]> git.saurik.com Git - apple/xnu.git/blobdiff - osfmk/kern/priority.c
xnu-3789.1.32.tar.gz
[apple/xnu.git] / osfmk / kern / priority.c
index 8ee162fa3158210c25ae0a7b9b54be02ab597b19..f4f5b1cc8a1a6e9b3159f8e372816575d9e43985 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000-2008 Apple Inc. All rights reserved.
+ * Copyright (c) 2000-2010 Apple Inc. All rights reserved.
  *
  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  * 
 /*
  */
 /*
- *     File:   clock_prim.c
+ *     File:   priority.c
  *     Author: Avadis Tevanian, Jr.
  *     Date:   1986
  *
- *     Clock primitives.
+ *     Priority related scheduler bits.
  */
 
 #include <mach/boolean.h>
 #include <kern/host.h>
 #include <kern/mach_param.h>
 #include <kern/sched.h>
+#include <sys/kdebug.h>
 #include <kern/spl.h>
 #include <kern/thread.h>
 #include <kern/processor.h>
+#include <kern/ledger.h>
 #include <machine/machparam.h>
+#include <kern/machine.h>
+
+#ifdef CONFIG_MACH_APPROXIMATE_TIME
+#include <machine/commpage.h>  /* for commpage_update_mach_approximate_time */
+#endif
+
+static void sched_update_thread_bucket(thread_t thread);
 
 /*
  *     thread_quantum_expire:
@@ -90,94 +99,255 @@ thread_quantum_expire(
        processor_t                     processor = p0;
        thread_t                        thread = p1;
        ast_t                           preempt;
+       uint64_t                        ctime;
+       int                                     urgency;
+       uint64_t                        ignore1, ignore2;
 
-       thread_lock(thread);
+       assert(processor == current_processor());
+       assert(thread == current_thread());
+
+       KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_QUANTUM_EXPIRED) | DBG_FUNC_START, 0, 0, 0, 0, 0);
+
+       SCHED_STATS_QUANTUM_TIMER_EXPIRATION(processor);
 
        /*
-        *      Check for fail-safe trip.
+        * We bill CPU time to both the individual thread and its task.
+        *
+        * Because this balance adjustment could potentially attempt to wake this very
+        * thread, we must credit the ledger before taking the thread lock. The ledger
+        * pointers are only manipulated by the thread itself at the ast boundary.
+        *
+        * TODO: This fails to account for the time between when the timer was armed and when it fired.
+        * It should be based on the system_timer and running a thread_timer_event operation here.
         */
-       if (!(thread->sched_mode & (TH_MODE_TIMESHARE|TH_MODE_PROMOTED))) {
-               uint64_t                        new_computation;
+       ledger_credit(thread->t_ledger, task_ledgers.cpu_time, thread->quantum_remaining);
+       ledger_credit(thread->t_threadledger, thread_ledgers.cpu_time, thread->quantum_remaining);
+#ifdef CONFIG_BANK
+       if (thread->t_bankledger) {
+               ledger_credit(thread->t_bankledger, bank_ledgers.cpu_time,
+                               (thread->quantum_remaining - thread->t_deduct_bank_ledger_time));
+       }
+       thread->t_deduct_bank_ledger_time = 0;
+#endif
 
-               new_computation = processor->quantum_end;
-               new_computation -= thread->computation_epoch;
-               if (new_computation + thread->computation_metered >
-                                                                                       max_unsafe_computation) {
+       ctime = mach_absolute_time();
 
-                       if (thread->sched_mode & TH_MODE_REALTIME) {
-                               thread->priority = DEPRESSPRI;
+#ifdef CONFIG_MACH_APPROXIMATE_TIME
+       commpage_update_mach_approximate_time(ctime);
+#endif
 
-                               thread->safe_mode |= TH_MODE_REALTIME;
-                               thread->sched_mode &= ~TH_MODE_REALTIME;
-                       }
+       thread_lock(thread);
 
-                       sched_share_incr();
+       /*
+        * We've run up until our quantum expiration, and will (potentially)
+        * continue without re-entering the scheduler, so update this now.
+        */
+       processor->last_dispatch = ctime;
+       thread->last_run_time = ctime;
 
-                       thread->safe_release = sched_tick + sched_safe_duration;
-                       thread->sched_mode |= (TH_MODE_FAILSAFE|TH_MODE_TIMESHARE);
+       /*
+        *      Check for fail-safe trip.
+        */
+       if ((thread->sched_mode == TH_MODE_REALTIME || thread->sched_mode == TH_MODE_FIXED) && 
+           !(thread->sched_flags & TH_SFLAG_PROMOTED_MASK) &&
+           !(thread->options & TH_OPT_SYSTEM_CRITICAL)) {
+               uint64_t new_computation;
+  
+               new_computation = ctime - thread->computation_epoch;
+               new_computation += thread->computation_metered;
+               if (new_computation > max_unsafe_computation) {
+                       KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_FAILSAFE)|DBG_FUNC_NONE,
+                                       (uintptr_t)thread->sched_pri, (uintptr_t)thread->sched_mode, 0, 0, 0);
+
+                       thread->safe_release = ctime + sched_safe_duration;
+
+                       sched_thread_mode_demote(thread, TH_SFLAG_FAILSAFE);
                }
        }
-               
+
        /*
         *      Recompute scheduled priority if appropriate.
         */
-       if (thread->sched_stamp != sched_tick)
-               update_priority(thread);
+       if (SCHED(can_update_priority)(thread))
+               SCHED(update_priority)(thread);
        else
-       if (thread->sched_mode & TH_MODE_TIMESHARE) {
-               register uint32_t       delta;
-
-               thread_timer_delta(thread, delta);
+               SCHED(lightweight_update_priority)(thread);
 
-               /*
-                *      Accumulate timesharing usage only
-                *      during contention for processor
-                *      resources.
-                */
-               if (thread->pri_shift < INT8_MAX)
-                       thread->sched_usage += delta;
-
-               thread->cpu_delta += delta;
-
-               /*
-                * Adjust the scheduled priority if
-                * the thread has not been promoted
-                * and is not depressed.
-                */
-               if (    !(thread->sched_mode & TH_MODE_PROMOTED)        &&
-                               !(thread->sched_mode & TH_MODE_ISDEPRESSED)             )
-                       compute_my_priority(thread);
-       }
+       if (thread->sched_mode != TH_MODE_REALTIME)
+               SCHED(quantum_expire)(thread);
 
        processor->current_pri = thread->sched_pri;
+       processor->current_thmode = thread->sched_mode;
+
+       /* Tell platform layer that we are still running this thread */
+       urgency = thread_get_urgency(thread, &ignore1, &ignore2);
+       machine_thread_going_on_core(thread, urgency, 0);
 
        /*
         *      This quantum is up, give this thread another.
         */
-       if (first_timeslice(processor))
-               processor->timeslice--;
+       processor->first_timeslice = FALSE;
 
        thread_quantum_init(thread);
-       processor->quantum_end += thread->current_quantum;
-       timer_call_enter1(&processor->quantum_timer,
-                                                       thread, processor->quantum_end);
+
+       /* Reload precise timing global policy to thread-local policy */
+       thread->precise_user_kernel_time = use_precise_user_kernel_time(thread);
+
+       /*
+        * Since non-precise user/kernel time doesn't update the state/thread timer
+        * during privilege transitions, synthesize an event now.
+        */
+       if (!thread->precise_user_kernel_time) {
+               timer_switch(PROCESSOR_DATA(processor, current_state),
+                                        ctime,
+                                        PROCESSOR_DATA(processor, current_state));
+               timer_switch(PROCESSOR_DATA(processor, thread_timer),
+                                        ctime,
+                                        PROCESSOR_DATA(processor, thread_timer));
+       }
+
+       processor->quantum_end = ctime + thread->quantum_remaining;
 
        /*
         *      Context switch check.
         */
-       if ((preempt = csw_check(thread, processor)) != AST_NONE)
+       if ((preempt = csw_check(processor, AST_QUANTUM)) != AST_NONE)
                ast_on(preempt);
-       else {
-               processor_set_t         pset = processor->processor_set;
 
-               pset_lock(pset);
+       thread_unlock(thread);
+
+       timer_call_enter1(&processor->quantum_timer, thread,
+           processor->quantum_end, TIMER_CALL_SYS_CRITICAL | TIMER_CALL_LOCAL);
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+       sched_timeshare_consider_maintenance(ctime);
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
+
+
+       KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_QUANTUM_EXPIRED) | DBG_FUNC_END, preempt, 0, 0, 0, 0);
+}
+
+/*
+ *     sched_set_thread_base_priority:
+ *
+ *     Set the base priority of the thread
+ *     and reset its scheduled priority.
+ *
+ *     This is the only path to change base_pri.
+ *
+ *     Called with the thread locked.
+ */
+void
+sched_set_thread_base_priority(thread_t thread, int priority)
+{
+       assert(priority >= MINPRI);
+
+       if (thread->sched_mode == TH_MODE_REALTIME)
+               assert(priority <= BASEPRI_RTQUEUES);
+       else
+               assert(priority < BASEPRI_RTQUEUES);
+
+       thread->base_pri = priority;
+
+       sched_update_thread_bucket(thread);
+
+       thread_recompute_sched_pri(thread, FALSE);
+}
+
+/*
+ *     thread_recompute_sched_pri:
+ *
+ *     Reset the scheduled priority of the thread
+ *     according to its base priority if the
+ *     thread has not been promoted or depressed.
+ *
+ *     This is the standard way to push base_pri changes into sched_pri,
+ *     or to recalculate the appropriate sched_pri after clearing
+ *     a promotion or depression.
+ *
+ *     Called at splsched with the thread locked.
+ */
+void
+thread_recompute_sched_pri(
+                           thread_t thread,
+                           boolean_t override_depress)
+{
+       int priority;
 
-               pset_pri_hint(pset, processor, processor->current_pri);
+       if (thread->sched_mode == TH_MODE_TIMESHARE)
+               priority = SCHED(compute_timeshare_priority)(thread);
+       else
+               priority = thread->base_pri;
 
-               pset_unlock(pset);
+       if ((!(thread->sched_flags & TH_SFLAG_PROMOTED_MASK)  || (priority > thread->sched_pri)) &&
+           (!(thread->sched_flags & TH_SFLAG_DEPRESSED_MASK) || override_depress)) {
+               set_sched_pri(thread, priority);
        }
+}
 
-       thread_unlock(thread);
+void
+sched_default_quantum_expire(thread_t thread __unused)
+{
+      /*
+       * No special behavior when a timeshare, fixed, or realtime thread
+       * uses up its entire quantum
+       */
+}
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+
+/*
+ *     lightweight_update_priority:
+ *
+ *     Update the scheduled priority for
+ *     a timesharing thread.
+ *
+ *     Only for use on the current thread.
+ *
+ *     Called with the thread locked.
+ */
+void
+lightweight_update_priority(thread_t thread)
+{
+       assert(thread->runq == PROCESSOR_NULL);
+       assert(thread == current_thread());
+
+       if (thread->sched_mode == TH_MODE_TIMESHARE) {
+               int priority;
+               uint32_t delta;
+
+               thread_timer_delta(thread, delta);
+
+               /*
+                *      Accumulate timesharing usage only
+                *      during contention for processor
+                *      resources.
+                */
+               if (thread->pri_shift < INT8_MAX)
+                       thread->sched_usage += delta;
+
+               thread->cpu_delta += delta;
+
+               priority = sched_compute_timeshare_priority(thread);
+
+               /*
+                * Adjust the scheduled priority like thread_recompute_sched_pri,
+                * except with the benefit of knowing the thread is on this core.
+                */
+               if ((!(thread->sched_flags & TH_SFLAG_PROMOTED_MASK)  || (priority > thread->sched_pri)) &&
+                   (!(thread->sched_flags & TH_SFLAG_DEPRESSED_MASK)) &&
+                   priority != thread->sched_pri) {
+
+                       thread->sched_pri = priority;
+
+                       KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY),
+                                             (uintptr_t)thread_tid(thread),
+                                             thread->base_pri,
+                                             thread->sched_pri,
+                                             0, /* eventually, 'reason' */
+                                             0);
+               }
+       }
 }
 
 /*
@@ -201,86 +371,43 @@ static struct shift_data  sched_decay_shifts[SCHED_DECAY_TICKS] = {
 };
 
 /*
- *     do_priority_computation:
+ *     sched_compute_timeshare_priority:
  *
  *     Calculate the timesharing priority based upon usage and load.
  */
-#define do_priority_computation(thread, pri)                                                   \
-       MACRO_BEGIN                                                                                                                     \
-       (pri) = (thread)->priority              /* start with base priority */          \
-           - ((thread)->sched_usage >> (thread)->pri_shift);                           \
-       if ((pri) < MINPRI_USER)                                                                                        \
-               (pri) = MINPRI_USER;                                                                                    \
-       else                                                                                                                            \
-       if ((pri) > MAXPRI_KERNEL)                                                                                      \
-               (pri) = MAXPRI_KERNEL;                                                                                  \
-       MACRO_END
+extern int sched_pri_decay_band_limit;
 
-/*
- *     set_priority:
- *
- *     Set the base priority of the thread
- *     and reset its scheduled priority.
- *
- *     Called with the thread locked.
- */
-void
-set_priority(
-       register thread_t       thread,
-       register int            priority)
-{
-       thread->priority = priority;
-       compute_priority(thread, FALSE);
-}
 
-/*
- *     compute_priority:
- *
- *     Reset the scheduled priority of the thread
- *     according to its base priority if the
- *     thread has not been promoted or depressed.
- *
- *     Called with the thread locked.
- */
-void
-compute_priority(
-       register thread_t       thread,
-       boolean_t                       override_depress)
+int
+sched_compute_timeshare_priority(thread_t thread)
 {
-       register int            priority;
+       /* start with base priority */
+       int priority = thread->base_pri - (thread->sched_usage >> thread->pri_shift);
 
-       if (    !(thread->sched_mode & TH_MODE_PROMOTED)                        &&
-                       (!(thread->sched_mode & TH_MODE_ISDEPRESSED)    ||
-                                override_depress                                                       )               ) {
-               if (thread->sched_mode & TH_MODE_TIMESHARE)
-                       do_priority_computation(thread, priority);
-               else
-                       priority = thread->priority;
+       if (priority < MINPRI_USER)
+               priority = MINPRI_USER;
+       else if (priority > MAXPRI_KERNEL)
+               priority = MAXPRI_KERNEL;
 
-               set_sched_pri(thread, priority);
-       }
+       return priority;
 }
 
+
 /*
- *     compute_my_priority:
- *
- *     Reset the scheduled priority for
- *     a timesharing thread.
+ *     can_update_priority
  *
- *     Only for use on the current thread
- *     if timesharing and not depressed.
+ *     Make sure we don't do re-dispatches more frequently than a scheduler tick.
  *
  *     Called with the thread locked.
  */
-void
-compute_my_priority(
-       register thread_t       thread)
+boolean_t
+can_update_priority(
+                                       thread_t        thread)
 {
-       register int            priority;
-
-       do_priority_computation(thread, priority);
-       assert(thread->runq == PROCESSOR_NULL);
-       thread->sched_pri = priority;
+       if (sched_tick == thread->sched_stamp)
+               return (FALSE);
+       else
+               return (TRUE);
 }
 
 /*
@@ -292,23 +419,26 @@ compute_my_priority(
  */
 void
 update_priority(
-       register thread_t       thread)
+       thread_t        thread)
 {
-       register unsigned       ticks;
-       register uint32_t       delta;
+       uint32_t ticks, delta;
 
        ticks = sched_tick - thread->sched_stamp;
        assert(ticks != 0);
+
        thread->sched_stamp += ticks;
-       thread->pri_shift = sched_pri_shift;
+
+       thread->pri_shift = sched_pri_shifts[thread->th_sched_bucket];
+
+       /* If requested, accelerate aging of sched_usage */
+       if (sched_decay_usage_age_factor > 1)
+               ticks *= sched_decay_usage_age_factor;
 
        /*
         *      Gather cpu usage data.
         */
        thread_timer_delta(thread, delta);
        if (ticks < SCHED_DECAY_TICKS) {
-               register struct shift_data      *shiftp;
-
                /*
                 *      Accumulate timesharing usage only
                 *      during contention for processor
@@ -320,25 +450,20 @@ update_priority(
                thread->cpu_usage += delta + thread->cpu_delta;
                thread->cpu_delta = 0;
 
-               shiftp = &sched_decay_shifts[ticks];
+               struct shift_data *shiftp = &sched_decay_shifts[ticks];
+
                if (shiftp->shift2 > 0) {
-                   thread->cpu_usage =
-                                               (thread->cpu_usage >> shiftp->shift1) +
-                                               (thread->cpu_usage >> shiftp->shift2);
-                   thread->sched_usage =
-                                               (thread->sched_usage >> shiftp->shift1) +
-                                               (thread->sched_usage >> shiftp->shift2);
-               }
-               else {
-                   thread->cpu_usage =
-                                               (thread->cpu_usage >> shiftp->shift1) -
-                                               (thread->cpu_usage >> -(shiftp->shift2));
-                   thread->sched_usage =
-                                               (thread->sched_usage >> shiftp->shift1) -
-                                               (thread->sched_usage >> -(shiftp->shift2));
+                       thread->cpu_usage =   (thread->cpu_usage >> shiftp->shift1) +
+                                             (thread->cpu_usage >> shiftp->shift2);
+                       thread->sched_usage = (thread->sched_usage >> shiftp->shift1) +
+                                             (thread->sched_usage >> shiftp->shift2);
+               } else {
+                       thread->cpu_usage =   (thread->cpu_usage >>   shiftp->shift1) -
+                                             (thread->cpu_usage >> -(shiftp->shift2));
+                       thread->sched_usage = (thread->sched_usage >>   shiftp->shift1) -
+                                             (thread->sched_usage >> -(shiftp->shift2));
                }
-       }
-       else {
+       } else {
                thread->cpu_usage = thread->cpu_delta = 0;
                thread->sched_usage = 0;
        }
@@ -346,43 +471,228 @@ update_priority(
        /*
         *      Check for fail-safe release.
         */
-       if (    (thread->sched_mode & TH_MODE_FAILSAFE)         &&
-                       thread->sched_stamp >= thread->safe_release             ) {
-               if (!(thread->safe_mode & TH_MODE_TIMESHARE)) {
-                       if (thread->safe_mode & TH_MODE_REALTIME) {
-                               thread->priority = BASEPRI_RTQUEUES;
+       if ((thread->sched_flags & TH_SFLAG_FAILSAFE) &&
+           mach_absolute_time() >= thread->safe_release) {
+               sched_thread_mode_undemote(thread, TH_SFLAG_FAILSAFE);
+       }
 
-                               thread->sched_mode |= TH_MODE_REALTIME;
-                       }
+       /*
+        *      Recompute scheduled priority if appropriate.
+        */
+       if (thread->sched_mode == TH_MODE_TIMESHARE) {
+               int priority = sched_compute_timeshare_priority(thread);
 
-                       thread->sched_mode &= ~TH_MODE_TIMESHARE;
+               /*
+                * Adjust the scheduled priority like thread_recompute_sched_pri,
+                * except without setting an AST.
+                */
+               if ((!(thread->sched_flags & TH_SFLAG_PROMOTED_MASK)  || (priority > thread->sched_pri)) &&
+                   (!(thread->sched_flags & TH_SFLAG_DEPRESSED_MASK)) &&
+                   priority != thread->sched_pri) {
 
-                       if ((thread->state & (TH_RUN|TH_IDLE)) == TH_RUN)
-                               sched_share_decr();
+                       boolean_t removed = thread_run_queue_remove(thread);
 
-                       if (!(thread->sched_mode & TH_MODE_ISDEPRESSED))
-                               set_sched_pri(thread, thread->priority);
-               }
+                       thread->sched_pri = priority;
+
+                       KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY),
+                                             (uintptr_t)thread_tid(thread),
+                                             thread->base_pri,
+                                             thread->sched_pri,
+                                             0, /* eventually, 'reason' */
+                                             0);
 
-               thread->safe_mode = 0;
-               thread->sched_mode &= ~TH_MODE_FAILSAFE;
+                       if (removed)
+                               thread_run_queue_reinsert(thread, SCHED_TAILQ);
+               }
        }
 
-       /*
-        *      Recompute scheduled priority if appropriate.
-        */
-       if (    (thread->sched_mode & TH_MODE_TIMESHARE)        &&
-                       !(thread->sched_mode & TH_MODE_PROMOTED)        &&
-                       !(thread->sched_mode & TH_MODE_ISDEPRESSED)             ) {
-               register int            new_pri;
+       return;
+}
 
-               do_priority_computation(thread, new_pri);
-               if (new_pri != thread->sched_pri) {
-                       boolean_t               removed = run_queue_remove(thread);
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
 
-                       thread->sched_pri = new_pri;
-                       if (removed)
-                               thread_setrun(thread, SCHED_TAILQ);
+
+/*
+ * TH_BUCKET_RUN is a count of *all* runnable non-idle threads.
+ * Each other bucket is a count of the runnable non-idle threads
+ * with that property.
+ */
+volatile uint32_t       sched_run_buckets[TH_BUCKET_MAX];
+
+static void
+sched_incr_bucket(sched_bucket_t bucket)
+{
+       assert(bucket >= TH_BUCKET_FIXPRI &&
+              bucket <= TH_BUCKET_SHARE_BG);
+
+       hw_atomic_add(&sched_run_buckets[bucket], 1);
+}
+
+static void
+sched_decr_bucket(sched_bucket_t bucket)
+{
+       assert(bucket >= TH_BUCKET_FIXPRI &&
+              bucket <= TH_BUCKET_SHARE_BG);
+
+       assert(sched_run_buckets[bucket] > 0);
+
+       hw_atomic_sub(&sched_run_buckets[bucket], 1);
+}
+
+/* TH_RUN & !TH_IDLE controls whether a thread has a run count */
+
+uint32_t
+sched_run_incr(thread_t thread)
+{
+       assert((thread->state & (TH_RUN|TH_IDLE)) == TH_RUN);
+
+       uint32_t new_count = hw_atomic_add(&sched_run_buckets[TH_BUCKET_RUN], 1);
+
+       sched_incr_bucket(thread->th_sched_bucket);
+
+       return new_count;
+}
+
+uint32_t
+sched_run_decr(thread_t thread)
+{
+       assert((thread->state & (TH_RUN|TH_IDLE)) != TH_RUN);
+
+       sched_decr_bucket(thread->th_sched_bucket);
+
+       uint32_t new_count = hw_atomic_sub(&sched_run_buckets[TH_BUCKET_RUN], 1);
+
+       return new_count;
+}
+
+static void
+sched_update_thread_bucket(thread_t thread)
+{
+       sched_bucket_t old_bucket = thread->th_sched_bucket;
+       sched_bucket_t new_bucket = TH_BUCKET_RUN;
+
+       switch (thread->sched_mode) {
+       case TH_MODE_FIXED:
+       case TH_MODE_REALTIME:
+               new_bucket = TH_BUCKET_FIXPRI;
+               break;
+
+       case TH_MODE_TIMESHARE:
+               if (thread->base_pri > BASEPRI_UTILITY)
+                       new_bucket = TH_BUCKET_SHARE_FG;
+               else if (thread->base_pri > MAXPRI_THROTTLE)
+                       new_bucket = TH_BUCKET_SHARE_UT;
+               else
+                       new_bucket = TH_BUCKET_SHARE_BG;
+               break;
+
+       default:
+               panic("unexpected mode: %d", thread->sched_mode);
+               break;
+       }
+
+       if (old_bucket != new_bucket) {
+               thread->th_sched_bucket = new_bucket;
+               thread->pri_shift = sched_pri_shifts[new_bucket];
+
+               if ((thread->state & (TH_RUN|TH_IDLE)) == TH_RUN) {
+                       sched_decr_bucket(old_bucket);
+                       sched_incr_bucket(new_bucket);
                }
        }
 }
+
+/*
+ * Set the thread's true scheduling mode
+ * Called with thread mutex and thread locked
+ * The thread has already been removed from the runqueue.
+ *
+ * (saved_mode is handled before this point)
+ */
+void
+sched_set_thread_mode(thread_t thread, sched_mode_t new_mode)
+{
+       assert(thread->runq == PROCESSOR_NULL);
+
+       switch (new_mode) {
+       case TH_MODE_FIXED:
+       case TH_MODE_REALTIME:
+       case TH_MODE_TIMESHARE:
+               break;
+
+       default:
+               panic("unexpected mode: %d", new_mode);
+               break;
+       }
+
+       thread->sched_mode = new_mode;
+
+       sched_update_thread_bucket(thread);
+}
+
+/*
+ * Demote the true scheduler mode to timeshare (called with the thread locked)
+ */
+void
+sched_thread_mode_demote(thread_t thread, uint32_t reason)
+{
+       assert(reason & TH_SFLAG_DEMOTED_MASK);
+       assert((thread->sched_flags & reason) != reason);
+
+       if (thread->policy_reset)
+               return;
+
+       if (thread->sched_flags & TH_SFLAG_DEMOTED_MASK) {
+               /* Another demotion reason is already active */
+               thread->sched_flags |= reason;
+               return;
+       }
+
+       assert(thread->saved_mode == TH_MODE_NONE);
+
+       boolean_t removed = thread_run_queue_remove(thread);
+
+       thread->sched_flags |= reason;
+
+       thread->saved_mode = thread->sched_mode;
+
+       sched_set_thread_mode(thread, TH_MODE_TIMESHARE);
+
+       thread_recompute_priority(thread);
+
+       if (removed)
+               thread_run_queue_reinsert(thread, SCHED_TAILQ);
+}
+
+/*
+ * Un-demote the true scheduler mode back to the saved mode (called with the thread locked)
+ */
+void
+sched_thread_mode_undemote(thread_t thread, uint32_t reason)
+{
+       assert(reason & TH_SFLAG_DEMOTED_MASK);
+       assert((thread->sched_flags & reason) == reason);
+       assert(thread->saved_mode != TH_MODE_NONE);
+       assert(thread->sched_mode == TH_MODE_TIMESHARE);
+       assert(thread->policy_reset == 0);
+
+       thread->sched_flags &= ~reason;
+
+       if (thread->sched_flags & TH_SFLAG_DEMOTED_MASK) {
+               /* Another demotion reason is still active */
+               return;
+       }
+
+       boolean_t removed = thread_run_queue_remove(thread);
+
+       sched_set_thread_mode(thread, thread->saved_mode);
+
+       thread->saved_mode = TH_MODE_NONE;
+
+       thread_recompute_priority(thread);
+
+       if (removed)
+               thread_run_queue_reinsert(thread, SCHED_TAILQ);
+}
+
+