X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/b0d623f7f2ae71ed96e60569f61f9a9a27016e80..c6bf4f310a33a9262d455ea4d3f0630b1255e3fe:/osfmk/kern/sched_prim.c diff --git a/osfmk/kern/sched_prim.c b/osfmk/kern/sched_prim.c index ca4a1354f..42e73b4f0 100644 --- a/osfmk/kern/sched_prim.c +++ b/osfmk/kern/sched_prim.c @@ -1,8 +1,8 @@ /* - * Copyright (c) 2000-2009 Apple Inc. All rights reserved. + * Copyright (c) 2000-2016 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ - * + * * This file contains Original Code and/or Modifications of Original Code * as defined in and that are subject to the Apple Public Source License * Version 2.0 (the 'License'). You may not use this file except in @@ -11,10 +11,10 @@ * unlawful or unlicensed copies of an Apple operating system, or to * circumvent, violate, or enable the circumvention or violation of, any * terms of an Apple operating system software license agreement. - * + * * Please obtain a copy of the License at * http://www.opensource.apple.com/apsl/ and read it before using this file. - * + * * The Original Code and all software distributed under the License are * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, @@ -22,34 +22,34 @@ * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. * Please see the License for the specific language governing rights and * limitations under the License. - * + * * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ /* * @OSF_FREE_COPYRIGHT@ */ -/* +/* * Mach Operating System * Copyright (c) 1991,1990,1989,1988,1987 Carnegie Mellon University * All Rights Reserved. - * + * * Permission to use, copy, modify and distribute this software and its * documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. - * + * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND FOR * ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. - * + * * Carnegie Mellon requests users of this software to return to - * + * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 - * + * * any improvements or extensions that they make and grant Carnegie Mellon * the rights to redistribute these changes. */ @@ -65,191 +65,323 @@ */ #include -#include - -#include #include #include #include #include +#include #include #include #include +#include +#include + +#include #include +#include #include #include #include #include +#include #include -#include #include #include #include +#if MONOTONIC +#include +#endif /* MONOTONIC */ #include #include #include #include +#include #include #include #include -#include +#include +#include +#include +#include +#include #include #include #include +#include #include +#include +#include #include - +#include +#include +#include #include +#include +#include + +int +rt_runq_count(processor_set_t pset) +{ + return atomic_load_explicit(&SCHED(rt_runq)(pset)->count, memory_order_relaxed); +} + +void +rt_runq_count_incr(processor_set_t pset) +{ + atomic_fetch_add_explicit(&SCHED(rt_runq)(pset)->count, 1, memory_order_relaxed); +} + +void +rt_runq_count_decr(processor_set_t pset) +{ + atomic_fetch_sub_explicit(&SCHED(rt_runq)(pset)->count, 1, memory_order_relaxed); +} + +#define DEFAULT_PREEMPTION_RATE 100 /* (1/s) */ +int default_preemption_rate = DEFAULT_PREEMPTION_RATE; + +#define DEFAULT_BG_PREEMPTION_RATE 400 /* (1/s) */ +int default_bg_preemption_rate = DEFAULT_BG_PREEMPTION_RATE; + +#define MAX_UNSAFE_QUANTA 800 +int max_unsafe_quanta = MAX_UNSAFE_QUANTA; + +#define MAX_POLL_QUANTA 2 +int max_poll_quanta = MAX_POLL_QUANTA; + +#define SCHED_POLL_YIELD_SHIFT 4 /* 1/16 */ +int sched_poll_yield_shift = SCHED_POLL_YIELD_SHIFT; + +uint64_t max_poll_computation; + +uint64_t max_unsafe_computation; +uint64_t sched_safe_duration; -struct run_queue rt_runq; -#define RT_RUNQ ((processor_t)-1) -decl_simple_lock_data(static,rt_lock); +#if defined(CONFIG_SCHED_TIMESHARE_CORE) -#define DEFAULT_PREEMPTION_RATE 100 /* (1/s) */ -int default_preemption_rate = DEFAULT_PREEMPTION_RATE; +uint32_t std_quantum; +uint32_t min_std_quantum; +uint32_t bg_quantum; -#define MAX_UNSAFE_QUANTA 800 -int max_unsafe_quanta = MAX_UNSAFE_QUANTA; +uint32_t std_quantum_us; +uint32_t bg_quantum_us; -#define MAX_POLL_QUANTA 2 -int max_poll_quanta = MAX_POLL_QUANTA; +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ -#define SCHED_POLL_YIELD_SHIFT 4 /* 1/16 */ -int sched_poll_yield_shift = SCHED_POLL_YIELD_SHIFT; +uint32_t thread_depress_time; +uint32_t default_timeshare_computation; +uint32_t default_timeshare_constraint; -uint64_t max_unsafe_computation; -uint32_t sched_safe_duration; -uint64_t max_poll_computation; +uint32_t max_rt_quantum; +uint32_t min_rt_quantum; -uint32_t std_quantum; -uint32_t min_std_quantum; +#if defined(CONFIG_SCHED_TIMESHARE_CORE) -uint32_t std_quantum_us; +unsigned sched_tick; +uint32_t sched_tick_interval; -uint32_t max_rt_quantum; -uint32_t min_rt_quantum; +/* Timeshare load calculation interval (15ms) */ +uint32_t sched_load_compute_interval_us = 15000; +uint64_t sched_load_compute_interval_abs; +static _Atomic uint64_t sched_load_compute_deadline; -uint32_t sched_cswtime; +uint32_t sched_pri_shifts[TH_BUCKET_MAX]; +uint32_t sched_fixed_shift; -unsigned sched_tick; -uint32_t sched_tick_interval; +uint32_t sched_decay_usage_age_factor = 1; /* accelerate 5/8^n usage aging */ -uint32_t sched_pri_shift = INT8_MAX; -uint32_t sched_fixed_shift; +/* Allow foreground to decay past default to resolve inversions */ +#define DEFAULT_DECAY_BAND_LIMIT ((BASEPRI_FOREGROUND - BASEPRI_DEFAULT) + 2) +int sched_pri_decay_band_limit = DEFAULT_DECAY_BAND_LIMIT; -uint32_t sched_run_count, sched_share_count; -uint32_t sched_load_average, sched_mach_factor; +/* Defaults for timer deadline profiling */ +#define TIMER_DEADLINE_TRACKING_BIN_1_DEFAULT 2000000 /* Timers with deadlines <= + * 2ms */ +#define TIMER_DEADLINE_TRACKING_BIN_2_DEFAULT 5000000 /* Timers with deadlines + * <= 5ms */ + +uint64_t timer_deadline_tracking_bin_1; +uint64_t timer_deadline_tracking_bin_2; + +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +thread_t sched_maintenance_thread; + +/* interrupts disabled lock to guard recommended cores state */ +decl_simple_lock_data(static, sched_recommended_cores_lock); +static uint64_t usercontrol_requested_recommended_cores = ALL_CORES_RECOMMENDED; +static void sched_update_recommended_cores(uint64_t recommended_cores); + +#if __arm__ || __arm64__ +static void sched_recommended_cores_maintenance(void); +uint64_t perfcontrol_failsafe_starvation_threshold; +extern char *proc_name_address(struct proc *p); +#endif /* __arm__ || __arm64__ */ + +uint64_t sched_one_second_interval; /* Forwards */ -static void load_shift_init(void) __attribute__((section("__TEXT, initcode"))); -static void preempt_pri_init(void) __attribute__((section("__TEXT, initcode"))); -static thread_t run_queue_dequeue( - run_queue_t runq, - integer_t options); +#if defined(CONFIG_SCHED_TIMESHARE_CORE) -static thread_t choose_thread( - processor_t processor, - int priority); +static void load_shift_init(void); +static void preempt_pri_init(void); -static thread_t thread_select_idle( - thread_t thread, - processor_t processor); +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ -static thread_t processor_idle( - thread_t thread, - processor_t processor); +thread_t processor_idle( + thread_t thread, + processor_t processor); -static thread_t steal_thread( - processor_set_t pset); +static ast_t +csw_check_locked( + thread_t thread, + processor_t processor, + processor_set_t pset, + ast_t check_reason); -static thread_t steal_processor_thread( - processor_t processor); +static void processor_setrun( + processor_t processor, + thread_t thread, + integer_t options); + +static void +sched_realtime_timebase_init(void); -static void thread_update_scan(void); +static void +sched_timer_deadline_tracking_init(void); -#if DEBUG +#if DEBUG extern int debug_task; #define TLOG(a, fmt, args...) if(debug_task & a) kprintf(fmt, ## args) #else #define TLOG(a, fmt, args...) do {} while (0) #endif -#if DEBUG -static -boolean_t thread_runnable( - thread_t thread); +static processor_t +thread_bind_internal( + thread_t thread, + processor_t processor); + +static void +sched_vm_group_maintenance(void); -#endif /*DEBUG*/ +#if defined(CONFIG_SCHED_TIMESHARE_CORE) +int8_t sched_load_shifts[NRQS]; +bitmap_t sched_preempt_pri[BITMAP_LEN(NRQS_MAX)]; +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ /* - * State machine - * - * states are combinations of: - * R running - * W waiting (or on wait queue) - * N non-interruptible - * O swapped out - * I being swapped in - * - * init action - * assert_wait thread_block clear_wait swapout swapin - * - * R RW, RWN R; setrun - - - * RN RWN RN; setrun - - - * - * RW W R - - * RWN WN RN - - * - * W R; setrun WO - * WN RN; setrun - - * - * RO - - R + * Statically allocate a buffer to hold the longest possible + * scheduler description string, as currently implemented. + * bsd/kern/kern_sysctl.c has a corresponding definition in bsd/ + * to export to userspace via sysctl(3). If either version + * changes, update the other. * + * Note that in addition to being an upper bound on the strings + * in the kernel, it's also an exact parameter to PE_get_default(), + * which interrogates the device tree on some platforms. That + * API requires the caller know the exact size of the device tree + * property, so we need both a legacy size (32) and the current size + * (48) to deal with old and new device trees. The device tree property + * is similarly padded to a fixed size so that the same kernel image + * can run on multiple devices with different schedulers configured + * in the device tree. */ +char sched_string[SCHED_STRING_MAX_LENGTH]; -int8_t sched_load_shifts[NRQS]; -int sched_preempt_pri[NRQBM]; +uint32_t sched_debug_flags = SCHED_DEBUG_FLAG_CHOOSE_PROCESSOR_TRACEPOINTS; + +/* Global flag which indicates whether Background Stepper Context is enabled */ +static int cpu_throttle_enabled = 1; void sched_init(void) +{ + kprintf("Scheduler: Default of %s\n", SCHED(sched_name)); + + if (!PE_parse_boot_argn("sched_pri_decay_limit", &sched_pri_decay_band_limit, sizeof(sched_pri_decay_band_limit))) { + /* No boot-args, check in device tree */ + if (!PE_get_default("kern.sched_pri_decay_limit", + &sched_pri_decay_band_limit, + sizeof(sched_pri_decay_band_limit))) { + /* Allow decay all the way to normal limits */ + sched_pri_decay_band_limit = DEFAULT_DECAY_BAND_LIMIT; + } + } + + kprintf("Setting scheduler priority decay band limit %d\n", sched_pri_decay_band_limit); + + if (PE_parse_boot_argn("sched_debug", &sched_debug_flags, sizeof(sched_debug_flags))) { + kprintf("Scheduler: Debug flags 0x%08x\n", sched_debug_flags); + } + strlcpy(sched_string, SCHED(sched_name), sizeof(sched_string)); + + cpu_quiescent_counter_init(); + + SCHED(init)(); + SCHED(rt_init)(&pset0); + sched_timer_deadline_tracking_init(); + + SCHED(pset_init)(&pset0); + SCHED(processor_init)(master_processor); +} + +void +sched_timebase_init(void) +{ + uint64_t abstime; + + clock_interval_to_absolutetime_interval(1, NSEC_PER_SEC, &abstime); + sched_one_second_interval = abstime; + + SCHED(timebase_init)(); + sched_realtime_timebase_init(); +} + +#if defined(CONFIG_SCHED_TIMESHARE_CORE) + +void +sched_timeshare_init(void) { /* * Calculate the timeslicing quantum * in us. */ - if (default_preemption_rate < 1) + if (default_preemption_rate < 1) { default_preemption_rate = DEFAULT_PREEMPTION_RATE; + } std_quantum_us = (1000 * 1000) / default_preemption_rate; printf("standard timeslicing quantum is %d us\n", std_quantum_us); - sched_safe_duration = (2 * max_unsafe_quanta / default_preemption_rate) * - (1 << SCHED_TICK_SHIFT); + if (default_bg_preemption_rate < 1) { + default_bg_preemption_rate = DEFAULT_BG_PREEMPTION_RATE; + } + bg_quantum_us = (1000 * 1000) / default_bg_preemption_rate; + + printf("standard background quantum is %d us\n", bg_quantum_us); load_shift_init(); preempt_pri_init(); - simple_lock_init(&rt_lock, 0); - run_queue_init(&rt_runq); sched_tick = 0; - ast_init(); } void -sched_timebase_init(void) +sched_timeshare_timebase_init(void) { - uint64_t abstime; - uint32_t shift; + uint64_t abstime; + uint32_t shift; /* standard timeslicing quantum */ clock_interval_to_absolutetime_interval( - std_quantum_us, NSEC_PER_USEC, &abstime); + std_quantum_us, NSEC_PER_USEC, &abstime); assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); std_quantum = (uint32_t)abstime; @@ -258,171 +390,249 @@ sched_timebase_init(void) assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); min_std_quantum = (uint32_t)abstime; - /* smallest rt computaton (50 us) */ - clock_interval_to_absolutetime_interval(50, NSEC_PER_USEC, &abstime); - assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); - min_rt_quantum = (uint32_t)abstime; - - /* maximum rt computation (50 ms) */ + /* quantum for background tasks */ clock_interval_to_absolutetime_interval( - 50, 1000*NSEC_PER_USEC, &abstime); + bg_quantum_us, NSEC_PER_USEC, &abstime); assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); - max_rt_quantum = (uint32_t)abstime; + bg_quantum = (uint32_t)abstime; /* scheduler tick interval */ clock_interval_to_absolutetime_interval(USEC_PER_SEC >> SCHED_TICK_SHIFT, - NSEC_PER_USEC, &abstime); + NSEC_PER_USEC, &abstime); assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); sched_tick_interval = (uint32_t)abstime; + /* timeshare load calculation interval & deadline initialization */ + clock_interval_to_absolutetime_interval(sched_load_compute_interval_us, NSEC_PER_USEC, &sched_load_compute_interval_abs); + sched_load_compute_deadline = sched_load_compute_interval_abs; + /* * Compute conversion factor from usage to * timesharing priorities with 5/8 ** n aging. */ abstime = (abstime * 5) / 3; - for (shift = 0; abstime > BASEPRI_DEFAULT; ++shift) + for (shift = 0; abstime > BASEPRI_DEFAULT; ++shift) { abstime >>= 1; + } sched_fixed_shift = shift; - max_unsafe_computation = max_unsafe_quanta * std_quantum; - max_poll_computation = max_poll_quanta * std_quantum; + for (uint32_t i = 0; i < TH_BUCKET_MAX; i++) { + sched_pri_shifts[i] = INT8_MAX; + } + + max_unsafe_computation = ((uint64_t)max_unsafe_quanta) * std_quantum; + sched_safe_duration = 2 * ((uint64_t)max_unsafe_quanta) * std_quantum; + + max_poll_computation = ((uint64_t)max_poll_quanta) * std_quantum; + thread_depress_time = 1 * std_quantum; + default_timeshare_computation = std_quantum / 2; + default_timeshare_constraint = std_quantum; + +#if __arm__ || __arm64__ + perfcontrol_failsafe_starvation_threshold = (2 * sched_tick_interval); +#endif /* __arm__ || __arm64__ */ } -/* - * Set up values for timeshare - * loading factors. - */ -static void -load_shift_init(void) +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +void +pset_rt_init(processor_set_t pset) { - int8_t k, *p = sched_load_shifts; - uint32_t i, j; + rt_lock_init(pset); - *p++ = INT8_MIN; *p++ = 0; + os_atomic_init(&pset->rt_runq.count, 0); + queue_init(&pset->rt_runq.queue); + memset(&pset->rt_runq.runq_stats, 0, sizeof pset->rt_runq.runq_stats); +} - for (i = j = 2, k = 1; i < NRQS; ++k) { - for (j <<= 1; i < j; ++i) - *p++ = k; +rt_queue_t +sched_rtglobal_runq(processor_set_t pset) +{ + (void)pset; + + return &pset0.rt_runq; +} + +void +sched_rtglobal_init(processor_set_t pset) +{ + if (pset == &pset0) { + return pset_rt_init(pset); } + + /* Only pset0 rt_runq is used, so make it easy to detect + * buggy accesses to others. + */ + memset(&pset->rt_runq, 0xfd, sizeof pset->rt_runq); +} + +void +sched_rtglobal_queue_shutdown(processor_t processor) +{ + (void)processor; } static void -preempt_pri_init(void) +sched_realtime_timebase_init(void) { - int i, *p = sched_preempt_pri; + uint64_t abstime; - for (i = BASEPRI_FOREGROUND + 1; i < MINPRI_KERNEL; ++i) - setbit(i, p); + /* smallest rt computaton (50 us) */ + clock_interval_to_absolutetime_interval(50, NSEC_PER_USEC, &abstime); + assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); + min_rt_quantum = (uint32_t)abstime; - for (i = BASEPRI_PREEMPT; i <= MAXPRI; ++i) - setbit(i, p); + /* maximum rt computation (50 ms) */ + clock_interval_to_absolutetime_interval( + 50, 1000 * NSEC_PER_USEC, &abstime); + assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); + max_rt_quantum = (uint32_t)abstime; } -/* - * Thread wait timer expiration. - */ void -thread_timer_expire( - void *p0, - __unused void *p1) +sched_check_spill(processor_set_t pset, thread_t thread) { - thread_t thread = p0; - spl_t s; + (void)pset; + (void)thread; - s = splsched(); - thread_lock(thread); - if (--thread->wait_timer_active == 0) { - if (thread->wait_timer_is_set) { - thread->wait_timer_is_set = FALSE; - clear_wait_internal(thread, THREAD_TIMED_OUT); - } - } - thread_unlock(thread); - splx(s); + return; +} + +bool +sched_thread_should_yield(processor_t processor, thread_t thread) +{ + (void)thread; + + return !SCHED(processor_queue_empty)(processor) || rt_runq_count(processor->processor_set) > 0; +} + +/* Default implementations of .steal_thread_enabled */ +bool +sched_steal_thread_DISABLED(processor_set_t pset) +{ + (void)pset; + return false; +} + +bool +sched_steal_thread_enabled(processor_set_t pset) +{ + return pset->node->pset_count > 1; } -#ifndef __LP64__ +#if defined(CONFIG_SCHED_TIMESHARE_CORE) /* - * thread_set_timer: - * - * Set a timer for the current thread, if the thread - * is ready to wait. Must be called between assert_wait() - * and thread_block(). + * Set up values for timeshare + * loading factors. */ -void -thread_set_timer( - uint32_t interval, - uint32_t scale_factor) +static void +load_shift_init(void) { - thread_t thread = current_thread(); - uint64_t deadline; - spl_t s; + int8_t k, *p = sched_load_shifts; + uint32_t i, j; - s = splsched(); - thread_lock(thread); - if ((thread->state & TH_WAIT) != 0) { - clock_interval_to_deadline(interval, scale_factor, &deadline); - if (!timer_call_enter(&thread->wait_timer, deadline)) - thread->wait_timer_active++; - thread->wait_timer_is_set = TRUE; + uint32_t sched_decay_penalty = 1; + + if (PE_parse_boot_argn("sched_decay_penalty", &sched_decay_penalty, sizeof(sched_decay_penalty))) { + kprintf("Overriding scheduler decay penalty %u\n", sched_decay_penalty); + } + + if (PE_parse_boot_argn("sched_decay_usage_age_factor", &sched_decay_usage_age_factor, sizeof(sched_decay_usage_age_factor))) { + kprintf("Overriding scheduler decay usage age factor %u\n", sched_decay_usage_age_factor); + } + + if (sched_decay_penalty == 0) { + /* + * There is no penalty for timeshare threads for using too much + * CPU, so set all load shifts to INT8_MIN. Even under high load, + * sched_pri_shift will be >INT8_MAX, and there will be no + * penalty applied to threads (nor will sched_usage be updated per + * thread). + */ + for (i = 0; i < NRQS; i++) { + sched_load_shifts[i] = INT8_MIN; + } + + return; + } + + *p++ = INT8_MIN; *p++ = 0; + + /* + * For a given system load "i", the per-thread priority + * penalty per quantum of CPU usage is ~2^k priority + * levels. "sched_decay_penalty" can cause more + * array entries to be filled with smaller "k" values + */ + for (i = 2, j = 1 << sched_decay_penalty, k = 1; i < NRQS; ++k) { + for (j <<= 1; (i < j) && (i < NRQS); ++i) { + *p++ = k; + } } - thread_unlock(thread); - splx(s); } -void -thread_set_timer_deadline( - uint64_t deadline) +static void +preempt_pri_init(void) { - thread_t thread = current_thread(); - spl_t s; + bitmap_t *p = sched_preempt_pri; - s = splsched(); - thread_lock(thread); - if ((thread->state & TH_WAIT) != 0) { - if (!timer_call_enter(&thread->wait_timer, deadline)) - thread->wait_timer_active++; - thread->wait_timer_is_set = TRUE; + for (int i = BASEPRI_FOREGROUND; i < MINPRI_KERNEL; ++i) { + bitmap_set(p, i); + } + + for (int i = BASEPRI_PREEMPT; i <= MAXPRI; ++i) { + bitmap_set(p, i); } - thread_unlock(thread); - splx(s); } +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +/* + * Thread wait timer expiration. + */ void -thread_cancel_timer(void) +thread_timer_expire( + void *p0, + __unused void *p1) { - thread_t thread = current_thread(); - spl_t s; + thread_t thread = p0; + spl_t s; + + assert_thread_magic(thread); s = splsched(); thread_lock(thread); - if (thread->wait_timer_is_set) { - if (timer_call_cancel(&thread->wait_timer)) - thread->wait_timer_active--; - thread->wait_timer_is_set = FALSE; + if (--thread->wait_timer_active == 0) { + if (thread->wait_timer_is_set) { + thread->wait_timer_is_set = FALSE; + clear_wait_internal(thread, THREAD_TIMED_OUT); + } } thread_unlock(thread); splx(s); } -#endif /* __LP64__ */ - /* * thread_unblock: * * Unblock thread on wake up. * - * Returns TRUE if the thread is still running. + * Returns TRUE if the thread should now be placed on the runqueue. * * Thread must be locked. + * + * Called at splsched(). */ boolean_t thread_unblock( - thread_t thread, - wait_result_t wresult) + thread_t thread, + wait_result_t wresult) { - boolean_t result = FALSE; + boolean_t ready_for_runq = FALSE; + thread_t cthread = current_thread(); + uint32_t new_run_count; + int old_thread_state; /* * Set wait_result. @@ -433,8 +643,9 @@ thread_unblock( * Cancel pending wait timer. */ if (thread->wait_timer_is_set) { - if (timer_call_cancel(&thread->wait_timer)) + if (timer_call_cancel(&thread->wait_timer)) { thread->wait_timer_active--; + } thread->wait_timer_is_set = FALSE; } @@ -442,56 +653,120 @@ thread_unblock( * Update scheduling state: not waiting, * set running. */ - thread->state &= ~(TH_WAIT|TH_UNINT); + old_thread_state = thread->state; + thread->state = (old_thread_state | TH_RUN) & + ~(TH_WAIT | TH_UNINT | TH_WAIT_REPORT); - if (!(thread->state & TH_RUN)) { - thread->state |= TH_RUN; + if ((old_thread_state & TH_RUN) == 0) { + uint64_t ctime = mach_approximate_time(); + thread->last_made_runnable_time = thread->last_basepri_change_time = ctime; + timer_start(&thread->runnable_timer, ctime); + + ready_for_runq = TRUE; - (*thread->sched_call)(SCHED_CALL_UNBLOCK, thread); + if (old_thread_state & TH_WAIT_REPORT) { + (*thread->sched_call)(SCHED_CALL_UNBLOCK, thread); + } + /* Update the runnable thread count */ + new_run_count = SCHED(run_count_incr)(thread); + } else { /* - * Update run counts. + * Either the thread is idling in place on another processor, + * or it hasn't finished context switching yet. */ - sched_run_incr(); - if (thread->sched_mode & TH_MODE_TIMESHARE) - sched_share_incr(); - } - else { + assert((thread->state & TH_IDLE) == 0); /* - * Signal if idling on another processor. + * The run count is only dropped after the context switch completes + * and the thread is still waiting, so we should not run_incr here */ - if (thread->state & TH_IDLE) { - processor_t processor = thread->last_processor; - - if (processor != current_processor()) - machine_signal_idle(processor); - } - - result = TRUE; + new_run_count = os_atomic_load(&sched_run_buckets[TH_BUCKET_RUN], relaxed); } + /* * Calculate deadline for real-time threads. */ - if (thread->sched_mode & TH_MODE_REALTIME) { - thread->realtime.deadline = mach_absolute_time(); - thread->realtime.deadline += thread->realtime.constraint; + if (thread->sched_mode == TH_MODE_REALTIME) { + uint64_t ctime; + + ctime = mach_absolute_time(); + thread->realtime.deadline = thread->realtime.constraint + ctime; } /* * Clear old quantum, fail-safe computation, etc. */ - thread->current_quantum = 0; + thread->quantum_remaining = 0; thread->computation_metered = 0; thread->reason = AST_NONE; + thread->block_hint = kThreadWaitNone; + + /* Obtain power-relevant interrupt and "platform-idle exit" statistics. + * We also account for "double hop" thread signaling via + * the thread callout infrastructure. + * DRK: consider removing the callout wakeup counters in the future + * they're present for verification at the moment. + */ + boolean_t aticontext, pidle; + ml_get_power_state(&aticontext, &pidle); + + if (__improbable(aticontext && !(thread_get_tag_internal(thread) & THREAD_TAG_CALLOUT))) { + DTRACE_SCHED2(iwakeup, struct thread *, thread, struct proc *, thread->task->bsd_info); + + uint64_t ttd = PROCESSOR_DATA(current_processor(), timer_call_ttd); + + if (ttd) { + if (ttd <= timer_deadline_tracking_bin_1) { + thread->thread_timer_wakeups_bin_1++; + } else if (ttd <= timer_deadline_tracking_bin_2) { + thread->thread_timer_wakeups_bin_2++; + } + } + + ledger_credit_thread(thread, thread->t_ledger, + task_ledgers.interrupt_wakeups, 1); + if (pidle) { + ledger_credit_thread(thread, thread->t_ledger, + task_ledgers.platform_idle_wakeups, 1); + } + } else if (thread_get_tag_internal(cthread) & THREAD_TAG_CALLOUT) { + /* TODO: what about an interrupt that does a wake taken on a callout thread? */ + if (cthread->callout_woken_from_icontext) { + ledger_credit_thread(thread, thread->t_ledger, + task_ledgers.interrupt_wakeups, 1); + thread->thread_callout_interrupt_wakeups++; + + if (cthread->callout_woken_from_platform_idle) { + ledger_credit_thread(thread, thread->t_ledger, + task_ledgers.platform_idle_wakeups, 1); + thread->thread_callout_platform_idle_wakeups++; + } + + cthread->callout_woke_thread = TRUE; + } + } + + if (thread_get_tag_internal(thread) & THREAD_TAG_CALLOUT) { + thread->callout_woken_from_icontext = aticontext; + thread->callout_woken_from_platform_idle = pidle; + thread->callout_woke_thread = FALSE; + } + +#if KPERF + if (ready_for_runq) { + kperf_make_runnable(thread, aticontext); + } +#endif /* KPERF */ - KERNEL_DEBUG_CONSTANT( - MACHDBG_CODE(DBG_MACH_SCHED,MACH_MAKE_RUNNABLE) | DBG_FUNC_NONE, - (uintptr_t)thread_tid(thread), thread->sched_pri, 0, 0, 0); + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_MAKE_RUNNABLE) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), thread->sched_pri, thread->wait_result, + sched_run_buckets[TH_BUCKET_RUN], 0); DTRACE_SCHED2(wakeup, struct thread *, thread, struct proc *, thread->task->bsd_info); - return (result); + return ready_for_runq; } /* @@ -501,27 +776,36 @@ thread_unblock( * Conditions: * thread lock held, IPC locks may be held. * thread must have been pulled from wait queue under same lock hold. - * Returns: + * thread must have been waiting + * Returns: * KERN_SUCCESS - Thread was set running - * KERN_NOT_WAITING - Thread was not waiting + * + * TODO: This should return void */ kern_return_t thread_go( - thread_t thread, - wait_result_t wresult) + thread_t thread, + wait_result_t wresult) { + assert_thread_magic(thread); + assert(thread->at_safe_point == FALSE); assert(thread->wait_event == NO_EVENT64); - assert(thread->wait_queue == WAIT_QUEUE_NULL); + assert(thread->waitq == NULL); - if ((thread->state & (TH_WAIT|TH_TERMINATE)) == TH_WAIT) { - if (!thread_unblock(thread, wresult)) - thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ); + assert(!(thread->state & (TH_TERMINATE | TH_TERMINATE2))); + assert(thread->state & TH_WAIT); - return (KERN_SUCCESS); + + if (thread_unblock(thread, wresult)) { +#if SCHED_TRACE_THREAD_WAKEUPS + backtrace(&thread->thread_wakeup_bt[0], + (sizeof(thread->thread_wakeup_bt) / sizeof(uintptr_t)), NULL); +#endif + thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ); } - return (KERN_NOT_WAITING); + return KERN_SUCCESS; } /* @@ -536,12 +820,17 @@ thread_go( __private_extern__ wait_result_t thread_mark_wait_locked( - thread_t thread, - wait_interrupt_t interruptible) + thread_t thread, + wait_interrupt_t interruptible_orig) { - boolean_t at_safe_point; + boolean_t at_safe_point; + wait_interrupt_t interruptible = interruptible_orig; - assert(thread == current_thread()); + if (thread->state & TH_IDLE) { + panic("Invalid attempt to wait while running the idle thread"); + } + + assert(!(thread->state & (TH_WAIT | TH_IDLE | TH_UNINT | TH_TERMINATE2 | TH_WAIT_REPORT))); /* * The thread may have certain types of interrupts/aborts masked @@ -549,27 +838,52 @@ thread_mark_wait_locked( * are OK, we have to honor mask settings (outer-scoped code may * not be able to handle aborts at the moment). */ - if (interruptible > (thread->options & TH_OPT_INTMASK)) + interruptible &= TH_OPT_INTMASK; + if (interruptible > (thread->options & TH_OPT_INTMASK)) { interruptible = thread->options & TH_OPT_INTMASK; + } at_safe_point = (interruptible == THREAD_ABORTSAFE); - if ( interruptible == THREAD_UNINT || - !(thread->sched_mode & TH_MODE_ABORT) || - (!at_safe_point && - (thread->sched_mode & TH_MODE_ABORTSAFELY))) { - - DTRACE_SCHED(sleep); + if (interruptible == THREAD_UNINT || + !(thread->sched_flags & TH_SFLAG_ABORT) || + (!at_safe_point && + (thread->sched_flags & TH_SFLAG_ABORTSAFELY))) { + if (!(thread->state & TH_TERMINATE)) { + DTRACE_SCHED(sleep); + } - thread->state |= (interruptible) ? TH_WAIT : (TH_WAIT | TH_UNINT); + int state_bits = TH_WAIT; + if (!interruptible) { + state_bits |= TH_UNINT; + } + if (thread->sched_call) { + wait_interrupt_t mask = THREAD_WAIT_NOREPORT_USER; + if (is_kerneltask(thread->task)) { + mask = THREAD_WAIT_NOREPORT_KERNEL; + } + if ((interruptible_orig & mask) == 0) { + state_bits |= TH_WAIT_REPORT; + } + } + thread->state |= state_bits; thread->at_safe_point = at_safe_point; - return (thread->wait_result = THREAD_WAITING); + + /* TODO: pass this through assert_wait instead, have + * assert_wait just take a struct as an argument */ + assert(!thread->block_hint); + thread->block_hint = thread->pending_block_hint; + thread->pending_block_hint = kThreadWaitNone; + + return thread->wait_result = THREAD_WAITING; + } else { + if (thread->sched_flags & TH_SFLAG_ABORTSAFELY) { + thread->sched_flags &= ~TH_SFLAG_ABORTED_MASK; + } } - else - if (thread->sched_mode & TH_MODE_ABORTSAFELY) - thread->sched_mode &= ~TH_MODE_ISABORTED; + thread->pending_block_hint = kThreadWaitNone; - return (thread->wait_result = THREAD_INTERRUPTED); + return thread->wait_result = THREAD_INTERRUPTED; } /* @@ -585,7 +899,7 @@ thread_mark_wait_locked( * Returns: * The old interrupt level for the thread. */ -__private_extern__ +__private_extern__ wait_interrupt_t thread_interrupt_level( wait_interrupt_t new_level) @@ -598,28 +912,6 @@ thread_interrupt_level( return result; } -/* - * Check to see if an assert wait is possible, without actually doing one. - * This is used by debug code in locks and elsewhere to verify that it is - * always OK to block when trying to take a blocking lock (since waiting - * for the actual assert_wait to catch the case may make it hard to detect - * this case. - */ -boolean_t -assert_wait_possible(void) -{ - - thread_t thread; - -#if DEBUG - if(debug_mode) return TRUE; /* Always succeed in debug mode */ -#endif - - thread = current_thread(); - - return (thread == NULL || wait_queue_assert_possible(thread)); -} - /* * assert_wait: * @@ -628,162 +920,239 @@ assert_wait_possible(void) */ wait_result_t assert_wait( - event_t event, - wait_interrupt_t interruptible) + event_t event, + wait_interrupt_t interruptible) { - register wait_queue_t wq; - register int index; + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); + } - assert(event != NO_EVENT); + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE, + VM_KERNEL_UNSLIDE_OR_PERM(event), 0, 0, 0, 0); - index = wait_hash(event); - wq = &wait_queues[index]; - return wait_queue_assert_wait(wq, event, interruptible, 0); + struct waitq *waitq; + waitq = global_eventq(event); + return waitq_assert_wait64(waitq, CAST_EVENT64_T(event), interruptible, TIMEOUT_WAIT_FOREVER); +} + +/* + * assert_wait_queue: + * + * Return the global waitq for the specified event + */ +struct waitq * +assert_wait_queue( + event_t event) +{ + return global_eventq(event); } wait_result_t assert_wait_timeout( - event_t event, - wait_interrupt_t interruptible, - uint32_t interval, - uint32_t scale_factor) + event_t event, + wait_interrupt_t interruptible, + uint32_t interval, + uint32_t scale_factor) { - thread_t thread = current_thread(); - wait_result_t wresult; - wait_queue_t wqueue; - uint64_t deadline; - spl_t s; + thread_t thread = current_thread(); + wait_result_t wresult; + uint64_t deadline; + spl_t s; + + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); + } - assert(event != NO_EVENT); - wqueue = &wait_queues[wait_hash(event)]; + struct waitq *waitq; + waitq = global_eventq(event); s = splsched(); - wait_queue_lock(wqueue); - thread_lock(thread); + waitq_lock(waitq); clock_interval_to_deadline(interval, scale_factor, &deadline); - wresult = wait_queue_assert_wait64_locked(wqueue, CAST_DOWN(event64_t, event), - interruptible, deadline, thread); - thread_unlock(thread); - wait_queue_unlock(wqueue); - splx(s); + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE, + VM_KERNEL_UNSLIDE_OR_PERM(event), interruptible, deadline, 0, 0); + + wresult = waitq_assert_wait64_locked(waitq, CAST_EVENT64_T(event), + interruptible, + TIMEOUT_URGENCY_SYS_NORMAL, + deadline, TIMEOUT_NO_LEEWAY, + thread); - return (wresult); + waitq_unlock(waitq); + splx(s); + return wresult; } wait_result_t -assert_wait_deadline( - event_t event, - wait_interrupt_t interruptible, - uint64_t deadline) +assert_wait_timeout_with_leeway( + event_t event, + wait_interrupt_t interruptible, + wait_timeout_urgency_t urgency, + uint32_t interval, + uint32_t leeway, + uint32_t scale_factor) { - thread_t thread = current_thread(); - wait_result_t wresult; - wait_queue_t wqueue; - spl_t s; + thread_t thread = current_thread(); + wait_result_t wresult; + uint64_t deadline; + uint64_t abstime; + uint64_t slop; + uint64_t now; + spl_t s; + + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); + } + + now = mach_absolute_time(); + clock_interval_to_absolutetime_interval(interval, scale_factor, &abstime); + deadline = now + abstime; - assert(event != NO_EVENT); - wqueue = &wait_queues[wait_hash(event)]; + clock_interval_to_absolutetime_interval(leeway, scale_factor, &slop); + + struct waitq *waitq; + waitq = global_eventq(event); s = splsched(); - wait_queue_lock(wqueue); - thread_lock(thread); + waitq_lock(waitq); - wresult = wait_queue_assert_wait64_locked(wqueue, CAST_DOWN(event64_t,event), - interruptible, deadline, thread); + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE, + VM_KERNEL_UNSLIDE_OR_PERM(event), interruptible, deadline, 0, 0); - thread_unlock(thread); - wait_queue_unlock(wqueue); - splx(s); + wresult = waitq_assert_wait64_locked(waitq, CAST_EVENT64_T(event), + interruptible, + urgency, deadline, slop, + thread); - return (wresult); + waitq_unlock(waitq); + splx(s); + return wresult; } -/* - * thread_sleep_fast_usimple_lock: - * - * Cause the current thread to wait until the specified event - * occurs. The specified simple_lock is unlocked before releasing - * the cpu and re-acquired as part of waking up. - * - * This is the simple lock sleep interface for components that use a - * faster version of simple_lock() than is provided by usimple_lock(). - */ -__private_extern__ wait_result_t -thread_sleep_fast_usimple_lock( - event_t event, - simple_lock_t lock, - wait_interrupt_t interruptible) +wait_result_t +assert_wait_deadline( + event_t event, + wait_interrupt_t interruptible, + uint64_t deadline) { - wait_result_t res; + thread_t thread = current_thread(); + wait_result_t wresult; + spl_t s; - res = assert_wait(event, interruptible); - if (res == THREAD_WAITING) { - simple_unlock(lock); - res = thread_block(THREAD_CONTINUE_NULL); - simple_lock(lock); + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); } - return res; -} + struct waitq *waitq; + waitq = global_eventq(event); + + s = splsched(); + waitq_lock(waitq); + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE, + VM_KERNEL_UNSLIDE_OR_PERM(event), interruptible, deadline, 0, 0); + + wresult = waitq_assert_wait64_locked(waitq, CAST_EVENT64_T(event), + interruptible, + TIMEOUT_URGENCY_SYS_NORMAL, deadline, + TIMEOUT_NO_LEEWAY, thread); + waitq_unlock(waitq); + splx(s); + return wresult; +} -/* - * thread_sleep_usimple_lock: - * - * Cause the current thread to wait until the specified event - * occurs. The specified usimple_lock is unlocked before releasing - * the cpu and re-acquired as part of waking up. - * - * This is the simple lock sleep interface for components where - * simple_lock() is defined in terms of usimple_lock(). - */ wait_result_t -thread_sleep_usimple_lock( - event_t event, - usimple_lock_t lock, - wait_interrupt_t interruptible) +assert_wait_deadline_with_leeway( + event_t event, + wait_interrupt_t interruptible, + wait_timeout_urgency_t urgency, + uint64_t deadline, + uint64_t leeway) { - wait_result_t res; + thread_t thread = current_thread(); + wait_result_t wresult; + spl_t s; - res = assert_wait(event, interruptible); - if (res == THREAD_WAITING) { - usimple_unlock(lock); - res = thread_block(THREAD_CONTINUE_NULL); - usimple_lock(lock); + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); } - return res; + + struct waitq *waitq; + waitq = global_eventq(event); + + s = splsched(); + waitq_lock(waitq); + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE, + VM_KERNEL_UNSLIDE_OR_PERM(event), interruptible, deadline, 0, 0); + + wresult = waitq_assert_wait64_locked(waitq, CAST_EVENT64_T(event), + interruptible, + urgency, deadline, leeway, + thread); + waitq_unlock(waitq); + splx(s); + return wresult; } /* - * thread_sleep_lock_write: + * thread_isoncpu: * - * Cause the current thread to wait until the specified event - * occurs. The specified (write) lock is unlocked before releasing - * the cpu. The (write) lock will be re-acquired before returning. + * Return TRUE if a thread is running on a processor such that an AST + * is needed to pull it out of userspace execution, or if executing in + * the kernel, bring to a context switch boundary that would cause + * thread state to be serialized in the thread PCB. + * + * Thread locked, returns the same way. While locked, fields + * like "state" cannot change. "runq" can change only from set to unset. */ -wait_result_t -thread_sleep_lock_write( - event_t event, - lock_t *lock, - wait_interrupt_t interruptible) +static inline boolean_t +thread_isoncpu(thread_t thread) { - wait_result_t res; + /* Not running or runnable */ + if (!(thread->state & TH_RUN)) { + return FALSE; + } - res = assert_wait(event, interruptible); - if (res == THREAD_WAITING) { - lock_write_done(lock); - res = thread_block(THREAD_CONTINUE_NULL); - lock_write(lock); + /* Waiting on a runqueue, not currently running */ + /* TODO: This is invalid - it can get dequeued without thread lock, but not context switched. */ + if (thread->runq != PROCESSOR_NULL) { + return FALSE; } - return res; + + /* + * Thread does not have a stack yet + * It could be on the stack alloc queue or preparing to be invoked + */ + if (!thread->kernel_stack) { + return FALSE; + } + + /* + * Thread must be running on a processor, or + * about to run, or just did run. In all these + * cases, an AST to the processor is needed + * to guarantee that the thread is kicked out + * of userspace and the processor has + * context switched (and saved register state). + */ + return TRUE; } /* * thread_stop: * * Force a preemption point for a thread and wait - * for it to stop running. Arbitrates access among + * for it to stop running on a CPU. If a stronger + * guarantee is requested, wait until no longer + * runnable. Arbitrates access among * multiple stop requests. (released by unstop) * * The thread must enter a wait state and stop via a @@ -793,10 +1162,12 @@ thread_sleep_lock_write( */ boolean_t thread_stop( - thread_t thread) + thread_t thread, + boolean_t until_not_runnable) { - wait_result_t wresult; - spl_t s = splsched(); + wait_result_t wresult; + spl_t s = splsched(); + boolean_t oncpu; wake_lock(thread); thread_lock(thread); @@ -809,11 +1180,13 @@ thread_stop( wake_unlock(thread); splx(s); - if (wresult == THREAD_WAITING) + if (wresult == THREAD_WAITING) { wresult = thread_block(THREAD_CONTINUE_NULL); + } - if (wresult != THREAD_AWAKENED) - return (FALSE); + if (wresult != THREAD_AWAKENED) { + return FALSE; + } s = splsched(); wake_lock(thread); @@ -822,11 +1195,15 @@ thread_stop( thread->state |= TH_SUSP; - while (thread->state & TH_RUN) { - processor_t processor = thread->last_processor; + while ((oncpu = thread_isoncpu(thread)) || + (until_not_runnable && (thread->state & TH_RUN))) { + processor_t processor; - if (processor != PROCESSOR_NULL && processor->active_thread == thread) + if (oncpu) { + assert(thread->state & TH_RUN); + processor = thread->chosen_processor; cause_ast_check(processor); + } thread->wake_active = TRUE; thread_unlock(thread); @@ -835,12 +1212,13 @@ thread_stop( wake_unlock(thread); splx(s); - if (wresult == THREAD_WAITING) + if (wresult == THREAD_WAITING) { wresult = thread_block(THREAD_CONTINUE_NULL); + } if (wresult != THREAD_AWAKENED) { thread_unstop(thread); - return (FALSE); + return FALSE; } s = splsched(); @@ -852,7 +1230,14 @@ thread_stop( wake_unlock(thread); splx(s); - return (TRUE); + /* + * We return with the thread unlocked. To prevent it from + * transitioning to a runnable state (or from TH_RUN to + * being on the CPU), the caller must ensure the thread + * is stopped via an external means (such as an AST) + */ + + return TRUE; } /* @@ -865,20 +1250,15 @@ thread_stop( */ void thread_unstop( - thread_t thread) + thread_t thread) { - spl_t s = splsched(); + spl_t s = splsched(); wake_lock(thread); thread_lock(thread); - if ((thread->state & (TH_RUN|TH_WAIT|TH_SUSP)) == TH_SUSP) { - thread->state &= ~TH_SUSP; - thread_unblock(thread, THREAD_AWAKENED); + assert((thread->state & (TH_RUN | TH_WAIT | TH_SUSP)) != TH_SUSP); - thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ); - } - else if (thread->state & TH_SUSP) { thread->state &= ~TH_SUSP; @@ -907,19 +1287,31 @@ thread_unstop( */ void thread_wait( - thread_t thread) + thread_t thread, + boolean_t until_not_runnable) { - wait_result_t wresult; - spl_t s = splsched(); + wait_result_t wresult; + boolean_t oncpu; + processor_t processor; + spl_t s = splsched(); wake_lock(thread); thread_lock(thread); - while (thread->state & TH_RUN) { - processor_t processor = thread->last_processor; - - if (processor != PROCESSOR_NULL && processor->active_thread == thread) + /* + * Wait until not running on a CPU. If stronger requirement + * desired, wait until not runnable. Assumption: if thread is + * on CPU, then TH_RUN is set, so we're not waiting in any case + * where the original, pure "TH_RUN" check would have let us + * finish. + */ + while ((oncpu = thread_isoncpu(thread)) || + (until_not_runnable && (thread->state & TH_RUN))) { + if (oncpu) { + assert(thread->state & TH_RUN); + processor = thread->chosen_processor; cause_ast_check(processor); + } thread->wake_active = TRUE; thread_unlock(thread); @@ -928,8 +1320,9 @@ thread_wait( wake_unlock(thread); splx(s); - if (wresult == THREAD_WAITING) + if (wresult == THREAD_WAITING) { thread_block(THREAD_CONTINUE_NULL); + } s = splsched(); wake_lock(thread); @@ -959,40 +1352,44 @@ thread_wait( */ __private_extern__ kern_return_t clear_wait_internal( - thread_t thread, - wait_result_t wresult) + thread_t thread, + wait_result_t wresult) { - wait_queue_t wq = thread->wait_queue; - int i = LockTimeOut; + uint32_t i = LockTimeOutUsec; + struct waitq *waitq = thread->waitq; do { - if (wresult == THREAD_INTERRUPTED && (thread->state & TH_UNINT)) - return (KERN_FAILURE); + if (wresult == THREAD_INTERRUPTED && (thread->state & TH_UNINT)) { + return KERN_FAILURE; + } - if (wq != WAIT_QUEUE_NULL) { - if (wait_queue_lock_try(wq)) { - wait_queue_pull_thread_locked(wq, thread, TRUE); - /* wait queue unlocked, thread still locked */ - } - else { + if (waitq != NULL) { + if (!waitq_pull_thread_locked(waitq, thread)) { thread_unlock(thread); delay(1); - + if (i > 0 && !machine_timeout_suspended()) { + i--; + } thread_lock(thread); - if (wq != thread->wait_queue) - return (KERN_NOT_WAITING); - + if (waitq != thread->waitq) { + return KERN_NOT_WAITING; + } continue; } } - return (thread_go(thread, wresult)); - } while (--i > 0); + /* TODO: Can we instead assert TH_TERMINATE is not set? */ + if ((thread->state & (TH_WAIT | TH_TERMINATE)) == TH_WAIT) { + return thread_go(thread, wresult); + } else { + return KERN_NOT_WAITING; + } + } while (i > 0); panic("clear_wait_internal: deadlock: thread=%p, wq=%p, cpu=%d\n", - thread, wq, cpu_number()); + thread, waitq, cpu_number()); - return (KERN_FAILURE); + return KERN_FAILURE; } @@ -1008,11 +1405,11 @@ clear_wait_internal( */ kern_return_t clear_wait( - thread_t thread, - wait_result_t result) + thread_t thread, + wait_result_t result) { kern_return_t ret; - spl_t s; + spl_t s; s = splsched(); thread_lock(thread); @@ -1032,25 +1429,88 @@ clear_wait( */ kern_return_t thread_wakeup_prim( - event_t event, - boolean_t one_thread, - wait_result_t result) + event_t event, + boolean_t one_thread, + wait_result_t result) +{ + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); + } + + struct waitq *wq = global_eventq(event); + + if (one_thread) { + return waitq_wakeup64_one(wq, CAST_EVENT64_T(event), result, WAITQ_ALL_PRIORITIES); + } else { + return waitq_wakeup64_all(wq, CAST_EVENT64_T(event), result, WAITQ_ALL_PRIORITIES); + } +} + +/* + * Wakeup a specified thread if and only if it's waiting for this event + */ +kern_return_t +thread_wakeup_thread( + event_t event, + thread_t thread) +{ + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); + } + + if (__improbable(thread == THREAD_NULL)) { + panic("%s() called with THREAD_NULL", __func__); + } + + struct waitq *wq = global_eventq(event); + + return waitq_wakeup64_thread(wq, CAST_EVENT64_T(event), thread, THREAD_AWAKENED); +} + +/* + * Wakeup a thread waiting on an event and promote it to a priority. + * + * Requires woken thread to un-promote itself when done. + */ +kern_return_t +thread_wakeup_one_with_pri( + event_t event, + int priority) +{ + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); + } + + struct waitq *wq = global_eventq(event); + + return waitq_wakeup64_one(wq, CAST_EVENT64_T(event), THREAD_AWAKENED, priority); +} + +/* + * Wakeup a thread waiting on an event, + * promote it to a priority, + * and return a reference to the woken thread. + * + * Requires woken thread to un-promote itself when done. + */ +thread_t +thread_wakeup_identify(event_t event, + int priority) { - register wait_queue_t wq; - register int index; + if (__improbable(event == NO_EVENT)) { + panic("%s() called with NO_EVENT", __func__); + } - index = wait_hash(event); - wq = &wait_queues[index]; - if (one_thread) - return (wait_queue_wakeup_one(wq, event, result)); - else - return (wait_queue_wakeup_all(wq, event, result)); + struct waitq *wq = global_eventq(event); + + return waitq_wakeup64_identify(wq, CAST_EVENT64_T(event), THREAD_AWAKENED, priority); } /* * thread_bind: * * Force the current thread to execute on the specified processor. + * Takes effect after the next thread_block(). * * Returns the previous binding. PROCESSOR_NULL means * not bound. @@ -1059,427 +1519,743 @@ thread_wakeup_prim( */ processor_t thread_bind( - processor_t processor) + processor_t processor) { - thread_t self = current_thread(); - processor_t prev; - spl_t s; + thread_t self = current_thread(); + processor_t prev; + spl_t s; s = splsched(); thread_lock(self); - prev = self->bound_processor; - self->bound_processor = processor; + prev = thread_bind_internal(self, processor); thread_unlock(self); splx(s); - return (prev); + return prev; } /* - * thread_select: + * thread_bind_internal: * - * Select a new thread for the current processor to execute. + * If the specified thread is not the current thread, and it is currently + * running on another CPU, a remote AST must be sent to that CPU to cause + * the thread to migrate to its bound processor. Otherwise, the migration + * will occur at the next quantum expiration or blocking point. * - * May select the current thread, which must be locked. + * When the thread is the current thread, and explicit thread_block() should + * be used to force the current processor to context switch away and + * let the thread migrate to the bound processor. + * + * Thread must be locked, and at splsched. */ -static thread_t -thread_select( - thread_t thread, - processor_t processor) + +static processor_t +thread_bind_internal( + thread_t thread, + processor_t processor) { - processor_set_t pset = processor->processor_set; - thread_t new_thread = THREAD_NULL; - boolean_t inactive_state; + processor_t prev; - do { - /* - * Update the priority. - */ - if (thread->sched_stamp != sched_tick) - update_priority(thread); + /* */ + assert(thread->sched_pri < BASEPRI_RTQUEUES); + /* A thread can't be bound if it's sitting on a (potentially incorrect) runqueue */ + assert(thread->runq == PROCESSOR_NULL); - processor->current_pri = thread->sched_pri; + KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_THREAD_BIND), thread_tid(thread), processor ? (uintptr_t)processor->cpu_id : (uintptr_t)-1, 0, 0, 0); - pset_lock(pset); + prev = thread->bound_processor; + thread->bound_processor = processor; + + return prev; +} + +/* + * thread_vm_bind_group_add: + * + * The "VM bind group" is a special mechanism to mark a collection + * of threads from the VM subsystem that, in general, should be scheduled + * with only one CPU of parallelism. To accomplish this, we initially + * bind all the threads to the master processor, which has the effect + * that only one of the threads in the group can execute at once, including + * preempting threads in the group that are a lower priority. Future + * mechanisms may use more dynamic mechanisms to prevent the collection + * of VM threads from using more CPU time than desired. + * + * The current implementation can result in priority inversions where + * compute-bound priority 95 or realtime threads that happen to have + * landed on the master processor prevent the VM threads from running. + * When this situation is detected, we unbind the threads for one + * scheduler tick to allow the scheduler to run the threads an + * additional CPUs, before restoring the binding (assuming high latency + * is no longer a problem). + */ - inactive_state = processor->state != PROCESSOR_SHUTDOWN && machine_cpu_is_inactive(processor->cpu_id); +/* + * The current max is provisioned for: + * vm_compressor_swap_trigger_thread (92) + * 2 x vm_pageout_iothread_internal (92) when vm_restricted_to_single_processor==TRUE + * vm_pageout_continue (92) + * memorystatus_thread (95) + */ +#define MAX_VM_BIND_GROUP_COUNT (5) +decl_simple_lock_data(static, sched_vm_group_list_lock); +static thread_t sched_vm_group_thread_list[MAX_VM_BIND_GROUP_COUNT]; +static int sched_vm_group_thread_count; +static boolean_t sched_vm_group_temporarily_unbound = FALSE; - simple_lock(&rt_lock); +void +thread_vm_bind_group_add(void) +{ + thread_t self = current_thread(); - /* - * Test to see if the current thread should continue - * to run on this processor. Must be runnable, and not - * bound to a different processor, nor be in the wrong - * processor set. - */ - if ( thread->state == TH_RUN && - (thread->sched_pri >= BASEPRI_RTQUEUES || - processor->processor_meta == PROCESSOR_META_NULL || - processor->processor_meta->primary == processor) && - (thread->bound_processor == PROCESSOR_NULL || - thread->bound_processor == processor) && - (thread->affinity_set == AFFINITY_SET_NULL || - thread->affinity_set->aset_pset == pset) ) { - if ( thread->sched_pri >= BASEPRI_RTQUEUES && - first_timeslice(processor) ) { - if (rt_runq.highq >= BASEPRI_RTQUEUES) { - register run_queue_t runq = &rt_runq; - register queue_t q; - - q = runq->queues + runq->highq; - if (((thread_t)q->next)->realtime.deadline < - processor->deadline) { - thread = (thread_t)q->next; - ((queue_entry_t)thread)->next->prev = q; - q->next = ((queue_entry_t)thread)->next; - thread->runq = PROCESSOR_NULL; - runq->count--; runq->urgency--; - assert(runq->urgency >= 0); - if (queue_empty(q)) { - if (runq->highq != IDLEPRI) - clrbit(MAXPRI - runq->highq, runq->bitmap); - runq->highq = MAXPRI - ffsbit(runq->bitmap); - } - } - } + thread_reference_internal(self); + self->options |= TH_OPT_SCHED_VM_GROUP; - simple_unlock(&rt_lock); + simple_lock(&sched_vm_group_list_lock, LCK_GRP_NULL); + assert(sched_vm_group_thread_count < MAX_VM_BIND_GROUP_COUNT); + sched_vm_group_thread_list[sched_vm_group_thread_count++] = self; + simple_unlock(&sched_vm_group_list_lock); - processor->deadline = thread->realtime.deadline; + thread_bind(master_processor); - pset_unlock(pset); + /* Switch to bound processor if not already there */ + thread_block(THREAD_CONTINUE_NULL); +} - return (thread); - } +static void +sched_vm_group_maintenance(void) +{ + uint64_t ctime = mach_absolute_time(); + uint64_t longtime = ctime - sched_tick_interval; + int i; + spl_t s; + boolean_t high_latency_observed = FALSE; + boolean_t runnable_and_not_on_runq_observed = FALSE; + boolean_t bind_target_changed = FALSE; + processor_t bind_target = PROCESSOR_NULL; + + /* Make sure nobody attempts to add new threads while we are enumerating them */ + simple_lock(&sched_vm_group_list_lock, LCK_GRP_NULL); - if (!inactive_state && rt_runq.highq < thread->sched_pri && - (new_thread = choose_thread(processor, thread->sched_pri)) == THREAD_NULL) { + s = splsched(); - simple_unlock(&rt_lock); + for (i = 0; i < sched_vm_group_thread_count; i++) { + thread_t thread = sched_vm_group_thread_list[i]; + assert(thread != THREAD_NULL); + thread_lock(thread); + if ((thread->state & (TH_RUN | TH_WAIT)) == TH_RUN) { + if (thread->runq != PROCESSOR_NULL && thread->last_made_runnable_time < longtime) { + high_latency_observed = TRUE; + } else if (thread->runq == PROCESSOR_NULL) { + /* There are some cases where a thread be transitiong that also fall into this case */ + runnable_and_not_on_runq_observed = TRUE; + } + } + thread_unlock(thread); - /* I am the highest priority runnable (non-idle) thread */ + if (high_latency_observed && runnable_and_not_on_runq_observed) { + /* All the things we are looking for are true, stop looking */ + break; + } + } - pset_pri_hint(pset, processor, processor->current_pri); + splx(s); - pset_count_hint(pset, processor, processor->runq.count); + if (sched_vm_group_temporarily_unbound) { + /* If we turned off binding, make sure everything is OK before rebinding */ + if (!high_latency_observed) { + /* rebind */ + bind_target_changed = TRUE; + bind_target = master_processor; + sched_vm_group_temporarily_unbound = FALSE; /* might be reset to TRUE if change cannot be completed */ + } + } else { + /* + * Check if we're in a bad state, which is defined by high + * latency with no core currently executing a thread. If a + * single thread is making progress on a CPU, that means the + * binding concept to reduce parallelism is working as + * designed. + */ + if (high_latency_observed && !runnable_and_not_on_runq_observed) { + /* unbind */ + bind_target_changed = TRUE; + bind_target = PROCESSOR_NULL; + sched_vm_group_temporarily_unbound = TRUE; + } + } - processor->deadline = UINT64_MAX; + if (bind_target_changed) { + s = splsched(); + for (i = 0; i < sched_vm_group_thread_count; i++) { + thread_t thread = sched_vm_group_thread_list[i]; + boolean_t removed; + assert(thread != THREAD_NULL); - pset_unlock(pset); + thread_lock(thread); + removed = thread_run_queue_remove(thread); + if (removed || ((thread->state & (TH_RUN | TH_WAIT)) == TH_WAIT)) { + thread_bind_internal(thread, bind_target); + } else { + /* + * Thread was in the middle of being context-switched-to, + * or was in the process of blocking. To avoid switching the bind + * state out mid-flight, defer the change if possible. + */ + if (bind_target == PROCESSOR_NULL) { + thread_bind_internal(thread, bind_target); + } else { + sched_vm_group_temporarily_unbound = TRUE; /* next pass will try again */ + } + } - return (thread); + if (removed) { + thread_run_queue_reinsert(thread, SCHED_PREEMPT | SCHED_TAILQ); } + thread_unlock(thread); } + splx(s); + } + + simple_unlock(&sched_vm_group_list_lock); +} - if (new_thread != THREAD_NULL || - (processor->runq.highq >= rt_runq.highq && - (new_thread = choose_thread(processor, MINPRI)) != THREAD_NULL)) { - simple_unlock(&rt_lock); +/* Invoked prior to idle entry to determine if, on SMT capable processors, an SMT + * rebalancing opportunity exists when a core is (instantaneously) idle, but + * other SMT-capable cores may be over-committed. TODO: some possible negatives: + * IPI thrash if this core does not remain idle following the load balancing ASTs + * Idle "thrash", when IPI issue is followed by idle entry/core power down + * followed by a wakeup shortly thereafter. + */ - if (!inactive_state) { - pset_pri_hint(pset, processor, new_thread->sched_pri); +#if (DEVELOPMENT || DEBUG) +int sched_smt_balance = 1; +#endif - pset_count_hint(pset, processor, processor->runq.count); - } +#if __SMP__ +/* Invoked with pset locked, returns with pset unlocked */ +void +sched_SMT_balance(processor_t cprocessor, processor_set_t cpset) +{ + processor_t ast_processor = NULL; - processor->deadline = UINT64_MAX; - pset_unlock(pset); +#if (DEVELOPMENT || DEBUG) + if (__improbable(sched_smt_balance == 0)) { + goto smt_balance_exit; + } +#endif - return (new_thread); - } + assert(cprocessor == current_processor()); + if (cprocessor->is_SMT == FALSE) { + goto smt_balance_exit; + } - if (rt_runq.count > 0) { - thread = run_queue_dequeue(&rt_runq, SCHED_HEADQ); - simple_unlock(&rt_lock); + processor_t sib_processor = cprocessor->processor_secondary ? cprocessor->processor_secondary : cprocessor->processor_primary; - processor->deadline = thread->realtime.deadline; - pset_unlock(pset); + /* Determine if both this processor and its sibling are idle, + * indicating an SMT rebalancing opportunity. + */ + if (sib_processor->state != PROCESSOR_IDLE) { + goto smt_balance_exit; + } - return (thread); + processor_t sprocessor; + + sched_ipi_type_t ipi_type = SCHED_IPI_NONE; + uint64_t running_secondary_map = (cpset->cpu_state_map[PROCESSOR_RUNNING] & + ~cpset->primary_map); + for (int cpuid = lsb_first(running_secondary_map); cpuid >= 0; cpuid = lsb_next(running_secondary_map, cpuid)) { + sprocessor = processor_array[cpuid]; + if ((sprocessor->processor_primary->state == PROCESSOR_RUNNING) && + (sprocessor->current_pri < BASEPRI_RTQUEUES)) { + ipi_type = sched_ipi_action(sprocessor, NULL, false, SCHED_IPI_EVENT_SMT_REBAL); + if (ipi_type != SCHED_IPI_NONE) { + assert(sprocessor != cprocessor); + ast_processor = sprocessor; + break; + } } + } - simple_unlock(&rt_lock); +smt_balance_exit: + pset_unlock(cpset); - processor->deadline = UINT64_MAX; + if (ast_processor) { + KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_SMT_BALANCE), ast_processor->cpu_id, ast_processor->state, ast_processor->processor_primary->state, 0, 0); + sched_ipi_perform(ast_processor, ipi_type); + } +} +#else +/* Invoked with pset locked, returns with pset unlocked */ +void +sched_SMT_balance(__unused processor_t cprocessor, processor_set_t cpset) +{ + pset_unlock(cpset); +} +#endif /* __SMP__ */ + +/* + * Called with pset locked, on a processor that is committing to run a new thread + * Will transition an idle or dispatching processor to running as it picks up + * the first new thread from the idle thread. + */ +static void +pset_commit_processor_to_new_thread(processor_set_t pset, processor_t processor, thread_t new_thread) +{ + if (processor->state == PROCESSOR_DISPATCHING || processor->state == PROCESSOR_IDLE) { + assert(current_thread() == processor->idle_thread); /* - * Set processor inactive based on - * indication from the platform code. + * Dispatching processor is now committed to running new_thread, + * so change its state to PROCESSOR_RUNNING. */ - if (inactive_state) { - if (processor->state == PROCESSOR_RUNNING) - remqueue(&pset->active_queue, (queue_entry_t)processor); - else - if (processor->state == PROCESSOR_IDLE) - remqueue(&pset->idle_queue, (queue_entry_t)processor); + pset_update_processor_state(pset, processor, PROCESSOR_RUNNING); + } else { + assert((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_SHUTDOWN)); + } - processor->state = PROCESSOR_INACTIVE; + processor_state_update_from_thread(processor, new_thread); +} - pset_unlock(pset); +static processor_t choose_processor_for_realtime_thread(processor_set_t pset); +static bool all_available_primaries_are_running_realtime_threads(processor_set_t pset); +static bool these_processors_are_running_realtime_threads(processor_set_t pset, uint64_t these_map); +static bool sched_ok_to_run_realtime_thread(processor_set_t pset, processor_t processor); +int sched_allow_rt_smt = 1; +int sched_avoid_cpu0 = 1; - return (processor->idle_thread); - } +/* + * thread_select: + * + * Select a new thread for the current processor to execute. + * + * May select the current thread, which must be locked. + */ +static thread_t +thread_select(thread_t thread, + processor_t processor, + ast_t *reason) +{ + processor_set_t pset = processor->processor_set; + thread_t new_thread = THREAD_NULL; - /* - * No runnable threads, attempt to steal - * from other processors. - */ - new_thread = steal_thread(pset); - if (new_thread != THREAD_NULL) - return (new_thread); + assert(processor == current_processor()); + assert((thread->state & (TH_RUN | TH_TERMINATE2)) == TH_RUN); + do { /* - * If other threads have appeared, shortcut - * around again. + * Update the priority. */ - if (processor->runq.count > 0 || rt_runq.count > 0) - continue; + if (SCHED(can_update_priority)(thread)) { + SCHED(update_priority)(thread); + } pset_lock(pset); - /* - * Nothing is runnable, so set this processor idle if it - * was running. - */ - if (processor->state == PROCESSOR_RUNNING) { - remqueue(&pset->active_queue, (queue_entry_t)processor); - processor->state = PROCESSOR_IDLE; + processor_state_update_from_thread(processor, thread); + +restart: + /* Acknowledge any pending IPIs here with pset lock held */ + bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id); + bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id); + +#if defined(CONFIG_SCHED_DEFERRED_AST) + bit_clear(pset->pending_deferred_AST_cpu_mask, processor->cpu_id); +#endif - if (processor->processor_meta == PROCESSOR_META_NULL || processor->processor_meta->primary == processor) { - enqueue_head(&pset->idle_queue, (queue_entry_t)processor); - pset->low_pri = pset->low_count = processor; + bool secondary_can_only_run_realtime_thread = false; + + assert(processor->state != PROCESSOR_OFF_LINE); + + if (!processor->is_recommended) { + /* + * The performance controller has provided a hint to not dispatch more threads, + * unless they are bound to us (and thus we are the only option + */ + if (!SCHED(processor_bound_count)(processor)) { + goto idle; } - else { - enqueue_head(&processor->processor_meta->idle_queue, (queue_entry_t)processor); + } else if (processor->processor_primary != processor) { + /* + * Should this secondary SMT processor attempt to find work? For pset runqueue systems, + * we should look for work only under the same conditions that choose_processor() + * would have assigned work, which is when all primary processors have been assigned work. + * + * An exception is that bound threads are dispatched to a processor without going through + * choose_processor(), so in those cases we should continue trying to dequeue work. + */ + if (!SCHED(processor_bound_count)(processor)) { + if ((pset->recommended_bitmask & pset->primary_map & pset->cpu_state_map[PROCESSOR_IDLE]) != 0) { + goto idle; + } - if (thread->sched_pri < BASEPRI_RTQUEUES) { - pset_unlock(pset); + /* + * TODO: What if a secondary core beat an idle primary to waking up from an IPI? + * Should it dequeue immediately, or spin waiting for the primary to wake up? + */ + + /* There are no idle primaries */ - return (processor->idle_thread); + if (processor->processor_primary->current_pri >= BASEPRI_RTQUEUES) { + bool secondary_can_run_realtime_thread = sched_allow_rt_smt && rt_runq_count(pset) && all_available_primaries_are_running_realtime_threads(pset); + if (!secondary_can_run_realtime_thread) { + goto idle; + } + secondary_can_only_run_realtime_thread = true; } } } - pset_unlock(pset); - /* - * Choose idle thread if fast idle is not possible. + * Test to see if the current thread should continue + * to run on this processor. Must not be attempting to wait, and not + * bound to a different processor, nor be in the wrong + * processor set, nor be forced to context switch by TH_SUSP. + * + * Note that there are never any RT threads in the regular runqueue. + * + * This code is very insanely tricky. */ - if ((thread->state & (TH_IDLE|TH_TERMINATE|TH_SUSP)) || !(thread->state & TH_WAIT) || thread->wake_active) - return (processor->idle_thread); + + /* i.e. not waiting, not TH_SUSP'ed */ + bool still_running = ((thread->state & (TH_TERMINATE | TH_IDLE | TH_WAIT | TH_RUN | TH_SUSP)) == TH_RUN); /* - * Perform idling activities directly without a - * context switch. Return dispatched thread, - * else check again for a runnable thread. + * Threads running on SMT processors are forced to context switch. Don't rebalance realtime threads. + * TODO: This should check if it's worth it to rebalance, i.e. 'are there any idle primary processors' + * + * + * A yielding thread shouldn't be forced to context switch. */ - new_thread = thread_select_idle(thread, processor); - } while (new_thread == THREAD_NULL); + bool is_yielding = (*reason & AST_YIELD) == AST_YIELD; - return (new_thread); -} + bool needs_smt_rebalance = !is_yielding && thread->sched_pri < BASEPRI_RTQUEUES && processor->processor_primary != processor; -/* - * thread_select_idle: - * - * Idle the processor using the current thread context. - * - * Called with thread locked, then dropped and relocked. - */ -static thread_t -thread_select_idle( - thread_t thread, - processor_t processor) -{ - thread_t new_thread; + bool affinity_mismatch = thread->affinity_set != AFFINITY_SET_NULL && thread->affinity_set->aset_pset != pset; - if (thread->sched_mode & TH_MODE_TIMESHARE) - sched_share_decr(); - sched_run_decr(); + bool bound_elsewhere = thread->bound_processor != PROCESSOR_NULL && thread->bound_processor != processor; - thread->state |= TH_IDLE; - processor->current_pri = IDLEPRI; + bool avoid_processor = !is_yielding && SCHED(avoid_processor_enabled) && SCHED(thread_avoid_processor)(processor, thread); - thread_unlock(thread); + if (still_running && !needs_smt_rebalance && !affinity_mismatch && !bound_elsewhere && !avoid_processor) { + /* + * This thread is eligible to keep running on this processor. + * + * RT threads with un-expired quantum stay on processor, + * unless there's a valid RT thread with an earlier deadline. + */ + if (thread->sched_pri >= BASEPRI_RTQUEUES && processor->first_timeslice) { + if (rt_runq_count(pset) > 0) { + rt_lock_lock(pset); + + if (rt_runq_count(pset) > 0) { + thread_t next_rt = qe_queue_first(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links); + + if (next_rt->realtime.deadline < processor->deadline && + (next_rt->bound_processor == PROCESSOR_NULL || + next_rt->bound_processor == processor)) { + /* The next RT thread is better, so pick it off the runqueue. */ + goto pick_new_rt_thread; + } + } - /* - * Switch execution timing to processor idle thread. - */ - processor->last_dispatch = mach_absolute_time(); - thread_timer_event(processor->last_dispatch, &processor->idle_thread->system_timer); - PROCESSOR_DATA(processor, kernel_timer) = &processor->idle_thread->system_timer; + rt_lock_unlock(pset); + } - /* - * Cancel the quantum timer while idling. - */ - timer_call_cancel(&processor->quantum_timer); - processor->timeslice = 0; + /* This is still the best RT thread to run. */ + processor->deadline = thread->realtime.deadline; - (*thread->sched_call)(SCHED_CALL_BLOCK, thread); + sched_update_pset_load_average(pset); - /* - * Enable interrupts and perform idling activities. No - * preemption due to TH_IDLE being set. - */ - spllo(); new_thread = processor_idle(thread, processor); + processor_t next_rt_processor = PROCESSOR_NULL; + sched_ipi_type_t next_rt_ipi_type = SCHED_IPI_NONE; - /* - * Return at splsched. - */ - (*thread->sched_call)(SCHED_CALL_UNBLOCK, thread); + if (rt_runq_count(pset) > 0) { + next_rt_processor = choose_processor_for_realtime_thread(pset); + if (next_rt_processor) { + if (next_rt_processor->state == PROCESSOR_IDLE) { + pset_update_processor_state(pset, next_rt_processor, PROCESSOR_DISPATCHING); + } + next_rt_ipi_type = sched_ipi_action(next_rt_processor, NULL, false, SCHED_IPI_EVENT_PREEMPT); + } + } + pset_unlock(pset); - thread_lock(thread); + if (next_rt_processor) { + sched_ipi_perform(next_rt_processor, next_rt_ipi_type); + } - /* - * If awakened, switch to thread timer and start a new quantum. - * Otherwise skip; we will context switch to another thread or return here. - */ - if (!(thread->state & TH_WAIT)) { - processor->last_dispatch = mach_absolute_time(); - thread_timer_event(processor->last_dispatch, &thread->system_timer); - PROCESSOR_DATA(processor, kernel_timer) = &thread->system_timer; + return thread; + } + + if ((rt_runq_count(pset) == 0) && + SCHED(processor_queue_has_priority)(processor, thread->sched_pri, TRUE) == FALSE) { + /* This thread is still the highest priority runnable (non-idle) thread */ + processor->deadline = UINT64_MAX; - thread_quantum_init(thread); + sched_update_pset_load_average(pset); + pset_unlock(pset); - processor->quantum_end = processor->last_dispatch + thread->current_quantum; - timer_call_enter1(&processor->quantum_timer, thread, processor->quantum_end); - processor->timeslice = 1; + return thread; + } + } else { + /* + * This processor must context switch. + * If it's due to a rebalance, we should aggressively find this thread a new home. + */ + if (needs_smt_rebalance || affinity_mismatch || bound_elsewhere || avoid_processor) { + *reason |= AST_REBALANCE; + } + } - thread->computation_epoch = processor->last_dispatch; - } + /* OK, so we're not going to run the current thread. Look at the RT queue. */ + bool ok_to_run_realtime_thread = sched_ok_to_run_realtime_thread(pset, processor); + if ((rt_runq_count(pset) > 0) && ok_to_run_realtime_thread) { + rt_lock_lock(pset); - thread->state &= ~TH_IDLE; + if ((rt_runq_count(pset) > 0) && ok_to_run_realtime_thread) { + thread_t next_rt = qe_queue_first(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links); - sched_run_incr(); - if (thread->sched_mode & TH_MODE_TIMESHARE) - sched_share_incr(); + if (__probable((next_rt->bound_processor == PROCESSOR_NULL || + (next_rt->bound_processor == processor)))) { +pick_new_rt_thread: + new_thread = qe_dequeue_head(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links); - return (new_thread); -} + new_thread->runq = PROCESSOR_NULL; + SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset)); + rt_runq_count_decr(pset); -/* - * choose_thread: - * - * Locate a thread to execute from the processor run queue - * and return it. Only choose a thread with greater or equal - * priority. - * - * Associated pset must be locked. Returns THREAD_NULL - * on failure. - */ -static thread_t -choose_thread( - processor_t processor, - int priority) -{ - run_queue_t rq = &processor->runq; - queue_t queue = rq->queues + rq->highq; - int pri = rq->highq, count = rq->count; - thread_t thread; - - while (count > 0 && pri >= priority) { - thread = (thread_t)queue_first(queue); - while (!queue_end(queue, (queue_entry_t)thread)) { - if (thread->bound_processor == PROCESSOR_NULL || - thread->bound_processor == processor) { - remqueue(queue, (queue_entry_t)thread); - - thread->runq = PROCESSOR_NULL; - rq->count--; - if (testbit(pri, sched_preempt_pri)) { - rq->urgency--; assert(rq->urgency >= 0); + processor->deadline = new_thread->realtime.deadline; + + pset_commit_processor_to_new_thread(pset, processor, new_thread); + + rt_lock_unlock(pset); + sched_update_pset_load_average(pset); + + processor_t ast_processor = PROCESSOR_NULL; + processor_t next_rt_processor = PROCESSOR_NULL; + sched_ipi_type_t ipi_type = SCHED_IPI_NONE; + sched_ipi_type_t next_rt_ipi_type = SCHED_IPI_NONE; + + if (processor->processor_secondary != NULL) { + processor_t sprocessor = processor->processor_secondary; + if ((sprocessor->state == PROCESSOR_RUNNING) || (sprocessor->state == PROCESSOR_DISPATCHING)) { + ipi_type = sched_ipi_action(sprocessor, NULL, false, SCHED_IPI_EVENT_SMT_REBAL); + ast_processor = sprocessor; + } + } + if (rt_runq_count(pset) > 0) { + next_rt_processor = choose_processor_for_realtime_thread(pset); + if (next_rt_processor) { + if (next_rt_processor->state == PROCESSOR_IDLE) { + pset_update_processor_state(pset, next_rt_processor, PROCESSOR_DISPATCHING); + } + next_rt_ipi_type = sched_ipi_action(next_rt_processor, NULL, false, SCHED_IPI_EVENT_PREEMPT); + } + } + pset_unlock(pset); + + if (ast_processor) { + sched_ipi_perform(ast_processor, ipi_type); + } + + if (next_rt_processor) { + sched_ipi_perform(next_rt_processor, next_rt_ipi_type); + } + + return new_thread; + } + } + + rt_lock_unlock(pset); + } + if (secondary_can_only_run_realtime_thread) { + goto idle; + } + + processor->deadline = UINT64_MAX; + + /* No RT threads, so let's look at the regular threads. */ + if ((new_thread = SCHED(choose_thread)(processor, MINPRI, *reason)) != THREAD_NULL) { + sched_update_pset_load_average(pset); + + pset_commit_processor_to_new_thread(pset, processor, new_thread); + + processor_t ast_processor = PROCESSOR_NULL; + sched_ipi_type_t ipi_type = SCHED_IPI_NONE; + + processor_t sprocessor = processor->processor_secondary; + if ((sprocessor != NULL) && (sprocessor->state == PROCESSOR_RUNNING)) { + if (thread_no_smt(new_thread)) { + ipi_type = sched_ipi_action(sprocessor, NULL, false, SCHED_IPI_EVENT_SMT_REBAL); + ast_processor = sprocessor; } - if (queue_empty(queue)) { - if (pri != IDLEPRI) - clrbit(MAXPRI - pri, rq->bitmap); - rq->highq = MAXPRI - ffsbit(rq->bitmap); + } + pset_unlock(pset); + + if (ast_processor) { + sched_ipi_perform(ast_processor, ipi_type); + } + return new_thread; + } + + if (processor->must_idle) { + processor->must_idle = false; + goto idle; + } + +#if __SMP__ + if (SCHED(steal_thread_enabled)(pset)) { + /* + * No runnable threads, attempt to steal + * from other processors. Returns with pset lock dropped. + */ + + if ((new_thread = SCHED(steal_thread)(pset)) != THREAD_NULL) { + /* + * Avoid taking the pset_lock unless it is necessary to change state. + * It's safe to read processor->state here, as only the current processor can change state + * from this point (interrupts are disabled and this processor is committed to run new_thread). + */ + if (processor->state == PROCESSOR_DISPATCHING || processor->state == PROCESSOR_IDLE) { + pset_lock(pset); + pset_commit_processor_to_new_thread(pset, processor, new_thread); + pset_unlock(pset); + } else { + assert((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_SHUTDOWN)); + processor_state_update_from_thread(processor, new_thread); } - return (thread); + return new_thread; } - count--; - thread = (thread_t)queue_next((queue_entry_t)thread); + /* + * If other threads have appeared, shortcut + * around again. + */ + if (!SCHED(processor_queue_empty)(processor) || (ok_to_run_realtime_thread && (rt_runq_count(pset) > 0))) { + continue; + } + + pset_lock(pset); + + /* Someone selected this processor while we had dropped the lock */ + if (bit_test(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id)) { + goto restart; + } } +#endif - queue--; pri--; - } +idle: + /* + * Nothing is runnable, so set this processor idle if it + * was running. + */ + if ((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_DISPATCHING)) { + pset_update_processor_state(pset, processor, PROCESSOR_IDLE); + processor_state_update_idle(processor); + } + +#if __SMP__ + /* Invoked with pset locked, returns with pset unlocked */ + SCHED(processor_balance)(processor, pset); +#else + pset_unlock(pset); +#endif - return (THREAD_NULL); + new_thread = processor->idle_thread; + } while (new_thread == THREAD_NULL); + + return new_thread; } /* - * Perform a context switch and start executing the new thread. + * thread_invoke * - * Returns FALSE on failure, and the thread is re-dispatched. + * Called at splsched with neither thread locked. * - * Called at splsched. + * Perform a context switch and start executing the new thread. + * + * Returns FALSE when the context switch didn't happen. + * The reference to the new thread is still consumed. + * + * "self" is what is currently running on the processor, + * "thread" is the new thread to context switch to + * (which may be the same thread in some cases) */ - -#define funnel_release_check(thread, debug) \ -MACRO_BEGIN \ - if ((thread)->funnel_state & TH_FN_OWNED) { \ - (thread)->funnel_state = TH_FN_REFUNNEL; \ - KERNEL_DEBUG(0x603242c | DBG_FUNC_NONE, \ - (thread)->funnel_lock, (debug), 0, 0, 0); \ - funnel_unlock((thread)->funnel_lock); \ - } \ -MACRO_END - -#define funnel_refunnel_check(thread, debug) \ -MACRO_BEGIN \ - if ((thread)->funnel_state & TH_FN_REFUNNEL) { \ - kern_return_t result = (thread)->wait_result; \ - \ - (thread)->funnel_state = 0; \ - KERNEL_DEBUG(0x6032428 | DBG_FUNC_NONE, \ - (thread)->funnel_lock, (debug), 0, 0, 0); \ - funnel_lock((thread)->funnel_lock); \ - KERNEL_DEBUG(0x6032430 | DBG_FUNC_NONE, \ - (thread)->funnel_lock, (debug), 0, 0, 0); \ - (thread)->funnel_state = TH_FN_OWNED; \ - (thread)->wait_result = result; \ - } \ -MACRO_END - static boolean_t thread_invoke( - register thread_t self, - register thread_t thread, - ast_t reason) + thread_t self, + thread_t thread, + ast_t reason) { - thread_continue_t continuation = self->continuation; - void *parameter = self->parameter; - processor_t processor; - - if (get_preemption_level() != 0) { + if (__improbable(get_preemption_level() != 0)) { int pl = get_preemption_level(); panic("thread_invoke: preemption_level %d, possible cause: %s", pl, (pl < 0 ? "unlocking an unlocked mutex or spinlock" : - "blocking while holding a spinlock, or within interrupt context")); + "blocking while holding a spinlock, or within interrupt context")); } + thread_continue_t continuation = self->continuation; + void *parameter = self->parameter; + processor_t processor; + + uint64_t ctime = mach_absolute_time(); + +#ifdef CONFIG_MACH_APPROXIMATE_TIME + commpage_update_mach_approximate_time(ctime); +#endif + +#if defined(CONFIG_SCHED_TIMESHARE_CORE) + if ((thread->state & TH_IDLE) == 0) { + sched_timeshare_consider_maintenance(ctime); + } +#endif + +#if MONOTONIC + mt_sched_update(self); +#endif /* MONOTONIC */ + + assert_thread_magic(self); assert(self == current_thread()); + assert(self->runq == PROCESSOR_NULL); + assert((self->state & (TH_RUN | TH_TERMINATE2)) == TH_RUN); - /* - * Mark thread interruptible. - */ thread_lock(thread); - thread->state &= ~TH_UNINT; -#if DEBUG - assert(thread_runnable(thread)); -#endif + assert_thread_magic(thread); + assert((thread->state & (TH_RUN | TH_WAIT | TH_UNINT | TH_TERMINATE | TH_TERMINATE2)) == TH_RUN); + assert(thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == current_processor()); + assert(thread->runq == PROCESSOR_NULL); + + /* Reload precise timing global policy to thread-local policy */ + thread->precise_user_kernel_time = use_precise_user_kernel_time(thread); + /* Update SFI class based on other factors */ + thread->sfi_class = sfi_thread_classify(thread); + + /* Update the same_pri_latency for the thread (used by perfcontrol callouts) */ + thread->same_pri_latency = ctime - thread->last_basepri_change_time; /* - * Allow time constraint threads to hang onto - * a stack. + * In case a base_pri update happened between the timestamp and + * taking the thread lock */ - if ((self->sched_mode & TH_MODE_REALTIME) && !self->reserved_stack) + if (ctime <= thread->last_basepri_change_time) { + thread->same_pri_latency = ctime - thread->last_made_runnable_time; + } + + /* Allow realtime threads to hang onto a stack. */ + if ((self->sched_mode == TH_MODE_REALTIME) && !self->reserved_stack) { self->reserved_stack = self->kernel_stack; + } + + /* Prepare for spin debugging */ +#if INTERRUPT_MASKED_DEBUG + ml_spin_debug_clear(thread); +#endif if (continuation != NULL) { if (!thread->kernel_stack) { @@ -1488,73 +2264,144 @@ thread_invoke( * check to see whether we can exchange it with * that of the other thread. */ - if (self->kernel_stack == self->reserved_stack && !thread->reserved_stack) + if (self->kernel_stack == self->reserved_stack && !thread->reserved_stack) { goto need_stack; + } /* * Context switch by performing a stack handoff. + * Requires both threads to be parked in a continuation. */ continuation = thread->continuation; parameter = thread->parameter; processor = current_processor(); processor->active_thread = thread; - processor->current_pri = thread->sched_pri; + processor_state_update_from_thread(processor, thread); + if (thread->last_processor != processor && thread->last_processor != NULL) { - if (thread->last_processor->processor_set != processor->processor_set) + if (thread->last_processor->processor_set != processor->processor_set) { thread->ps_switch++; + } thread->p_switch++; } thread->last_processor = processor; thread->c_switch++; ast_context(thread); + thread_unlock(thread); self->reason = reason; - processor->last_dispatch = mach_absolute_time(); - thread_timer_event(processor->last_dispatch, &thread->system_timer); + processor->last_dispatch = ctime; + self->last_run_time = ctime; + processor_timer_switch_thread(ctime, &thread->system_timer); + timer_update(&thread->runnable_timer, ctime); PROCESSOR_DATA(processor, kernel_timer) = &thread->system_timer; - - KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_STACK_HANDOFF)|DBG_FUNC_NONE, - self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0); + + /* + * Since non-precise user/kernel time doesn't update the state timer + * during privilege transitions, synthesize an event now. + */ + if (!thread->precise_user_kernel_time) { + timer_update(PROCESSOR_DATA(processor, current_state), ctime); + } + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_STACK_HANDOFF) | DBG_FUNC_NONE, + self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0); + + if ((thread->chosen_processor != processor) && (thread->chosen_processor != PROCESSOR_NULL)) { + SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MOVED) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), (uintptr_t)thread->chosen_processor->cpu_id, 0, 0, 0); + } DTRACE_SCHED2(off__cpu, struct thread *, thread, struct proc *, thread->task->bsd_info); - TLOG(1, "thread_invoke: calling machine_stack_handoff\n"); - machine_stack_handoff(self, thread); + SCHED_STATS_CSW(processor, self->reason, self->sched_pri, thread->sched_pri); + +#if KPERF + kperf_off_cpu(self); +#endif /* KPERF */ + + /* + * This is where we actually switch thread identity, + * and address space if required. However, register + * state is not switched - this routine leaves the + * stack and register state active on the current CPU. + */ + TLOG(1, "thread_invoke: calling stack_handoff\n"); + stack_handoff(self, thread); + + /* 'self' is now off core */ + assert(thread == current_thread_volatile()); DTRACE_SCHED(on__cpu); +#if KPERF + kperf_on_cpu(thread, continuation, NULL); +#endif /* KPERF */ + thread_dispatch(self, thread); +#if KASAN + /* Old thread's stack has been moved to the new thread, so explicitly + * unpoison it. */ + kasan_unpoison_stack(thread->kernel_stack, kernel_stack_size); +#endif + thread->continuation = thread->parameter = NULL; counter(c_thread_invoke_hits++); - funnel_refunnel_check(thread, 2); - (void) spllo(); + boolean_t enable_interrupts = TRUE; + + /* idle thread needs to stay interrupts-disabled */ + if ((thread->state & TH_IDLE)) { + enable_interrupts = FALSE; + } assert(continuation); - call_continuation(continuation, parameter, thread->wait_result); + call_continuation(continuation, parameter, + thread->wait_result, enable_interrupts); /*NOTREACHED*/ - } - else if (thread == self) { + } else if (thread == self) { /* same thread but with continuation */ ast_context(self); counter(++c_thread_invoke_same); + thread_unlock(self); +#if KPERF + kperf_on_cpu(thread, continuation, NULL); +#endif /* KPERF */ + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED) | DBG_FUNC_NONE, + self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0); + +#if KASAN + /* stack handoff to self - no thread_dispatch(), so clear the stack + * and free the fakestack directly */ + kasan_fakestack_drop(self); + kasan_fakestack_gc(self); + kasan_unpoison_stack(self->kernel_stack, kernel_stack_size); +#endif + self->continuation = self->parameter = NULL; - funnel_refunnel_check(self, 3); - (void) spllo(); + boolean_t enable_interrupts = TRUE; + + /* idle thread needs to stay interrupts-disabled */ + if ((self->state & TH_IDLE)) { + enable_interrupts = FALSE; + } - call_continuation(continuation, parameter, self->wait_result); + call_continuation(continuation, parameter, + self->wait_result, enable_interrupts); /*NOTREACHED*/ } - } - else { + } else { /* * Check that the other thread has a stack */ @@ -1564,14 +2411,18 @@ need_stack: counter(c_thread_invoke_misses++); thread_unlock(thread); thread_stack_enqueue(thread); - return (FALSE); + return FALSE; } - } - else if (thread == self) { + } else if (thread == self) { ast_context(self); counter(++c_thread_invoke_same); thread_unlock(self); - return (TRUE); + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED) | DBG_FUNC_NONE, + self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0); + + return TRUE; } } @@ -1580,59 +2431,199 @@ need_stack: */ processor = current_processor(); processor->active_thread = thread; - processor->current_pri = thread->sched_pri; + processor_state_update_from_thread(processor, thread); + if (thread->last_processor != processor && thread->last_processor != NULL) { - if (thread->last_processor->processor_set != processor->processor_set) + if (thread->last_processor->processor_set != processor->processor_set) { thread->ps_switch++; + } thread->p_switch++; } thread->last_processor = processor; thread->c_switch++; ast_context(thread); + thread_unlock(thread); counter(c_thread_invoke_csw++); - assert(self->runq == PROCESSOR_NULL); self->reason = reason; - processor->last_dispatch = mach_absolute_time(); - thread_timer_event(processor->last_dispatch, &thread->system_timer); + processor->last_dispatch = ctime; + self->last_run_time = ctime; + processor_timer_switch_thread(ctime, &thread->system_timer); + timer_update(&thread->runnable_timer, ctime); PROCESSOR_DATA(processor, kernel_timer) = &thread->system_timer; - KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED,MACH_SCHED) | DBG_FUNC_NONE, - self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0); + /* + * Since non-precise user/kernel time doesn't update the state timer + * during privilege transitions, synthesize an event now. + */ + if (!thread->precise_user_kernel_time) { + timer_update(PROCESSOR_DATA(processor, current_state), ctime); + } + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED) | DBG_FUNC_NONE, + self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0); + + if ((thread->chosen_processor != processor) && (thread->chosen_processor != NULL)) { + SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MOVED) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), (uintptr_t)thread->chosen_processor->cpu_id, 0, 0, 0); + } DTRACE_SCHED2(off__cpu, struct thread *, thread, struct proc *, thread->task->bsd_info); + SCHED_STATS_CSW(processor, self->reason, self->sched_pri, thread->sched_pri); + +#if KPERF + kperf_off_cpu(self); +#endif /* KPERF */ + /* * This is where we actually switch register context, * and address space if required. We will next run * as a result of a subsequent context switch. + * + * Once registers are switched and the processor is running "thread", + * the stack variables and non-volatile registers will contain whatever + * was there the last time that thread blocked. No local variables should + * be used after this point, except for the special case of "thread", which + * the platform layer returns as the previous thread running on the processor + * via the function call ABI as a return register, and "self", which may have + * been stored on the stack or a non-volatile register, but a stale idea of + * what was on the CPU is newly-accurate because that thread is again + * running on the CPU. + * + * If one of the threads is using a continuation, thread_continue + * is used to stitch up its context. + * + * If we are invoking a thread which is resuming from a continuation, + * the CPU will invoke thread_continue next. + * + * If the current thread is parking in a continuation, then its state + * won't be saved and the stack will be discarded. When the stack is + * re-allocated, it will be configured to resume from thread_continue. */ + assert(continuation == self->continuation); thread = machine_switch_context(self, continuation, thread); - TLOG(1,"thread_invoke: returning machine_switch_context: self %p continuation %p thread %p\n", self, continuation, thread); + assert(self == current_thread_volatile()); + TLOG(1, "thread_invoke: returning machine_switch_context: self %p continuation %p thread %p\n", self, continuation, thread); + + assert(continuation == NULL && self->continuation == NULL); DTRACE_SCHED(on__cpu); - /* - * We have been resumed and are set to run. - */ +#if KPERF + kperf_on_cpu(self, NULL, __builtin_frame_address(0)); +#endif /* KPERF */ + + /* We have been resumed and are set to run. */ thread_dispatch(thread, self); - if (continuation) { - self->continuation = self->parameter = NULL; + return TRUE; +} + +#if defined(CONFIG_SCHED_DEFERRED_AST) +/* + * pset_cancel_deferred_dispatch: + * + * Cancels all ASTs that we can cancel for the given processor set + * if the current processor is running the last runnable thread in the + * system. + * + * This function assumes the current thread is runnable. This must + * be called with the pset unlocked. + */ +static void +pset_cancel_deferred_dispatch( + processor_set_t pset, + processor_t processor) +{ + processor_t active_processor = NULL; + uint32_t sampled_sched_run_count; + + pset_lock(pset); + sampled_sched_run_count = os_atomic_load(&sched_run_buckets[TH_BUCKET_RUN], relaxed); - funnel_refunnel_check(self, 3); - (void) spllo(); + /* + * If we have emptied the run queue, and our current thread is runnable, we + * should tell any processors that are still DISPATCHING that they will + * probably not have any work to do. In the event that there are no + * pending signals that we can cancel, this is also uninteresting. + * + * In the unlikely event that another thread becomes runnable while we are + * doing this (sched_run_count is atomically updated, not guarded), the + * codepath making it runnable SHOULD (a dangerous word) need the pset lock + * in order to dispatch it to a processor in our pset. So, the other + * codepath will wait while we squash all cancelable ASTs, get the pset + * lock, and then dispatch the freshly runnable thread. So this should be + * correct (we won't accidentally have a runnable thread that hasn't been + * dispatched to an idle processor), if not ideal (we may be restarting the + * dispatch process, which could have some overhead). + */ - call_continuation(continuation, parameter, self->wait_result); - /*NOTREACHED*/ + if ((sampled_sched_run_count == 1) && (pset->pending_deferred_AST_cpu_mask)) { + uint64_t dispatching_map = (pset->cpu_state_map[PROCESSOR_DISPATCHING] & + pset->pending_deferred_AST_cpu_mask & + ~pset->pending_AST_URGENT_cpu_mask); + for (int cpuid = lsb_first(dispatching_map); cpuid >= 0; cpuid = lsb_next(dispatching_map, cpuid)) { + active_processor = processor_array[cpuid]; + /* + * If a processor is DISPATCHING, it could be because of + * a cancelable signal. + * + * IF the processor is not our + * current processor (the current processor should not + * be DISPATCHING, so this is a bit paranoid), AND there + * is a cancelable signal pending on the processor, AND + * there is no non-cancelable signal pending (as there is + * no point trying to backtrack on bringing the processor + * up if a signal we cannot cancel is outstanding), THEN + * it should make sense to roll back the processor state + * to the IDLE state. + * + * If the racey nature of this approach (as the signal + * will be arbitrated by hardware, and can fire as we + * roll back state) results in the core responding + * despite being pushed back to the IDLE state, it + * should be no different than if the core took some + * interrupt while IDLE. + */ + if (active_processor != processor) { + /* + * Squash all of the processor state back to some + * reasonable facsimile of PROCESSOR_IDLE. + */ + + processor_state_update_idle(active_processor); + active_processor->deadline = UINT64_MAX; + pset_update_processor_state(pset, active_processor, PROCESSOR_IDLE); + bit_clear(pset->pending_deferred_AST_cpu_mask, active_processor->cpu_id); + machine_signal_idle_cancel(active_processor); + } + } } - return (TRUE); + pset_unlock(pset); +} +#else +/* We don't support deferred ASTs; everything is candycanes and sunshine. */ +#endif + +static void +thread_csw_callout( + thread_t old, + thread_t new, + uint64_t timestamp) +{ + perfcontrol_event event = (new->state & TH_IDLE) ? IDLE : CONTEXT_SWITCH; + uint64_t same_pri_latency = (new->state & TH_IDLE) ? 0 : new->same_pri_latency; + machine_switch_perfcontrol_context(event, timestamp, 0, + same_pri_latency, old, new); } + /* * thread_dispatch: * @@ -1641,143 +2632,360 @@ need_stack: * special actions. Update quantum for other thread and begin * the quantum for ourselves. * + * "thread" is the old thread that we have switched away from. + * "self" is the new current thread that we have context switched to + * * Called at splsched. + * */ void thread_dispatch( - thread_t thread, - thread_t self) + thread_t thread, + thread_t self) { - processor_t processor = self->last_processor; + processor_t processor = self->last_processor; + + assert(processor == current_processor()); + assert(self == current_thread_volatile()); + assert(thread != self); if (thread != THREAD_NULL) { + /* + * Do the perfcontrol callout for context switch. + * The reason we do this here is: + * - thread_dispatch() is called from various places that are not + * the direct context switch path for eg. processor shutdown etc. + * So adding the callout here covers all those cases. + * - We want this callout as early as possible to be close + * to the timestamp taken in thread_invoke() + * - We want to avoid holding the thread lock while doing the + * callout + * - We do not want to callout if "thread" is NULL. + */ + thread_csw_callout(thread, self, processor->last_dispatch); + +#if KASAN + if (thread->continuation != NULL) { + /* + * Thread has a continuation and the normal stack is going away. + * Unpoison the stack and mark all fakestack objects as unused. + */ + kasan_fakestack_drop(thread); + if (thread->kernel_stack) { + kasan_unpoison_stack(thread->kernel_stack, kernel_stack_size); + } + } + + /* + * Free all unused fakestack objects. + */ + kasan_fakestack_gc(thread); +#endif + /* * If blocked at a continuation, discard * the stack. */ - if (thread->continuation != NULL && thread->kernel_stack != 0) + if (thread->continuation != NULL && thread->kernel_stack != 0) { stack_free(thread); + } + + if (thread->state & TH_IDLE) { + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_DISPATCH) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), 0, thread->state, + sched_run_buckets[TH_BUCKET_RUN], 0); + } else { + int64_t consumed; + int64_t remainder = 0; + + if (processor->quantum_end > processor->last_dispatch) { + remainder = processor->quantum_end - + processor->last_dispatch; + } + + consumed = thread->quantum_remaining - remainder; + + if ((thread->reason & AST_LEDGER) == 0) { + /* + * Bill CPU time to both the task and + * the individual thread. + */ + ledger_credit_thread(thread, thread->t_ledger, + task_ledgers.cpu_time, consumed); + ledger_credit_thread(thread, thread->t_threadledger, + thread_ledgers.cpu_time, consumed); + if (thread->t_bankledger) { + ledger_credit_thread(thread, thread->t_bankledger, + bank_ledgers.cpu_time, + (consumed - thread->t_deduct_bank_ledger_time)); + } + thread->t_deduct_bank_ledger_time = 0; + } - if (!(thread->state & TH_IDLE)) { wake_lock(thread); thread_lock(thread); /* - * Compute remainder of current quantum. + * Apply a priority floor if the thread holds a kernel resource + * Do this before checking starting_pri to avoid overpenalizing + * repeated rwlock blockers. + */ + if (__improbable(thread->rwlock_count != 0)) { + lck_rw_set_promotion_locked(thread); + } + + boolean_t keep_quantum = processor->first_timeslice; + + /* + * Treat a thread which has dropped priority since it got on core + * as having expired its quantum. */ - if ( first_timeslice(processor) && - processor->quantum_end > processor->last_dispatch ) - thread->current_quantum = (uint32_t)(processor->quantum_end - processor->last_dispatch); - else - thread->current_quantum = 0; + if (processor->starting_pri > thread->sched_pri) { + keep_quantum = FALSE; + } + + /* Compute remainder of current quantum. */ + if (keep_quantum && + processor->quantum_end > processor->last_dispatch) { + thread->quantum_remaining = (uint32_t)remainder; + } else { + thread->quantum_remaining = 0; + } - if (thread->sched_mode & TH_MODE_REALTIME) { + if (thread->sched_mode == TH_MODE_REALTIME) { /* * Cancel the deadline if the thread has * consumed the entire quantum. */ - if (thread->current_quantum == 0) { + if (thread->quantum_remaining == 0) { thread->realtime.deadline = UINT64_MAX; - thread->reason |= AST_QUANTUM; } - } - else { + } else { +#if defined(CONFIG_SCHED_TIMESHARE_CORE) /* * For non-realtime threads treat a tiny * remaining quantum as an expired quantum * but include what's left next time. */ - if (thread->current_quantum < min_std_quantum) { + if (thread->quantum_remaining < min_std_quantum) { thread->reason |= AST_QUANTUM; - thread->current_quantum += std_quantum; + thread->quantum_remaining += SCHED(initial_quantum_size)(thread); } +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ } /* * If we are doing a direct handoff then * take the remainder of the quantum. */ - if ((thread->reason & (AST_HANDOFF|AST_QUANTUM)) == AST_HANDOFF) { - self->current_quantum = thread->current_quantum; + if ((thread->reason & (AST_HANDOFF | AST_QUANTUM)) == AST_HANDOFF) { + self->quantum_remaining = thread->quantum_remaining; thread->reason |= AST_QUANTUM; - thread->current_quantum = 0; + thread->quantum_remaining = 0; + } else { +#if defined(CONFIG_SCHED_MULTIQ) + if (SCHED(sched_groups_enabled) && + thread->sched_group == self->sched_group) { + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_QUANTUM_HANDOFF), + self->reason, (uintptr_t)thread_tid(thread), + self->quantum_remaining, thread->quantum_remaining, 0); + + self->quantum_remaining = thread->quantum_remaining; + thread->quantum_remaining = 0; + /* Don't set AST_QUANTUM here - old thread might still want to preempt someone else */ + } +#endif /* defined(CONFIG_SCHED_MULTIQ) */ } thread->computation_metered += (processor->last_dispatch - thread->computation_epoch); if (!(thread->state & TH_WAIT)) { /* - * Still running. + * Still runnable. */ - if (thread->reason & AST_QUANTUM) - thread_setrun(thread, SCHED_TAILQ); - else - if (thread->reason & AST_PREEMPT) - thread_setrun(thread, SCHED_HEADQ); - else - thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ); + thread->last_made_runnable_time = thread->last_basepri_change_time = processor->last_dispatch; + + machine_thread_going_off_core(thread, FALSE, processor->last_dispatch, TRUE); + + ast_t reason = thread->reason; + sched_options_t options = SCHED_NONE; + + if (reason & AST_REBALANCE) { + options |= SCHED_REBALANCE; + if (reason & AST_QUANTUM) { + /* + * Having gone to the trouble of forcing this thread off a less preferred core, + * we should force the preferable core to reschedule immediately to give this + * thread a chance to run instead of just sitting on the run queue where + * it may just be stolen back by the idle core we just forced it off. + * But only do this at the end of a quantum to prevent cascading effects. + */ + options |= SCHED_PREEMPT; + } + } - thread->reason = AST_NONE; + if (reason & AST_QUANTUM) { + options |= SCHED_TAILQ; + } else if (reason & AST_PREEMPT) { + options |= SCHED_HEADQ; + } else { + options |= (SCHED_PREEMPT | SCHED_TAILQ); + } + + thread_setrun(thread, options); + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_DISPATCH) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), thread->reason, thread->state, + sched_run_buckets[TH_BUCKET_RUN], 0); + + if (thread->wake_active) { + thread->wake_active = FALSE; + thread_unlock(thread); + + thread_wakeup(&thread->wake_active); + } else { + thread_unlock(thread); + } - thread_unlock(thread); wake_unlock(thread); - } - else { + } else { /* * Waiting. */ - thread->state &= ~TH_RUN; + boolean_t should_terminate = FALSE; + uint32_t new_run_count; + int thread_state = thread->state; + + /* Only the first call to thread_dispatch + * after explicit termination should add + * the thread to the termination queue + */ + if ((thread_state & (TH_TERMINATE | TH_TERMINATE2)) == TH_TERMINATE) { + should_terminate = TRUE; + thread_state |= TH_TERMINATE2; + } + + timer_stop(&thread->runnable_timer, processor->last_dispatch); + + thread_state &= ~TH_RUN; + thread->state = thread_state; - if (thread->sched_mode & TH_MODE_TIMESHARE) - sched_share_decr(); - sched_run_decr(); + thread->last_made_runnable_time = thread->last_basepri_change_time = THREAD_NOT_RUNNABLE; + thread->chosen_processor = PROCESSOR_NULL; + + new_run_count = SCHED(run_count_decr)(thread); + +#if CONFIG_SCHED_SFI + if (thread->reason & AST_SFI) { + thread->wait_sfi_begin_time = processor->last_dispatch; + } +#endif + machine_thread_going_off_core(thread, should_terminate, processor->last_dispatch, FALSE); + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_DISPATCH) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), thread->reason, thread_state, + new_run_count, 0); + + if (thread_state & TH_WAIT_REPORT) { + (*thread->sched_call)(SCHED_CALL_BLOCK, thread); + } if (thread->wake_active) { thread->wake_active = FALSE; thread_unlock(thread); thread_wakeup(&thread->wake_active); - } - else + } else { thread_unlock(thread); + } wake_unlock(thread); - (*thread->sched_call)(SCHED_CALL_BLOCK, thread); - - if (thread->state & TH_TERMINATE) + if (should_terminate) { thread_terminate_enqueue(thread); + } } } } + int urgency = THREAD_URGENCY_NONE; + uint64_t latency = 0; + + /* Update (new) current thread and reprogram quantum timer */ + thread_lock(self); + if (!(self->state & TH_IDLE)) { + uint64_t arg1, arg2; + +#if CONFIG_SCHED_SFI + ast_t new_ast; + + new_ast = sfi_thread_needs_ast(self, NULL); + + if (new_ast != AST_NONE) { + ast_on(new_ast); + } +#endif + + assertf(processor->last_dispatch >= self->last_made_runnable_time, + "Non-monotonic time? dispatch at 0x%llx, runnable at 0x%llx", + processor->last_dispatch, self->last_made_runnable_time); + + assert(self->last_made_runnable_time <= self->last_basepri_change_time); + + latency = processor->last_dispatch - self->last_made_runnable_time; + assert(latency >= self->same_pri_latency); + + urgency = thread_get_urgency(self, &arg1, &arg2); + + thread_tell_urgency(urgency, arg1, arg2, latency, self); + /* * Get a new quantum if none remaining. */ - if (self->current_quantum == 0) + if (self->quantum_remaining == 0) { thread_quantum_init(self); + } /* * Set up quantum timer and timeslice. */ - processor->quantum_end = (processor->last_dispatch + self->current_quantum); - timer_call_enter1(&processor->quantum_timer, self, processor->quantum_end); + processor->quantum_end = processor->last_dispatch + self->quantum_remaining; + timer_call_quantum_timer_enter(&processor->quantum_timer, self, + processor->quantum_end, processor->last_dispatch); - processor->timeslice = 1; + processor->first_timeslice = TRUE; + } else { + timer_call_quantum_timer_cancel(&processor->quantum_timer); + processor->first_timeslice = FALSE; - self->computation_epoch = processor->last_dispatch; + thread_tell_urgency(THREAD_URGENCY_NONE, 0, 0, 0, self); } - else { - timer_call_cancel(&processor->quantum_timer); - processor->timeslice = 0; - } -} -#include + assert(self->block_hint == kThreadWaitNone); + self->computation_epoch = processor->last_dispatch; + self->reason = AST_NONE; + processor->starting_pri = self->sched_pri; -uint32_t kdebug_thread_block = 0; + thread_unlock(self); + + machine_thread_going_on_core(self, urgency, latency, self->same_pri_latency, + processor->last_dispatch); +#if defined(CONFIG_SCHED_DEFERRED_AST) + /* + * TODO: Can we state that redispatching our old thread is also + * uninteresting? + */ + if ((os_atomic_load(&sched_run_buckets[TH_BUCKET_RUN], relaxed) == 1) && !(self->state & TH_IDLE)) { + pset_cancel_deferred_dispatch(processor->processor_set, processor); + } +#endif +} /* * thread_block_reason: @@ -1790,57 +2998,59 @@ uint32_t kdebug_thread_block = 0; * thread resumes, it will execute the continuation function * on a new kernel stack. */ -counter(mach_counter_t c_thread_block_calls = 0;) - +counter(mach_counter_t c_thread_block_calls = 0; ) + wait_result_t thread_block_reason( - thread_continue_t continuation, - void *parameter, - ast_t reason) + thread_continue_t continuation, + void *parameter, + ast_t reason) { - register thread_t self = current_thread(); - register processor_t processor; - register thread_t new_thread; - spl_t s; + thread_t self = current_thread(); + processor_t processor; + thread_t new_thread; + spl_t s; counter(++c_thread_block_calls); s = splsched(); - if (!(reason & AST_PREEMPT)) - funnel_release_check(self, 2); - processor = current_processor(); /* If we're explicitly yielding, force a subsequent quantum */ - if (reason & AST_YIELD) - processor->timeslice = 0; + if (reason & AST_YIELD) { + processor->first_timeslice = FALSE; + } /* We're handling all scheduling AST's */ ast_off(AST_SCHEDULING); +#if PROC_REF_DEBUG + if ((continuation != NULL) && (self->task != kernel_task)) { + if (uthread_get_proc_refcount(self->uthread) != 0) { + panic("thread_block_reason with continuation uthread %p with uu_proc_refcount != 0", self->uthread); + } + } +#endif + self->continuation = continuation; self->parameter = parameter; - if (kdebug_thread_block && kdebug_enable && self->state != TH_RUN) { - uint32_t bt[8]; - - OSBacktrace((void **)&bt[0], 8); - - KERNEL_DEBUG_CONSTANT(0x140004c | DBG_FUNC_START, bt[0], bt[1], bt[2], bt[3], 0); - KERNEL_DEBUG_CONSTANT(0x140004c | DBG_FUNC_END, bt[4], bt[5], bt[6], bt[7], 0); + if (self->state & ~(TH_RUN | TH_IDLE)) { + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_BLOCK), + reason, VM_KERNEL_UNSLIDE(continuation), 0, 0, 0); } do { thread_lock(self); - new_thread = thread_select(self, processor); + new_thread = thread_select(self, processor, &reason); thread_unlock(self); } while (!thread_invoke(self, new_thread, reason)); - funnel_refunnel_check(self, 5); splx(s); - return (self->wait_result); + return self->wait_result; } /* @@ -1850,15 +3060,15 @@ thread_block_reason( */ wait_result_t thread_block( - thread_continue_t continuation) + thread_continue_t continuation) { return thread_block_reason(continuation, NULL, AST_NONE); } wait_result_t thread_block_parameter( - thread_continue_t continuation, - void *parameter) + thread_continue_t continuation, + void *parameter) { return thread_block_reason(continuation, parameter, AST_NONE); } @@ -1875,30 +3085,32 @@ thread_block_parameter( */ int thread_run( - thread_t self, - thread_continue_t continuation, - void *parameter, - thread_t new_thread) + thread_t self, + thread_continue_t continuation, + void *parameter, + thread_t new_thread) { - ast_t handoff = AST_HANDOFF; + ast_t reason = AST_NONE; - funnel_release_check(self, 3); + if ((self->state & TH_IDLE) == 0) { + reason = AST_HANDOFF; + } self->continuation = continuation; self->parameter = parameter; - while (!thread_invoke(self, new_thread, handoff)) { - processor_t processor = current_processor(); + while (!thread_invoke(self, new_thread, reason)) { + /* the handoff failed, so we have to fall back to the normal block path */ + processor_t processor = current_processor(); + + reason = AST_NONE; thread_lock(self); - new_thread = thread_select(self, processor); + new_thread = thread_select(self, processor, &reason); thread_unlock(self); - handoff = AST_NONE; } - funnel_refunnel_check(self, 6); - - return (self->wait_result); + return self->wait_result; } /* @@ -1906,34 +3118,71 @@ thread_run( * * Called at splsched when a thread first receives * a new stack after a continuation. + * + * Called with THREAD_NULL as the old thread when + * invoked by machine_load_context. */ void thread_continue( - register thread_t thread) + thread_t thread) { - register thread_t self = current_thread(); - register thread_continue_t continuation; - register void *parameter; + thread_t self = current_thread(); + thread_continue_t continuation; + void *parameter; DTRACE_SCHED(on__cpu); continuation = self->continuation; parameter = self->parameter; + assert(continuation != NULL); + +#if KPERF + kperf_on_cpu(self, continuation, NULL); +#endif + thread_dispatch(thread, self); self->continuation = self->parameter = NULL; - funnel_refunnel_check(self, 4); +#if INTERRUPT_MASKED_DEBUG + /* Reset interrupt-masked spin debugging timeout */ + ml_spin_debug_clear(self); +#endif + + TLOG(1, "thread_continue: calling call_continuation\n"); + + boolean_t enable_interrupts = TRUE; - if (thread != THREAD_NULL) - (void)spllo(); + /* bootstrap thread, idle thread need to stay interrupts-disabled */ + if (thread == THREAD_NULL || (self->state & TH_IDLE)) { + enable_interrupts = FALSE; + } - TLOG(1, "thread_continue: calling call_continuation \n"); - call_continuation(continuation, parameter, self->wait_result); + call_continuation(continuation, parameter, self->wait_result, enable_interrupts); /*NOTREACHED*/ } +void +thread_quantum_init(thread_t thread) +{ + if (thread->sched_mode == TH_MODE_REALTIME) { + thread->quantum_remaining = thread->realtime.computation; + } else { + thread->quantum_remaining = SCHED(initial_quantum_size)(thread); + } +} + +uint32_t +sched_timeshare_initial_quantum_size(thread_t thread) +{ + if ((thread != THREAD_NULL) && thread->th_sched_bucket == TH_BUCKET_SHARE_BG) { + return bg_quantum; + } else { + return std_quantum; + } +} + /* * run_queue_init: * @@ -1941,17 +3190,16 @@ thread_continue( */ void run_queue_init( - run_queue_t rq) + run_queue_t rq) { - int i; - - rq->highq = IDLEPRI; - for (i = 0; i < NRQBM; i++) + rq->highq = NOPRI; + for (u_int i = 0; i < BITMAP_LEN(NRQS); i++) { rq->bitmap[i] = 0; - setbit(MAXPRI - IDLEPRI, rq->bitmap); + } rq->urgency = rq->count = 0; - for (i = 0; i < NRQS; i++) - queue_init(&rq->queues[i]); + for (int i = 0; i < NRQS; i++) { + circle_queue_init(&rq->queues[i]); + } } /* @@ -1960,453 +3208,972 @@ run_queue_init( * Perform a dequeue operation on a run queue, * and return the resulting thread. * - * The run queue must be locked (see run_queue_remove() + * The run queue must be locked (see thread_run_queue_remove() * for more info), and not empty. */ -static thread_t +thread_t run_queue_dequeue( - run_queue_t rq, - integer_t options) + run_queue_t rq, + sched_options_t options) { - thread_t thread; - queue_t queue = rq->queues + rq->highq; + thread_t thread; + circle_queue_t queue = &rq->queues[rq->highq]; if (options & SCHED_HEADQ) { - thread = (thread_t)queue->next; - ((queue_entry_t)thread)->next->prev = queue; - queue->next = ((queue_entry_t)thread)->next; - } - else { - thread = (thread_t)queue->prev; - ((queue_entry_t)thread)->prev->next = queue; - queue->prev = ((queue_entry_t)thread)->prev; + thread = cqe_dequeue_head(queue, struct thread, runq_links); + } else { + thread = cqe_dequeue_tail(queue, struct thread, runq_links); } + assert(thread != THREAD_NULL); + assert_thread_magic(thread); + thread->runq = PROCESSOR_NULL; + SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count); rq->count--; - if (testbit(rq->highq, sched_preempt_pri)) { + if (SCHED(priority_is_urgent)(rq->highq)) { rq->urgency--; assert(rq->urgency >= 0); } - if (queue_empty(queue)) { - if (rq->highq != IDLEPRI) - clrbit(MAXPRI - rq->highq, rq->bitmap); - rq->highq = MAXPRI - ffsbit(rq->bitmap); + if (circle_queue_empty(queue)) { + bitmap_clear(rq->bitmap, rq->highq); + rq->highq = bitmap_first(rq->bitmap, NRQS); } - return (thread); + return thread; } /* - * realtime_queue_insert: + * run_queue_enqueue: * - * Enqueue a thread for realtime execution. + * Perform a enqueue operation on a run queue. + * + * The run queue must be locked (see thread_run_queue_remove() + * for more info). */ -static boolean_t -realtime_queue_insert( - thread_t thread) +boolean_t +run_queue_enqueue( + run_queue_t rq, + thread_t thread, + sched_options_t options) { - run_queue_t rq = &rt_runq; - queue_t queue = rq->queues + thread->sched_pri; - uint64_t deadline = thread->realtime.deadline; - boolean_t preempt = FALSE; + circle_queue_t queue = &rq->queues[thread->sched_pri]; + boolean_t result = FALSE; - simple_lock(&rt_lock); + assert_thread_magic(thread); - if (queue_empty(queue)) { - enqueue_tail(queue, (queue_entry_t)thread); + if (circle_queue_empty(queue)) { + circle_enqueue_tail(queue, &thread->runq_links); - setbit(MAXPRI - thread->sched_pri, rq->bitmap); - if (thread->sched_pri > rq->highq) + rq_bitmap_set(rq->bitmap, thread->sched_pri); + if (thread->sched_pri > rq->highq) { rq->highq = thread->sched_pri; - preempt = TRUE; + result = TRUE; + } + } else { + if (options & SCHED_TAILQ) { + circle_enqueue_tail(queue, &thread->runq_links); + } else { + circle_enqueue_head(queue, &thread->runq_links); + } + } + if (SCHED(priority_is_urgent)(thread->sched_pri)) { + rq->urgency++; } - else { - register thread_t entry = (thread_t)queue_first(queue); + SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count); + rq->count++; - while (TRUE) { - if ( queue_end(queue, (queue_entry_t)entry) || - deadline < entry->realtime.deadline ) { - entry = (thread_t)queue_prev((queue_entry_t)entry); - break; - } + return result; +} - entry = (thread_t)queue_next((queue_entry_t)entry); - } +/* + * run_queue_remove: + * + * Remove a specific thread from a runqueue. + * + * The run queue must be locked. + */ +void +run_queue_remove( + run_queue_t rq, + thread_t thread) +{ + circle_queue_t queue = &rq->queues[thread->sched_pri]; - if ((queue_entry_t)entry == queue) - preempt = TRUE; + assert(thread->runq != PROCESSOR_NULL); + assert_thread_magic(thread); - insque((queue_entry_t)thread, (queue_entry_t)entry); + circle_dequeue(queue, &thread->runq_links); + SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count); + rq->count--; + if (SCHED(priority_is_urgent)(thread->sched_pri)) { + rq->urgency--; assert(rq->urgency >= 0); } - thread->runq = RT_RUNQ; - rq->count++; rq->urgency++; - - simple_unlock(&rt_lock); + if (circle_queue_empty(queue)) { + /* update run queue status */ + bitmap_clear(rq->bitmap, thread->sched_pri); + rq->highq = bitmap_first(rq->bitmap, NRQS); + } - return (preempt); + thread->runq = PROCESSOR_NULL; } /* - * realtime_setrun: + * run_queue_peek * - * Dispatch a thread for realtime execution. + * Peek at the runq and return the highest + * priority thread from the runq. * - * Thread must be locked. Associated pset must - * be locked, and is returned unlocked. + * The run queue must be locked. */ -static void -realtime_setrun( - processor_t processor, - thread_t thread) +thread_t +run_queue_peek( + run_queue_t rq) +{ + if (rq->count > 0) { + circle_queue_t queue = &rq->queues[rq->highq]; + thread_t thread = cqe_queue_first(queue, struct thread, runq_links); + assert_thread_magic(thread); + return thread; + } else { + return THREAD_NULL; + } +} + +/* Assumes RT lock is not held, and acquires splsched/rt_lock itself */ +void +sched_rtglobal_runq_scan(sched_update_scan_context_t scan_context) { - processor_set_t pset = processor->processor_set; + spl_t s; + thread_t thread; - /* - * Dispatch directly onto idle processor. - */ - if (processor->state == PROCESSOR_IDLE) { - remqueue(&pset->idle_queue, (queue_entry_t)processor); - enqueue_tail(&pset->active_queue, (queue_entry_t)processor); + processor_set_t pset = &pset0; - processor->next_thread = thread; - processor->deadline = thread->realtime.deadline; - processor->state = PROCESSOR_DISPATCHING; - pset_unlock(pset); + s = splsched(); + rt_lock_lock(pset); - if (processor != current_processor()) - machine_signal_idle(processor); - return; + qe_foreach_element_safe(thread, &pset->rt_runq.queue, runq_links) { + if (thread->last_made_runnable_time < scan_context->earliest_rt_make_runnable_time) { + scan_context->earliest_rt_make_runnable_time = thread->last_made_runnable_time; + } } - if (realtime_queue_insert(thread)) { - if (processor == current_processor()) - ast_on(AST_PREEMPT | AST_URGENT); - else - cause_ast_check(processor); - } + rt_lock_unlock(pset); + splx(s); +} - pset_unlock(pset); +int64_t +sched_rtglobal_runq_count_sum(void) +{ + return pset0.rt_runq.runq_stats.count_sum; } /* - * processor_enqueue: - * - * Enqueue thread on a processor run queue. Thread must be locked, - * and not already be on a run queue. - * - * Returns TRUE if a preemption is indicated based on the state - * of the run queue. + * realtime_queue_insert: * - * The run queue must be locked (see run_queue_remove() - * for more info). + * Enqueue a thread for realtime execution. */ static boolean_t -processor_enqueue( - processor_t processor, - thread_t thread, - integer_t options) -{ - run_queue_t rq = &processor->runq; - queue_t queue = rq->queues + thread->sched_pri; - boolean_t result = FALSE; - - if (queue_empty(queue)) { - enqueue_tail(queue, (queue_entry_t)thread); +realtime_queue_insert(processor_t processor, processor_set_t pset, thread_t thread) +{ + queue_t queue = &SCHED(rt_runq)(pset)->queue; + uint64_t deadline = thread->realtime.deadline; + boolean_t preempt = FALSE; - setbit(MAXPRI - thread->sched_pri, rq->bitmap); - if (thread->sched_pri > rq->highq) { - rq->highq = thread->sched_pri; - result = TRUE; + rt_lock_lock(pset); + + if (queue_empty(queue)) { + enqueue_tail(queue, &thread->runq_links); + preempt = TRUE; + } else { + /* Insert into rt_runq in thread deadline order */ + queue_entry_t iter; + qe_foreach(iter, queue) { + thread_t iter_thread = qe_element(iter, struct thread, runq_links); + assert_thread_magic(iter_thread); + + if (deadline < iter_thread->realtime.deadline) { + if (iter == queue_first(queue)) { + preempt = TRUE; + } + insque(&thread->runq_links, queue_prev(iter)); + break; + } else if (iter == queue_last(queue)) { + enqueue_tail(queue, &thread->runq_links); + break; + } } } - else - if (options & SCHED_TAILQ) - enqueue_tail(queue, (queue_entry_t)thread); - else - enqueue_head(queue, (queue_entry_t)thread); thread->runq = processor; - if (testbit(thread->sched_pri, sched_preempt_pri)) - rq->urgency++; - rq->count++; + SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset)); + rt_runq_count_incr(pset); - return (result); + rt_lock_unlock(pset); + + return preempt; } /* - * processor_setrun: + * realtime_setrun: * - * Dispatch a thread for execution on a - * processor. + * Dispatch a thread for realtime execution. * * Thread must be locked. Associated pset must * be locked, and is returned unlocked. */ static void -processor_setrun( - processor_t processor, - thread_t thread, - integer_t options) +realtime_setrun( + processor_t processor, + thread_t thread) { - processor_set_t pset = processor->processor_set; - ast_t preempt; + processor_set_t pset = processor->processor_set; + pset_assert_locked(pset); + ast_t preempt; - /* - * Dispatch directly onto idle processor. - */ - if (processor->state == PROCESSOR_IDLE) { - remqueue(&pset->idle_queue, (queue_entry_t)processor); - enqueue_tail(&pset->active_queue, (queue_entry_t)processor); + sched_ipi_type_t ipi_type = SCHED_IPI_NONE; - processor->next_thread = thread; - processor->deadline = UINT64_MAX; - processor->state = PROCESSOR_DISPATCHING; - pset_unlock(pset); + thread->chosen_processor = processor; - if (processor != current_processor()) - machine_signal_idle(processor); - return; - } + /* */ + assert(thread->bound_processor == PROCESSOR_NULL); - /* - * Set preemption mode. - */ - if (testbit(thread->sched_pri, sched_preempt_pri)) + if (processor->current_pri < BASEPRI_RTQUEUES) { + preempt = (AST_PREEMPT | AST_URGENT); + } else if (thread->realtime.deadline < processor->deadline) { preempt = (AST_PREEMPT | AST_URGENT); - else - if (thread->sched_mode & TH_MODE_TIMESHARE && thread->sched_pri < thread->priority) + } else { preempt = AST_NONE; - else - preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE; + } - if (!processor_enqueue(processor, thread, options)) - preempt = AST_NONE; + realtime_queue_insert(processor, pset, thread); + ipi_type = SCHED_IPI_NONE; if (preempt != AST_NONE) { - if (processor == current_processor()) { - if (csw_check(processor) != AST_NONE) + if (processor->state == PROCESSOR_IDLE) { + processor_state_update_from_thread(processor, thread); + processor->deadline = thread->realtime.deadline; + pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING); + if (processor == current_processor()) { ast_on(preempt); + } else { + ipi_type = sched_ipi_action(processor, thread, true, SCHED_IPI_EVENT_PREEMPT); + } + } else if (processor->state == PROCESSOR_DISPATCHING) { + if ((processor->current_pri < thread->sched_pri) || (processor->deadline > thread->realtime.deadline)) { + processor_state_update_from_thread(processor, thread); + processor->deadline = thread->realtime.deadline; + } + } else { + if (processor == current_processor()) { + ast_on(preempt); + + if ((preempt & AST_URGENT) == AST_URGENT) { + bit_set(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id); + } + + if ((preempt & AST_PREEMPT) == AST_PREEMPT) { + bit_set(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id); + } + } else { + ipi_type = sched_ipi_action(processor, thread, false, SCHED_IPI_EVENT_PREEMPT); + } } - else - if ( (processor->state == PROCESSOR_RUNNING || - processor->state == PROCESSOR_SHUTDOWN) && - thread->sched_pri >= processor->current_pri ) { - cause_ast_check(processor); - } - } - else - if ( processor->state == PROCESSOR_SHUTDOWN && - thread->sched_pri >= processor->current_pri ) { - cause_ast_check(processor); + } else { + /* Selected processor was too busy, just keep thread enqueued and let other processors drain it naturally. */ } pset_unlock(pset); + sched_ipi_perform(processor, ipi_type); } -#define next_pset(p) (((p)->pset_list != PROCESSOR_SET_NULL)? (p)->pset_list: (p)->node->psets) -/* - * choose_next_pset: - * - * Return the next sibling pset containing - * available processors. - * - * Returns the original pset if none other is - * suitable. - */ -static processor_set_t -choose_next_pset( - processor_set_t pset) +sched_ipi_type_t +sched_ipi_deferred_policy(processor_set_t pset, processor_t dst, + __unused sched_ipi_event_t event) { - processor_set_t nset = pset; - - do { - nset = next_pset(nset); - } while (nset->processor_count < 1 && nset != pset); - - return (nset); +#if defined(CONFIG_SCHED_DEFERRED_AST) + if (!bit_test(pset->pending_deferred_AST_cpu_mask, dst->cpu_id)) { + return SCHED_IPI_DEFERRED; + } +#else /* CONFIG_SCHED_DEFERRED_AST */ + panic("Request for deferred IPI on an unsupported platform; pset: %p CPU: %d", pset, dst->cpu_id); +#endif /* CONFIG_SCHED_DEFERRED_AST */ + return SCHED_IPI_NONE; } -/* - * choose_processor: - * - * Choose a processor for the thread, beginning at - * the pset. - * - * Returns a processor, possibly from a different pset. - * - * The thread must be locked. The pset must be locked, - * and the resulting pset is locked on return. - */ -static processor_t -choose_processor( - processor_set_t pset, - thread_t thread) +sched_ipi_type_t +sched_ipi_action(processor_t dst, thread_t thread, boolean_t dst_idle, sched_ipi_event_t event) { - processor_set_t nset, cset = pset; - processor_t processor = thread->last_processor; - processor_meta_t pmeta = PROCESSOR_META_NULL; + sched_ipi_type_t ipi_type = SCHED_IPI_NONE; + assert(dst != NULL); - /* - * Prefer the last processor, when appropriate. - */ - if (processor != PROCESSOR_NULL) { - if (thread->sched_pri < BASEPRI_RTQUEUES && processor->processor_meta != PROCESSOR_META_NULL && - processor->processor_meta->primary->state == PROCESSOR_IDLE) - processor = processor->processor_meta->primary; + processor_set_t pset = dst->processor_set; + if (current_processor() == dst) { + return SCHED_IPI_NONE; + } - if (processor->processor_set != pset || processor->state == PROCESSOR_INACTIVE || - processor->state == PROCESSOR_SHUTDOWN || processor->state == PROCESSOR_OFF_LINE) - processor = PROCESSOR_NULL; - else - if (processor->state == PROCESSOR_IDLE || - (thread->sched_pri > BASEPRI_DEFAULT && processor->current_pri < thread->sched_pri)) - return (processor); + if (bit_test(pset->pending_AST_URGENT_cpu_mask, dst->cpu_id)) { + return SCHED_IPI_NONE; } - /* - * Iterate through the processor sets to locate - * an appropriate processor. - */ - do { + ipi_type = SCHED(ipi_policy)(dst, thread, dst_idle, event); + switch (ipi_type) { + case SCHED_IPI_NONE: + return SCHED_IPI_NONE; +#if defined(CONFIG_SCHED_DEFERRED_AST) + case SCHED_IPI_DEFERRED: + bit_set(pset->pending_deferred_AST_cpu_mask, dst->cpu_id); + break; +#endif /* CONFIG_SCHED_DEFERRED_AST */ + default: + bit_set(pset->pending_AST_URGENT_cpu_mask, dst->cpu_id); + bit_set(pset->pending_AST_PREEMPT_cpu_mask, dst->cpu_id); + break; + } + return ipi_type; +} + +sched_ipi_type_t +sched_ipi_policy(processor_t dst, thread_t thread, boolean_t dst_idle, sched_ipi_event_t event) +{ + sched_ipi_type_t ipi_type = SCHED_IPI_NONE; + boolean_t deferred_ipi_supported = false; + processor_set_t pset = dst->processor_set; + +#if defined(CONFIG_SCHED_DEFERRED_AST) + deferred_ipi_supported = true; +#endif /* CONFIG_SCHED_DEFERRED_AST */ + + switch (event) { + case SCHED_IPI_EVENT_SPILL: + case SCHED_IPI_EVENT_SMT_REBAL: + case SCHED_IPI_EVENT_REBALANCE: + case SCHED_IPI_EVENT_BOUND_THR: /* - * Choose an idle processor. + * The spill, SMT rebalance, rebalance and the bound thread + * scenarios use immediate IPIs always. */ - if (!queue_empty(&cset->idle_queue)) - return ((processor_t)queue_first(&cset->idle_queue)); + ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE; + break; + case SCHED_IPI_EVENT_PREEMPT: + /* In the preemption case, use immediate IPIs for RT threads */ + if (thread && (thread->sched_pri >= BASEPRI_RTQUEUES)) { + ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE; + break; + } - if (thread->sched_pri >= BASEPRI_RTQUEUES) { - /* - * For an RT thread, iterate through active processors, first fit. - */ - processor = (processor_t)queue_first(&cset->active_queue); - while (!queue_end(&cset->active_queue, (queue_entry_t)processor)) { - if (thread->sched_pri > processor->current_pri || - thread->realtime.deadline < processor->deadline) - return (processor); - - if (pmeta == PROCESSOR_META_NULL) { - if (processor->processor_meta != PROCESSOR_META_NULL && - !queue_empty(&processor->processor_meta->idle_queue)) - pmeta = processor->processor_meta; - } + /* + * For Non-RT threads preemption, + * If the core is active, use immediate IPIs. + * If the core is idle, use deferred IPIs if supported; otherwise immediate IPI. + */ + if (deferred_ipi_supported && dst_idle) { + return sched_ipi_deferred_policy(pset, dst, event); + } + ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE; + break; + default: + panic("Unrecognized scheduler IPI event type %d", event); + } + assert(ipi_type != SCHED_IPI_NONE); + return ipi_type; +} - processor = (processor_t)queue_next((queue_entry_t)processor); - } +void +sched_ipi_perform(processor_t dst, sched_ipi_type_t ipi) +{ + switch (ipi) { + case SCHED_IPI_NONE: + break; + case SCHED_IPI_IDLE: + machine_signal_idle(dst); + break; + case SCHED_IPI_IMMEDIATE: + cause_ast_check(dst); + break; + case SCHED_IPI_DEFERRED: + machine_signal_idle_deferred(dst); + break; + default: + panic("Unrecognized scheduler IPI type: %d", ipi); + } +} - if (pmeta != PROCESSOR_META_NULL) - return ((processor_t)queue_first(&pmeta->idle_queue)); +#if defined(CONFIG_SCHED_TIMESHARE_CORE) - processor = PROCESSOR_NULL; - } - else { - /* - * Check any hinted processors in the processor set if available. - */ - if (cset->low_pri != PROCESSOR_NULL && cset->low_pri->state != PROCESSOR_INACTIVE && - cset->low_pri->state != PROCESSOR_SHUTDOWN && cset->low_pri->state != PROCESSOR_OFF_LINE && - (processor == PROCESSOR_NULL || - (thread->sched_pri > BASEPRI_DEFAULT && cset->low_pri->current_pri < thread->sched_pri))) { - processor = cset->low_pri; +boolean_t +priority_is_urgent(int priority) +{ + return bitmap_test(sched_preempt_pri, priority) ? TRUE : FALSE; +} + +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +/* + * processor_setrun: + * + * Dispatch a thread for execution on a + * processor. + * + * Thread must be locked. Associated pset must + * be locked, and is returned unlocked. + */ +static void +processor_setrun( + processor_t processor, + thread_t thread, + integer_t options) +{ + processor_set_t pset = processor->processor_set; + pset_assert_locked(pset); + ast_t preempt; + enum { eExitIdle, eInterruptRunning, eDoNothing } ipi_action = eDoNothing; + + sched_ipi_type_t ipi_type = SCHED_IPI_NONE; + + thread->chosen_processor = processor; + + /* + * Set preemption mode. + */ +#if defined(CONFIG_SCHED_DEFERRED_AST) + /* TODO: Do we need to care about urgency (see rdar://problem/20136239)? */ +#endif + if (SCHED(priority_is_urgent)(thread->sched_pri) && thread->sched_pri > processor->current_pri) { + preempt = (AST_PREEMPT | AST_URGENT); + } else if (processor->active_thread && thread_eager_preemption(processor->active_thread)) { + preempt = (AST_PREEMPT | AST_URGENT); + } else if ((thread->sched_mode == TH_MODE_TIMESHARE) && (thread->sched_pri < thread->base_pri)) { + if (SCHED(priority_is_urgent)(thread->base_pri) && thread->sched_pri > processor->current_pri) { + preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE; + } else { + preempt = AST_NONE; + } + } else { + preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE; + } + + if ((options & (SCHED_PREEMPT | SCHED_REBALANCE)) == (SCHED_PREEMPT | SCHED_REBALANCE)) { + /* + * Having gone to the trouble of forcing this thread off a less preferred core, + * we should force the preferable core to reschedule immediately to give this + * thread a chance to run instead of just sitting on the run queue where + * it may just be stolen back by the idle core we just forced it off. + */ + preempt |= AST_PREEMPT; + } + + SCHED(processor_enqueue)(processor, thread, options); + sched_update_pset_load_average(pset); + + if (preempt != AST_NONE) { + if (processor->state == PROCESSOR_IDLE) { + processor_state_update_from_thread(processor, thread); + processor->deadline = UINT64_MAX; + pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING); + ipi_action = eExitIdle; + } else if (processor->state == PROCESSOR_DISPATCHING) { + if (processor->current_pri < thread->sched_pri) { + processor_state_update_from_thread(processor, thread); + processor->deadline = UINT64_MAX; } - else - if (cset->low_count != PROCESSOR_NULL && cset->low_count->state != PROCESSOR_INACTIVE && - cset->low_count->state != PROCESSOR_SHUTDOWN && cset->low_count->state != PROCESSOR_OFF_LINE && - (processor == PROCESSOR_NULL || (thread->sched_pri <= BASEPRI_DEFAULT && - cset->low_count->runq.count < processor->runq.count))) { - processor = cset->low_count; + } else if ((processor->state == PROCESSOR_RUNNING || + processor->state == PROCESSOR_SHUTDOWN) && + (thread->sched_pri >= processor->current_pri)) { + ipi_action = eInterruptRunning; + } + } else { + /* + * New thread is not important enough to preempt what is running, but + * special processor states may need special handling + */ + if (processor->state == PROCESSOR_SHUTDOWN && + thread->sched_pri >= processor->current_pri) { + ipi_action = eInterruptRunning; + } else if (processor->state == PROCESSOR_IDLE) { + processor_state_update_from_thread(processor, thread); + processor->deadline = UINT64_MAX; + pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING); + + ipi_action = eExitIdle; + } + } + + if (ipi_action != eDoNothing) { + if (processor == current_processor()) { + if ((preempt = csw_check_locked(processor->active_thread, processor, pset, AST_NONE)) != AST_NONE) { + ast_on(preempt); } - /* - * Otherwise, choose an available processor in the set. - */ - if (processor == PROCESSOR_NULL) { - processor = (processor_t)dequeue_head(&cset->active_queue); - if (processor != PROCESSOR_NULL) - enqueue_tail(&cset->active_queue, (queue_entry_t)processor); + if ((preempt & AST_URGENT) == AST_URGENT) { + bit_set(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id); + } else { + bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id); } - if (processor != PROCESSOR_NULL && pmeta == PROCESSOR_META_NULL) { - if (processor->processor_meta != PROCESSOR_META_NULL && - !queue_empty(&processor->processor_meta->idle_queue)) - pmeta = processor->processor_meta; + if ((preempt & AST_PREEMPT) == AST_PREEMPT) { + bit_set(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id); + } else { + bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id); } + } else { + sched_ipi_event_t event = (options & SCHED_REBALANCE) ? SCHED_IPI_EVENT_REBALANCE : SCHED_IPI_EVENT_PREEMPT; + ipi_type = sched_ipi_action(processor, thread, (ipi_action == eExitIdle), event); } + } + pset_unlock(pset); + sched_ipi_perform(processor, ipi_type); +} - /* - * Move onto the next processor set. - */ - nset = next_pset(cset); +/* + * choose_next_pset: + * + * Return the next sibling pset containing + * available processors. + * + * Returns the original pset if none other is + * suitable. + */ +static processor_set_t +choose_next_pset( + processor_set_t pset) +{ + processor_set_t nset = pset; - if (nset != pset) { - pset_unlock(cset); + do { + nset = next_pset(nset); + } while (nset->online_processor_count < 1 && nset != pset); + + return nset; +} + +/* + * choose_processor: + * + * Choose a processor for the thread, beginning at + * the pset. Accepts an optional processor hint in + * the pset. + * + * Returns a processor, possibly from a different pset. + * + * The thread must be locked. The pset must be locked, + * and the resulting pset is locked on return. + */ +processor_t +choose_processor( + processor_set_t starting_pset, + processor_t processor, + thread_t thread) +{ + processor_set_t pset = starting_pset; + processor_set_t nset; + + assert(thread->sched_pri <= BASEPRI_RTQUEUES); + + /* + * Prefer the hinted processor, when appropriate. + */ + + /* Fold last processor hint from secondary processor to its primary */ + if (processor != PROCESSOR_NULL) { + processor = processor->processor_primary; + } - cset = nset; - pset_lock(cset); + /* + * Only consult platform layer if pset is active, which + * it may not be in some cases when a multi-set system + * is going to sleep. + */ + if (pset->online_processor_count) { + if ((processor == PROCESSOR_NULL) || (processor->processor_set == pset && processor->state == PROCESSOR_IDLE)) { + processor_t mc_processor = machine_choose_processor(pset, processor); + if (mc_processor != PROCESSOR_NULL) { + processor = mc_processor->processor_primary; + } + } + } + + /* + * At this point, we may have a processor hint, and we may have + * an initial starting pset. If the hint is not in the pset, or + * if the hint is for a processor in an invalid state, discard + * the hint. + */ + if (processor != PROCESSOR_NULL) { + if (processor->processor_set != pset) { + processor = PROCESSOR_NULL; + } else if (!processor->is_recommended) { + processor = PROCESSOR_NULL; + } else if ((thread->sched_pri >= BASEPRI_RTQUEUES) && !sched_ok_to_run_realtime_thread(pset, processor)) { + processor = PROCESSOR_NULL; + } else { + switch (processor->state) { + case PROCESSOR_START: + case PROCESSOR_SHUTDOWN: + case PROCESSOR_OFF_LINE: + /* + * Hint is for a processor that cannot support running new threads. + */ + processor = PROCESSOR_NULL; + break; + case PROCESSOR_IDLE: + /* + * Hint is for an idle processor. Assume it is no worse than any other + * idle processor. The platform layer had an opportunity to provide + * the "least cost idle" processor above. + */ + return processor; + case PROCESSOR_RUNNING: + case PROCESSOR_DISPATCHING: + /* + * Hint is for an active CPU. This fast-path allows + * realtime threads to preempt non-realtime threads + * to regain their previous executing processor. + */ + if ((thread->sched_pri >= BASEPRI_RTQUEUES) && + (processor->current_pri < BASEPRI_RTQUEUES)) { + return processor; + } + + /* Otherwise, use hint as part of search below */ + break; + default: + processor = PROCESSOR_NULL; + break; + } } - } while (nset != pset); + } /* - * Make sure that we pick a running processor, - * and that the correct processor set is locked. + * Iterate through the processor sets to locate + * an appropriate processor. Seed results with + * a last-processor hint, if available, so that + * a search must find something strictly better + * to replace it. + * + * A primary/secondary pair of SMT processors are + * "unpaired" if the primary is busy but its + * corresponding secondary is idle (so the physical + * core has full use of its resources). */ + + integer_t lowest_priority = MAXPRI + 1; + integer_t lowest_secondary_priority = MAXPRI + 1; + integer_t lowest_unpaired_primary_priority = MAXPRI + 1; + integer_t lowest_idle_secondary_priority = MAXPRI + 1; + integer_t lowest_count = INT_MAX; + uint64_t furthest_deadline = 1; + processor_t lp_processor = PROCESSOR_NULL; + processor_t lp_unpaired_primary_processor = PROCESSOR_NULL; + processor_t lp_idle_secondary_processor = PROCESSOR_NULL; + processor_t lp_paired_secondary_processor = PROCESSOR_NULL; + processor_t lc_processor = PROCESSOR_NULL; + processor_t fd_processor = PROCESSOR_NULL; + + if (processor != PROCESSOR_NULL) { + /* All other states should be enumerated above. */ + assert(processor->state == PROCESSOR_RUNNING || processor->state == PROCESSOR_DISPATCHING); + + lowest_priority = processor->current_pri; + lp_processor = processor; + + if (processor->current_pri >= BASEPRI_RTQUEUES) { + furthest_deadline = processor->deadline; + fd_processor = processor; + } + + lowest_count = SCHED(processor_runq_count)(processor); + lc_processor = processor; + } + do { - if (pmeta != PROCESSOR_META_NULL) { - if (cset != pmeta->primary->processor_set) { - pset_unlock(cset); + int cpuid; - cset = pmeta->primary->processor_set; - pset_lock(cset); + if (thread->sched_pri >= BASEPRI_RTQUEUES) { + processor = choose_processor_for_realtime_thread(pset); + if (processor) { + return processor; } + } else { + /* + * Choose an idle processor, in pset traversal order + */ - if (!queue_empty(&pmeta->idle_queue)) - return ((processor_t)queue_first(&pmeta->idle_queue)); + uint64_t idle_primary_map = (pset->cpu_state_map[PROCESSOR_IDLE] & + pset->primary_map & + pset->recommended_bitmask); - pmeta = PROCESSOR_META_NULL; + /* there shouldn't be a pending AST if the processor is idle */ + assert((idle_primary_map & pset->pending_AST_URGENT_cpu_mask) == 0); + + cpuid = lsb_first(idle_primary_map); + if (cpuid >= 0) { + processor = processor_array[cpuid]; + return processor; + } } /* - * If we haven't been able to choose a processor, - * pick the boot processor and return it. + * Otherwise, enumerate active and idle processors to find primary candidates + * with lower priority/etc. */ - if (processor == PROCESSOR_NULL) { - processor = master_processor; + uint64_t active_map = ((pset->cpu_state_map[PROCESSOR_RUNNING] | pset->cpu_state_map[PROCESSOR_DISPATCHING]) & + pset->recommended_bitmask & + ~pset->pending_AST_URGENT_cpu_mask); + + if (SCHED(priority_is_urgent)(thread->sched_pri) == FALSE) { + active_map &= ~pset->pending_AST_PREEMPT_cpu_mask; + } + + active_map = bit_ror64(active_map, (pset->last_chosen + 1)); + for (int rotid = lsb_first(active_map); rotid >= 0; rotid = lsb_next(active_map, rotid)) { + cpuid = ((rotid + pset->last_chosen + 1) & 63); + processor = processor_array[cpuid]; + + integer_t cpri = processor->current_pri; + processor_t primary = processor->processor_primary; + if (primary != processor) { + /* If primary is running a NO_SMT thread, don't choose its secondary */ + if (!((primary->state == PROCESSOR_RUNNING) && processor_active_thread_no_smt(primary))) { + if (cpri < lowest_secondary_priority) { + lowest_secondary_priority = cpri; + lp_paired_secondary_processor = processor; + } + } + } else { + if (cpri < lowest_priority) { + lowest_priority = cpri; + lp_processor = processor; + } + } + + if ((cpri >= BASEPRI_RTQUEUES) && (processor->deadline > furthest_deadline)) { + furthest_deadline = processor->deadline; + fd_processor = processor; + } + + integer_t ccount = SCHED(processor_runq_count)(processor); + if (ccount < lowest_count) { + lowest_count = ccount; + lc_processor = processor; + } + } + + /* + * For SMT configs, these idle secondary processors must have active primary. Otherwise + * the idle primary would have short-circuited the loop above + */ + uint64_t idle_secondary_map = (pset->cpu_state_map[PROCESSOR_IDLE] & + ~pset->primary_map & + pset->recommended_bitmask); + + /* there shouldn't be a pending AST if the processor is idle */ + assert((idle_secondary_map & pset->pending_AST_URGENT_cpu_mask) == 0); + assert((idle_secondary_map & pset->pending_AST_PREEMPT_cpu_mask) == 0); + + for (cpuid = lsb_first(idle_secondary_map); cpuid >= 0; cpuid = lsb_next(idle_secondary_map, cpuid)) { + processor = processor_array[cpuid]; + + processor_t cprimary = processor->processor_primary; + + integer_t primary_pri = cprimary->current_pri; + + /* + * TODO: This should also make the same decisions + * as secondary_can_run_realtime_thread + * + * TODO: Keep track of the pending preemption priority + * of the primary to make this more accurate. + */ + + /* If the primary is running a no-smt thread, then don't choose its secondary */ + if (cprimary->state == PROCESSOR_RUNNING && + processor_active_thread_no_smt(cprimary)) { + continue; + } + + /* + * Find the idle secondary processor with the lowest priority primary + * + * We will choose this processor as a fallback if we find no better + * primary to preempt. + */ + if (primary_pri < lowest_idle_secondary_priority) { + lp_idle_secondary_processor = processor; + lowest_idle_secondary_priority = primary_pri; + } + + /* Find the the lowest priority active primary with idle secondary */ + if (primary_pri < lowest_unpaired_primary_priority) { + /* If the primary processor is offline or starting up, it's not a candidate for this path */ + if (cprimary->state != PROCESSOR_RUNNING && + cprimary->state != PROCESSOR_DISPATCHING) { + continue; + } + + if (!cprimary->is_recommended) { + continue; + } + + /* if the primary is pending preemption, don't try to re-preempt it */ + if (bit_test(pset->pending_AST_URGENT_cpu_mask, cprimary->cpu_id)) { + continue; + } + + if (SCHED(priority_is_urgent)(thread->sched_pri) == FALSE && + bit_test(pset->pending_AST_PREEMPT_cpu_mask, cprimary->cpu_id)) { + continue; + } + + lowest_unpaired_primary_priority = primary_pri; + lp_unpaired_primary_processor = cprimary; + } + } + + /* + * We prefer preempting a primary processor over waking up its secondary. + * The secondary will then be woken up by the preempted thread. + */ + if (thread->sched_pri > lowest_unpaired_primary_priority) { + pset->last_chosen = lp_unpaired_primary_processor->cpu_id; + return lp_unpaired_primary_processor; + } + + /* + * We prefer preempting a lower priority active processor over directly + * waking up an idle secondary. + * The preempted thread will then find the idle secondary. + */ + if (thread->sched_pri > lowest_priority) { + pset->last_chosen = lp_processor->cpu_id; + return lp_processor; + } + + if (thread->sched_pri >= BASEPRI_RTQUEUES) { /* - * Check that the correct processor set is - * returned locked. + * For realtime threads, the most important aspect is + * scheduling latency, so we will pick an active + * secondary processor in this pset, or preempt + * another RT thread with a further deadline before + * going to the next pset. */ - if (cset != processor->processor_set) { - pset_unlock(cset); - cset = processor->processor_set; - pset_lock(cset); + if (sched_allow_rt_smt && (thread->sched_pri > lowest_secondary_priority)) { + pset->last_chosen = lp_paired_secondary_processor->cpu_id; + return lp_paired_secondary_processor; + } + + if (thread->realtime.deadline < furthest_deadline) { + return fd_processor; } + } - return (processor); + /* + * lc_processor is used to indicate the best processor set run queue + * on which to enqueue a thread when all available CPUs are busy with + * higher priority threads, so try to make sure it is initialized. + */ + if (lc_processor == PROCESSOR_NULL) { + cpumap_t available_map = ((pset->cpu_state_map[PROCESSOR_IDLE] | + pset->cpu_state_map[PROCESSOR_RUNNING] | + pset->cpu_state_map[PROCESSOR_DISPATCHING]) & + pset->recommended_bitmask); + cpuid = lsb_first(available_map); + if (cpuid >= 0) { + lc_processor = processor_array[cpuid]; + lowest_count = SCHED(processor_runq_count)(lc_processor); + } } /* - * Check that the processor set for the chosen - * processor is locked. + * Move onto the next processor set. + * + * If all primary processors in this pset are running a higher + * priority thread, move on to next pset. Only when we have + * exhausted the search for primary processors do we + * fall back to secondaries. */ - if (cset != processor->processor_set) { - pset_unlock(cset); + nset = next_pset(pset); + + if (nset != starting_pset) { + pset_unlock(pset); + + pset = nset; + pset_lock(pset); + } + } while (nset != starting_pset); + + /* + * Make sure that we pick a running processor, + * and that the correct processor set is locked. + * Since we may have unlocked the candidate processor's + * pset, it may have changed state. + * + * All primary processors are running a higher priority + * thread, so the only options left are enqueuing on + * the secondary processor that would perturb the least priority + * primary, or the least busy primary. + */ + boolean_t fallback_processor = false; + do { + /* lowest_priority is evaluated in the main loops above */ + if (lp_idle_secondary_processor != PROCESSOR_NULL) { + processor = lp_idle_secondary_processor; + lp_idle_secondary_processor = PROCESSOR_NULL; + } else if (lp_paired_secondary_processor != PROCESSOR_NULL) { + processor = lp_paired_secondary_processor; + lp_paired_secondary_processor = PROCESSOR_NULL; + } else if (lc_processor != PROCESSOR_NULL) { + processor = lc_processor; + lc_processor = PROCESSOR_NULL; + } else { + /* + * All processors are executing higher priority threads, and + * the lowest_count candidate was not usable. + * + * For AMP platforms running the clutch scheduler always + * return a processor from the requested pset to allow the + * thread to be enqueued in the correct runq. For non-AMP + * platforms, simply return the master_processor. + */ + fallback_processor = true; +#if CONFIG_SCHED_CLUTCH && __AMP__ + processor = processor_array[lsb_first(starting_pset->primary_map)]; +#else /* CONFIG_SCHED_CLUTCH && __AMP__ */ + processor = master_processor; +#endif /* CONFIG_SCHED_CLUTCH && __AMP__ */ + } - cset = processor->processor_set; - pset_lock(cset); + /* + * Check that the correct processor set is + * returned locked. + */ + if (pset != processor->processor_set) { + pset_unlock(pset); + pset = processor->processor_set; + pset_lock(pset); } /* - * We must verify that the chosen processor is still available. + * We must verify that the chosen processor is still available. + * The cases where we pick the master_processor or the fallback + * processor are execptions, since we may need enqueue a thread + * on its runqueue if this is the last remaining processor + * during pset shutdown. + * + * would really help here since it + * gets rid of the weird last processor SHUTDOWN case where + * the pset is still schedulable. */ - if (processor->state == PROCESSOR_INACTIVE || - processor->state == PROCESSOR_SHUTDOWN || processor->state == PROCESSOR_OFF_LINE) + if (processor != master_processor && (fallback_processor == false) && (processor->state == PROCESSOR_SHUTDOWN || processor->state == PROCESSOR_OFF_LINE)) { processor = PROCESSOR_NULL; + } } while (processor == PROCESSOR_NULL); - return (processor); + pset->last_chosen = processor->cpu_id; + return processor; } /* @@ -2420,24 +4187,27 @@ choose_processor( */ void thread_setrun( - thread_t thread, - integer_t options) + thread_t thread, + sched_options_t options) { - processor_t processor; - processor_set_t pset; + processor_t processor; + processor_set_t pset; + + assert((thread->state & (TH_RUN | TH_WAIT | TH_UNINT | TH_TERMINATE | TH_TERMINATE2)) == TH_RUN); + assert(thread->runq == PROCESSOR_NULL); -#if DEBUG - assert(thread_runnable(thread)); -#endif - /* * Update priority if needed. */ - if (thread->sched_stamp != sched_tick) - update_priority(thread); + if (SCHED(can_update_priority)(thread)) { + SCHED(update_priority)(thread); + } + + thread->sfi_class = sfi_thread_classify(thread); assert(thread->runq == PROCESSOR_NULL); +#if __SMP__ if (thread->bound_processor == PROCESSOR_NULL) { /* * Unbound case. @@ -2449,53 +4219,48 @@ thread_setrun( pset = thread->affinity_set->aset_pset; pset_lock(pset); - processor = choose_processor(pset, thread); - } - else - if (thread->last_processor != PROCESSOR_NULL) { + processor = SCHED(choose_processor)(pset, PROCESSOR_NULL, thread); + pset = processor->processor_set; + + SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), (uintptr_t)-1, processor->cpu_id, processor->state, 0); + } else if (thread->last_processor != PROCESSOR_NULL) { /* * Simple (last processor) affinity case. */ processor = thread->last_processor; pset = processor->processor_set; pset_lock(pset); + processor = SCHED(choose_processor)(pset, processor, thread); + pset = processor->processor_set; - /* - * Choose a different processor in certain cases. - */ - if (thread->sched_pri >= BASEPRI_RTQUEUES) { - /* - * If the processor is executing an RT thread with - * an earlier deadline, choose another. - */ - if (thread->sched_pri <= processor->current_pri || - thread->realtime.deadline >= processor->deadline) - processor = choose_processor(pset, thread); - } - else - processor = choose_processor(pset, thread); - } - else { + SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), thread->last_processor->cpu_id, processor->cpu_id, processor->state, 0); + } else { /* * No Affinity case: * * Utilitize a per task hint to spread threads * among the available processor sets. */ - task_t task = thread->task; + task_t task = thread->task; pset = task->pset_hint; - if (pset == PROCESSOR_SET_NULL) + if (pset == PROCESSOR_SET_NULL) { pset = current_processor()->processor_set; + } pset = choose_next_pset(pset); pset_lock(pset); - processor = choose_processor(pset, thread); - task->pset_hint = processor->processor_set; + processor = SCHED(choose_processor)(pset, PROCESSOR_NULL, thread); + pset = processor->processor_set; + task->pset_hint = pset; + + SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), (uintptr_t)-1, processor->cpu_id, processor->state, 0); } - } - else { + } else { /* * Bound case: * @@ -2504,158 +4269,233 @@ thread_setrun( processor = thread->bound_processor; pset = processor->processor_set; pset_lock(pset); + + SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE, + (uintptr_t)thread_tid(thread), (uintptr_t)-2, processor->cpu_id, processor->state, 0); } +#else /* !__SMP__ */ + /* Only one processor to choose */ + assert(thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == master_processor); + processor = master_processor; + pset = processor->processor_set; + pset_lock(pset); +#endif /* !__SMP__ */ /* - * Dispatch the thread on the choosen processor. + * Dispatch the thread on the chosen processor. + * TODO: This should be based on sched_mode, not sched_pri */ - if (thread->sched_pri >= BASEPRI_RTQUEUES) + if (thread->sched_pri >= BASEPRI_RTQUEUES) { realtime_setrun(processor, thread); - else + } else { processor_setrun(processor, thread, options); + } + /* pset is now unlocked */ + if (thread->bound_processor == PROCESSOR_NULL) { + SCHED(check_spill)(pset, thread); + } } processor_set_t task_choose_pset( - task_t task) + task_t task) { - processor_set_t pset = task->pset_hint; + processor_set_t pset = task->pset_hint; - if (pset != PROCESSOR_SET_NULL) + if (pset != PROCESSOR_SET_NULL) { pset = choose_next_pset(pset); + } - return (pset); + return pset; } /* - * processor_queue_shutdown: - * - * Shutdown a processor run queue by - * re-dispatching non-bound threads. + * Check for a preemption point in + * the current context. * - * Associated pset must be locked, and is - * returned unlocked. + * Called at splsched with thread locked. */ -void -processor_queue_shutdown( - processor_t processor) -{ - processor_set_t pset = processor->processor_set; - run_queue_t rq = &processor->runq; - queue_t queue = rq->queues + rq->highq; - int pri = rq->highq, count = rq->count; - thread_t next, thread; - queue_head_t tqueue; - - queue_init(&tqueue); - - while (count > 0) { - thread = (thread_t)queue_first(queue); - while (!queue_end(queue, (queue_entry_t)thread)) { - next = (thread_t)queue_next((queue_entry_t)thread); - - if (thread->bound_processor == PROCESSOR_NULL) { - remqueue(queue, (queue_entry_t)thread); - - thread->runq = PROCESSOR_NULL; - rq->count--; - if (testbit(pri, sched_preempt_pri)) { - rq->urgency--; assert(rq->urgency >= 0); - } - if (queue_empty(queue)) { - if (pri != IDLEPRI) - clrbit(MAXPRI - pri, rq->bitmap); - rq->highq = MAXPRI - ffsbit(rq->bitmap); - } +ast_t +csw_check( + thread_t thread, + processor_t processor, + ast_t check_reason) +{ + processor_set_t pset = processor->processor_set; - enqueue_tail(&tqueue, (queue_entry_t)thread); - } - count--; + assert(thread == processor->active_thread); - thread = next; - } + pset_lock(pset); - queue--; pri--; - } + processor_state_update_from_thread(processor, thread); - pset_unlock(pset); + ast_t preempt = csw_check_locked(thread, processor, pset, check_reason); - while ((thread = (thread_t)dequeue_head(&tqueue)) != THREAD_NULL) { - thread_lock(thread); + /* Acknowledge the IPI if we decided not to preempt */ - thread_setrun(thread, SCHED_TAILQ); + if ((preempt & AST_URGENT) == 0) { + bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id); + } - thread_unlock(thread); + if ((preempt & AST_PREEMPT) == 0) { + bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id); } + + pset_unlock(pset); + + return preempt; } /* - * Check for a preemption point in - * the current context. - * - * Called at splsched. + * Check for preemption at splsched with + * pset and thread locked */ ast_t -csw_check( - processor_t processor) +csw_check_locked( + thread_t thread, + processor_t processor, + processor_set_t pset, + ast_t check_reason) { - ast_t result = AST_NONE; - run_queue_t runq; + ast_t result; + + if (processor->first_timeslice) { + if (rt_runq_count(pset) > 0) { + return check_reason | AST_PREEMPT | AST_URGENT; + } + } else { + if (rt_runq_count(pset) > 0) { + if (BASEPRI_RTQUEUES > processor->current_pri) { + return check_reason | AST_PREEMPT | AST_URGENT; + } else { + return check_reason | AST_PREEMPT; + } + } + } - if (first_timeslice(processor)) { - runq = &rt_runq; - if (runq->highq >= BASEPRI_RTQUEUES) - return (AST_PREEMPT | AST_URGENT); +#if __SMP__ + /* + * If the current thread is running on a processor that is no longer recommended, + * urgently preempt it, at which point thread_select() should + * try to idle the processor and re-dispatch the thread to a recommended processor. + */ + if (!processor->is_recommended) { + return check_reason | AST_PREEMPT | AST_URGENT; + } +#endif - if (runq->highq > processor->current_pri) { - if (runq->urgency > 0) - return (AST_PREEMPT | AST_URGENT); + result = SCHED(processor_csw_check)(processor); + if (result != AST_NONE) { + return check_reason | result | (thread_eager_preemption(thread) ? AST_URGENT : AST_NONE); + } - result |= AST_PREEMPT; - } +#if __SMP__ + /* + * Same for avoid-processor + * + * TODO: Should these set AST_REBALANCE? + */ + if (SCHED(avoid_processor_enabled) && SCHED(thread_avoid_processor)(processor, thread)) { + return check_reason | AST_PREEMPT; + } - runq = &processor->runq; - if (runq->highq > processor->current_pri) { - if (runq->urgency > 0) - return (AST_PREEMPT | AST_URGENT); + /* + * Even though we could continue executing on this processor, a + * secondary SMT core should try to shed load to another primary core. + * + * TODO: Should this do the same check that thread_select does? i.e. + * if no bound threads target this processor, and idle primaries exist, preempt + * The case of RT threads existing is already taken care of above + */ - result |= AST_PREEMPT; - } + if (processor->current_pri < BASEPRI_RTQUEUES && + processor->processor_primary != processor) { + return check_reason | AST_PREEMPT; } - else { - runq = &rt_runq; - if (runq->highq >= processor->current_pri) { - if (runq->urgency > 0) - return (AST_PREEMPT | AST_URGENT); +#endif - result |= AST_PREEMPT; - } + if (thread->state & TH_SUSP) { + return check_reason | AST_PREEMPT; + } + +#if CONFIG_SCHED_SFI + /* + * Current thread may not need to be preempted, but maybe needs + * an SFI wait? + */ + result = sfi_thread_needs_ast(thread, NULL); + if (result != AST_NONE) { + return check_reason | result; + } +#endif - runq = &processor->runq; - if (runq->highq >= processor->current_pri) { - if (runq->urgency > 0) - return (AST_PREEMPT | AST_URGENT); + return AST_NONE; +} - result |= AST_PREEMPT; - } +/* + * Handle preemption IPI or IPI in response to setting an AST flag + * Triggered by cause_ast_check + * Called at splsched + */ +void +ast_check(processor_t processor) +{ + if (processor->state != PROCESSOR_RUNNING && + processor->state != PROCESSOR_SHUTDOWN) { + return; } - if (result != AST_NONE) - return (result); + thread_t thread = processor->active_thread; + + assert(thread == current_thread()); + + thread_lock(thread); + + /* + * Propagate thread ast to processor. + * (handles IPI in response to setting AST flag) + */ + ast_propagate(thread); + + /* + * Stash the old urgency and perfctl values to find out if + * csw_check updates them. + */ + thread_urgency_t old_urgency = processor->current_urgency; + perfcontrol_class_t old_perfctl_class = processor->current_perfctl_class; + + ast_t preempt; + + if ((preempt = csw_check(thread, processor, AST_NONE)) != AST_NONE) { + ast_on(preempt); + } + + if (old_urgency != processor->current_urgency) { + /* + * Urgency updates happen with the thread lock held (ugh). + * TODO: This doesn't notice QoS changes... + */ + uint64_t urgency_param1, urgency_param2; - if (processor->current_pri < BASEPRI_RTQUEUES && processor->processor_meta != PROCESSOR_META_NULL && - processor->processor_meta->primary != processor) - return (AST_PREEMPT); + thread_urgency_t urgency = thread_get_urgency(thread, &urgency_param1, &urgency_param2); + thread_tell_urgency(urgency, urgency_param1, urgency_param2, 0, thread); + } - if (machine_cpu_is_inactive(processor->cpu_id)) - return (AST_PREEMPT); + thread_unlock(thread); - if (processor->active_thread->state & TH_SUSP) - return (AST_PREEMPT); + if (old_perfctl_class != processor->current_perfctl_class) { + /* + * We updated the perfctl class of this thread from another core. + * Let CLPC know that the currently running thread has a new + * class. + */ - return (AST_NONE); + machine_switch_perfcontrol_state_update(PERFCONTROL_ATTR_UPDATE, + mach_approximate_time(), 0, thread); + } } + /* * set_sched_pri: * @@ -2667,233 +4507,342 @@ csw_check( */ void set_sched_pri( - thread_t thread, - int priority) + thread_t thread, + int new_priority, + set_sched_pri_options_t options) { - boolean_t removed = run_queue_remove(thread); + bool is_current_thread = (thread == current_thread()); + bool removed_from_runq = false; + bool lazy_update = ((options & SETPRI_LAZY) == SETPRI_LAZY); + + int old_priority = thread->sched_pri; + + /* If we're already at this priority, no need to mess with the runqueue */ + if (new_priority == old_priority) { +#if CONFIG_SCHED_CLUTCH + /* For the first thread in the system, the priority is correct but + * th_sched_bucket is still TH_BUCKET_RUN. Since the clutch + * scheduler relies on the bucket being set for all threads, update + * its bucket here. + */ + if (thread->th_sched_bucket == TH_BUCKET_RUN) { + assert(is_current_thread); + SCHED(update_thread_bucket)(thread); + } +#endif /* CONFIG_SCHED_CLUTCH */ - thread->sched_pri = priority; - if (removed) - thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ); - else - if (thread->state & TH_RUN) { - processor_t processor = thread->last_processor; + return; + } + + if (is_current_thread) { + assert(thread->state & TH_RUN); + assert(thread->runq == PROCESSOR_NULL); + } else { + removed_from_runq = thread_run_queue_remove(thread); + } + + thread->sched_pri = new_priority; + +#if CONFIG_SCHED_CLUTCH + /* + * Since for the clutch scheduler, the thread's bucket determines its runq + * in the hierarchy it is important to update the bucket when the thread + * lock is held and the thread has been removed from the runq hierarchy. + */ + SCHED(update_thread_bucket)(thread); + +#endif /* CONFIG_SCHED_CLUTCH */ + + KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY), + (uintptr_t)thread_tid(thread), + thread->base_pri, + thread->sched_pri, + thread->sched_usage, + 0); + + if (removed_from_runq) { + thread_run_queue_reinsert(thread, SCHED_PREEMPT | SCHED_TAILQ); + } else if (is_current_thread) { + processor_t processor = thread->last_processor; + assert(processor == current_processor()); - if (thread == current_thread()) { - ast_t preempt; + thread_urgency_t old_urgency = processor->current_urgency; + + /* + * When dropping in priority, check if the thread no longer belongs on core. + * If a thread raises its own priority, don't aggressively rebalance it. + * + * + * csw_check does a processor_state_update_from_thread, but + * we should do our own if we're being lazy. + */ + if (!lazy_update && new_priority < old_priority) { + ast_t preempt; - processor->current_pri = priority; - if ((preempt = csw_check(processor)) != AST_NONE) + if ((preempt = csw_check(thread, processor, AST_NONE)) != AST_NONE) { ast_on(preempt); + } + } else { + processor_state_update_from_thread(processor, thread); + } + + /* + * set_sched_pri doesn't alter RT params. We expect direct base priority/QoS + * class alterations from user space to occur relatively infrequently, hence + * those are lazily handled. QoS classes have distinct priority bands, and QoS + * inheritance is expected to involve priority changes. + */ + if (processor->current_urgency != old_urgency) { + uint64_t urgency_param1, urgency_param2; + + thread_urgency_t new_urgency = thread_get_urgency(thread, + &urgency_param1, &urgency_param2); + + thread_tell_urgency(new_urgency, urgency_param1, + urgency_param2, 0, thread); } - else - if ( processor != PROCESSOR_NULL && - processor->active_thread == thread ) + + /* TODO: only call this if current_perfctl_class changed */ + uint64_t ctime = mach_approximate_time(); + machine_thread_going_on_core(thread, processor->current_urgency, 0, 0, ctime); + } else if (thread->state & TH_RUN) { + processor_t processor = thread->last_processor; + + if (!lazy_update && + processor != PROCESSOR_NULL && + processor != current_processor() && + processor->active_thread == thread) { cause_ast_check(processor); + } } } -#if 0 - -static void -run_queue_check( - run_queue_t rq, - thread_t thread) +/* + * thread_run_queue_remove_for_handoff + * + * Pull a thread or its (recursive) push target out of the runqueue + * so that it is ready for thread_run() + * + * Called at splsched + * + * Returns the thread that was pulled or THREAD_NULL if no thread could be pulled. + * This may be different than the thread that was passed in. + */ +thread_t +thread_run_queue_remove_for_handoff(thread_t thread) { - queue_t q; - queue_entry_t qe; - - if (rq != thread->runq) - panic("run_queue_check: thread runq"); + thread_t pulled_thread = THREAD_NULL; - if (thread->sched_pri > MAXPRI || thread->sched_pri < MINPRI) - panic("run_queue_check: thread sched_pri"); + thread_lock(thread); - q = &rq->queues[thread->sched_pri]; - qe = queue_first(q); - while (!queue_end(q, qe)) { - if (qe == (queue_entry_t)thread) - return; + /* + * Check that the thread is not bound + * to a different processor, and that realtime + * is not involved. + * + * Next, pull it off its run queue. If it + * doesn't come, it's not eligible. + */ - qe = queue_next(qe); + processor_t processor = current_processor(); + if (processor->current_pri < BASEPRI_RTQUEUES && thread->sched_pri < BASEPRI_RTQUEUES && + (thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == processor)) { + if (thread_run_queue_remove(thread)) { + pulled_thread = thread; + } } - panic("run_queue_check: end"); -} + thread_unlock(thread); -#endif /* DEBUG */ + return pulled_thread; +} /* - * run_queue_remove: + * thread_run_queue_remove: * - * Remove a thread from a current run queue and + * Remove a thread from its current run queue and * return TRUE if successful. * * Thread must be locked. + * + * If thread->runq is PROCESSOR_NULL, the thread will not re-enter the + * run queues because the caller locked the thread. Otherwise + * the thread is on a run queue, but could be chosen for dispatch + * and removed by another processor under a different lock, which + * will set thread->runq to PROCESSOR_NULL. + * + * Hence the thread select path must not rely on anything that could + * be changed under the thread lock after calling this function, + * most importantly thread->sched_pri. */ boolean_t -run_queue_remove( - thread_t thread) +thread_run_queue_remove( + thread_t thread) { - processor_t processor = thread->runq; + boolean_t removed = FALSE; + processor_t processor = thread->runq; - /* - * If processor is PROCESSOR_NULL, the thread will stay out of the - * run queues because the caller locked the thread. Otherwise - * the thread is on a run queue, but could be chosen for dispatch - * and removed. - */ - if (processor != PROCESSOR_NULL) { - void * rqlock; - run_queue_t rq; + if ((thread->state & (TH_RUN | TH_WAIT)) == TH_WAIT) { + /* Thread isn't runnable */ + assert(thread->runq == PROCESSOR_NULL); + return FALSE; + } + if (processor == PROCESSOR_NULL) { /* - * The processor run queues are locked by the - * processor set. Real-time priorities use a - * global queue with a dedicated lock. + * The thread is either not on the runq, + * or is in the midst of being removed from the runq. + * + * runq is set to NULL under the pset lock, not the thread + * lock, so the thread may still be in the process of being dequeued + * from the runq. It will wait in invoke for the thread lock to be + * dropped. */ - if (thread->sched_pri < BASEPRI_RTQUEUES) { - rqlock = &processor->processor_set->sched_lock; - rq = &processor->runq; - } - else { - rqlock = &rt_lock; rq = &rt_runq; - } - simple_lock(rqlock); + return FALSE; + } - if (processor == thread->runq) { - /* - * Thread is on a run queue and we have a lock on - * that run queue. - */ - remqueue(&rq->queues[0], (queue_entry_t)thread); - rq->count--; - if (testbit(thread->sched_pri, sched_preempt_pri)) { - rq->urgency--; assert(rq->urgency >= 0); - } + if (thread->sched_pri < BASEPRI_RTQUEUES) { + return SCHED(processor_queue_remove)(processor, thread); + } - if (queue_empty(rq->queues + thread->sched_pri)) { - /* update run queue status */ - if (thread->sched_pri != IDLEPRI) - clrbit(MAXPRI - thread->sched_pri, rq->bitmap); - rq->highq = MAXPRI - ffsbit(rq->bitmap); - } + processor_set_t pset = processor->processor_set; - thread->runq = PROCESSOR_NULL; - } - else { - /* - * The thread left the run queue before we could - * lock the run queue. - */ - assert(thread->runq == PROCESSOR_NULL); - processor = PROCESSOR_NULL; - } + rt_lock_lock(pset); + + if (thread->runq != PROCESSOR_NULL) { + /* + * Thread is on the RT run queue and we have a lock on + * that run queue. + */ - simple_unlock(rqlock); + remqueue(&thread->runq_links); + SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset)); + rt_runq_count_decr(pset); + + thread->runq = PROCESSOR_NULL; + + removed = TRUE; } - return (processor != PROCESSOR_NULL); + rt_lock_unlock(pset); + + return removed; } /* - * steal_processor_thread: + * Put the thread back where it goes after a thread_run_queue_remove * - * Locate a thread to steal from the processor and - * return it. + * Thread must have been removed under the same thread lock hold * - * Associated pset must be locked. Returns THREAD_NULL - * on failure. + * thread locked, at splsched */ -static thread_t -steal_processor_thread( - processor_t processor) -{ - run_queue_t rq = &processor->runq; - queue_t queue = rq->queues + rq->highq; - int pri = rq->highq, count = rq->count; - thread_t thread; - - while (count > 0) { - thread = (thread_t)queue_first(queue); - while (!queue_end(queue, (queue_entry_t)thread)) { - if (thread->bound_processor == PROCESSOR_NULL) { - remqueue(queue, (queue_entry_t)thread); - - thread->runq = PROCESSOR_NULL; - rq->count--; - if (testbit(pri, sched_preempt_pri)) { - rq->urgency--; assert(rq->urgency >= 0); - } - if (queue_empty(queue)) { - if (pri != IDLEPRI) - clrbit(MAXPRI - pri, rq->bitmap); - rq->highq = MAXPRI - ffsbit(rq->bitmap); - } - - return (thread); - } - count--; +void +thread_run_queue_reinsert(thread_t thread, sched_options_t options) +{ + assert(thread->runq == PROCESSOR_NULL); + assert(thread->state & (TH_RUN)); - thread = (thread_t)queue_next((queue_entry_t)thread); - } + thread_setrun(thread, options); +} - queue--; pri--; +void +sys_override_cpu_throttle(boolean_t enable_override) +{ + if (enable_override) { + cpu_throttle_enabled = 0; + } else { + cpu_throttle_enabled = 1; } - - return (THREAD_NULL); } -/* - * Locate and steal a thread, beginning - * at the pset. - * - * The pset must be locked, and is returned - * unlocked. - * - * Returns the stolen thread, or THREAD_NULL on - * failure. - */ -static thread_t -steal_thread( - processor_set_t pset) +thread_urgency_t +thread_get_urgency(thread_t thread, uint64_t *arg1, uint64_t *arg2) { - processor_set_t nset, cset = pset; - processor_t processor; - thread_t thread; - - do { - processor = (processor_t)queue_first(&cset->active_queue); - while (!queue_end(&cset->active_queue, (queue_entry_t)processor)) { - if (processor->runq.count > 0) { - thread = steal_processor_thread(processor); - if (thread != THREAD_NULL) { - remqueue(&cset->active_queue, (queue_entry_t)processor); - enqueue_tail(&cset->active_queue, (queue_entry_t)processor); + uint64_t urgency_param1 = 0, urgency_param2 = 0; - pset_unlock(cset); + thread_urgency_t urgency; - return (thread); - } - } + if (thread == NULL || (thread->state & TH_IDLE)) { + urgency_param1 = 0; + urgency_param2 = 0; - processor = (processor_t)queue_next((queue_entry_t)processor); - } + urgency = THREAD_URGENCY_NONE; + } else if (thread->sched_mode == TH_MODE_REALTIME) { + urgency_param1 = thread->realtime.period; + urgency_param2 = thread->realtime.deadline; - nset = next_pset(cset); + urgency = THREAD_URGENCY_REAL_TIME; + } else if (cpu_throttle_enabled && + (thread->sched_pri <= MAXPRI_THROTTLE) && + (thread->base_pri <= MAXPRI_THROTTLE)) { + /* + * Threads that are running at low priority but are not + * tagged with a specific QoS are separated out from + * the "background" urgency. Performance management + * subsystem can decide to either treat these threads + * as normal threads or look at other signals like thermal + * levels for optimal power/perf tradeoffs for a platform. + */ + boolean_t thread_lacks_qos = (proc_get_effective_thread_policy(thread, TASK_POLICY_QOS) == THREAD_QOS_UNSPECIFIED); //thread_has_qos_policy(thread); + boolean_t task_is_suppressed = (proc_get_effective_task_policy(thread->task, TASK_POLICY_SUP_ACTIVE) == 0x1); - if (nset != pset) { - pset_unlock(cset); + /* + * Background urgency applied when thread priority is + * MAXPRI_THROTTLE or lower and thread is not promoted + * and thread has a QoS specified + */ + urgency_param1 = thread->sched_pri; + urgency_param2 = thread->base_pri; - cset = nset; - pset_lock(cset); + if (thread_lacks_qos && !task_is_suppressed) { + urgency = THREAD_URGENCY_LOWPRI; + } else { + urgency = THREAD_URGENCY_BACKGROUND; } - } while (nset != pset); + } else { + /* For otherwise unclassified threads, report throughput QoS parameters */ + urgency_param1 = proc_get_effective_thread_policy(thread, TASK_POLICY_THROUGH_QOS); + urgency_param2 = proc_get_effective_task_policy(thread->task, TASK_POLICY_THROUGH_QOS); + urgency = THREAD_URGENCY_NORMAL; + } + + if (arg1 != NULL) { + *arg1 = urgency_param1; + } + if (arg2 != NULL) { + *arg2 = urgency_param2; + } - pset_unlock(cset); + return urgency; +} + +perfcontrol_class_t +thread_get_perfcontrol_class(thread_t thread) +{ + /* Special case handling */ + if (thread->state & TH_IDLE) { + return PERFCONTROL_CLASS_IDLE; + } + if (thread->task == kernel_task) { + return PERFCONTROL_CLASS_KERNEL; + } + if (thread->sched_mode == TH_MODE_REALTIME) { + return PERFCONTROL_CLASS_REALTIME; + } - return (THREAD_NULL); + /* perfcontrol_class based on base_pri */ + if (thread->base_pri <= MAXPRI_THROTTLE) { + return PERFCONTROL_CLASS_BACKGROUND; + } else if (thread->base_pri <= BASEPRI_UTILITY) { + return PERFCONTROL_CLASS_UTILITY; + } else if (thread->base_pri <= BASEPRI_DEFAULT) { + return PERFCONTROL_CLASS_NONUI; + } else if (thread->base_pri <= BASEPRI_FOREGROUND) { + return PERFCONTROL_CLASS_UI; + } else { + return PERFCONTROL_CLASS_ABOVEUI; + } } /* @@ -2903,164 +4852,205 @@ steal_thread( * * Returns a the next thread to execute if dispatched directly. */ -static thread_t + +#if 0 +#define IDLE_KERNEL_DEBUG_CONSTANT(...) KERNEL_DEBUG_CONSTANT(__VA_ARGS__) +#else +#define IDLE_KERNEL_DEBUG_CONSTANT(...) do { } while(0) +#endif + +thread_t processor_idle( - thread_t thread, - processor_t processor) + thread_t thread, + processor_t processor) { - processor_set_t pset = processor->processor_set; - thread_t new_thread; - int state; + processor_set_t pset = processor->processor_set; (void)splsched(); - KERNEL_DEBUG_CONSTANT( - MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_START, (uintptr_t)thread_tid(thread), 0, 0, 0, 0); + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_START, + (uintptr_t)thread_tid(thread), 0, 0, 0, 0); - timer_switch(&PROCESSOR_DATA(processor, system_state), - mach_absolute_time(), &PROCESSOR_DATA(processor, idle_state)); + SCHED_STATS_CPU_IDLE_START(processor); + + uint64_t ctime = mach_absolute_time(); + + timer_switch(&PROCESSOR_DATA(processor, system_state), ctime, &PROCESSOR_DATA(processor, idle_state)); PROCESSOR_DATA(processor, current_state) = &PROCESSOR_DATA(processor, idle_state); - while (processor->next_thread == THREAD_NULL && processor->runq.count == 0 && rt_runq.count == 0 && - (thread == THREAD_NULL || ((thread->state & (TH_WAIT|TH_SUSP)) == TH_WAIT && !thread->wake_active))) { - machine_idle(); + cpu_quiescent_counter_leave(ctime); - (void)splsched(); + while (1) { + /* + * Ensure that updates to my processor and pset state, + * made by the IPI source processor before sending the IPI, + * are visible on this processor now (even though we don't + * take the pset lock yet). + */ + atomic_thread_fence(memory_order_acquire); - if (processor->state == PROCESSOR_INACTIVE && !machine_cpu_is_inactive(processor->cpu_id)) + if (processor->state != PROCESSOR_IDLE) { break; - } - - timer_switch(&PROCESSOR_DATA(processor, idle_state), - mach_absolute_time(), &PROCESSOR_DATA(processor, system_state)); - PROCESSOR_DATA(processor, current_state) = &PROCESSOR_DATA(processor, system_state); + } + if (bit_test(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id)) { + break; + } +#if defined(CONFIG_SCHED_DEFERRED_AST) + if (bit_test(pset->pending_deferred_AST_cpu_mask, processor->cpu_id)) { + break; + } +#endif + if (processor->is_recommended && (processor->processor_primary == processor)) { + if (rt_runq_count(pset)) { + break; + } + } else { + if (SCHED(processor_bound_count)(processor)) { + break; + } + } - pset_lock(pset); + IDLE_KERNEL_DEBUG_CONSTANT( + MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_NONE, (uintptr_t)thread_tid(thread), rt_runq_count(pset), SCHED(processor_runq_count)(processor), -1, 0); - state = processor->state; - if (state == PROCESSOR_DISPATCHING) { - /* - * Commmon case -- cpu dispatched. - */ - new_thread = processor->next_thread; - processor->next_thread = THREAD_NULL; - processor->state = PROCESSOR_RUNNING; + machine_track_platform_idle(TRUE); - if ( processor->runq.highq > new_thread->sched_pri || - (rt_runq.highq > 0 && rt_runq.highq >= new_thread->sched_pri) ) { - processor->deadline = UINT64_MAX; + machine_idle(); + /* returns with interrupts enabled */ - pset_unlock(pset); + machine_track_platform_idle(FALSE); - thread_lock(new_thread); - thread_setrun(new_thread, SCHED_HEADQ); - thread_unlock(new_thread); + (void)splsched(); - KERNEL_DEBUG_CONSTANT( - MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_END, (uintptr_t)thread_tid(thread), state, 0, 0, 0); - - return (THREAD_NULL); + /* + * Check if we should call sched_timeshare_consider_maintenance() here. + * The CPU was woken out of idle due to an interrupt and we should do the + * call only if the processor is still idle. If the processor is non-idle, + * the threads running on the processor would do the call as part of + * context swithing. + */ + if (processor->state == PROCESSOR_IDLE) { + sched_timeshare_consider_maintenance(mach_absolute_time()); } - pset_unlock(pset); - - KERNEL_DEBUG_CONSTANT( - MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_END, (uintptr_t)thread_tid(thread), state, (uintptr_t)thread_tid(new_thread), 0, 0); + IDLE_KERNEL_DEBUG_CONSTANT( + MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_NONE, (uintptr_t)thread_tid(thread), rt_runq_count(pset), SCHED(processor_runq_count)(processor), -2, 0); - return (new_thread); + if (!SCHED(processor_queue_empty)(processor)) { + /* Secondary SMT processors respond to directed wakeups + * exclusively. Some platforms induce 'spurious' SMT wakeups. + */ + if (processor->processor_primary == processor) { + break; + } + } } - else - if (state == PROCESSOR_IDLE) { - remqueue(&pset->idle_queue, (queue_entry_t)processor); - processor->state = PROCESSOR_RUNNING; - enqueue_tail(&pset->active_queue, (queue_entry_t)processor); - } - else - if (state == PROCESSOR_INACTIVE) { - processor->state = PROCESSOR_RUNNING; - enqueue_tail(&pset->active_queue, (queue_entry_t)processor); - } - else - if (state == PROCESSOR_SHUTDOWN) { - /* - * Going off-line. Force a - * reschedule. - */ - if ((new_thread = processor->next_thread) != THREAD_NULL) { - processor->next_thread = THREAD_NULL; - processor->deadline = UINT64_MAX; + ctime = mach_absolute_time(); - pset_unlock(pset); + timer_switch(&PROCESSOR_DATA(processor, idle_state), ctime, &PROCESSOR_DATA(processor, system_state)); + PROCESSOR_DATA(processor, current_state) = &PROCESSOR_DATA(processor, system_state); - thread_lock(new_thread); - thread_setrun(new_thread, SCHED_HEADQ); - thread_unlock(new_thread); + cpu_quiescent_counter_join(ctime); - KERNEL_DEBUG_CONSTANT( - MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_END, (uintptr_t)thread_tid(thread), state, 0, 0, 0); + ast_t reason = AST_NONE; - return (THREAD_NULL); - } - } + /* We're handling all scheduling AST's */ + ast_off(AST_SCHEDULING); - pset_unlock(pset); + /* + * thread_select will move the processor from dispatching to running, + * or put it in idle if there's nothing to do. + */ + thread_t current_thread = current_thread(); + + thread_lock(current_thread); + thread_t new_thread = thread_select(current_thread, processor, &reason); + thread_unlock(current_thread); - KERNEL_DEBUG_CONSTANT( - MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_END, (uintptr_t)thread_tid(thread), state, 0, 0, 0); + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_END, + (uintptr_t)thread_tid(thread), processor->state, (uintptr_t)thread_tid(new_thread), reason, 0); - return (THREAD_NULL); + return new_thread; } /* * Each processor has a dedicated thread which * executes the idle loop when there is no suitable * previous context. + * + * This continuation is entered with interrupts disabled. */ void -idle_thread(void) +idle_thread(__assert_only void* parameter, + __unused wait_result_t result) { - processor_t processor = current_processor(); - thread_t new_thread; + assert(ml_get_interrupts_enabled() == FALSE); + assert(parameter == NULL); + + processor_t processor = current_processor(); + + /* + * Ensure that anything running in idle context triggers + * preemption-disabled checks. + */ + disable_preemption(); + + /* + * Enable interrupts temporarily to handle any pending interrupts + * or IPIs before deciding to sleep + */ + spllo(); + + thread_t new_thread = processor_idle(THREAD_NULL, processor); + /* returns with interrupts disabled */ + + enable_preemption(); - new_thread = processor_idle(THREAD_NULL, processor); if (new_thread != THREAD_NULL) { - thread_run(processor->idle_thread, (thread_continue_t)idle_thread, NULL, new_thread); + thread_run(processor->idle_thread, + idle_thread, NULL, new_thread); /*NOTREACHED*/ } - thread_block((thread_continue_t)idle_thread); + thread_block(idle_thread); /*NOTREACHED*/ } kern_return_t idle_thread_create( - processor_t processor) + processor_t processor) { - kern_return_t result; - thread_t thread; - spl_t s; + kern_return_t result; + thread_t thread; + spl_t s; + char name[MAXTHREADNAMESIZE]; + + result = kernel_thread_create(idle_thread, NULL, MAXPRI_KERNEL, &thread); + if (result != KERN_SUCCESS) { + return result; + } - result = kernel_thread_create((thread_continue_t)idle_thread, NULL, MAXPRI_KERNEL, &thread); - if (result != KERN_SUCCESS) - return (result); + snprintf(name, sizeof(name), "idle #%d", processor->cpu_id); + thread_set_thread_name(thread, name); s = splsched(); thread_lock(thread); thread->bound_processor = processor; processor->idle_thread = thread; - thread->sched_pri = thread->priority = IDLEPRI; + thread->sched_pri = thread->base_pri = IDLEPRI; thread->state = (TH_RUN | TH_IDLE); + thread->options |= TH_OPT_IDLE_THREAD; thread_unlock(thread); splx(s); thread_deallocate(thread); - return (KERN_SUCCESS); + return KERN_SUCCESS; } -static uint64_t sched_tick_deadline; - /* * sched_startup: * @@ -3071,112 +5061,246 @@ static uint64_t sched_tick_deadline; void sched_startup(void) { - kern_return_t result; - thread_t thread; + kern_return_t result; + thread_t thread; + + simple_lock_init(&sched_vm_group_list_lock, 0); + +#if __arm__ || __arm64__ + simple_lock_init(&sched_recommended_cores_lock, 0); +#endif /* __arm__ || __arm64__ */ - result = kernel_thread_start_priority((thread_continue_t)sched_tick_thread, NULL, MAXPRI_KERNEL, &thread); - if (result != KERN_SUCCESS) + result = kernel_thread_start_priority((thread_continue_t)sched_init_thread, + (void *)SCHED(maintenance_continuation), MAXPRI_KERNEL, &thread); + if (result != KERN_SUCCESS) { panic("sched_startup"); + } thread_deallocate(thread); + assert_thread_magic(thread); + /* - * Yield to the sched_tick_thread while it times - * a series of context switches back. It stores - * the baseline value in sched_cswtime. + * Yield to the sched_init_thread once, to + * initialize our own thread after being switched + * back to. * * The current thread is the only other thread * active at this point. */ - while (sched_cswtime == 0) - thread_block(THREAD_CONTINUE_NULL); + thread_block(THREAD_CONTINUE_NULL); +} - thread_daemon_init(); +#if __arm64__ +static _Atomic uint64_t sched_perfcontrol_callback_deadline; +#endif /* __arm64__ */ + + +#if defined(CONFIG_SCHED_TIMESHARE_CORE) + +static volatile uint64_t sched_maintenance_deadline; +static uint64_t sched_tick_last_abstime; +static uint64_t sched_tick_delta; +uint64_t sched_tick_max_delta; - thread_call_initialize(); -} /* - * sched_tick_thread: + * sched_init_thread: * * Perform periodic bookkeeping functions about ten * times per second. */ -static void -sched_tick_continue(void) +void +sched_timeshare_maintenance_continue(void) { - uint64_t abstime = mach_absolute_time(); + uint64_t sched_tick_ctime, late_time; + + struct sched_update_scan_context scan_context = { + .earliest_bg_make_runnable_time = UINT64_MAX, + .earliest_normal_make_runnable_time = UINT64_MAX, + .earliest_rt_make_runnable_time = UINT64_MAX + }; + + sched_tick_ctime = mach_absolute_time(); + + if (__improbable(sched_tick_last_abstime == 0)) { + sched_tick_last_abstime = sched_tick_ctime; + late_time = 0; + sched_tick_delta = 1; + } else { + late_time = sched_tick_ctime - sched_tick_last_abstime; + sched_tick_delta = late_time / sched_tick_interval; + /* Ensure a delta of 1, since the interval could be slightly + * smaller than the sched_tick_interval due to dispatch + * latencies. + */ + sched_tick_delta = MAX(sched_tick_delta, 1); - sched_tick++; + /* In the event interrupt latencies or platform + * idle events that advanced the timebase resulted + * in periods where no threads were dispatched, + * cap the maximum "tick delta" at SCHED_TICK_MAX_DELTA + * iterations. + */ + sched_tick_delta = MIN(sched_tick_delta, SCHED_TICK_MAX_DELTA); + + sched_tick_last_abstime = sched_tick_ctime; + sched_tick_max_delta = MAX(sched_tick_delta, sched_tick_max_delta); + } + + KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_MAINTENANCE) | DBG_FUNC_START, + sched_tick_delta, late_time, 0, 0, 0); + + /* Add a number of pseudo-ticks corresponding to the elapsed interval + * This could be greater than 1 if substantial intervals where + * all processors are idle occur, which rarely occurs in practice. + */ + + sched_tick += sched_tick_delta; + + update_vm_info(); /* * Compute various averages. */ - compute_averages(); + compute_averages(sched_tick_delta); /* * Scan the run queues for threads which - * may need to be updated. + * may need to be updated, and find the earliest runnable thread on the runqueue + * to report its latency. + */ + SCHED(thread_update_scan)(&scan_context); + + SCHED(rt_runq_scan)(&scan_context); + + uint64_t ctime = mach_absolute_time(); + + uint64_t bg_max_latency = (ctime > scan_context.earliest_bg_make_runnable_time) ? + ctime - scan_context.earliest_bg_make_runnable_time : 0; + + uint64_t default_max_latency = (ctime > scan_context.earliest_normal_make_runnable_time) ? + ctime - scan_context.earliest_normal_make_runnable_time : 0; + + uint64_t realtime_max_latency = (ctime > scan_context.earliest_rt_make_runnable_time) ? + ctime - scan_context.earliest_rt_make_runnable_time : 0; + + machine_max_runnable_latency(bg_max_latency, default_max_latency, realtime_max_latency); + + /* + * Check to see if the special sched VM group needs attention. */ - thread_update_scan(); + sched_vm_group_maintenance(); - clock_deadline_for_periodic_event(sched_tick_interval, abstime, - &sched_tick_deadline); +#if __arm__ || __arm64__ + /* Check to see if the recommended cores failsafe is active */ + sched_recommended_cores_maintenance(); +#endif /* __arm__ || __arm64__ */ - assert_wait_deadline((event_t)sched_tick_thread, THREAD_UNINT, sched_tick_deadline); - thread_block((thread_continue_t)sched_tick_continue); + +#if DEBUG || DEVELOPMENT +#if __x86_64__ +#include + /* Check for long-duration interrupts */ + mp_interrupt_watchdog(); +#endif /* __x86_64__ */ +#endif /* DEBUG || DEVELOPMENT */ + + KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_MAINTENANCE) | DBG_FUNC_END, + sched_pri_shifts[TH_BUCKET_SHARE_FG], sched_pri_shifts[TH_BUCKET_SHARE_BG], + sched_pri_shifts[TH_BUCKET_SHARE_UT], sched_pri_shifts[TH_BUCKET_SHARE_DF], 0); + + assert_wait((event_t)sched_timeshare_maintenance_continue, THREAD_UNINT); + thread_block((thread_continue_t)sched_timeshare_maintenance_continue); /*NOTREACHED*/ } +static uint64_t sched_maintenance_wakeups; + /* - * Time a series of context switches to determine - * a baseline. Toss the high and low and return - * the one-way value. + * Determine if the set of routines formerly driven by a maintenance timer + * must be invoked, based on a deadline comparison. Signals the scheduler + * maintenance thread on deadline expiration. Must be invoked at an interval + * lower than the "sched_tick_interval", currently accomplished by + * invocation via the quantum expiration timer and at context switch time. + * Performance matters: this routine reuses a timestamp approximating the + * current absolute time received from the caller, and should perform + * no more than a comparison against the deadline in the common case. */ -static uint32_t -time_cswitch(void) +void +sched_timeshare_consider_maintenance(uint64_t ctime) { - uint32_t new, hi, low, accum; - uint64_t abstime; - int i, tries = 7; + cpu_quiescent_counter_checkin(ctime); - accum = hi = low = 0; - for (i = 0; i < tries; ++i) { - abstime = mach_absolute_time(); - thread_block(THREAD_CONTINUE_NULL); + uint64_t deadline = sched_maintenance_deadline; - new = (uint32_t)(mach_absolute_time() - abstime); + if (__improbable(ctime >= deadline)) { + if (__improbable(current_thread() == sched_maintenance_thread)) { + return; + } + OSMemoryBarrier(); - if (i == 0) - accum = hi = low = new; - else { - if (new < low) - low = new; - else - if (new > hi) - hi = new; - accum += new; + uint64_t ndeadline = ctime + sched_tick_interval; + + if (__probable(os_atomic_cmpxchg(&sched_maintenance_deadline, deadline, ndeadline, seq_cst))) { + thread_wakeup((event_t)sched_timeshare_maintenance_continue); + sched_maintenance_wakeups++; + } + } + +#if !CONFIG_SCHED_CLUTCH + /* + * Only non-clutch schedulers use the global load calculation EWMA algorithm. For clutch + * scheduler, the load is maintained at the thread group and bucket level. + */ + uint64_t load_compute_deadline = os_atomic_load_wide(&sched_load_compute_deadline, relaxed); + + if (__improbable(load_compute_deadline && ctime >= load_compute_deadline)) { + uint64_t new_deadline = 0; + if (os_atomic_cmpxchg(&sched_load_compute_deadline, load_compute_deadline, new_deadline, relaxed)) { + compute_sched_load(); + new_deadline = ctime + sched_load_compute_interval_abs; + os_atomic_store_wide(&sched_load_compute_deadline, new_deadline, relaxed); } } +#endif /* CONFIG_SCHED_CLUTCH */ - return ((accum - hi - low) / (2 * (tries - 2))); +#if __arm64__ + uint64_t perf_deadline = os_atomic_load(&sched_perfcontrol_callback_deadline, relaxed); + + if (__improbable(perf_deadline && ctime >= perf_deadline)) { + /* CAS in 0, if success, make callback. Otherwise let the next context switch check again. */ + if (os_atomic_cmpxchg(&sched_perfcontrol_callback_deadline, perf_deadline, 0, relaxed)) { + machine_perfcontrol_deadline_passed(perf_deadline); + } + } +#endif /* __arm64__ */ } +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + void -sched_tick_thread(void) +sched_init_thread(void (*continuation)(void)) { - sched_cswtime = time_cswitch(); + thread_block(THREAD_CONTINUE_NULL); + + thread_t thread = current_thread(); - sched_tick_deadline = mach_absolute_time(); + thread_set_thread_name(thread, "sched_maintenance_thread"); + + sched_maintenance_thread = thread; + + continuation(); - sched_tick_continue(); /*NOTREACHED*/ } +#if defined(CONFIG_SCHED_TIMESHARE_CORE) + /* * thread_update_scan / runq_scan: * - * Scan the run queues to account for timesharing threads + * Scan the run queues to account for timesharing threads * which need to be updated. * * Scanner runs in two passes. Pass one squirrels likely @@ -3189,114 +5313,205 @@ sched_tick_thread(void) * disabling preemption for long periods. */ -#define THREAD_UPDATE_SIZE 128 +#define THREAD_UPDATE_SIZE 128 -static thread_t thread_update_array[THREAD_UPDATE_SIZE]; -static int thread_update_count = 0; +static thread_t thread_update_array[THREAD_UPDATE_SIZE]; +static uint32_t thread_update_count = 0; + +/* Returns TRUE if thread was added, FALSE if thread_update_array is full */ +boolean_t +thread_update_add_thread(thread_t thread) +{ + if (thread_update_count == THREAD_UPDATE_SIZE) { + return FALSE; + } + + thread_update_array[thread_update_count++] = thread; + thread_reference_internal(thread); + return TRUE; +} + +void +thread_update_process_threads(void) +{ + assert(thread_update_count <= THREAD_UPDATE_SIZE); + + for (uint32_t i = 0; i < thread_update_count; i++) { + thread_t thread = thread_update_array[i]; + assert_thread_magic(thread); + thread_update_array[i] = THREAD_NULL; + + spl_t s = splsched(); + thread_lock(thread); + if (!(thread->state & (TH_WAIT)) && thread->sched_stamp != sched_tick) { + SCHED(update_priority)(thread); + } + thread_unlock(thread); + splx(s); + + thread_deallocate(thread); + } + + thread_update_count = 0; +} /* * Scan a runq for candidate threads. * * Returns TRUE if retry is needed. */ -static boolean_t +boolean_t runq_scan( - run_queue_t runq) -{ - register int count; - register queue_t q; - register thread_t thread; - - if ((count = runq->count) > 0) { - q = runq->queues + runq->highq; - while (count > 0) { - queue_iterate(q, thread, thread_t, links) { - if ( thread->sched_stamp != sched_tick && - (thread->sched_mode & TH_MODE_TIMESHARE) ) { - if (thread_update_count == THREAD_UPDATE_SIZE) - return (TRUE); - - thread_update_array[thread_update_count++] = thread; - thread_reference_internal(thread); + run_queue_t runq, + sched_update_scan_context_t scan_context) +{ + int count = runq->count; + int queue_index; + + assert(count >= 0); + + if (count == 0) { + return FALSE; + } + + for (queue_index = bitmap_first(runq->bitmap, NRQS); + queue_index >= 0; + queue_index = bitmap_next(runq->bitmap, queue_index)) { + thread_t thread; + circle_queue_t queue = &runq->queues[queue_index]; + + cqe_foreach_element(thread, queue, runq_links) { + assert(count > 0); + assert_thread_magic(thread); + + if (thread->sched_stamp != sched_tick && + thread->sched_mode == TH_MODE_TIMESHARE) { + if (thread_update_add_thread(thread) == FALSE) { + return TRUE; } + } - count--; + if (cpu_throttle_enabled && ((thread->sched_pri <= MAXPRI_THROTTLE) && (thread->base_pri <= MAXPRI_THROTTLE))) { + if (thread->last_made_runnable_time < scan_context->earliest_bg_make_runnable_time) { + scan_context->earliest_bg_make_runnable_time = thread->last_made_runnable_time; + } + } else { + if (thread->last_made_runnable_time < scan_context->earliest_normal_make_runnable_time) { + scan_context->earliest_normal_make_runnable_time = thread->last_made_runnable_time; + } } + count--; + } + } + + return FALSE; +} + +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +boolean_t +thread_eager_preemption(thread_t thread) +{ + return (thread->sched_flags & TH_SFLAG_EAGERPREEMPT) != 0; +} + +void +thread_set_eager_preempt(thread_t thread) +{ + spl_t x; + processor_t p; + ast_t ast = AST_NONE; + + x = splsched(); + p = current_processor(); + + thread_lock(thread); + thread->sched_flags |= TH_SFLAG_EAGERPREEMPT; - q--; + if (thread == current_thread()) { + ast = csw_check(thread, p, AST_NONE); + thread_unlock(thread); + if (ast != AST_NONE) { + (void) thread_block_reason(THREAD_CONTINUE_NULL, NULL, ast); + } + } else { + p = thread->last_processor; + + if (p != PROCESSOR_NULL && p->state == PROCESSOR_RUNNING && + p->active_thread == thread) { + cause_ast_check(p); } + + thread_unlock(thread); } - return (FALSE); + splx(x); } -static void -thread_update_scan(void) +void +thread_clear_eager_preempt(thread_t thread) { - boolean_t restart_needed = FALSE; - processor_t processor = processor_list; - processor_set_t pset; - thread_t thread; - spl_t s; + spl_t x; - do { - do { - pset = processor->processor_set; + x = splsched(); + thread_lock(thread); - s = splsched(); - pset_lock(pset); + thread->sched_flags &= ~TH_SFLAG_EAGERPREEMPT; - restart_needed = runq_scan(&processor->runq); + thread_unlock(thread); + splx(x); +} - pset_unlock(pset); - splx(s); +/* + * Scheduling statistics + */ +void +sched_stats_handle_csw(processor_t processor, int reasons, int selfpri, int otherpri) +{ + struct processor_sched_statistics *stats; + boolean_t to_realtime = FALSE; - if (restart_needed) - break; + stats = &processor->processor_data.sched_stats; + stats->csw_count++; - thread = processor->idle_thread; - if (thread != THREAD_NULL && thread->sched_stamp != sched_tick) { - if (thread_update_count == THREAD_UPDATE_SIZE) { - restart_needed = TRUE; - break; - } + if (otherpri >= BASEPRI_REALTIME) { + stats->rt_sched_count++; + to_realtime = TRUE; + } - thread_update_array[thread_update_count++] = thread; - thread_reference_internal(thread); - } - } while ((processor = processor->processor_list) != NULL); + if ((reasons & AST_PREEMPT) != 0) { + stats->preempt_count++; - /* - * Ok, we now have a collection of candidates -- fix them. - */ - while (thread_update_count > 0) { - thread = thread_update_array[--thread_update_count]; - thread_update_array[thread_update_count] = THREAD_NULL; + if (selfpri >= BASEPRI_REALTIME) { + stats->preempted_rt_count++; + } - s = splsched(); - thread_lock(thread); - if ( !(thread->state & (TH_WAIT|TH_SUSP)) && - thread->sched_stamp != sched_tick ) - update_priority(thread); - thread_unlock(thread); - splx(s); + if (to_realtime) { + stats->preempted_by_rt_count++; + } + } +} + +void +sched_stats_handle_runq_change(struct runq_stats *stats, int old_count) +{ + uint64_t timestamp = mach_absolute_time(); - thread_deallocate(thread); - } - } while (restart_needed); + stats->count_sum += (timestamp - stats->last_change_timestamp) * old_count; + stats->last_change_timestamp = timestamp; } - + /* - * Just in case someone doesn't use the macro + * For calls from assembly code */ -#undef thread_wakeup +#undef thread_wakeup void thread_wakeup( - event_t x); + event_t x); void thread_wakeup( - event_t x) + event_t x) { thread_wakeup_with_result(x, THREAD_AWAKENED); } @@ -3304,46 +5519,791 @@ thread_wakeup( boolean_t preemption_enabled(void) { - return (get_preemption_level() == 0 && ml_get_interrupts_enabled()); + return get_preemption_level() == 0 && ml_get_interrupts_enabled(); } -#if DEBUG -static boolean_t -thread_runnable( - thread_t thread) +static void +sched_timer_deadline_tracking_init(void) +{ + nanoseconds_to_absolutetime(TIMER_DEADLINE_TRACKING_BIN_1_DEFAULT, &timer_deadline_tracking_bin_1); + nanoseconds_to_absolutetime(TIMER_DEADLINE_TRACKING_BIN_2_DEFAULT, &timer_deadline_tracking_bin_2); +} + +#if __arm__ || __arm64__ + +uint32_t perfcontrol_requested_recommended_cores = ALL_CORES_RECOMMENDED; +uint32_t perfcontrol_requested_recommended_core_count = MAX_CPUS; +bool perfcontrol_failsafe_active = false; +bool perfcontrol_sleep_override = false; + +uint64_t perfcontrol_failsafe_maintenance_runnable_time; +uint64_t perfcontrol_failsafe_activation_time; +uint64_t perfcontrol_failsafe_deactivation_time; + +/* data covering who likely caused it and how long they ran */ +#define FAILSAFE_NAME_LEN 33 /* (2*MAXCOMLEN)+1 from size of p_name */ +char perfcontrol_failsafe_name[FAILSAFE_NAME_LEN]; +int perfcontrol_failsafe_pid; +uint64_t perfcontrol_failsafe_tid; +uint64_t perfcontrol_failsafe_thread_timer_at_start; +uint64_t perfcontrol_failsafe_thread_timer_last_seen; +uint32_t perfcontrol_failsafe_recommended_at_trigger; + +/* + * Perf controller calls here to update the recommended core bitmask. + * If the failsafe is active, we don't immediately apply the new value. + * Instead, we store the new request and use it after the failsafe deactivates. + * + * If the failsafe is not active, immediately apply the update. + * + * No scheduler locks are held, no other locks are held that scheduler might depend on, + * interrupts are enabled + * + * currently prototype is in osfmk/arm/machine_routines.h + */ +void +sched_perfcontrol_update_recommended_cores(uint32_t recommended_cores) +{ + assert(preemption_enabled()); + + spl_t s = splsched(); + simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL); + + perfcontrol_requested_recommended_cores = recommended_cores; + perfcontrol_requested_recommended_core_count = __builtin_popcountll(recommended_cores); + + if ((perfcontrol_failsafe_active == false) && (perfcontrol_sleep_override == false)) { + sched_update_recommended_cores(perfcontrol_requested_recommended_cores & usercontrol_requested_recommended_cores); + } else { + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_REC_CORES_FAILSAFE) | DBG_FUNC_NONE, + perfcontrol_requested_recommended_cores, + sched_maintenance_thread->last_made_runnable_time, 0, 0, 0); + } + + simple_unlock(&sched_recommended_cores_lock); + splx(s); +} + +void +sched_override_recommended_cores_for_sleep(void) +{ + spl_t s = splsched(); + simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL); + + if (perfcontrol_sleep_override == false) { + perfcontrol_sleep_override = true; + sched_update_recommended_cores(ALL_CORES_RECOMMENDED); + } + + simple_unlock(&sched_recommended_cores_lock); + splx(s); +} + +void +sched_restore_recommended_cores_after_sleep(void) +{ + spl_t s = splsched(); + simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL); + + if (perfcontrol_sleep_override == true) { + perfcontrol_sleep_override = false; + sched_update_recommended_cores(perfcontrol_requested_recommended_cores & usercontrol_requested_recommended_cores); + } + + simple_unlock(&sched_recommended_cores_lock); + splx(s); +} + +/* + * Consider whether we need to activate the recommended cores failsafe + * + * Called from quantum timer interrupt context of a realtime thread + * No scheduler locks are held, interrupts are disabled + */ +void +sched_consider_recommended_cores(uint64_t ctime, thread_t cur_thread) +{ + /* + * Check if a realtime thread is starving the system + * and bringing up non-recommended cores would help + * + * TODO: Is this the correct check for recommended == possible cores? + * TODO: Validate the checks without the relevant lock are OK. + */ + + if (__improbable(perfcontrol_failsafe_active == TRUE)) { + /* keep track of how long the responsible thread runs */ + + simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL); + + if (perfcontrol_failsafe_active == TRUE && + cur_thread->thread_id == perfcontrol_failsafe_tid) { + perfcontrol_failsafe_thread_timer_last_seen = timer_grab(&cur_thread->user_timer) + + timer_grab(&cur_thread->system_timer); + } + + simple_unlock(&sched_recommended_cores_lock); + + /* we're already trying to solve the problem, so bail */ + return; + } + + /* The failsafe won't help if there are no more processors to enable */ + if (__probable(perfcontrol_requested_recommended_core_count >= processor_count)) { + return; + } + + uint64_t too_long_ago = ctime - perfcontrol_failsafe_starvation_threshold; + + /* Use the maintenance thread as our canary in the coal mine */ + thread_t m_thread = sched_maintenance_thread; + + /* If it doesn't look bad, nothing to see here */ + if (__probable(m_thread->last_made_runnable_time >= too_long_ago)) { + return; + } + + /* It looks bad, take the lock to be sure */ + thread_lock(m_thread); + + if (m_thread->runq == PROCESSOR_NULL || + (m_thread->state & (TH_RUN | TH_WAIT)) != TH_RUN || + m_thread->last_made_runnable_time >= too_long_ago) { + /* + * Maintenance thread is either on cpu or blocked, and + * therefore wouldn't benefit from more cores + */ + thread_unlock(m_thread); + return; + } + + uint64_t maintenance_runnable_time = m_thread->last_made_runnable_time; + + thread_unlock(m_thread); + + /* + * There are cores disabled at perfcontrol's recommendation, but the + * system is so overloaded that the maintenance thread can't run. + * That likely means that perfcontrol can't run either, so it can't fix + * the recommendation. We have to kick in a failsafe to keep from starving. + * + * When the maintenance thread has been starved for too long, + * ignore the recommendation from perfcontrol and light up all the cores. + * + * TODO: Consider weird states like boot, sleep, or debugger + */ + + simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL); + + if (perfcontrol_failsafe_active == TRUE) { + simple_unlock(&sched_recommended_cores_lock); + return; + } + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_REC_CORES_FAILSAFE) | DBG_FUNC_START, + perfcontrol_requested_recommended_cores, maintenance_runnable_time, 0, 0, 0); + + perfcontrol_failsafe_active = TRUE; + perfcontrol_failsafe_activation_time = mach_absolute_time(); + perfcontrol_failsafe_maintenance_runnable_time = maintenance_runnable_time; + perfcontrol_failsafe_recommended_at_trigger = perfcontrol_requested_recommended_cores; + + /* Capture some data about who screwed up (assuming that the thread on core is at fault) */ + task_t task = cur_thread->task; + perfcontrol_failsafe_pid = task_pid(task); + strlcpy(perfcontrol_failsafe_name, proc_name_address(task->bsd_info), sizeof(perfcontrol_failsafe_name)); + + perfcontrol_failsafe_tid = cur_thread->thread_id; + + /* Blame the thread for time it has run recently */ + uint64_t recent_computation = (ctime - cur_thread->computation_epoch) + cur_thread->computation_metered; + + uint64_t last_seen = timer_grab(&cur_thread->user_timer) + timer_grab(&cur_thread->system_timer); + + /* Compute the start time of the bad behavior in terms of the thread's on core time */ + perfcontrol_failsafe_thread_timer_at_start = last_seen - recent_computation; + perfcontrol_failsafe_thread_timer_last_seen = last_seen; + + /* Ignore the previously recommended core configuration */ + sched_update_recommended_cores(ALL_CORES_RECOMMENDED); + + simple_unlock(&sched_recommended_cores_lock); +} + +/* + * Now that our bacon has been saved by the failsafe, consider whether to turn it off + * + * Runs in the context of the maintenance thread, no locks held + */ +static void +sched_recommended_cores_maintenance(void) +{ + /* Common case - no failsafe, nothing to be done here */ + if (__probable(perfcontrol_failsafe_active == FALSE)) { + return; + } + + uint64_t ctime = mach_absolute_time(); + + boolean_t print_diagnostic = FALSE; + char p_name[FAILSAFE_NAME_LEN] = ""; + + spl_t s = splsched(); + simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL); + + /* Check again, under the lock, to avoid races */ + if (perfcontrol_failsafe_active == FALSE) { + goto out; + } + + /* + * Ensure that the other cores get another few ticks to run some threads + * If we don't have this hysteresis, the maintenance thread is the first + * to run, and then it immediately kills the other cores + */ + if ((ctime - perfcontrol_failsafe_activation_time) < perfcontrol_failsafe_starvation_threshold) { + goto out; + } + + /* Capture some diagnostic state under the lock so we can print it out later */ + + int pid = perfcontrol_failsafe_pid; + uint64_t tid = perfcontrol_failsafe_tid; + + uint64_t thread_usage = perfcontrol_failsafe_thread_timer_last_seen - + perfcontrol_failsafe_thread_timer_at_start; + uint32_t rec_cores_before = perfcontrol_failsafe_recommended_at_trigger; + uint32_t rec_cores_after = perfcontrol_requested_recommended_cores; + uint64_t failsafe_duration = ctime - perfcontrol_failsafe_activation_time; + strlcpy(p_name, perfcontrol_failsafe_name, sizeof(p_name)); + + print_diagnostic = TRUE; + + /* Deactivate the failsafe and reinstate the requested recommendation settings */ + + perfcontrol_failsafe_deactivation_time = ctime; + perfcontrol_failsafe_active = FALSE; + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_REC_CORES_FAILSAFE) | DBG_FUNC_END, + perfcontrol_requested_recommended_cores, failsafe_duration, 0, 0, 0); + + sched_update_recommended_cores(perfcontrol_requested_recommended_cores & usercontrol_requested_recommended_cores); + +out: + simple_unlock(&sched_recommended_cores_lock); + splx(s); + + if (print_diagnostic) { + uint64_t failsafe_duration_ms = 0, thread_usage_ms = 0; + + absolutetime_to_nanoseconds(failsafe_duration, &failsafe_duration_ms); + failsafe_duration_ms = failsafe_duration_ms / NSEC_PER_MSEC; + + absolutetime_to_nanoseconds(thread_usage, &thread_usage_ms); + thread_usage_ms = thread_usage_ms / NSEC_PER_MSEC; + + printf("recommended core failsafe kicked in for %lld ms " + "likely due to %s[%d] thread 0x%llx spending " + "%lld ms on cpu at realtime priority - " + "new recommendation: 0x%x -> 0x%x\n", + failsafe_duration_ms, p_name, pid, tid, thread_usage_ms, + rec_cores_before, rec_cores_after); + } +} + +#endif /* __arm__ || __arm64__ */ + +kern_return_t +sched_processor_enable(processor_t processor, boolean_t enable) { - return ((thread->state & (TH_RUN|TH_WAIT)) == TH_RUN); + assert(preemption_enabled()); + + spl_t s = splsched(); + simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL); + + if (enable) { + bit_set(usercontrol_requested_recommended_cores, processor->cpu_id); + } else { + bit_clear(usercontrol_requested_recommended_cores, processor->cpu_id); + } + +#if __arm__ || __arm64__ + if ((perfcontrol_failsafe_active == false) && (perfcontrol_sleep_override == false)) { + sched_update_recommended_cores(perfcontrol_requested_recommended_cores & usercontrol_requested_recommended_cores); + } else { + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED, MACH_REC_CORES_FAILSAFE) | DBG_FUNC_NONE, + perfcontrol_requested_recommended_cores, + sched_maintenance_thread->last_made_runnable_time, 0, 0, 0); + } +#else /* __arm__ || __arm64__ */ + sched_update_recommended_cores(usercontrol_requested_recommended_cores); +#endif /* !__arm__ || __arm64__ */ + + simple_unlock(&sched_recommended_cores_lock); + splx(s); + + return KERN_SUCCESS; } -#endif /* DEBUG */ -#if MACH_KDB -#include -#define printf kdbprintf -void db_sched(void); + +/* + * Apply a new recommended cores mask to the processors it affects + * Runs after considering failsafes and such + * + * Iterate over processors and update their ->is_recommended field. + * If a processor is running, we let it drain out at its next + * quantum expiration or blocking point. If a processor is idle, there + * may be more work for it to do, so IPI it. + * + * interrupts disabled, sched_recommended_cores_lock is held + */ +static void +sched_update_recommended_cores(uint64_t recommended_cores) +{ + processor_set_t pset, nset; + processor_t processor; + uint64_t needs_exit_idle_mask = 0x0; + uint32_t avail_count; + + processor = processor_list; + pset = processor->processor_set; + + KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_UPDATE_REC_CORES) | DBG_FUNC_START, + recommended_cores, +#if __arm__ || __arm64__ + perfcontrol_failsafe_active, 0, 0); +#else /* __arm__ || __arm64__ */ + 0, 0, 0); +#endif /* ! __arm__ || __arm64__ */ + + if (__builtin_popcountll(recommended_cores) == 0) { + bit_set(recommended_cores, master_processor->cpu_id); /* add boot processor or we hang */ + } + + boolean_t pset_newly_recommended = false; + + /* First set recommended cores */ + pset_lock(pset); + avail_count = 0; + do { + nset = processor->processor_set; + if (nset != pset) { + pset_unlock(pset); + pset = nset; + pset_newly_recommended = false; + pset_lock(pset); + } + + if (bit_test(recommended_cores, processor->cpu_id)) { + processor->is_recommended = TRUE; + if (bit_first(pset->recommended_bitmask) == -1) { + pset_newly_recommended = true; + } + bit_set(pset->recommended_bitmask, processor->cpu_id); + + if (processor->state == PROCESSOR_IDLE) { + if (processor != current_processor()) { + bit_set(needs_exit_idle_mask, processor->cpu_id); + } + } + if (processor->state != PROCESSOR_OFF_LINE) { + avail_count++; + } + if (pset_newly_recommended) { + SCHED(pset_made_schedulable)(processor, pset, false); + } + } + } while ((processor = processor->processor_list) != NULL); + pset_unlock(pset); + + /* Now shutdown not recommended cores */ + processor = processor_list; + pset = processor->processor_set; + + pset_lock(pset); + do { + nset = processor->processor_set; + if (nset != pset) { + pset_unlock(pset); + pset = nset; + pset_lock(pset); + } + + if (!bit_test(recommended_cores, processor->cpu_id)) { + sched_ipi_type_t ipi_type = SCHED_IPI_NONE; + + processor->is_recommended = FALSE; + bit_clear(pset->recommended_bitmask, processor->cpu_id); + + if ((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_DISPATCHING)) { + ipi_type = SCHED_IPI_IMMEDIATE; + } + SCHED(processor_queue_shutdown)(processor); + /* pset unlocked */ + + SCHED(rt_queue_shutdown)(processor); + + if (ipi_type != SCHED_IPI_NONE) { + if (processor == current_processor()) { + ast_on(AST_PREEMPT); + } else { + sched_ipi_perform(processor, ipi_type); + } + } + + pset_lock(pset); + } + } while ((processor = processor->processor_list) != NULL); + + processor_avail_count_user = avail_count; +#if defined(__x86_64__) + commpage_update_active_cpus(); +#endif + + pset_unlock(pset); + + /* Issue all pending IPIs now that the pset lock has been dropped */ + for (int cpuid = lsb_first(needs_exit_idle_mask); cpuid >= 0; cpuid = lsb_next(needs_exit_idle_mask, cpuid)) { + processor = processor_array[cpuid]; + machine_signal_idle(processor); + } + + KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_UPDATE_REC_CORES) | DBG_FUNC_END, + needs_exit_idle_mask, 0, 0, 0); +} + +void +thread_set_options(uint32_t thopt) +{ + spl_t x; + thread_t t = current_thread(); + + x = splsched(); + thread_lock(t); + + t->options |= thopt; + + thread_unlock(t); + splx(x); +} void -db_sched(void) -{ - iprintf("Scheduling Statistics:\n"); - db_indent += 2; - iprintf("Thread invocations: csw %d same %d\n", - c_thread_invoke_csw, c_thread_invoke_same); -#if MACH_COUNTERS - iprintf("Thread block: calls %d\n", - c_thread_block_calls); - iprintf("Idle thread:\n\thandoff %d block %d\n", - c_idle_thread_handoff, - c_idle_thread_block); - iprintf("Sched thread blocks: %d\n", c_sched_thread_block); -#endif /* MACH_COUNTERS */ - db_indent -= 2; -} - -#include -void db_show_thread_log(void); +thread_set_pending_block_hint(thread_t thread, block_hint_t block_hint) +{ + thread->pending_block_hint = block_hint; +} + +uint32_t +qos_max_parallelism(int qos, uint64_t options) +{ + return SCHED(qos_max_parallelism)(qos, options); +} + +uint32_t +sched_qos_max_parallelism(__unused int qos, uint64_t options) +{ + host_basic_info_data_t hinfo; + mach_msg_type_number_t count = HOST_BASIC_INFO_COUNT; + /* Query the machine layer for core information */ + __assert_only kern_return_t kret = host_info(host_self(), HOST_BASIC_INFO, + (host_info_t)&hinfo, &count); + assert(kret == KERN_SUCCESS); + + if (options & QOS_PARALLELISM_COUNT_LOGICAL) { + return hinfo.logical_cpu; + } else { + return hinfo.physical_cpu; + } +} + +int sched_allow_NO_SMT_threads = 1; +bool +thread_no_smt(thread_t thread) +{ +#if DEBUG || DEVELOPMENT + return sched_allow_NO_SMT_threads && (thread->bound_processor == PROCESSOR_NULL) && ((thread->sched_flags & TH_SFLAG_NO_SMT) || (thread->task->t_flags & TF_NO_SMT)); +#else + return sched_allow_NO_SMT_threads && (thread->bound_processor == PROCESSOR_NULL) && (thread->sched_flags & TH_SFLAG_NO_SMT); +#endif +} + +bool +processor_active_thread_no_smt(processor_t processor) +{ + return sched_allow_NO_SMT_threads && !processor->current_is_bound && processor->current_is_NO_SMT; +} + +#if __arm64__ + +/* + * Set up or replace old timer with new timer + * + * Returns true if canceled old timer, false if it did not + */ +boolean_t +sched_perfcontrol_update_callback_deadline(uint64_t new_deadline) +{ + /* + * Exchange deadline for new deadline, if old deadline was nonzero, + * then I cancelled the callback, otherwise I didn't + */ + + return os_atomic_xchg(&sched_perfcontrol_callback_deadline, new_deadline, + relaxed) != 0; +} + +#endif /* __arm64__ */ void -db_show_thread_log(void) +sched_update_pset_load_average(processor_set_t pset) { +#if CONFIG_SCHED_CLUTCH + int non_rt_load = sched_clutch_root_count(&pset->pset_clutch_root); +#else /* CONFIG_SCHED_CLUTCH */ + int non_rt_load = pset->pset_runq.count; +#endif /* CONFIG_SCHED_CLUTCH */ + + int load = ((bit_count(pset->cpu_state_map[PROCESSOR_RUNNING]) + non_rt_load + rt_runq_count(pset)) << PSET_LOAD_NUMERATOR_SHIFT); + int new_load_average = (pset->load_average + load) >> 1; + + pset->load_average = new_load_average; + +#if (DEVELOPMENT || DEBUG) +#if __AMP__ + if (pset->pset_cluster_type == PSET_AMP_P) { + KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_PSET_LOAD_AVERAGE) | DBG_FUNC_NONE, sched_get_pset_load_average(pset), (bit_count(pset->cpu_state_map[PROCESSOR_RUNNING]) + pset->pset_runq.count + rt_runq_count(pset))); + } +#endif +#endif +} + +/* pset is locked */ +static processor_t +choose_processor_for_realtime_thread(processor_set_t pset) +{ +#if defined(__x86_64__) + bool avoid_cpu0 = sched_avoid_cpu0 && bit_test(pset->cpu_bitmask, 0); +#else + const bool avoid_cpu0 = false; +#endif + + uint64_t cpu_map = (pset->cpu_bitmask & pset->recommended_bitmask & ~pset->pending_AST_URGENT_cpu_mask); + if (avoid_cpu0) { + cpu_map = bit_ror64(cpu_map, 1); + } + + for (int rotid = lsb_first(cpu_map); rotid >= 0; rotid = lsb_next(cpu_map, rotid)) { + int cpuid = avoid_cpu0 ? ((rotid + 1) & 63) : rotid; + + processor_t processor = processor_array[cpuid]; + + if (processor->processor_primary != processor) { + continue; + } + + if (processor->state == PROCESSOR_IDLE) { + return processor; + } + + if ((processor->state != PROCESSOR_RUNNING) && (processor->state != PROCESSOR_DISPATCHING)) { + continue; + } + + if (processor->current_pri >= BASEPRI_RTQUEUES) { + continue; + } + + return processor; + } + + if (!sched_allow_rt_smt) { + return PROCESSOR_NULL; + } + + /* Consider secondary processors */ + if (avoid_cpu0) { + /* Also avoid cpu1 */ + cpu_map = bit_ror64(cpu_map, 1); + } + for (int rotid = lsb_first(cpu_map); rotid >= 0; rotid = lsb_next(cpu_map, rotid)) { + int cpuid = avoid_cpu0 ? ((rotid + 2) & 63) : rotid; + + processor_t processor = processor_array[cpuid]; + + if (processor->processor_primary == processor) { + continue; + } + + if (processor->state == PROCESSOR_IDLE) { + return processor; + } + + if ((processor->state != PROCESSOR_RUNNING) && (processor->state != PROCESSOR_DISPATCHING)) { + continue; + } + + if (processor->current_pri >= BASEPRI_RTQUEUES) { + continue; + } + + return processor; + } + + return PROCESSOR_NULL; +} + +/* pset is locked */ +static bool +all_available_primaries_are_running_realtime_threads(processor_set_t pset) +{ + return these_processors_are_running_realtime_threads(pset, pset->primary_map); +} + +/* pset is locked */ +static bool +these_processors_are_running_realtime_threads(processor_set_t pset, uint64_t these_map) +{ + uint64_t cpu_map = (pset->cpu_bitmask & pset->recommended_bitmask) & these_map; + + for (int cpuid = lsb_first(cpu_map); cpuid >= 0; cpuid = lsb_next(cpu_map, cpuid)) { + processor_t processor = processor_array[cpuid]; + + if (processor->state == PROCESSOR_IDLE) { + return false; + } + + if (processor->state == PROCESSOR_DISPATCHING) { + return false; + } + + if (processor->state != PROCESSOR_RUNNING) { + /* + * All other processor states are considered unavailable to run + * realtime threads. In particular, we prefer an available secondary + * processor over the risk of leaving a realtime thread on the run queue + * while waiting for a processor in PROCESSOR_START state, + * which should anyway be a rare case. + */ + continue; + } + + if (processor->current_pri < BASEPRI_RTQUEUES) { + return false; + } + } + + return true; +} + +static bool +sched_ok_to_run_realtime_thread(processor_set_t pset, processor_t processor) +{ + bool ok_to_run_realtime_thread = true; +#if defined(__x86_64__) + if (sched_avoid_cpu0 && processor->cpu_id == 0) { + ok_to_run_realtime_thread = these_processors_are_running_realtime_threads(pset, pset->primary_map & ~0x1); + } else if (sched_avoid_cpu0 && (processor->cpu_id == 1) && processor->is_SMT) { + ok_to_run_realtime_thread = sched_allow_rt_smt && these_processors_are_running_realtime_threads(pset, ~0x2); + } else if (processor->processor_primary != processor) { + ok_to_run_realtime_thread = sched_allow_rt_smt && all_available_primaries_are_running_realtime_threads(pset); + } +#else + (void)pset; + (void)processor; +#endif + return ok_to_run_realtime_thread; +} + +void +sched_pset_made_schedulable(__unused processor_t processor, processor_set_t pset, boolean_t drop_lock) +{ + if (drop_lock) { + pset_unlock(pset); + } +} + +void +thread_set_no_smt(bool set) +{ + thread_t thread = current_thread(); + + spl_t s = splsched(); + thread_lock(thread); + if (set) { + thread->sched_flags |= TH_SFLAG_NO_SMT; + } else { + thread->sched_flags &= ~TH_SFLAG_NO_SMT; + } + thread_unlock(thread); + splx(s); +} + +bool +thread_get_no_smt(void) +{ + return current_thread()->sched_flags & TH_SFLAG_NO_SMT; +} + +#if DEBUG || DEVELOPMENT +extern void sysctl_task_set_no_smt(char no_smt); +void +sysctl_task_set_no_smt(char no_smt) +{ + thread_t thread = current_thread(); + task_t task = thread->task; + + if (no_smt == '1') { + task->t_flags |= TF_NO_SMT; + } else { + task->t_flags &= ~TF_NO_SMT; + } +} + +extern char sysctl_task_get_no_smt(void); +char +sysctl_task_get_no_smt(void) +{ + thread_t thread = current_thread(); + task_t task = thread->task; + + if (task->t_flags & TF_NO_SMT) { + return '1'; + } + return '0'; +} +#endif /* DEVELOPMENT || DEBUG */ + + +__private_extern__ void +thread_bind_cluster_type(char cluster_type) +{ +#if __AMP__ + thread_t thread = current_thread(); + + spl_t s = splsched(); + thread_lock(thread); + thread->sched_flags &= ~(TH_SFLAG_ECORE_ONLY | TH_SFLAG_PCORE_ONLY); + switch (cluster_type) { + case 'e': + case 'E': + thread->sched_flags |= TH_SFLAG_ECORE_ONLY; + break; + case 'p': + case 'P': + thread->sched_flags |= TH_SFLAG_PCORE_ONLY; + break; + default: + break; + } + thread_unlock(thread); + splx(s); + + thread_block(THREAD_CONTINUE_NULL); +#else /* __AMP__ */ + (void)cluster_type; +#endif /* __AMP__ */ } -#endif /* MACH_KDB */