempt_thresh:      Priority threshold for preemption and remote IPIs.
  184  */
  185 static int sched_interact = SCHED_INTERACT_THRESH;
  186 static int realstathz;
  187 static int tickincr;
  188 static int sched_slice;
  189 #ifdef PREEMPTION
  190 #ifdef FULL_PREEMPTION
  191 static int preempt_thresh = PRI_MAX_IDLE;
  192 #else
  193 static int preempt_thresh = PRI_MIN_KERN;
  194 #endif
  195 #else 
  196 static int preempt_thresh = 0;
  197 #endif
  198 
  199 /*
  200  * tdq - per processor runqs and statistics.  All fields are protected by the
  201  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
  202  * locking in sched_pickcpu();
  203  */
  204 struct tdq {
  205         struct mtx      *tdq_lock;              /* Pointer to group lock. */
  206         struct runq     tdq_realtime;           /* real-time run queue. */
  207         struct runq     tdq_timeshare;          /* timeshare run queue. */
  208         struct runq     tdq_idle;               /* Queue of IDLE threads. */
  209         int             tdq_load;               /* Aggregate load. */
  210         u_char          tdq_idx;                /* Current insert index. */
  211         u_char          tdq_ridx;               /* Current removal index. */
  212 #ifdef SMP
  213         u_char          tdq_lowpri;             /* Lowest priority thread. */
  214         int             tdq_transferable;       /* Transferable thread count. */
  215         LIST_ENTRY(tdq) tdq_siblings;           /* Next in tdq group. */
  216         struct tdq_group *tdq_group;            /* Our processor group. */
  217 #else
  218         int             tdq_sysload;            /* For loadavg, !ITHD load. */
  219 #endif
  220 } __aligned(64);
  221 
  222 
  223 #ifdef SMP
  224 /*
  225  * tdq groups are groups of processors which can cheaply share threads.  When
  226  * one processor in the group goes idle it will check the runqs of the other
  227  * processors in its group prior to halting and waiting for an interrupt.
  228  * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
  229  * In a numa environment we'd want an idle bitmap per group and a two tiered
  230  * load balancer.
  231  */
  232 struct tdq_group {
  233         struct mtx      tdg_lock;       /* Protects all fields below. */
  234         int             tdg_cpus;       /* Count of CPUs in this tdq group. */
  235         cpumask_t       tdg_cpumask;    /* Mask of cpus in this group. */
  236         cpumask_t       tdg_idlemask;   /* Idle cpus in this group. */
  237         cpumask_t       tdg_mask;       /* Bit mask for first cpu. */
  238         int             tdg_load;       /* Total load of this group. */
  239         int     tdg_transferable;       /* Transferable load of this group. */
  240         LIST_HEAD(, tdq) tdg_members;   /* Linked list of all members. */
  241         char            tdg_name[16];   /* lock name. */
  242 } __aligned(64);
  243 
  244 #define SCHED_AFFINITY_DEFAULT  (max(1, hz / 300))
  245 #define SCHED_AFFINITY(ts)      ((ts)->ts_rltick > ticks - affinity)
  246 
  247 /*
  248  * Run-time tunables.
  249  */
  250 static int rebalance = 1;
  251 static int balance_interval = 128;      /* Default set in sched_initticks(). */
  252 static int pick_pri = 1;
  253 static int affinity;
  254 static int tryself = 1;
  255 static int steal_htt = 1;
  256 static int steal_idle = 1;
  257 static int steal_thresh = 2;
  258 static int topology = 0;
  259 
  260 /*
  261  * One thread queue per processor.
  262  */
  263 static volatile cpumask_t tdq_idle;
  264 static int tdg_maxid;
  265 static struct tdq       tdq_cpu[MAXCPU];
  266 static struct tdq_group tdq_groups[MAXCPU];
  267 static struct tdq       *balance_tdq;
  268 static int balance_group_ticks;
  269 static int balance_ticks;
  270 
  271 #define TDQ_SELF()      (&tdq_cpu[PCPU_GET(cpuid)])
  272 #define TDQ_CPU(x)      (&tdq_cpu[(x)])
  273 #define TDQ_ID(x)       ((int)((x) - tdq_cpu))
  274 #define TDQ_GROUP(x)    (&tdq_groups[(x)])
  275 #define TDG_ID(x)       ((int)((x) - tdq_groups))
  276 #else   /* !SMP */
  277 static struct tdq       tdq_cpu;
  278 static struct mtx       tdq_lock;
  279 
  280 #define TDQ_ID(x)       (0)
  281 #define TDQ_SELF()      (&tdq_cpu)
  282 #define TDQ_CPU(x)      (&tdq_cpu)
  283 #endif
  284 
  285 #define TDQ_LOCK_ASSERT(t, type)        mtx_assert(TDQ_LOCKPTR((t)), (type))
  286 #define TDQ_LOCK(t)             mtx_lock_spin(TDQ_LOCKPTR((t)))
  287 #define TDQ_LOCK_FLAGS(t, f)    mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
  288 #define TDQ_UNLOCK(t)           mtx_unlock_spin(TDQ_LOCKPTR((t)))
  289 #define TDQ_LOCKPTR(t)          ((t)->tdq_lock)
  290 
  291 static void sched_priority(struct thread *);
  292 static void sched_thread_priority(struct thread *, u_char);
  293 static int sched_interact_score(struct thread *);
  294 static void sched_interact_update(struct thread *);
  295 static void sched_interact_fork(struct thread *);
  296 static void sched_pctcpu_update(struct td_sched *);
  297 
  298 /* Operations on per processor queues */
  299 static struct td_sched * tdq_choose(struct tdq *);
  300 static void tdq_setup(struct tdq *);
  301 static void tdq_load_add(struct tdq *, struct td_sched *);
  302 static void tdq_load_rem(struct tdq *, struct td_sched *);
  303 static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int);
  304 static __inline void tdq_runq_rem(struct tdq *, struct td_sched *);
  305 void tdq_print(int cpu);
  306 static void runq_print(struct runq *rq);
  307 static void tdq_add(struct tdq *, struct thread *, int);
  308 #ifdef SMP
  309 static void tdq_move(struct tdq *, struct tdq *);
  310 static int tdq_idled(struct tdq *);
  311 static void tdq_notify(struct td_sched *);
  312 static struct td_sched *tdq_steal(struct tdq *, int);
  313 static struct td_sched *runq_steal(struct runq *, int);
  314 static int sched_pickcpu(struct thread *, int);
  315 static void sched_balance(void);
  316 static void sched_balance_groups(void);
  317 static void sched_balance_group(struct tdq_group *);
  318 static void sched_balance_pair(struct tdq *, struct tdq *);
  319 static inline struct tdq *sched_setcpu(struct td_sched *, int, int);
  320 static inline struct mtx *thread_block_switch(struct thread *);
  321 static inline void thread_unblock_switch(struct thread *, struct mtx *);
  322 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
  323 #endif
  324 
  325 static void sched_setup(void *dummy);
  326 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
  327 
  328 static void sched_initticks(void *dummy);
  329 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
  330     NULL);
  331 
  332 /*
  333  * Print the threads waiting on a run-queue.
  334  */
  335 static void
  336 runq_print(struct runq *rq)
  337 {
  338         struct rqhead *rqh;
  339         struct td_sched *ts;
  340         int pri;
  341         int j;
  342         int i;
  343 
  344         for (i = 0; i < RQB_LEN; i++) {
  345                 printf("\t\trunq bits %d 0x%zx\n",
  346                     i, rq->rq_status.rqb_bits[i]);
  347                 for (j = 0; j < RQB_BPW; j++)
  348                         if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
  349                                 pri = j + (i << RQB_L2BPW);
  350                                 rqh = &rq->rq_queues[pri];
  351                                 TAILQ_FOREACH(ts, rqh, ts_procq) {
  352                                         printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
  353                                             ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri);
  354                                 }
  355                         }
  356         }
  357 }
  358 
  359 /*
  360  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
  361  */
  362 void
  363 tdq_print(int cpu)
  364 {
  365         struct tdq *tdq;
  366 
  367         tdq = TDQ_CPU(cpu);
  368 
  369         printf("tdq %d:\n", TDQ_ID(tdq));
  370         printf("\tlockptr         %p\n", TDQ_LOCKPTR(tdq));
  371         printf("\tload:           %d\n", tdq->tdq_load);
  372         printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
  373         printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
  374         printf("\trealtime runq:\n");
  375         runq_print(&tdq->tdq_realtime);
  376         printf("\ttimeshare runq:\n");
  377         runq_print(&tdq->tdq_timeshare);
  378         printf("\tidle runq:\n");
  379         runq_print(&tdq->tdq_idle);
  380 #ifdef SMP
  381         printf("\tload transferable: %d\n", tdq->tdq_transferable);
  382         printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
  383         printf("\tgroup:             %d\n", TDG_ID(tdq->tdq_group));
  384         printf("\tLock name:         %s\n", tdq->tdq_group->tdg_name);
  385 #endif
  386 }
  387 
  388 #define TS_RQ_PPQ       (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
  389 /*
  390  * Add a thread to the actual run-queue.  Keeps transferable counts up to
  391  * date with what is actually on the run-queue.  Selects the correct
  392  * queue position for timeshare threads.
  393  */
  394 static __inline void
  395 tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags)
  396 {
  397         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  398         THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
  399 #ifdef SMP
  400         if (THREAD_CAN_MIGRATE(ts->ts_thread)) {
  401                 tdq->tdq_transferable++;
  402                 tdq->tdq_group->tdg_transferable++;
  403                 ts->ts_flags |= TSF_XFERABLE;
  404         }
  405 #endif
  406         if (ts->ts_runq == &tdq->tdq_timeshare) {
  407                 u_char pri;
  408 
  409                 pri = ts->ts_thread->td_priority;
  410                 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
  411                         ("Invalid priority %d on timeshare runq", pri));
  412                 /*
  413                  * This queue contains only priorities between MIN and MAX
  414                  * realtime.  Use the whole queue to represent these values.
  415                  */
  416                 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
  417                         pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
  418                         pri = (pri + tdq->tdq_idx) % RQ_NQS;
  419                         /*
  420                          * This effectively shortens the queue by one so we
  421                          * can have a one slot difference between idx and
  422                          * ridx while we wait for threads to drain.
  423                          */
  424                         if (tdq->tdq_ridx != tdq->tdq_idx &&
  425                             pri == tdq->tdq_ridx)
  426                                 pri = (unsigned char)(pri - 1) % RQ_NQS;
  427                 } else
  428                         pri = tdq->tdq_ridx;
  429                 runq_add_pri(ts->ts_runq, ts, pri, flags);
  430         } else
  431                 runq_add(ts->ts_runq, ts, flags);
  432 }
  433 
  434 /* 
  435  * Remove a thread from a run-queue.  This typically happens when a thread
  436  * is selected to run.  Running threads are not on the queue and the
  437  * transferable count does not reflect them.
  438  */
  439 static __inline void
  440 tdq_runq_rem(struct tdq *tdq, struct td_sched *ts)
  441 {
  442         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  443         KASSERT(ts->ts_runq != NULL,
  444             ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread));
  445 #ifdef SMP
  446         if (ts->ts_flags & TSF_XFERABLE) {
  447                 tdq->tdq_transferable--;
  448                 tdq->tdq_group->tdg_transferable--;
  449                 ts->ts_flags &= ~TSF_XFERABLE;
  450         }
  451 #endif
  452         if (ts->ts_runq == &tdq->tdq_timeshare) {
  453                 if (tdq->tdq_idx != tdq->tdq_ridx)
  454                         runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx);
  455                 else
  456                         runq_remove_idx(ts->ts_runq, ts, NULL);
  457                 /*
  458                  * For timeshare threads we update the priority here so
  459                  * the priority reflects the time we've been sleeping.
  460                  */
  461                 ts->ts_ltick = ticks;
  462                 sched_pctcpu_update(ts);
  463                 sched_priority(ts->ts_thread);
  464         } else
  465                 runq_remove(ts->ts_runq, ts);
  466 }
  467 
  468 /*
  469  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
  470  * for this thread to the referenced thread queue.
  471  */
  472 static void
  473 tdq_load_add(struct tdq *tdq, struct td_sched *ts)
  474 {
  475         int class;
  476 
  477         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  478         THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
  479         class = PRI_BASE(ts->ts_thread->td_pri_class);
  480         tdq->tdq_load++;
  481         CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load);
  482         if (class != PRI_ITHD &&
  483             (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
  484 #ifdef SMP
  485                 tdq->tdq_group->tdg_load++;
  486 #else
  487                 tdq->tdq_sysload++;
  488 #endif
  489 }
  490 
  491 /*
  492  * Remove the load from a thread that is transitioning to a sleep state or
  493  * exiting.
  494  */
  495 static void
  496 tdq_load_rem(struct tdq *tdq, struct td_sched *ts)
  497 {
  498         int class;
  499 
  500         THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
  501         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  502         class = PRI_BASE(ts->ts_thread->td_pri_class);
  503         if (class != PRI_ITHD &&
  504             (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
  505 #ifdef SMP
  506                 tdq->tdq_group->tdg_load--;
  507 #else
  508                 tdq->tdq_sysload--;
  509 #endif
  510         KASSERT(tdq->tdq_load != 0,
  511             ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
  512         tdq->tdq_load--;
  513         CTR1(KTR_SCHED, "load: %d", tdq->tdq_load);
  514         ts->ts_runq = NULL;
  515 }
  516 
  517 #ifdef SMP
  518 /*
  519  * sched_balance is a simple CPU load balancing algorithm.  It operates by
  520  * finding the least loaded and most loaded cpu and equalizing their load
  521  * by migrating some processes.
  522  *
  523  * Dealing only with two CPUs at a time has two advantages.  Firstly, most
  524  * installations will only have 2 cpus.  Secondly, load balancing too much at
  525  * once can have an unpleasant effect on the system.  The scheduler rarely has
  526  * enough information to make perfect decisions.  So this algorithm chooses
  527  * simplicity and more gradual effects on load in larger systems.
  528  *
  529  */
  530 static void
  531 sched_balance()
  532 {
  533         struct tdq_group *high;
  534         struct tdq_group *low;
  535         struct tdq_group *tdg;
  536         struct tdq *tdq;
  537         int cnt;
  538         int i;
  539 
  540         /*
  541          * Select a random time between .5 * balance_interval and
  542          * 1.5 * balance_interval.
  543          */
  544         balance_ticks = max(balance_interval / 2, 1);
  545         balance_ticks += random() % balance_interval;
  546         if (smp_started == 0 || rebalance == 0)
  547                 return;
  548         tdq = TDQ_SELF();
  549         TDQ_UNLOCK(tdq);
  550         low = high = NULL;
  551         i = random() % (tdg_maxid + 1);
  552         for (cnt = 0; cnt <= tdg_maxid; cnt++) {
  553                 tdg = TDQ_GROUP(i);
  554                 /*
  555                  * Find the CPU with the highest load that has some
  556                  * threads to transfer.
  557                  */
  558                 if ((high == NULL || tdg->tdg_load > high->tdg_load)
  559                     && tdg->tdg_transferable)
  560                         high = tdg;
  561                 if (low == NULL || tdg->tdg_load < low->tdg_load)
  562                         low = tdg;
  563                 if (++i > tdg_maxid)
  564                         i = 0;
  565         }
  566         if (low != NULL && high != NULL && high != low)
  567                 sched_balance_pair(LIST_FIRST(&high->tdg_members),
  568                     LIST_FIRST(&low->tdg_members));
  569         TDQ_LOCK(tdq);
  570 }
  571 
  572 /*
  573  * Balance load between CPUs in a group.  Will only migrate within the group.
  574  */
  575 static void
  576 sched_balance_groups()
  577 {
  578         struct tdq *tdq;
  579         int i;
  580 
  581         /*
  582          * Select a random time between .5 * balance_interval and
  583          * 1.5 * balance_interval.
  584          */
  585         balance_group_ticks = max(balance_interval / 2, 1);
  586         balance_group_ticks += random() % balance_interval;
  587         if (smp_started == 0 || rebalance == 0)
  588                 return;
  589         tdq = TDQ_SELF();
  590         TDQ_UNLOCK(tdq);
  591         for (i = 0; i <= tdg_maxid; i++)
  592                 sched_balance_group(TDQ_GROUP(i));
  593         TDQ_LOCK(tdq);
  594 }
  595 
  596 /*
  597  * Finds the greatest imbalance between two tdqs in a group.
  598  */
  599 static void
  600 sched_balance_group(struct tdq_group *tdg)
  601 {
  602         struct tdq *tdq;
  603         struct tdq *high;
  604         struct tdq *low;
  605         int load;
  606 
  607         if (tdg->tdg_transferable == 0)
  608                 return;
  609         low = NULL;
  610         high = NULL;
  611         LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
  612                 load = tdq->tdq_load;
  613                 if (high == NULL || load > high->tdq_load)
  614                         high = tdq;
  615                 if (low == NULL || load < low->tdq_load)
  616                         low = tdq;
  617         }
  618         if (high != NULL && low != NULL && high != low)
  619                 sched_balance_pair(high, low);
  620 }
  621 
  622 /*
  623  * Lock two thread queues using their address to maintain lock order.
  624  */
  625 static void
  626 tdq_lock_pair(struct tdq *one, struct tdq *two)
  627 {
  628         if (one < two) {
  629                 TDQ_LOCK(one);
  630                 TDQ_LOCK_FLAGS(two, MTX_DUPOK);
  631         } else {
  632                 TDQ_LOCK(two);
  633                 TDQ_LOCK_FLAGS(one, MTX_DUPOK);
  634         }
  635 }
  636 
  637 /*
  638  * Unlock two thread queues.  Order is not important here.
  639  */
  640 static void
  641 tdq_unlock_pair(struct tdq *one, struct tdq *two)
  642 {
  643         TDQ_UNLOCK(one);
  644         TDQ_UNLOCK(two);
  645 }
  646 
  647 /*
  648  * Transfer load between two imbalanced thread queues.
  649  */
  650 static void
  651 sched_balance_pair(struct tdq *high, struct tdq *low)
  652 {
  653         int transferable;
  654         int high_load;
  655         int low_load;
  656         int move;
  657         int diff;
  658         int i;
  659 
  660         tdq_lock_pair(high, low);
  661         /*
  662          * If we're transfering within a group we have to use this specific
  663          * tdq's transferable count, otherwise we can steal from other members
  664          * of the group.
  665          */
  666         if (high->tdq_group == low->tdq_group) {
  667                 transferable = high->tdq_transferable;
  668                 high_load = high->tdq_load;
  669                 low_load = low->tdq_load;
  670         } else {
  671                 transferable = high->tdq_group->tdg_transferable;
  672                 high_load = high->tdq_group->tdg_load;
  673                 low_load = low->tdq_group->tdg_load;
  674         }
  675         /*
  676          * Determine what the imbalance is and then adjust that to how many
  677          * threads we actually have to give up (transferable).
  678          */
  679         if (transferable != 0) {
  680                 diff = high_load - low_load;
  681                 move = diff / 2;
  682                 if (diff & 0x1)
  683                         move++;
  684                 move = min(move, transferable);
  685                 for (i = 0; i < move; i++)
  686                         tdq_move(high, low);
  687                 /*
  688                  * IPI the target cpu to force it to reschedule with the new
  689                  * workload.
  690                  */
  691                 ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT);
  692         }
  693         tdq_unlock_pair(high, low);
  694         return;
  695 }
  696 
  697 /*
  698  * Move a thread from one thread queue to another.
  699  */
  700 static void
  701 tdq_move(struct tdq *from, struct tdq *to)
  702 {
  703         struct td_sched *ts;
  704         struct thread *td;
  705         struct tdq *tdq;
  706         int cpu;
  707 
  708         TDQ_LOCK_ASSERT(from, MA_OWNED);
  709         TDQ_LOCK_ASSERT(to, MA_OWNED);
  710 
  711         tdq = from;
  712         cpu = TDQ_ID(to);
  713         ts = tdq_steal(tdq, cpu);
  714         if (ts == NULL) {
  715                 struct tdq_group *tdg;
  716 
  717                 tdg = tdq->tdq_group;
  718                 LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
  719                         if (tdq == from || tdq->tdq_transferable == 0)
  720                                 continue;
  721                         ts = tdq_steal(tdq, cpu);
  722                         break;
  723                 }
  724                 if (ts == NULL)
  725                         return;
  726         }
  727         if (tdq == to)
  728                 return;
  729         td = ts->ts_thread;
  730         /*
  731          * Although the run queue is locked the thread may be blocked.  Lock
  732          * it to clear this and acquire the run-queue lock.
  733          */
  734         thread_lock(td);
  735         /* Drop recursive lock on from acquired via thread_lock(). */
  736         TDQ_UNLOCK(from);
  737         sched_rem(td);
  738         ts->ts_cpu = cpu;
  739         td->td_lock = TDQ_LOCKPTR(to);
  740         tdq_add(to, td, SRQ_YIELDING);
  741 }
  742 
  743 /*
  744  * This tdq has idled.  Try to steal a thread from another cpu and switch
  745  * to it.
  746  */
  747 static int
  748 tdq_idled(struct tdq *tdq)
  749 {
  750         struct tdq_group *tdg;
  751         struct tdq *steal;
  752         int highload;
  753         int highcpu;
  754         int cpu;
  755 
  756         if (smp_started == 0 || steal_idle == 0)
  757                 return (1);
  758         /* We don't want to be preempted while we're iterating over tdqs */
  759         spinlock_enter();
  760         tdg = tdq->tdq_group;
  761         /*
  762          * If we're in a cpu group, try and steal threads from another cpu in
  763          * the group before idling.  In a HTT group all cpus share the same
  764          * run-queue lock, however, we still need a recursive lock to
  765          * call tdq_move().
  766          */
  767         if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) {
  768                 TDQ_LOCK(tdq);
  769                 LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) {
  770                         if (steal == tdq || steal->tdq_transferable == 0)
  771                                 continue;
  772                         TDQ_LOCK(steal);
  773                         goto steal;
  774                 }
  775                 TDQ_UNLOCK(tdq);
  776         }
  777         /*
  778          * Find the least loaded CPU with a transferable thread and attempt
  779          * to steal it.  We make a lockless pass and then verify that the
  780          * thread is still available after locking.
  781          */
  782         for (;;) {
  783                 highcpu = 0;
  784                 highload = 0;
  785                 for (cpu = 0; cpu <= mp_maxid; cpu++) {
  786                         if (CPU_ABSENT(cpu))
  787                                 continue;
  788                         steal = TDQ_CPU(cpu);
  789                         if (steal->tdq_transferable == 0)
  790                                 continue;
  791                         if (steal->tdq_load < highload)
  792                                 continue;
  793                         highload = steal->tdq_load;
  794                         highcpu = cpu;
  795                 }
  796                 if (highload < steal_thresh)
  797                         break;
  798                 steal = TDQ_CPU(highcpu);
  799                 if (steal == tdq)
  800                         break;
  801                 tdq_lock_pair(tdq, steal);
  802                 if (steal->tdq_load >= steal_thresh && steal->tdq_transferable)
  803                         goto steal;
  804                 tdq_unlock_pair(tdq, steal);
  805         }
  806         spinlock_exit();
  807         return (1);
  808 steal:
  809         spinlock_exit();
  810         tdq_move(steal, tdq);
  811         TDQ_UNLOCK(steal);
  812         mi_switch(SW_VOL, NULL);
  813         thread_unlock(curthread);
  814 
  815         return (0);
  816 }
  817 
  818 /*
  819  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
  820  */
  821 static void
  822 tdq_notify(struct td_sched *ts)
  823 {
  824         struct thread *ctd;
  825         struct pcpu *pcpu;
  826         int cpri;
  827         int pri;
  828         int cpu;
  829 
  830         cpu = ts->ts_cpu;
  831         pri = ts->ts_thread->td_priority;
  832         pcpu = pcpu_find(cpu);
  833         ctd = pcpu->pc_curthread;
  834         cpri = ctd->td_priority;
  835 
  836         /*
  837          * If our priority is not better than the current priority there is
  838          * nothing to do.
  839          */
  840         if (pri > cpri)
  841                 return;
  842         /*
  843          * Always IPI idle.
  844          */
  845         if (cpri > PRI_MIN_IDLE)
  846                 goto sendipi;
  847         /*
  848          * If we're realtime or better and there is timeshare or worse running
  849          * send an IPI.
  850          */
  851         if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
  852                 goto sendipi;
  853         /*
  854          * Otherwise only IPI if we exceed the threshold.
  855          */
  856         if (pri > preempt_thresh)
  857                 return;
  858 sendipi:
  859         ipi_selected(1 << cpu, IPI_PREEMPT);
  860 }
  861 
  862 /*
  863  * Steals load from a timeshare queue.  Honors the rotating queue head
  864  * index.
  865  */
  866 static struct td_sched *
  867 runq_steal_from(struct runq *rq, int cpu, u_char start)
  868 {
  869         struct td_sched *ts;
  870         struct rqbits *rqb;
  871         struct rqhead *rqh;
  872         int first;
  873         int bit;
  874         int pri;
  875         int i;
  876 
  877         rqb = &rq->rq_status;
  878         bit = start & (RQB_BPW -1);
  879         pri = 0;
  880         first = 0;
  881 again:
  882         for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
  883                 if (rqb->rqb_bits[i] == 0)
  884                         continue;
  885                 if (bit != 0) {
  886                         for (pri = bit; pri < RQB_BPW; pri++)
  887                                 if (rqb->rqb_bits[i] & (1ul << pri))
  888                                         break;
  889                         if (pri >= RQB_BPW)
  890                                 continue;
  891                 } else
  892                         pri = RQB_FFS(rqb->rqb_bits[i]);
  893                 pri += (i << RQB_L2BPW);
  894                 rqh = &rq->rq_queues[pri];
  895                 TAILQ_FOREACH(ts, rqh, ts_procq) {
  896                         if (first && THREAD_CAN_MIGRATE(ts->ts_thread) &&
  897                             THREAD_CAN_SCHED(ts->ts_thread, cpu))
  898                                 return (ts);
  899                         first = 1;
  900                 }
  901         }
  902         if (start != 0) {
  903                 start = 0;
  904                 goto again;
  905         }
  906 
  907         return (NULL);
  908 }
  909 
  910 /*
  911  * Steals load from a standard linear queue.
  912  */
  913 static struct td_sched *
  914 runq_steal(struct runq *rq, int cpu)
  915 {
  916         struct rqhead *rqh;
  917         struct rqbits *rqb;
  918         struct td_sched *ts;
  919         int word;
  920         int bit;
  921 
  922         rqb = &rq->rq_status;
  923         for (word = 0; word < RQB_LEN; word++) {
  924                 if (rqb->rqb_bits[word] == 0)
  925                         continue;
  926                 for (bit = 0; bit < RQB_BPW; bit++) {
  927                         if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
  928                                 continue;
  929                         rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
  930                         TAILQ_FOREACH(ts, rqh, ts_procq)
  931                                 if (THREAD_CAN_MIGRATE(ts->ts_thread) &&
  932                                     THREAD_CAN_SCHED(ts->ts_thread, cpu))
  933                                         return (ts);
  934                 }
  935         }
  936         return (NULL);
  937 }
  938 
  939 /*
  940  * Attempt to steal a thread in priority order from a thread queue.
  941  */
  942 static struct td_sched *
  943 tdq_steal(struct tdq *tdq, int cpu)
  944 {
  945         struct td_sched *ts;
  946 
  947         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
  948         if ((ts = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
  949                 return (ts);
  950         if ((ts = runq_steal_from(&tdq->tdq_timeshare,
  951             cpu, tdq->tdq_ridx)) != NULL)
  952                 return (ts);
  953         return (runq_steal(&tdq->tdq_idle, cpu));
  954 }
  955 
  956 /*
  957  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
  958  * current lock and returns with the assigned queue locked.
  959  */
  960 static inline struct tdq *
  961 sched_setcpu(struct td_sched *ts, int cpu, int flags)
  962 {
  963         struct thread *td;
  964         struct tdq *tdq;
  965 
  966         THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
  967 
  968         tdq = TDQ_CPU(cpu);
  969         td = ts->ts_thread;
  970         ts->ts_cpu = cpu;
  971 
  972         /* If the lock matches just return the queue. */
  973         if (td->td_lock == TDQ_LOCKPTR(tdq))
  974                 return (tdq);
  975 #ifdef notyet
  976         /*
  977          * If the thread isn't running its lockptr is a
  978          * turnstile or a sleepqueue.  We can just lock_set without
  979          * blocking.
  980          */
  981         if (TD_CAN_RUN(td)) {
  982                 TDQ_LOCK(tdq);
  983                 thread_lock_set(td, TDQ_LOCKPTR(tdq));
  984                 return (tdq);
  985         }
  986 #endif
  987         /*
  988          * The hard case, migration, we need to block the thread first to
  989          * prevent order reversals with other cpus locks.
  990          */
  991         thread_lock_block(td);
  992         TDQ_LOCK(tdq);
  993         thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
  994         return (tdq);
  995 }
  996 
  997 /*
  998  * Find the thread queue running the lowest priority thread.
  999  */
 1000 static int
 1001 tdq_lowestpri(struct thread *td)
 1002 {
 1003         struct tdq *tdq;
 1004         int lowpri;
 1005         int lowcpu;
 1006         int lowload;
 1007         int load;
 1008         int cpu;
 1009         int pri;
 1010 
 1011         lowload = 0;
 1012         lowpri = lowcpu = 0;
 1013         for (cpu = 0; cpu <= mp_maxid; cpu++) {
 1014                 if (CPU_ABSENT(cpu))
 1015                         continue;
 1016                 if (!THREAD_CAN_SCHED(td, cpu))
 1017                         continue;
 1018                 tdq = TDQ_CPU(cpu);
 1019                 pri = tdq->tdq_lowpri;
 1020                 load = TDQ_CPU(cpu)->tdq_load;
 1021                 CTR4(KTR_ULE,
 1022                     "cpu %d pri %d lowcpu %d lowpri %d",
 1023                     cpu, pri, lowcpu, lowpri);
 1024                 if (pri < lowpri)
 1025                         continue;
 1026                 if (lowpri && lowpri == pri && load > lowload)
 1027                         continue;
 1028                 lowpri = pri;
 1029                 lowcpu = cpu;
 1030                 lowload = load;
 1031         }
 1032 
 1033         return (lowcpu);
 1034 }
 1035 
 1036 /*
 1037  * Find the thread queue with the least load.
 1038  */
 1039 static int
 1040 tdq_lowestload(struct thread *td)
 1041 {
 1042         struct tdq *tdq;
 1043         int lowload;
 1044         int lowpri;
 1045         int lowcpu;
 1046         int load;
 1047         int cpu;
 1048         int pri;
 1049 
 1050         lowcpu = 0;
 1051         lowload = TDQ_CPU(0)->tdq_load;
 1052         lowpri = TDQ_CPU(0)->tdq_lowpri;
 1053         for (cpu = 1; cpu <= mp_maxid; cpu++) {
 1054                 if (CPU_ABSENT(cpu))
 1055                         continue;
 1056                 if (!THREAD_CAN_SCHED(td, cpu))
 1057                         continue;
 1058                 tdq = TDQ_CPU(cpu);
 1059                 load = tdq->tdq_load;
 1060                 pri = tdq->tdq_lowpri;
 1061                 CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d",
 1062                     cpu, load, lowcpu, lowload);
 1063                 if (load > lowload)
 1064                         continue;
 1065                 if (load == lowload && pri < lowpri)
 1066                         continue;
 1067                 lowcpu = cpu;
 1068                 lowload = load;
 1069                 lowpri = pri;
 1070         }
 1071 
 1072         return (lowcpu);
 1073 }
 1074 
 1075 /*
 1076  * Pick the destination cpu for sched_add().  Respects affinity and makes
 1077  * a determination based on load or priority of available processors.
 1078  */
 1079 static int
 1080 sched_pickcpu(struct thread *td, int flags)
 1081 {
 1082         struct tdq *tdq;
 1083         struct td_sched *ts;
 1084         cpumask_t mask;
 1085         int self;
 1086         int pri;
 1087         int cpu;
 1088 
 1089         self = PCPU_GET(cpuid);
 1090         ts = td->td_sched;
 1091         if (smp_started == 0)
 1092                 return (self);
 1093         /*
 1094          * Don't migrate a running thread from sched_switch().
 1095          */
 1096         if (flags & SRQ_OURSELF) {
 1097                 CTR1(KTR_ULE, "YIELDING %d",
 1098                     curthread->td_priority);
 1099                 return (self);
 1100         }
 1101         pri = ts->ts_thread->td_priority;
 1102         cpu = ts->ts_cpu;
 1103         if (THREAD_CAN_SCHED(td, cpu)) {
 1104                 /*
 1105                  * Regardless of affinity, if the last cpu is idle
 1106                  * send it there.
 1107                  */
 1108                 tdq = TDQ_CPU(cpu);
 1109                 if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
 1110                         CTR5(KTR_ULE,
 1111                     "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d",
 1112                             ts->ts_cpu, ts->ts_rltick, ticks, pri,
 1113                             tdq->tdq_lowpri);
 1114                         return (ts->ts_cpu);
 1115                 }
 1116                 /*
 1117                  * If we have affinity, try to place it on the cpu we
 1118                  * last ran on.
 1119                  */
 1120                 if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) {
 1121                         CTR5(KTR_ULE,
 1122                     "affinity for %d, ltick %d ticks %d pri %d curthread %d",
 1123                             ts->ts_cpu, ts->ts_rltick, ticks, pri,
 1124                             tdq->tdq_lowpri);
 1125                         return (ts->ts_cpu);
 1126                 }
 1127         }
 1128 
 1129         /*
 1130          * Look for an idle group.
 1131          */
 1132         CTR1(KTR_ULE, "tdq_idle %X", tdq_idle);
 1133         mask = tdq_idle;
 1134         while ((cpu = ffs(mask)) != 0) {
 1135                 --cpu;
 1136                 if (THREAD_CAN_SCHED(td, cpu))
 1137                         return (cpu);
 1138                 mask &= ~(1 << cpu);
 1139         }
 1140         /*
 1141          * If there are no idle cores see if we can run the thread locally.
 1142          * This may improve locality among sleepers and wakers when there
 1143          * is shared data.
 1144          */
 1145         if (tryself && THREAD_CAN_SCHED(td, self) &&
 1146             pri < curthread->td_priority) {
 1147                 CTR1(KTR_ULE, "tryself %d",
 1148                     curthread->td_priority);
 1149                 return (self);
 1150         }
 1151         /*
 1152          * Now search for the cpu running the lowest priority thread with
 1153          * the least load.
 1154          */
 1155         if (pick_pri)
 1156                 cpu = tdq_lowestpri(td);
 1157         else
 1158                 cpu = tdq_lowestload(td);
 1159         return (cpu);
 1160 }
 1161 
 1162 #endif  /* SMP */
 1163 
 1164 /*
 1165  * Pick the highest priority task we have and return it.
 1166  */
 1167 static struct td_sched *
 1168 tdq_choose(struct tdq *tdq)
 1169 {
 1170         struct td_sched *ts;
 1171 
 1172         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 1173         ts = runq_choose(&tdq->tdq_realtime);
 1174         if (ts != NULL)
 1175                 return (ts);
 1176         ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
 1177         if (ts != NULL) {
 1178                 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE,
 1179                     ("tdq_choose: Invalid priority on timeshare queue %d",
 1180                     ts->ts_thread->td_priority));
 1181                 return (ts);
 1182         }
 1183 
 1184         ts = runq_choose(&tdq->tdq_idle);
 1185         if (ts != NULL) {
 1186                 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE,
 1187                     ("tdq_choose: Invalid priority on idle queue %d",
 1188                     ts->ts_thread->td_priority));
 1189                 return (ts);
 1190         }
 1191 
 1192         return (NULL);
 1193 }
 1194 
 1195 /*
 1196  * Initialize a thread queue.
 1197  */
 1198 static void
 1199 tdq_setup(struct tdq *tdq)
 1200 {
 1201 
 1202         if (bootverbose)
 1203                 printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
 1204         runq_init(&tdq->tdq_realtime);
 1205         runq_init(&tdq->tdq_timeshare);
 1206         runq_init(&tdq->tdq_idle);
 1207         tdq->tdq_load = 0;
 1208 }
 1209 
 1210 #ifdef SMP
 1211 static void
 1212 tdg_setup(struct tdq_group *tdg)
 1213 {
 1214         if (bootverbose)
 1215                 printf("ULE: setup cpu group %d\n", TDG_ID(tdg));
 1216         snprintf(tdg->tdg_name, sizeof(tdg->tdg_name),
 1217             "sched lock %d", (int)TDG_ID(tdg));
 1218         mtx_init(&tdg->tdg_lock, tdg->tdg_name, "sched lock",
 1219             MTX_SPIN | MTX_RECURSE);
 1220         LIST_INIT(&tdg->tdg_members);
 1221         tdg->tdg_load = 0;
 1222         tdg->tdg_transferable = 0;
 1223         tdg->tdg_cpus = 0;
 1224         tdg->tdg_mask = 0;
 1225         tdg->tdg_cpumask = 0;
 1226         tdg->tdg_idlemask = 0;
 1227 }
 1228 
 1229 static void
 1230 tdg_add(struct tdq_group *tdg, struct tdq *tdq)
 1231 {
 1232         if (tdg->tdg_mask == 0)
 1233                 tdg->tdg_mask |= 1 << TDQ_ID(tdq);
 1234         tdg->tdg_cpumask |= 1 << TDQ_ID(tdq);
 1235         tdg->tdg_cpus++;
 1236         tdq->tdq_group = tdg;
 1237         tdq->tdq_lock = &tdg->tdg_lock;
 1238         LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings);
 1239         if (bootverbose)
 1240                 printf("ULE: adding cpu %d to group %d: cpus %d mask 0x%X\n",
 1241                     TDQ_ID(tdq), TDG_ID(tdg), tdg->tdg_cpus, tdg->tdg_cpumask);
 1242 }
 1243 
 1244 static void
 1245 sched_setup_topology(void)
 1246 {
 1247         struct tdq_group *tdg;
 1248         struct cpu_group *cg;
 1249         int balance_groups;
 1250         struct tdq *tdq;
 1251         int i;
 1252         int j;
 1253 
 1254         topology = 1;
 1255         balance_groups = 0;
 1256         for (i = 0; i < smp_topology->ct_count; i++) {
 1257                 cg = &smp_topology->ct_group[i];
 1258                 tdg = &tdq_groups[i];
 1259                 /*
 1260                  * Initialize the group.
 1261                  */
 1262                 tdg_setup(tdg);
 1263                 /*
 1264                  * Find all of the group members and add them.
 1265                  */
 1266                 for (j = 0; j < MAXCPU; j++) { 
 1267                         if ((cg->cg_mask & (1 << j)) != 0) {
 1268                                 tdq = TDQ_CPU(j);
 1269                                 tdq_setup(tdq);
 1270                                 tdg_add(tdg, tdq);
 1271                         }
 1272                 }
 1273                 if (tdg->tdg_cpus > 1)
 1274                         balance_groups = 1;
 1275         }
 1276         tdg_maxid = smp_topology->ct_count - 1;
 1277         if (balance_groups)
 1278                 sched_balance_groups();
 1279 }
 1280 
 1281 static void
 1282 sched_setup_smp(void)
 1283 {
 1284         struct tdq_group *tdg;
 1285         struct tdq *tdq;
 1286         int cpus;
 1287         int i;
 1288 
 1289         for (cpus = 0, i = 0; i < MAXCPU; i++) {
 1290                 if (CPU_ABSENT(i))
 1291                         continue;
 1292                 tdq = &tdq_cpu[i];
 1293                 tdg = &tdq_groups[i];
 1294                 /*
 1295                  * Setup a tdq group with one member.
 1296                  */
 1297                 tdg_setup(tdg);
 1298                 tdq_setup(tdq);
 1299                 tdg_add(tdg, tdq);
 1300                 cpus++;
 1301         }
 1302         tdg_maxid = cpus - 1;
 1303 }
 1304 
 1305 /*
 1306  * Fake a topology with one group containing all CPUs.
 1307  */
 1308 static void
 1309 sched_fake_topo(void)
 1310 {
 1311 #ifdef SCHED_FAKE_TOPOLOGY
 1312         static struct cpu_top top;
 1313         static struct cpu_group group;
 1314 
 1315         top.ct_count = 1;
 1316         top.ct_group = &group;
 1317         group.cg_mask = all_cpus;
 1318         group.cg_count = mp_ncpus;
 1319         group.cg_children = 0;
 1320         smp_topology = &top;
 1321 #endif
 1322 }
 1323 #endif
 1324 
 1325 /*
 1326  * Setup the thread queues and initialize the topology based on MD
 1327  * information.
 1328  */
 1329 static void
 1330 sched_setup(void *dummy)
 1331 {
 1332         struct tdq *tdq;
 1333 
 1334         tdq = TDQ_SELF();
 1335 #ifdef SMP
 1336         sched_fake_topo();
 1337         /*
 1338          * Setup tdqs based on a topology configuration or vanilla SMP based
 1339          * on mp_maxid.
 1340          */
 1341         if (smp_topology == NULL)
 1342                 sched_setup_smp();
 1343         else 
 1344                 sched_setup_topology();
 1345         balance_tdq = tdq;
 1346         sched_balance();
 1347 #else
 1348         tdq_setup(tdq);
 1349         mtx_init(&tdq_lock, "sched lock", "sched lock", MTX_SPIN | MTX_RECURSE);
 1350         tdq->tdq_lock = &tdq_lock;
 1351 #endif
 1352         /*
 1353          * To avoid divide-by-zero, we set realstathz a dummy value
 1354          * in case which sched_clock() called before sched_initticks().
 1355          */
 1356         realstathz = hz;
 1357         sched_slice = (realstathz/10);  /* ~100ms */
 1358         tickincr = 1 << SCHED_TICK_SHIFT;
 1359 
 1360         /* Add thread0's load since it's running. */
 1361         TDQ_LOCK(tdq);
 1362         thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
 1363         tdq_load_add(tdq, &td_sched0);
 1364         TDQ_UNLOCK(tdq);
 1365 }
 1366 
 1367 /*
 1368  * This routine determines the tickincr 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 
 1379         /*
 1380          * tickincr is shifted out by 10 to avoid rounding errors due to
 1381          * hz not being evenly divisible by stathz on all platforms.
 1382          */
 1383         incr = (hz << SCHED_TICK_SHIFT) / realstathz;
 1384         /*
 1385          * This does not work for values of stathz that are more than
 1386          * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
 1387          */
 1388         if (incr == 0)
 1389                 incr = 1;
 1390         tickincr = incr;
 1391 #ifdef SMP
 1392         /*
 1393          * Set the default balance interval now that we know
 1394          * what realstathz is.
 1395          */
 1396         balance_interval = realstathz;
 1397         /*
 1398          * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4. 
 1399          * This prevents excess thrashing on large machines and excess idle 
 1400          * on smaller machines.
 1401          */
 1402         steal_thresh = min(fls(mp_ncpus) - 1, 3);
 1403         affinity = SCHED_AFFINITY_DEFAULT;
 1404 #endif
 1405 }
 1406 
 1407 
 1408 /*
 1409  * This is the core of the interactivity algorithm.  Determines a score based
 1410  * on past behavior.  It is the ratio of sleep time to run time scaled to
 1411  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
 1412  * differs from the cpu usage because it does not account for time spent
 1413  * waiting on a run-queue.  Would be prettier if we had floating point.
 1414  */
 1415 static int
 1416 sched_interact_score(struct thread *td)
 1417 {
 1418         struct td_sched *ts;
 1419         int div;
 1420 
 1421         ts = td->td_sched;
 1422         /*
 1423          * The score is only needed if this is likely to be an interactive
 1424          * task.  Don't go through the expense of computing it if there's
 1425          * no chance.
 1426          */
 1427         if (sched_interact <= SCHED_INTERACT_HALF &&
 1428                 ts->ts_runtime >= ts->ts_slptime)
 1429                         return (SCHED_INTERACT_HALF);
 1430 
 1431         if (ts->ts_runtime > ts->ts_slptime) {
 1432                 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
 1433                 return (SCHED_INTERACT_HALF +
 1434                     (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
 1435         }
 1436         if (ts->ts_slptime > ts->ts_runtime) {
 1437                 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
 1438                 return (ts->ts_runtime / div);
 1439         }
 1440         /* runtime == slptime */
 1441         if (ts->ts_runtime)
 1442                 return (SCHED_INTERACT_HALF);
 1443 
 1444         /*
 1445          * This can happen if slptime and runtime are 0.
 1446          */
 1447         return (0);
 1448 
 1449 }
 1450 
 1451 /*
 1452  * Scale the scheduling priority according to the "interactivity" of this
 1453  * process.
 1454  */
 1455 static void
 1456 sched_priority(struct thread *td)
 1457 {
 1458         int score;
 1459         int pri;
 1460 
 1461         if (td->td_pri_class != PRI_TIMESHARE)
 1462                 return;
 1463         /*
 1464          * If the score is interactive we place the thread in the realtime
 1465          * queue with a priority that is less than kernel and interrupt
 1466          * priorities.  These threads are not subject to nice restrictions.
 1467          *
 1468          * Scores greater than this are placed on the normal timeshare queue
 1469          * where the priority is partially decided by the most recent cpu
 1470          * utilization and the rest is decided by nice value.
 1471          *
 1472          * The nice value of the process has a linear effect on the calculated
 1473          * score.  Negative nice values make it easier for a thread to be
 1474          * considered interactive.
 1475          */
 1476         score = imax(0, sched_interact_score(td) - td->td_proc->p_nice);
 1477         if (score < sched_interact) {
 1478                 pri = PRI_MIN_REALTIME;
 1479                 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
 1480                     * score;
 1481                 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
 1482                     ("sched_priority: invalid interactive priority %d score %d",
 1483                     pri, score));
 1484         } else {
 1485                 pri = SCHED_PRI_MIN;
 1486                 if (td->td_sched->ts_ticks)
 1487                         pri += SCHED_PRI_TICKS(td->td_sched);
 1488                 pri += SCHED_PRI_NICE(td->td_proc->p_nice);
 1489                 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
 1490                     ("sched_priority: invalid priority %d: nice %d, " 
 1491                     "ticks %d ftick %d ltick %d tick pri %d",
 1492                     pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
 1493                     td->td_sched->ts_ftick, td->td_sched->ts_ltick,
 1494                     SCHED_PRI_TICKS(td->td_sched)));
 1495         }
 1496         sched_user_prio(td, pri);
 1497 
 1498         return;
 1499 }
 1500 
 1501 /*
 1502  * This routine enforces a maximum limit on the amount of scheduling history
 1503  * kept.  It is called after either the slptime or runtime is adjusted.  This
 1504  * function is ugly due to integer math.
 1505  */
 1506 static void
 1507 sched_interact_update(struct thread *td)
 1508 {
 1509         struct td_sched *ts;
 1510         u_int sum;
 1511 
 1512         ts = td->td_sched;
 1513         sum = ts->ts_runtime + ts->ts_slptime;
 1514         if (sum < SCHED_SLP_RUN_MAX)
 1515                 return;
 1516         /*
 1517          * This only happens from two places:
 1518          * 1) We have added an unusual amount of run time from fork_exit.
 1519          * 2) We have added an unusual amount of sleep time from sched_sleep().
 1520          */
 1521         if (sum > SCHED_SLP_RUN_MAX * 2) {
 1522                 if (ts->ts_runtime > ts->ts_slptime) {
 1523                         ts->ts_runtime = SCHED_SLP_RUN_MAX;
 1524                         ts->ts_slptime = 1;
 1525                 } else {
 1526                         ts->ts_slptime = SCHED_SLP_RUN_MAX;
 1527                         ts->ts_runtime = 1;
 1528                 }
 1529                 return;
 1530         }
 1531         /*
 1532          * If we have exceeded by more than 1/5th then the algorithm below
 1533          * will not bring us back into range.  Dividing by two here forces
 1534          * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
 1535          */
 1536         if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
 1537                 ts->ts_runtime /= 2;
 1538                 ts->ts_slptime /= 2;
 1539                 return;
 1540         }
 1541         ts->ts_runtime = (ts->ts_runtime / 5) * 4;
 1542         ts->ts_slptime = (ts->ts_slptime / 5) * 4;
 1543 }
 1544 
 1545 /*
 1546  * Scale back the interactivity history when a child thread is created.  The
 1547  * history is inherited from the parent but the thread may behave totally
 1548  * differently.  For example, a shell spawning a compiler process.  We want
 1549  * to learn that the compiler is behaving badly very quickly.
 1550  */
 1551 static void
 1552 sched_interact_fork(struct thread *td)
 1553 {
 1554         int ratio;
 1555         int sum;
 1556 
 1557         sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
 1558         if (sum > SCHED_SLP_RUN_FORK) {
 1559                 ratio = sum / SCHED_SLP_RUN_FORK;
 1560                 td->td_sched->ts_runtime /= ratio;
 1561                 td->td_sched->ts_slptime /= ratio;
 1562         }
 1563 }
 1564 
 1565 /*
 1566  * Called from proc0_init() to setup the scheduler fields.
 1567  */
 1568 void
 1569 schedinit(void)
 1570 {
 1571 
 1572         /*
 1573          * Set up the scheduler specific parts of proc0.
 1574          */
 1575         proc0.p_sched = NULL; /* XXX */
 1576         thread0.td_sched = &td_sched0;
 1577         td_sched0.ts_ltick = ticks;
 1578         td_sched0.ts_ftick = ticks;
 1579         td_sched0.ts_thread = &thread0;
 1580 }
 1581 
 1582 /*
 1583  * This is only somewhat accurate since given many processes of the same
 1584  * priority they will switch when their slices run out, which will be
 1585  * at most sched_slice stathz ticks.
 1586  */
 1587 int
 1588 sched_rr_interval(void)
 1589 {
 1590 
 1591         /* Convert sched_slice to hz */
 1592         return (hz/(realstathz/sched_slice));
 1593 }
 1594 
 1595 /*
 1596  * Update the percent cpu tracking information when it is requested or
 1597  * the total history exceeds the maximum.  We keep a sliding history of
 1598  * tick counts that slowly decays.  This is less precise than the 4BSD
 1599  * mechanism since it happens with less regular and frequent events.
 1600  */
 1601 static void
 1602 sched_pctcpu_update(struct td_sched *ts)
 1603 {
 1604 
 1605         if (ts->ts_ticks == 0)
 1606                 return;
 1607         if (ticks - (hz / 10) < ts->ts_ltick &&
 1608             SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
 1609                 return;
 1610         /*
 1611          * Adjust counters and watermark for pctcpu calc.
 1612          */
 1613         if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
 1614                 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
 1615                             SCHED_TICK_TARG;
 1616         else
 1617                 ts->ts_ticks = 0;
 1618         ts->ts_ltick = ticks;
 1619         ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
 1620 }
 1621 
 1622 /*
 1623  * Adjust the priority of a thread.  Move it to the appropriate run-queue
 1624  * if necessary.  This is the back-end for several priority related
 1625  * functions.
 1626  */
 1627 static void
 1628 sched_thread_priority(struct thread *td, u_char prio)
 1629 {
 1630         struct td_sched *ts;
 1631 
 1632         CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
 1633             td, td->td_proc->p_comm, td->td_priority, prio, curthread,
 1634             curthread->td_proc->p_comm);
 1635         ts = td->td_sched;
 1636         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1637         if (td->td_priority == prio)
 1638                 return;
 1639 
 1640         if (TD_ON_RUNQ(td) && prio < td->td_priority) {
 1641                 /*
 1642                  * If the priority has been elevated due to priority
 1643                  * propagation, we may have to move ourselves to a new
 1644                  * queue.  This could be optimized to not re-add in some
 1645                  * cases.
 1646                  */
 1647                 sched_rem(td);
 1648                 td->td_priority = prio;
 1649                 sched_add(td, SRQ_BORROWING);
 1650         } else {
 1651 #ifdef SMP
 1652                 struct tdq *tdq;
 1653 
 1654                 tdq = TDQ_CPU(ts->ts_cpu);
 1655                 if (prio < tdq->tdq_lowpri)
 1656                         tdq->tdq_lowpri = prio;
 1657 #endif
 1658                 td->td_priority = prio;
 1659         }
 1660 }
 1661 
 1662 /*
 1663  * Update a thread's priority when it is lent another thread's
 1664  * priority.
 1665  */
 1666 void
 1667 sched_lend_prio(struct thread *td, u_char prio)
 1668 {
 1669 
 1670         td->td_flags |= TDF_BORROWING;
 1671         sched_thread_priority(td, prio);
 1672 }
 1673 
 1674 /*
 1675  * Restore a thread's priority when priority propagation is
 1676  * over.  The prio argument is the minimum priority the thread
 1677  * needs to have to satisfy other possible priority lending
 1678  * requests.  If the thread's regular priority is less
 1679  * important than prio, the thread will keep a priority boost
 1680  * of prio.
 1681  */
 1682 void
 1683 sched_unlend_prio(struct thread *td, u_char prio)
 1684 {
 1685         u_char base_pri;
 1686 
 1687         if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
 1688             td->td_base_pri <= PRI_MAX_TIMESHARE)
 1689                 base_pri = td->td_user_pri;
 1690         else
 1691                 base_pri = td->td_base_pri;
 1692         if (prio >= base_pri) {
 1693                 td->td_flags &= ~TDF_BORROWING;
 1694                 sched_thread_priority(td, base_pri);
 1695         } else
 1696                 sched_lend_prio(td, prio);
 1697 }
 1698 
 1699 /*
 1700  * Standard entry for setting the priority to an absolute value.
 1701  */
 1702 void
 1703 sched_prio(struct thread *td, u_char prio)
 1704 {
 1705         u_char oldprio;
 1706 
 1707         /* First, update the base priority. */
 1708         td->td_base_pri = prio;
 1709 
 1710         /*
 1711          * If the thread is borrowing another thread's priority, don't
 1712          * ever lower the priority.
 1713          */
 1714         if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
 1715                 return;
 1716 
 1717         /* Change the real priority. */
 1718         oldprio = td->td_priority;
 1719         sched_thread_priority(td, prio);
 1720 
 1721         /*
 1722          * If the thread is on a turnstile, then let the turnstile update
 1723          * its state.
 1724          */
 1725         if (TD_ON_LOCK(td) && oldprio != prio)
 1726                 turnstile_adjust(td, oldprio);
 1727 }
 1728 
 1729 /*
 1730  * Set the base user priority, does not effect current running priority.
 1731  */
 1732 void
 1733 sched_user_prio(struct thread *td, u_char prio)
 1734 {
 1735         u_char oldprio;
 1736 
 1737         td->td_base_user_pri = prio;
 1738         if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
 1739                 return;
 1740         oldprio = td->td_user_pri;
 1741         td->td_user_pri = prio;
 1742 }
 1743 
 1744 void
 1745 sched_lend_user_prio(struct thread *td, u_char prio)
 1746 {
 1747         u_char oldprio;
 1748 
 1749         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1750         td->td_flags |= TDF_UBORROWING;
 1751         oldprio = td->td_user_pri;
 1752         td->td_user_pri = prio;
 1753 }
 1754 
 1755 void
 1756 sched_unlend_user_prio(struct thread *td, u_char prio)
 1757 {
 1758         u_char base_pri;
 1759 
 1760         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1761         base_pri = td->td_base_user_pri;
 1762         if (prio >= base_pri) {
 1763                 td->td_flags &= ~TDF_UBORROWING;
 1764                 sched_user_prio(td, base_pri);
 1765         } else {
 1766                 sched_lend_user_prio(td, prio);
 1767         }
 1768 }
 1769 
 1770 /*
 1771  * Add the thread passed as 'newtd' to the run queue before selecting
 1772  * the next thread to run.  This is only used for KSE.
 1773  */
 1774 static void
 1775 sched_switchin(struct tdq *tdq, struct thread *td)
 1776 {
 1777 #ifdef SMP
 1778         spinlock_enter();
 1779         TDQ_UNLOCK(tdq);
 1780         thread_lock(td);
 1781         spinlock_exit();
 1782         sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING);
 1783 #else
 1784         td->td_lock = TDQ_LOCKPTR(tdq);
 1785 #endif
 1786         tdq_add(tdq, td, SRQ_YIELDING);
 1787         MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 1788 }
 1789 
 1790 /*
 1791  * Block a thread for switching.  Similar to thread_block() but does not
 1792  * bump the spin count.
 1793  */
 1794 static inline struct mtx *
 1795 thread_block_switch(struct thread *td)
 1796 {
 1797         struct mtx *lock;
 1798 
 1799         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1800         lock = td->td_lock;
 1801         td->td_lock = &blocked_lock;
 1802         mtx_unlock_spin(lock);
 1803 
 1804         return (lock);
 1805 }
 1806 
 1807 /*
 1808  * Handle migration from sched_switch().  This happens only for
 1809  * cpu binding.
 1810  */
 1811 static struct mtx *
 1812 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
 1813 {
 1814         struct tdq *tdn;
 1815 
 1816         tdn = TDQ_CPU(td->td_sched->ts_cpu);
 1817 #ifdef SMP
 1818         /*
 1819          * Do the lock dance required to avoid LOR.  We grab an extra
 1820          * spinlock nesting to prevent preemption while we're
 1821          * not holding either run-queue lock.
 1822          */
 1823         spinlock_enter();
 1824         thread_block_switch(td);        /* This releases the lock on tdq. */
 1825 
 1826         /*
 1827          * Acquire both run-queue locks before placing the thread on the new
 1828          * run-queue to avoid deadlocks created by placing a thread with a
 1829          * blocked lock on the run-queue of a remote processor.  The deadlock
 1830          * occurs when a third processor attempts to lock the two queues in
 1831          * question while the target processor is spinning with its own
 1832          * run-queue lock held while waiting for the blocked lock to clear.
 1833          */
 1834         if (TDQ_LOCKPTR(tdn) == TDQ_LOCKPTR(tdq)) {
 1835                 TDQ_LOCK(tdq);
 1836                 tdq_add(tdn, td, flags);
 1837                 tdq_notify(td->td_sched);
 1838         } else {
 1839                 tdq_lock_pair(tdn, tdq);
 1840                 tdq_add(tdn, td, flags);
 1841                 tdq_notify(td->td_sched);
 1842                 TDQ_UNLOCK(tdn);
 1843         }
 1844         spinlock_exit();
 1845 #endif
 1846         return (TDQ_LOCKPTR(tdn));
 1847 }
 1848 
 1849 /*
 1850  * Release a thread that was blocked with thread_block_switch().
 1851  */
 1852 static inline void
 1853 thread_unblock_switch(struct thread *td, struct mtx *mtx)
 1854 {
 1855         atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
 1856             (uintptr_t)mtx);
 1857 }
 1858 
 1859 /*
 1860  * Switch threads.  This function has to handle threads coming in while
 1861  * blocked for some reason, running, or idle.  It also must deal with
 1862  * migrating a thread from one queue to another as running threads may
 1863  * be assigned elsewhere via binding.
 1864  */
 1865 void
 1866 sched_switch(struct thread *td, struct thread *newtd, int flags)
 1867 {
 1868         struct tdq *tdq;
 1869         struct td_sched *ts;
 1870         struct mtx *mtx;
 1871         int srqflag;
 1872         int cpuid;
 1873 
 1874         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1875 
 1876         cpuid = PCPU_GET(cpuid);
 1877         tdq = TDQ_CPU(cpuid);
 1878         ts = td->td_sched;
 1879         mtx = td->td_lock;
 1880 #ifdef SMP
 1881         ts->ts_rltick = ticks;
 1882         if (newtd && newtd->td_priority < tdq->tdq_lowpri)
 1883                 tdq->tdq_lowpri = newtd->td_priority;
 1884 #endif
 1885         td->td_lastcpu = td->td_oncpu;
 1886         td->td_oncpu = NOCPU;
 1887         td->td_flags &= ~TDF_NEEDRESCHED;
 1888         td->td_owepreempt = 0;
 1889         /*
 1890          * The lock pointer in an idle thread should never change.  Reset it
 1891          * to CAN_RUN as well.
 1892          */
 1893         if (TD_IS_IDLETHREAD(td)) {
 1894                 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 1895                 TD_SET_CAN_RUN(td);
 1896         } else if (TD_IS_RUNNING(td)) {
 1897                 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 1898                 tdq_load_rem(tdq, ts);
 1899                 srqflag = (flags & SW_PREEMPT) ?
 1900                     SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
 1901                     SRQ_OURSELF|SRQ_YIELDING;
 1902                 if (ts->ts_cpu == cpuid)
 1903                         tdq_add(tdq, td, srqflag);
 1904                 else
 1905                         mtx = sched_switch_migrate(tdq, td, srqflag);
 1906         } else {
 1907                 /* This thread must be going to sleep. */
 1908                 TDQ_LOCK(tdq);
 1909                 mtx = thread_block_switch(td);
 1910                 tdq_load_rem(tdq, ts);
 1911         }
 1912         /*
 1913          * We enter here with the thread blocked and assigned to the
 1914          * appropriate cpu run-queue or sleep-queue and with the current
 1915          * thread-queue locked.
 1916          */
 1917         TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
 1918         /*
 1919          * If KSE assigned a new thread just add it here and let choosethread
 1920          * select the best one.
 1921          */
 1922         if (newtd != NULL)
 1923                 sched_switchin(tdq, newtd);
 1924         newtd = choosethread();
 1925         /*
 1926          * Call the MD code to switch contexts if necessary.
 1927          */
 1928         if (td != newtd) {
 1929 #ifdef  HWPMC_HOOKS
 1930                 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 1931                         PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
 1932 #endif
 1933                 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
 1934 
 1935 #ifdef KDTRACE_HOOKS
 1936                 /*
 1937                  * If DTrace has set the active vtime enum to anything
 1938                  * other than INACTIVE (0), then it should have set the
 1939                  * function to call.
 1940                  */
 1941                 if (dtrace_vtime_active)
 1942                         (*dtrace_vtime_switch_func)(newtd);
 1943 #endif
 1944                 cpu_switch(td, newtd, mtx);
 1945                 /*
 1946                  * We may return from cpu_switch on a different cpu.  However,
 1947                  * we always return with td_lock pointing to the current cpu's
 1948                  * run queue lock.
 1949                  */
 1950                 cpuid = PCPU_GET(cpuid);
 1951                 tdq = TDQ_CPU(cpuid);
 1952 #ifdef  HWPMC_HOOKS
 1953                 if (PMC_PROC_IS_USING_PMCS(td->td_proc))
 1954                         PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
 1955 #endif
 1956         } else
 1957                 thread_unblock_switch(td, mtx);
 1958         /*
 1959          * Assert that all went well and return.
 1960          */
 1961 #ifdef SMP
 1962         /* We should always get here with the lowest priority td possible */
 1963         tdq->tdq_lowpri = td->td_priority;
 1964 #endif
 1965         TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
 1966         MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 1967         td->td_oncpu = cpuid;
 1968 }
 1969 
 1970 /*
 1971  * Adjust thread priorities as a result of a nice request.
 1972  */
 1973 void
 1974 sched_nice(struct proc *p, int nice)
 1975 {
 1976         struct thread *td;
 1977 
 1978         PROC_LOCK_ASSERT(p, MA_OWNED);
 1979         PROC_SLOCK_ASSERT(p, MA_OWNED);
 1980 
 1981         p->p_nice = nice;
 1982         FOREACH_THREAD_IN_PROC(p, td) {
 1983                 thread_lock(td);
 1984                 sched_priority(td);
 1985                 sched_prio(td, td->td_base_user_pri);
 1986                 thread_unlock(td);
 1987         }
 1988 }
 1989 
 1990 /*
 1991  * Record the sleep time for the interactivity scorer.
 1992  */
 1993 void
 1994 sched_sleep(struct thread *td)
 1995 {
 1996 
 1997         THREAD_LOCK_ASSERT(td, MA_OWNED);
 1998 
 1999         td->td_slptick = ticks;
 2000 }
 2001 
 2002 /*
 2003  * Schedule a thread to resume execution and record how long it voluntarily
 2004  * slept.  We also update the pctcpu, interactivity, and priority.
 2005  */
 2006 void
 2007 sched_wakeup(struct thread *td)
 2008 {
 2009         struct td_sched *ts;
 2010         int slptick;
 2011 
 2012         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2013         ts = td->td_sched;
 2014         /*
 2015          * If we slept for more than a tick update our interactivity and
 2016          * priority.
 2017          */
 2018         slptick = td->td_slptick;
 2019         td->td_slptick = 0;
 2020         if (slptick && slptick != ticks) {
 2021                 u_int hzticks;
 2022 
 2023                 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
 2024                 ts->ts_slptime += hzticks;
 2025                 sched_interact_update(td);
 2026                 sched_pctcpu_update(ts);
 2027                 sched_priority(td);
 2028         }
 2029         /* Reset the slice value after we sleep. */
 2030         ts->ts_slice = sched_slice;
 2031         sched_add(td, SRQ_BORING);
 2032 }
 2033 
 2034 /*
 2035  * Penalize the parent for creating a new child and initialize the child's
 2036  * priority.
 2037  */
 2038 void
 2039 sched_fork(struct thread *td, struct thread *child)
 2040 {
 2041         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2042         sched_fork_thread(td, child);
 2043         /*
 2044          * Penalize the parent and child for forking.
 2045          */
 2046         sched_interact_fork(child);
 2047         sched_priority(child);
 2048         td->td_sched->ts_runtime += tickincr;
 2049         sched_interact_update(td);
 2050         sched_priority(td);
 2051 }
 2052 
 2053 /*
 2054  * Fork a new thread, may be within the same process.
 2055  */
 2056 void
 2057 sched_fork_thread(struct thread *td, struct thread *child)
 2058 {
 2059         struct td_sched *ts;
 2060         struct td_sched *ts2;
 2061 
 2062         /*
 2063          * Initialize child.
 2064          */
 2065         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2066         sched_newthread(child);
 2067         child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
 2068         child->td_cpuset = cpuset_ref(td->td_cpuset);
 2069         ts = td->td_sched;
 2070         ts2 = child->td_sched;
 2071         ts2->ts_cpu = ts->ts_cpu;
 2072         ts2->ts_runq = NULL;
 2073         /*
 2074          * Grab our parents cpu estimation information and priority.
 2075          */
 2076         ts2->ts_ticks = ts->ts_ticks;
 2077         ts2->ts_ltick = ts->ts_ltick;
 2078         ts2->ts_ftick = ts->ts_ftick;
 2079         child->td_user_pri = td->td_user_pri;
 2080         child->td_base_user_pri = td->td_base_user_pri;
 2081         /*
 2082          * And update interactivity score.
 2083          */
 2084         ts2->ts_slptime = ts->ts_slptime;
 2085         ts2->ts_runtime = ts->ts_runtime;
 2086         ts2->ts_slice = 1;      /* Attempt to quickly learn interactivity. */
 2087 }
 2088 
 2089 /*
 2090  * Adjust the priority class of a thread.
 2091  */
 2092 void
 2093 sched_class(struct thread *td, int class)
 2094 {
 2095 
 2096         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2097         if (td->td_pri_class == class)
 2098                 return;
 2099 
 2100 #ifdef SMP
 2101         /*
 2102          * On SMP if we're on the RUNQ we must adjust the transferable
 2103          * count because could be changing to or from an interrupt
 2104          * class.
 2105          */
 2106         if (TD_ON_RUNQ(td)) {
 2107                 struct tdq *tdq;
 2108 
 2109                 tdq = TDQ_CPU(td->td_sched->ts_cpu);
 2110                 if (THREAD_CAN_MIGRATE(td)) {
 2111                         tdq->tdq_transferable--;
 2112                         tdq->tdq_group->tdg_transferable--;
 2113                 }
 2114                 td->td_pri_class = class;
 2115                 if (THREAD_CAN_MIGRATE(td)) {
 2116                         tdq->tdq_transferable++;
 2117                         tdq->tdq_group->tdg_transferable++;
 2118                 }
 2119         }
 2120 #endif
 2121         td->td_pri_class = class;
 2122 }
 2123 
 2124 /*
 2125  * Return some of the child's priority and interactivity to the parent.
 2126  */
 2127 void
 2128 sched_exit(struct proc *p, struct thread *child)
 2129 {
 2130         struct thread *td;
 2131         
 2132         CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
 2133             child, child->td_proc->p_comm, child->td_priority);
 2134 
 2135         PROC_SLOCK_ASSERT(p, MA_OWNED);
 2136         td = FIRST_THREAD_IN_PROC(p);
 2137         sched_exit_thread(td, child);
 2138 }
 2139 
 2140 /*
 2141  * Penalize another thread for the time spent on this one.  This helps to
 2142  * worsen the priority and interactivity of processes which schedule batch
 2143  * jobs such as make.  This has little effect on the make process itself but
 2144  * causes new processes spawned by it to receive worse scores immediately.
 2145  */
 2146 void
 2147 sched_exit_thread(struct thread *td, struct thread *child)
 2148 {
 2149 
 2150         CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
 2151             child, child->td_proc->p_comm, child->td_priority);
 2152 
 2153 #ifdef KSE
 2154         /*
 2155          * KSE forks and exits so often that this penalty causes short-lived
 2156          * threads to always be non-interactive.  This causes mozilla to
 2157          * crawl under load.
 2158          */
 2159         if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc)
 2160                 return;
 2161 #endif
 2162         /*
 2163          * Give the child's runtime to the parent without returning the
 2164          * sleep time as a penalty to the parent.  This causes shells that
 2165          * launch expensive things to mark their children as expensive.
 2166          */
 2167         thread_lock(td);
 2168         td->td_sched->ts_runtime += child->td_sched->ts_runtime;
 2169         sched_interact_update(td);
 2170         sched_priority(td);
 2171         thread_unlock(td);
 2172 }
 2173 
 2174 /*
 2175  * Fix priorities on return to user-space.  Priorities may be elevated due
 2176  * to static priorities in msleep() or similar.
 2177  */
 2178 void
 2179 sched_userret(struct thread *td)
 2180 {
 2181         /*
 2182          * XXX we cheat slightly on the locking here to avoid locking in  
 2183          * the usual case.  Setting td_priority here is essentially an
 2184          * incomplete workaround for not setting it properly elsewhere.
 2185          * Now that some interrupt handlers are threads, not setting it
 2186          * properly elsewhere can clobber it in the window between setting
 2187          * it here and returning to user mode, so don't waste time setting
 2188          * it perfectly here.
 2189          */
 2190         KASSERT((td->td_flags & TDF_BORROWING) == 0,
 2191             ("thread with borrowed priority returning to userland"));
 2192         if (td->td_priority != td->td_user_pri) {
 2193                 thread_lock(td);
 2194                 td->td_priority = td->td_user_pri;
 2195                 td->td_base_pri = td->td_user_pri;
 2196                 thread_unlock(td);
 2197         }
 2198 }
 2199 
 2200 /*
 2201  * Handle a stathz tick.  This is really only relevant for timeshare
 2202  * threads.
 2203  */
 2204 void
 2205 sched_clock(struct thread *td)
 2206 {
 2207         struct tdq *tdq;
 2208         struct td_sched *ts;
 2209 
 2210         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2211         tdq = TDQ_SELF();
 2212 #ifdef SMP
 2213         /*
 2214          * We run the long term load balancer infrequently on the first cpu.
 2215          */
 2216         if (balance_tdq == tdq) {
 2217                 if (balance_ticks && --balance_ticks == 0)
 2218                         sched_balance();
 2219                 if (balance_group_ticks && --balance_group_ticks == 0)
 2220                         sched_balance_groups();
 2221         }
 2222 #endif
 2223         /*
 2224          * Advance the insert index once for each tick to ensure that all
 2225          * threads get a chance to run.
 2226          */
 2227         if (tdq->tdq_idx == tdq->tdq_ridx) {
 2228                 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
 2229                 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
 2230                         tdq->tdq_ridx = tdq->tdq_idx;
 2231         }
 2232         ts = td->td_sched;
 2233         if (td->td_pri_class & PRI_FIFO_BIT)
 2234                 return;
 2235         if (td->td_pri_class == PRI_TIMESHARE) {
 2236                 /*
 2237                  * We used a tick; charge it to the thread so
 2238                  * that we can compute our interactivity.
 2239                  */
 2240                 td->td_sched->ts_runtime += tickincr;
 2241                 sched_interact_update(td);
 2242         }
 2243         /*
 2244          * We used up one time slice.
 2245          */
 2246         if (--ts->ts_slice > 0)
 2247                 return;
 2248         /*
 2249          * We're out of time, recompute priorities and requeue.
 2250          */
 2251         sched_priority(td);
 2252         td->td_flags |= TDF_NEEDRESCHED;
 2253 }
 2254 
 2255 /*
 2256  * Called once per hz tick.  Used for cpu utilization information.  This
 2257  * is easier than trying to scale based on stathz.
 2258  */
 2259 void
 2260 sched_tick(void)
 2261 {
 2262         struct td_sched *ts;
 2263 
 2264         ts = curthread->td_sched;
 2265         /*
 2266          * Ticks is updated asynchronously on a single cpu.  Check here to
 2267          * avoid incrementing ts_ticks multiple times in a single tick.
 2268          */
 2269         if (ts->ts_ltick == ticks)
 2270                 return;
 2271         /* Adjust ticks for pctcpu */
 2272         ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
 2273         ts->ts_ltick = ticks;
 2274         /*
 2275          * Update if we've exceeded our desired tick threshhold by over one
 2276          * second.
 2277          */
 2278         if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
 2279                 sched_pctcpu_update(ts);
 2280 }
 2281 
 2282 /*
 2283  * Return whether the current CPU has runnable tasks.  Used for in-kernel
 2284  * cooperative idle threads.
 2285  */
 2286 int
 2287 sched_runnable(void)
 2288 {
 2289         struct tdq *tdq;
 2290         int load;
 2291 
 2292         load = 1;
 2293 
 2294         tdq = TDQ_SELF();
 2295         if ((curthread->td_flags & TDF_IDLETD) != 0) {
 2296                 if (tdq->tdq_load > 0)
 2297                         goto out;
 2298         } else
 2299                 if (tdq->tdq_load - 1 > 0)
 2300                         goto out;
 2301         load = 0;
 2302 out:
 2303         return (load);
 2304 }
 2305 
 2306 /*
 2307  * Choose the highest priority thread to run.  The thread is removed from
 2308  * the run-queue while running however the load remains.  For SMP we set
 2309  * the tdq in the global idle bitmask if it idles here.
 2310  */
 2311 struct thread *
 2312 sched_choose(void)
 2313 {
 2314 #ifdef SMP
 2315         struct tdq_group *tdg;
 2316 #endif
 2317         struct td_sched *ts;
 2318         struct tdq *tdq;
 2319 
 2320         tdq = TDQ_SELF();
 2321         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 2322         ts = tdq_choose(tdq);
 2323         if (ts) {
 2324                 tdq_runq_rem(tdq, ts);
 2325                 return (ts->ts_thread);
 2326         }
 2327 #ifdef SMP
 2328         /*
 2329          * We only set the idled bit when all of the cpus in the group are
 2330          * idle.  Otherwise we could get into a situation where a thread bounces
 2331          * back and forth between two idle cores on seperate physical CPUs.
 2332          */
 2333         tdg = tdq->tdq_group;
 2334         tdg->tdg_idlemask |= PCPU_GET(cpumask);
 2335         if (tdg->tdg_idlemask == tdg->tdg_cpumask)
 2336                 atomic_set_int(&tdq_idle, tdg->tdg_mask);
 2337         tdq->tdq_lowpri = PRI_MAX_IDLE;
 2338 #endif
 2339         return (PCPU_GET(idlethread));
 2340 }
 2341 
 2342 /*
 2343  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
 2344  * we always request it once we exit a critical section.
 2345  */
 2346 static inline void
 2347 sched_setpreempt(struct thread *td)
 2348 {
 2349         struct thread *ctd;
 2350         int cpri;
 2351         int pri;
 2352 
 2353         ctd = curthread;
 2354         pri = td->td_priority;
 2355         cpri = ctd->td_priority;
 2356         if (td->td_priority < ctd->td_priority)
 2357                 curthread->td_flags |= TDF_NEEDRESCHED;
 2358         if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
 2359                 return;
 2360         /*
 2361          * Always preempt IDLE threads.  Otherwise only if the preempting
 2362          * thread is an ithread.
 2363          */
 2364         if (pri > preempt_thresh && cpri < PRI_MIN_IDLE)
 2365                 return;
 2366         ctd->td_owepreempt = 1;
 2367         return;
 2368 }
 2369 
 2370 /*
 2371  * Add a thread to a thread queue.  Initializes priority, slice, runq, and
 2372  * add it to the appropriate queue.  This is the internal function called
 2373  * when the tdq is predetermined.
 2374  */
 2375 void
 2376 tdq_add(struct tdq *tdq, struct thread *td, int flags)
 2377 {
 2378         struct td_sched *ts;
 2379         int class;
 2380 #ifdef SMP
 2381         int cpumask;
 2382 #endif
 2383 
 2384         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 2385         KASSERT((td->td_inhibitors == 0),
 2386             ("sched_add: trying to run inhibited thread"));
 2387         KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
 2388             ("sched_add: bad thread state"));
 2389         KASSERT(td->td_flags & TDF_INMEM,
 2390             ("sched_add: thread swapped out"));
 2391 
 2392         ts = td->td_sched;
 2393         class = PRI_BASE(td->td_pri_class);
 2394         TD_SET_RUNQ(td);
 2395         if (ts->ts_slice == 0)
 2396                 ts->ts_slice = sched_slice;
 2397         /*
 2398          * Pick the run queue based on priority.
 2399          */
 2400         if (td->td_priority <= PRI_MAX_REALTIME)
 2401                 ts->ts_runq = &tdq->tdq_realtime;
 2402         else if (td->td_priority <= PRI_MAX_TIMESHARE)
 2403                 ts->ts_runq = &tdq->tdq_timeshare;
 2404         else
 2405                 ts->ts_runq = &tdq->tdq_idle;
 2406 #ifdef SMP
 2407         cpumask = 1 << ts->ts_cpu;
 2408         /*
 2409          * If we had been idle, clear our bit in the group and potentially
 2410          * the global bitmap.
 2411          */
 2412         if ((class != PRI_IDLE && class != PRI_ITHD) &&
 2413             (tdq->tdq_group->tdg_idlemask & cpumask) != 0) {
 2414                 /*
 2415                  * Check to see if our group is unidling, and if so, remove it
 2416                  * from the global idle mask.
 2417                  */
 2418                 if (tdq->tdq_group->tdg_idlemask ==
 2419                     tdq->tdq_group->tdg_cpumask)
 2420                         atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask);
 2421                 /*
 2422                  * Now remove ourselves from the group specific idle mask.
 2423                  */
 2424                 tdq->tdq_group->tdg_idlemask &= ~cpumask;
 2425         }
 2426         if (td->td_priority < tdq->tdq_lowpri)
 2427                 tdq->tdq_lowpri = td->td_priority;
 2428 #endif
 2429         tdq_runq_add(tdq, ts, flags);
 2430         tdq_load_add(tdq, ts);
 2431 }
 2432 
 2433 /*
 2434  * Select the target thread queue and add a thread to it.  Request
 2435  * preemption or IPI a remote processor if required.
 2436  */
 2437 void
 2438 sched_add(struct thread *td, int flags)
 2439 {
 2440         struct td_sched *ts;
 2441         struct tdq *tdq;
 2442 #ifdef SMP
 2443         int cpuid;
 2444         int cpu;
 2445 #endif
 2446         CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
 2447             td, td->td_proc->p_comm, td->td_priority, curthread,
 2448             curthread->td_proc->p_comm);
 2449         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2450         ts = td->td_sched;
 2451         /*
 2452          * Recalculate the priority before we select the target cpu or
 2453          * run-queue.
 2454          */
 2455         if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
 2456                 sched_priority(td);
 2457 #ifdef SMP
 2458         cpuid = PCPU_GET(cpuid);
 2459         /*
 2460          * Pick the destination cpu and if it isn't ours transfer to the
 2461          * target cpu.
 2462          */
 2463         if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td) &&
 2464             curthread->td_intr_nesting_level)
 2465                 ts->ts_cpu = cpuid;
 2466         if (!THREAD_CAN_MIGRATE(td))
 2467                 cpu = ts->ts_cpu;
 2468         else
 2469                 cpu = sched_pickcpu(td, flags);
 2470         tdq = sched_setcpu(ts, cpu, flags);
 2471         tdq_add(tdq, td, flags);
 2472         if (cpu != cpuid) {
 2473                 tdq_notify(ts);
 2474                 return;
 2475         }
 2476 #else
 2477         tdq = TDQ_SELF();
 2478         TDQ_LOCK(tdq);
 2479         /*
 2480          * Now that the thread is moving to the run-queue, set the lock
 2481          * to the scheduler's lock.
 2482          */
 2483         thread_lock_set(td, TDQ_LOCKPTR(tdq));
 2484         tdq_add(tdq, td, flags);
 2485 #endif
 2486         if (!(flags & SRQ_YIELDING))
 2487                 sched_setpreempt(td);
 2488 }
 2489 
 2490 /*
 2491  * Remove a thread from a run-queue without running it.  This is used
 2492  * when we're stealing a thread from a remote queue.  Otherwise all threads
 2493  * exit by calling sched_exit_thread() and sched_throw() themselves.
 2494  */
 2495 void
 2496 sched_rem(struct thread *td)
 2497 {
 2498         struct tdq *tdq;
 2499         struct td_sched *ts;
 2500 
 2501         CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
 2502             td, td->td_proc->p_comm, td->td_priority, curthread,
 2503             curthread->td_proc->p_comm);
 2504         ts = td->td_sched;
 2505         tdq = TDQ_CPU(ts->ts_cpu);
 2506         TDQ_LOCK_ASSERT(tdq, MA_OWNED);
 2507         MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 2508         KASSERT(TD_ON_RUNQ(td),
 2509             ("sched_rem: thread not on run queue"));
 2510         tdq_runq_rem(tdq, ts);
 2511         tdq_load_rem(tdq, ts);
 2512         TD_SET_CAN_RUN(td);
 2513 }
 2514 
 2515 /*
 2516  * Fetch cpu utilization information.  Updates on demand.
 2517  */
 2518 fixpt_t
 2519 sched_pctcpu(struct thread *td)
 2520 {
 2521         fixpt_t pctcpu;
 2522         struct td_sched *ts;
 2523 
 2524         pctcpu = 0;
 2525         ts = td->td_sched;
 2526         if (ts == NULL)
 2527                 return (0);
 2528 
 2529         thread_lock(td);
 2530         if (ts->ts_ticks) {
 2531                 int rtick;
 2532 
 2533                 sched_pctcpu_update(ts);
 2534                 /* How many rtick per second ? */
 2535                 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
 2536                 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
 2537         }
 2538         thread_unlock(td);
 2539 
 2540         return (pctcpu);
 2541 }
 2542 
 2543 /*
 2544  * Enforce affinity settings for a thread.  Called after adjustments to
 2545  * cpumask.
 2546  */
 2547 void
 2548 sched_affinity(struct thread *td)
 2549 {
 2550 #ifdef SMP
 2551         struct td_sched *ts;
 2552         int cpu;
 2553 
 2554         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2555         ts = td->td_sched;
 2556         if (THREAD_CAN_SCHED(td, ts->ts_cpu))
 2557                 return;
 2558         if (TD_ON_RUNQ(td)) {
 2559                 sched_rem(td);
 2560                 sched_add(td, SRQ_BORING);
 2561                 return;
 2562         }
 2563         if (!TD_IS_RUNNING(td))
 2564                 return;
 2565         td->td_flags |= TDF_NEEDRESCHED;
 2566         if (!THREAD_CAN_MIGRATE(td))
 2567                 return;
 2568         /*
 2569          * Assign the new cpu and force a switch before returning to
 2570          * userspace.  If the target thread is not running locally send
 2571          * an ipi to force the issue.
 2572          */
 2573         cpu = ts->ts_cpu;
 2574         ts->ts_cpu = sched_pickcpu(td, 0);
 2575         if (cpu != PCPU_GET(cpuid))
 2576                 ipi_selected(1 << cpu, IPI_PREEMPT);
 2577 #endif
 2578 }
 2579 
 2580 /*
 2581  * Bind a thread to a target cpu.
 2582  */
 2583 void
 2584 sched_bind(struct thread *td, int cpu)
 2585 {
 2586         struct td_sched *ts;
 2587 
 2588         THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
 2589         ts = td->td_sched;
 2590         if (ts->ts_flags & TSF_BOUND)
 2591                 sched_unbind(td);
 2592         ts->ts_flags |= TSF_BOUND;
 2593 #ifdef SMP
 2594         sched_pin();
 2595         if (PCPU_GET(cpuid) == cpu)
 2596                 return;
 2597         ts->ts_cpu = cpu;
 2598         /* When we return from mi_switch we'll be on the correct cpu. */
 2599         mi_switch(SW_VOL, NULL);
 2600 #endif
 2601 }
 2602 
 2603 /*
 2604  * Release a bound thread.
 2605  */
 2606 void
 2607 sched_unbind(struct thread *td)
 2608 {
 2609         struct td_sched *ts;
 2610 
 2611         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2612         ts = td->td_sched;
 2613         if ((ts->ts_flags & TSF_BOUND) == 0)
 2614                 return;
 2615         ts->ts_flags &= ~TSF_BOUND;
 2616 #ifdef SMP
 2617         sched_unpin();
 2618 #endif
 2619 }
 2620 
 2621 int
 2622 sched_is_bound(struct thread *td)
 2623 {
 2624         THREAD_LOCK_ASSERT(td, MA_OWNED);
 2625         return (td->td_sched->ts_flags & TSF_BOUND);
 2626 }
 2627 
 2628 /*
 2629  * Basic yield call.
 2630  */
 2631 void
 2632 sched_relinquish(struct thread *td)
 2633 {
 2634         thread_lock(td);
 2635         SCHED_STAT_INC(switch_relinquish);
 2636         mi_switch(SW_VOL, NULL);
 2637         thread_unlock(td);
 2638 }
 2639 
 2640 /*
 2641  * Return the total system load.
 2642  */
 2643 int
 2644 sched_load(void)
 2645 {
 2646 #ifdef SMP
 2647         int total;
 2648         int i;
 2649 
 2650         total = 0;
 2651         for (i = 0; i <= tdg_maxid; i++)
 2652                 total += TDQ_GROUP(i)->tdg_load;
 2653         return (total);
 2654 #else
 2655         return (TDQ_SELF()->tdq_sysload);
 2656 #endif
 2657 }
 2658 
 2659 int
 2660 sched_sizeof_proc(void)
 2661 {
 2662         return (sizeof(struct proc));
 2663 }
 2664 
 2665 int
 2666 sched_sizeof_thread(void)
 2667 {
 2668         return (sizeof(struct thread) + sizeof(struct td_sched));
 2669 }
 2670 
 2671 /*
 2672  * The actual idle process.
 2673  */
 2674 void
 2675 sched_idletd(void *dummy)
 2676 {
 2677         struct thread *td;
 2678         struct tdq *tdq;
 2679 
 2680         td = curthread;
 2681         tdq = TDQ_SELF();
 2682         mtx_assert(&Giant, MA_NOTOWNED);
 2683         /* ULE relies on preemption for idle interruption. */
 2684         for (;;) {
 2685 #ifdef SMP
 2686                 if (tdq_idled(tdq))
 2687                         cpu_idle();
 2688 #else
 2689                 cpu_idle();
 2690 #endif
 2691         }
 2692 }
 2693 
 2694 /*
 2695  * A CPU is entering for the first time or a thread is exiting.
 2696  */
 2697 void
 2698 sched_throw(struct thread *td)
 2699 {
 2700         struct thread *newtd;
 2701         struct tdq *tdq;
 2702 
 2703         tdq = TDQ_SELF();
 2704         if (td == NULL) {
 2705                 /* Correct spinlock nesting and acquire the correct lock. */
 2706                 TDQ_LOCK(tdq);
 2707                 spinlock_exit();
 2708         } else {
 2709                 MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 2710                 tdq_load_rem(tdq, td->td_sched);
 2711         }
 2712         KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
 2713         newtd = choosethread();
 2714         TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
 2715         PCPU_SET(switchtime, cpu_ticks());
 2716         PCPU_SET(switchticks, ticks);
 2717         cpu_throw(td, newtd);           /* doesn't return */
 2718 }
 2719 
 2720 /*
 2721  * This is called from fork_exit().  Just acquire the correct locks and
 2722  * let fork do the rest of the work.
 2723  */
 2724 void
 2725 sched_fork_exit(struct thread *td)
 2726 {
 2727         struct td_sched *ts;
 2728         struct tdq *tdq;
 2729         int cpuid;
 2730 
 2731         /*
 2732          * Finish setting up thread glue so that it begins execution in a
 2733          * non-nested critical section with the scheduler lock held.
 2734          */
 2735         cpuid = PCPU_GET(cpuid);
 2736         tdq = TDQ_CPU(cpuid);
 2737         ts = td->td_sched;
 2738         if (TD_IS_IDLETHREAD(td))
 2739                 td->td_lock = TDQ_LOCKPTR(tdq);
 2740         MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
 2741         td->td_oncpu = cpuid;
 2742         TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
 2743 }
 2744 
 2745 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0,
 2746     "Scheduler");
 2747 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
 2748     "Scheduler name");
 2749 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
 2750     "Slice size for timeshare threads");
 2751 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
 2752      "Interactivity score threshold");
 2753 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
 2754      0,"Min priority for preemption, lower priorities have greater precedence");
 2755 #ifdef SMP
 2756 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0,
 2757     "Pick the target cpu based on priority rather than load.");
 2758 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
 2759     "Number of hz ticks to keep thread affinity for");
 2760 SYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, "");
 2761 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
 2762     "Enables the long-term load balancer");
 2763 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
 2764     &balance_interval, 0,
 2765     "Average frequency in stathz ticks to run the long-term balancer");
 2766 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
 2767     "Steals work from another hyper-threaded core on idle");
 2768 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
 2769     "Attempts to steal work from other cores before idling");
 2770 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
 2771     "Minimum load on remote cpu before we'll steal");
 2772 SYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0,
 2773     "True when a topology has been specified by the MD code.");
 2774 #endif
 2775 
 2776 /* ps compat.  All cpu percentages from ULE are weighted. */
 2777 static int ccpu = 0;
 2778 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
 2779 
 2780 
 2781 #define KERN_SWITCH_INCLUDE 1
 2782 #include "kern/kern_switch.c"

Cache object: 254ec17633132e3838cc21ced311353a


[ 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.