X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/1c79356b52d46aa6b508fb032f5ae709b1f2897b..743345f9a4b36f7e2f9ba37691e70c50baecb56e:/osfmk/kern/sched_prim.c?ds=sidebyside diff --git a/osfmk/kern/sched_prim.c b/osfmk/kern/sched_prim.c index ae49e69ca..2b2a98d68 100644 --- a/osfmk/kern/sched_prim.c +++ b/osfmk/kern/sched_prim.c @@ -1,23 +1,29 @@ /* - * Copyright (c) 2000 Apple Computer, Inc. All rights reserved. + * Copyright (c) 2000-2016 Apple Inc. All rights reserved. * - * @APPLE_LICENSE_HEADER_START@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * - * The contents of this file constitute Original Code as defined in and - * are subject to the Apple Public Source License Version 1.1 (the - * "License"). You may not use this file except in compliance with the - * License. Please obtain a copy of the License at - * http://www.apple.com/publicsource and read it before using this file. + * 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 + * compliance with the License. The rights granted to you under the License + * may not be used to create, or enable the creation or redistribution of, + * 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. * - * This Original Code and all software distributed under the License are - * distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, EITHER + * 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, * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE OR NON-INFRINGEMENT. Please see the - * License for the specific language governing rights and limitations - * under the License. + * 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_LICENSE_HEADER_END@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ /* * @OSF_FREE_COPYRIGHT@ @@ -59,23 +65,30 @@ */ #include -#include -#include -#include -#include -#include -#include +#include #include +#include +#include +#include + #include #include -#include +#include +#include + +#ifdef CONFIG_MACH_APPROXIMATE_TIME +#include +#endif + +#include +#include #include #include #include #include -#include -#include +#include +#include #include #include #include @@ -83,457 +96,756 @@ #include #include #include +#include #include #include #include -#include +#include +#include +#include +#include + #include #include #include -#include -#include -#include -#include /*** ??? fix so this can be removed ***/ + +#include + #include +#include +#include -#if TASK_SWAPPER -#include -extern int task_swap_on; -#endif /* TASK_SWAPPER */ +#include -extern int hz; +struct rt_queue rt_runq; + +uintptr_t sched_thread_on_rt_queue = (uintptr_t)0xDEAFBEE0; + +/* Lock RT runq, must be done with interrupts disabled (under splsched()) */ +#if __SMP__ +decl_simple_lock_data(static,rt_lock); +#define rt_lock_init() simple_lock_init(&rt_lock, 0) +#define rt_lock_lock() simple_lock(&rt_lock) +#define rt_lock_unlock() simple_unlock(&rt_lock) +#else +#define rt_lock_init() do { } while(0) +#define rt_lock_lock() do { } while(0) +#define rt_lock_unlock() do { } while(0) +#endif -#define DEFAULT_PREEMPTION_RATE 100 /* (1/s) */ +#define DEFAULT_PREEMPTION_RATE 100 /* (1/s) */ int default_preemption_rate = DEFAULT_PREEMPTION_RATE; -#define NO_KERNEL_PREEMPT 0 -#define KERNEL_PREEMPT 1 -int kernel_preemption_mode = KERNEL_PREEMPT; +#define DEFAULT_BG_PREEMPTION_RATE 400 /* (1/s) */ +int default_bg_preemption_rate = DEFAULT_BG_PREEMPTION_RATE; -int min_quantum; -natural_t min_quantum_ms; +#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; + +#if defined(CONFIG_SCHED_TIMESHARE_CORE) + +uint32_t std_quantum; +uint32_t min_std_quantum; +uint32_t bg_quantum; + +uint32_t std_quantum_us; +uint32_t bg_quantum_us; + +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +uint32_t thread_depress_time; +uint32_t default_timeshare_computation; +uint32_t default_timeshare_constraint; + +uint32_t max_rt_quantum; +uint32_t min_rt_quantum; + +#if defined(CONFIG_SCHED_TIMESHARE_CORE) unsigned sched_tick; +uint32_t sched_tick_interval; + +uint32_t sched_pri_shifts[TH_BUCKET_MAX]; +uint32_t sched_fixed_shift; + +uint32_t sched_decay_usage_age_factor = 1; /* accelerate 5/8^n usage aging */ + +/* 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; + +/* 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; -#if SIMPLE_CLOCK -int sched_usec; -#endif /* SIMPLE_CLOCK */ +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +thread_t sched_maintenance_thread; + + +uint64_t sched_one_second_interval; /* Forwards */ -void thread_continue(thread_t); -void wait_queues_init(void); +#if defined(CONFIG_SCHED_TIMESHARE_CORE) + +static void load_shift_init(void); +static void preempt_pri_init(void); + +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +static thread_t thread_select( + thread_t thread, + processor_t processor, + ast_t reason); + +#if CONFIG_SCHED_IDLE_IN_PLACE +static thread_t thread_select_idle( + thread_t thread, + processor_t processor); +#endif -void set_pri( - thread_t thread, - int pri, - int resched); +thread_t processor_idle( + thread_t thread, + processor_t processor); -thread_t choose_pset_thread( - processor_t myprocessor, - processor_set_t pset); +ast_t +csw_check_locked( processor_t processor, + processor_set_t pset, + ast_t check_reason); -thread_t choose_thread( - processor_t myprocessor); +static void processor_setrun( + processor_t processor, + thread_t thread, + integer_t options); -int run_queue_enqueue( - run_queue_t runq, - thread_t thread, - boolean_t tail); +static void +sched_realtime_init(void); -void idle_thread_continue(void); -void do_thread_scan(void); +static void +sched_realtime_timebase_init(void); -void clear_wait_internal( - thread_t thread, - int result); +static void +sched_timer_deadline_tracking_init(void); #if DEBUG -void dump_run_queues( - run_queue_t rq); -void dump_run_queue_struct( - run_queue_t rq); -void dump_processor( - processor_t p); -void dump_processor_set( - processor_set_t ps); - -void checkrq( - run_queue_t rq, - char *msg); - -void thread_check( - thread_t thread, - run_queue_t runq); -#endif /*DEBUG*/ - -boolean_t thread_runnable( - thread_t thread); +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 -/* - * 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 - * - */ +static processor_t +thread_bind_internal( + thread_t thread, + processor_t processor); + +static void +sched_vm_group_maintenance(void); + +#if defined(CONFIG_SCHED_TIMESHARE_CORE) +int8_t sched_load_shifts[NRQS]; +bitmap_t sched_preempt_pri[BITMAP_LEN(NRQS)]; +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +const struct sched_dispatch_table *sched_current_dispatch = NULL; /* - * Waiting protocols and implementation: - * - * Each thread may be waiting for exactly one event; this event - * is set using assert_wait(). That thread may be awakened either - * by performing a thread_wakeup_prim() on its event, - * or by directly waking that thread up with clear_wait(). + * 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. * - * The implementation of wait events uses a hash table. Each - * bucket is queue of threads having the same hash function - * value; the chain for the queue (linked list) is the run queue - * field. [It is not possible to be waiting and runnable at the - * same time.] - * - * Locks on both the thread and on the hash buckets govern the - * wait event field and the queue chain field. Because wakeup - * operations only have the event as an argument, the event hash - * bucket must be locked before any thread. - * - * Scheduling operations may also occur at interrupt level; therefore, - * interrupts below splsched() must be prevented when holding - * thread or hash bucket locks. - * - * The wait event hash table declarations are as follows: + * 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]; -#define NUMQUEUES 59 - -struct wait_queue wait_queues[NUMQUEUES]; +uint32_t sched_debug_flags; -#define wait_hash(event) \ - ((((int)(event) < 0)? ~(int)(event): (int)(event)) % NUMQUEUES) +/* Global flag which indicates whether Background Stepper Context is enabled */ +static int cpu_throttle_enabled = 1; void sched_init(void) +{ + char sched_arg[SCHED_STRING_MAX_LENGTH] = { '\0' }; + + /* Check for runtime selection of the scheduler algorithm */ + if (!PE_parse_boot_argn("sched", sched_arg, sizeof (sched_arg))) { + /* If no boot-args override, look in device tree */ + if (!PE_get_default("kern.sched", sched_arg, + SCHED_STRING_MAX_LENGTH)) { + sched_arg[0] = '\0'; + } + } + + + 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 (strlen(sched_arg) > 0) { + if (0) { + /* Allow pattern below */ +#if defined(CONFIG_SCHED_TRADITIONAL) + } else if (0 == strcmp(sched_arg, sched_traditional_dispatch.sched_name)) { + sched_current_dispatch = &sched_traditional_dispatch; + } else if (0 == strcmp(sched_arg, sched_traditional_with_pset_runqueue_dispatch.sched_name)) { + sched_current_dispatch = &sched_traditional_with_pset_runqueue_dispatch; +#endif +#if defined(CONFIG_SCHED_PROTO) + } else if (0 == strcmp(sched_arg, sched_proto_dispatch.sched_name)) { + sched_current_dispatch = &sched_proto_dispatch; +#endif +#if defined(CONFIG_SCHED_GRRR) + } else if (0 == strcmp(sched_arg, sched_grrr_dispatch.sched_name)) { + sched_current_dispatch = &sched_grrr_dispatch; +#endif +#if defined(CONFIG_SCHED_MULTIQ) + } else if (0 == strcmp(sched_arg, sched_multiq_dispatch.sched_name)) { + sched_current_dispatch = &sched_multiq_dispatch; + } else if (0 == strcmp(sched_arg, sched_dualq_dispatch.sched_name)) { + sched_current_dispatch = &sched_dualq_dispatch; +#endif + } else { +#if defined(CONFIG_SCHED_TRADITIONAL) + printf("Unrecognized scheduler algorithm: %s\n", sched_arg); + printf("Scheduler: Using instead: %s\n", sched_traditional_with_pset_runqueue_dispatch.sched_name); + sched_current_dispatch = &sched_traditional_with_pset_runqueue_dispatch; +#else + panic("Unrecognized scheduler algorithm: %s", sched_arg); +#endif + } + kprintf("Scheduler: Runtime selection of %s\n", SCHED(sched_name)); + } else { +#if defined(CONFIG_SCHED_MULTIQ) + sched_current_dispatch = &sched_multiq_dispatch; +#elif defined(CONFIG_SCHED_TRADITIONAL) + sched_current_dispatch = &sched_traditional_with_pset_runqueue_dispatch; +#elif defined(CONFIG_SCHED_PROTO) + sched_current_dispatch = &sched_proto_dispatch; +#elif defined(CONFIG_SCHED_GRRR) + sched_current_dispatch = &sched_grrr_dispatch; +#else +#error No default scheduler implementation +#endif + kprintf("Scheduler: Default of %s\n", SCHED(sched_name)); + } + + strlcpy(sched_string, SCHED(sched_name), sizeof(sched_string)); + + if (PE_parse_boot_argn("sched_debug", &sched_debug_flags, sizeof(sched_debug_flags))) { + kprintf("Scheduler: Debug flags 0x%08x\n", sched_debug_flags); + } + + SCHED(init)(); + sched_realtime_init(); + ast_init(); + 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 minimum quantum - * in ticks. + * Calculate the timeslicing quantum + * in us. */ if (default_preemption_rate < 1) default_preemption_rate = DEFAULT_PREEMPTION_RATE; - min_quantum = hz / default_preemption_rate; + std_quantum_us = (1000 * 1000) / default_preemption_rate; - /* - * Round up result (4/5) to an - * integral number of ticks. - */ - if (((hz * 10) / default_preemption_rate) - (min_quantum * 10) >= 5) - min_quantum++; - if (min_quantum < 1) - min_quantum = 1; + printf("standard timeslicing quantum is %d us\n", std_quantum_us); - min_quantum_ms = (1000 / hz) * min_quantum; + if (default_bg_preemption_rate < 1) + default_bg_preemption_rate = DEFAULT_BG_PREEMPTION_RATE; + bg_quantum_us = (1000 * 1000) / default_bg_preemption_rate; - printf("scheduling quantum is %d ms\n", min_quantum_ms); + printf("standard background quantum is %d us\n", bg_quantum_us); - wait_queues_init(); - pset_sys_bootstrap(); /* initialize processor mgmt. */ - processor_action(); + load_shift_init(); + preempt_pri_init(); sched_tick = 0; -#if SIMPLE_CLOCK - sched_usec = 0; -#endif /* SIMPLE_CLOCK */ - ast_init(); - sf_init(); } void -wait_queues_init(void) +sched_timeshare_timebase_init(void) { - register int i; + uint64_t abstime; + uint32_t shift; + + /* standard timeslicing quantum */ + clock_interval_to_absolutetime_interval( + std_quantum_us, NSEC_PER_USEC, &abstime); + assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); + std_quantum = (uint32_t)abstime; + + /* smallest remaining quantum (250 us) */ + clock_interval_to_absolutetime_interval(250, NSEC_PER_USEC, &abstime); + assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); + min_std_quantum = (uint32_t)abstime; + + /* quantum for background tasks */ + clock_interval_to_absolutetime_interval( + bg_quantum_us, NSEC_PER_USEC, &abstime); + assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); + bg_quantum = (uint32_t)abstime; + + /* scheduler tick interval */ + clock_interval_to_absolutetime_interval(USEC_PER_SEC >> SCHED_TICK_SHIFT, + NSEC_PER_USEC, &abstime); + assert((abstime >> 32) == 0 && (uint32_t)abstime != 0); + sched_tick_interval = (uint32_t)abstime; + + /* + * Compute conversion factor from usage to + * timesharing priorities with 5/8 ** n aging. + */ + abstime = (abstime * 5) / 3; + for (shift = 0; abstime > BASEPRI_DEFAULT; ++shift) + abstime >>= 1; + sched_fixed_shift = shift; + + 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; - for (i = 0; i < NUMQUEUES; i++) { - wait_queue_init(&wait_queues[i], SYNC_POLICY_FIFO); - } } -/* - * Thread timeout routine, called when timer expires. - */ -void -thread_timer_expire( - timer_call_param_t p0, - timer_call_param_t p1) +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +static void +sched_realtime_init(void) { - thread_t thread = p0; - spl_t s; + rt_lock_init(); + + rt_runq.count = 0; + queue_init(&rt_runq.queue); +} + +static void +sched_realtime_timebase_init(void) +{ + uint64_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) */ + 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; - s = splsched(); - wake_lock(thread); - if ( thread->wait_timer_is_set && - !timer_call_is_delayed(&thread->wait_timer, NULL) ) { - thread->wait_timer_active--; - thread->wait_timer_is_set = FALSE; - thread_lock(thread); - if (thread->active) - clear_wait_internal(thread, THREAD_TIMED_OUT); - thread_unlock(thread); - } - else - if (--thread->wait_timer_active == 0) - thread_wakeup_one(&thread->wait_timer_active); - wake_unlock(thread); - splx(s); } +#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( - natural_t interval, - natural_t scale_factor) +static void +load_shift_init(void) { - thread_t thread = current_thread(); - AbsoluteTime deadline; - spl_t s; + int8_t k, *p = sched_load_shifts; + uint32_t i, j; - s = splsched(); - wake_lock(thread); - thread_lock(thread); - if ((thread->state & TH_WAIT) != 0) { - clock_interval_to_deadline(interval, scale_factor, &deadline); - timer_call_enter(&thread->wait_timer, deadline); - assert(!thread->wait_timer_is_set); - 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); } - thread_unlock(thread); - wake_unlock(thread); - splx(s); + + 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; + } +} + +static void +preempt_pri_init(void) +{ + bitmap_t *p = sched_preempt_pri; + + 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); } +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +/* + * Thread wait timer expiration. + */ void -thread_set_timer_deadline( - AbsoluteTime deadline) +thread_timer_expire( + void *p0, + __unused void *p1) { - thread_t thread = current_thread(); + thread_t thread = p0; spl_t s; + assert_thread_magic(thread); + s = splsched(); - wake_lock(thread); thread_lock(thread); - if ((thread->state & TH_WAIT) != 0) { - timer_call_enter(&thread->wait_timer, deadline); - assert(!thread->wait_timer_is_set); - thread->wait_timer_active++; - thread->wait_timer_is_set = TRUE; + 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); - wake_unlock(thread); splx(s); } -void -thread_cancel_timer(void) +/* + * thread_unblock: + * + * Unblock thread on wake up. + * + * 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 = current_thread(); - spl_t s; + boolean_t ready_for_runq = FALSE; + thread_t cthread = current_thread(); + uint32_t new_run_count; - s = splsched(); - wake_lock(thread); + /* + * Set wait_result. + */ + thread->wait_result = wresult; + + /* + * Cancel pending wait timer. + */ if (thread->wait_timer_is_set) { if (timer_call_cancel(&thread->wait_timer)) thread->wait_timer_active--; thread->wait_timer_is_set = FALSE; } - wake_unlock(thread); - splx(s); -} -/* - * thread_depress_timeout: - * - * Timeout routine for priority depression. - */ -void -thread_depress_timeout( - thread_call_param_t p0, - thread_call_param_t p1) -{ - thread_t thread = p0; - sched_policy_t *policy; - spl_t s; + /* + * Update scheduling state: not waiting, + * set running. + */ + thread->state &= ~(TH_WAIT|TH_UNINT); - s = splsched(); - thread_lock(thread); - policy = policy_id_to_sched_policy(thread->policy); - thread_unlock(thread); - splx(s); + if (!(thread->state & TH_RUN)) { + thread->state |= TH_RUN; + thread->last_made_runnable_time = mach_approximate_time(); - if (policy != SCHED_POLICY_NULL) - policy->sp_ops.sp_thread_depress_timeout(policy, thread); + ready_for_runq = TRUE; - thread_deallocate(thread); -} + (*thread->sched_call)(SCHED_CALL_UNBLOCK, thread); -/* - * Set up thread timeout element when thread is created. - */ -void -thread_timer_setup( - thread_t thread) -{ - timer_call_setup(&thread->wait_timer, thread_timer_expire, thread); - thread->wait_timer_is_set = FALSE; - thread->wait_timer_active = 1; - thread->ref_count++; + /* Update the runnable thread count */ + new_run_count = sched_run_incr(thread); + } else { + /* + * Either the thread is idling in place on another processor, + * or it hasn't finished context switching yet. + */ +#if CONFIG_SCHED_IDLE_IN_PLACE + if (thread->state & TH_IDLE) { + processor_t processor = thread->last_processor; - thread_call_setup(&thread->depress_timer, thread_depress_timeout, thread); -} + if (processor != current_processor()) + machine_signal_idle(processor); + } +#else + assert((thread->state & TH_IDLE) == 0); +#endif + /* + * The run count is only dropped after the context switch completes + * and the thread is still waiting, so we should not run_incr here + */ + new_run_count = sched_run_buckets[TH_BUCKET_RUN]; + } -void -thread_timer_terminate(void) -{ - thread_t thread = current_thread(); - spl_t s; - s = splsched(); - wake_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; + /* + * Calculate deadline for real-time threads. + */ + if (thread->sched_mode == TH_MODE_REALTIME) { + uint64_t ctime; + + ctime = mach_absolute_time(); + thread->realtime.deadline = thread->realtime.constraint + ctime; } - thread->wait_timer_active--; + /* + * Clear old quantum, fail-safe computation, etc. + */ + thread->quantum_remaining = 0; + thread->computation_metered = 0; + thread->reason = AST_NONE; + + /* 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); - while (thread->wait_timer_active > 0) { - assert_wait((event_t)&thread->wait_timer_active, THREAD_UNINT); - wake_unlock(thread); - splx(s); + if (__improbable(aticontext && !(thread_get_tag_internal(thread) & THREAD_TAG_CALLOUT))) { + ledger_credit(thread->t_ledger, task_ledgers.interrupt_wakeups, 1); + DTRACE_SCHED2(iwakeup, struct thread *, thread, struct proc *, thread->task->bsd_info); - thread_block((void (*)(void)) 0); + uint64_t ttd = PROCESSOR_DATA(current_processor(), timer_call_ttd); - s = splsched(); - wake_lock(thread); + 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++; + } + + if (pidle) { + ledger_credit(thread->t_ledger, task_ledgers.platform_idle_wakeups, 1); + } + + } else if (thread_get_tag_internal(cthread) & THREAD_TAG_CALLOUT) { + if (cthread->callout_woken_from_icontext) { + ledger_credit(thread->t_ledger, task_ledgers.interrupt_wakeups, 1); + thread->thread_callout_interrupt_wakeups++; + if (cthread->callout_woken_from_platform_idle) { + ledger_credit(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; } - wake_unlock(thread); - splx(s); + 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); - thread_deallocate(thread); + DTRACE_SCHED2(wakeup, struct thread *, thread, struct proc *, thread->task->bsd_info); + + return (ready_for_runq); } /* - * Routine: thread_go_locked + * Routine: thread_go * Purpose: - * Start a thread running. + * Unblock and dispatch thread. * Conditions: * thread lock held, IPC locks may be held. * thread must have been pulled from wait queue under same lock hold. + * thread must have been waiting + * Returns: + * KERN_SUCCESS - Thread was set running + * + * TODO: This should return void */ -void -thread_go_locked( - thread_t thread, - int result) +kern_return_t +thread_go( + thread_t thread, + wait_result_t wresult) { - int state; - sched_policy_t *policy; - sf_return_t sfr; + assert_thread_magic(thread); assert(thread->at_safe_point == FALSE); - assert(thread->wait_event == NO_EVENT); - assert(thread->wait_queue == WAIT_QUEUE_NULL); - - if (thread->state & TH_WAIT) { - - thread->state &= ~(TH_WAIT|TH_UNINT); - if (!(thread->state & TH_RUN)) { - thread->state |= TH_RUN; -#if THREAD_SWAPPER - if (thread->state & TH_SWAPPED_OUT) - thread_swapin(thread->top_act, FALSE); - else -#endif /* THREAD_SWAPPER */ - { - policy = &sched_policy[thread->policy]; - sfr = policy->sp_ops.sp_thread_unblock(policy, thread); - assert(sfr == SF_SUCCESS); - } - } - thread->wait_result = result; - } + assert(thread->wait_event == NO_EVENT64); + assert(thread->waitq == NULL); - - /* - * The next few lines are a major hack. Hopefully this will get us - * around all of the scheduling framework hooha. We can't call - * sp_thread_unblock yet because we could still be finishing up the - * durn two stage block on another processor and thread_setrun - * could be called by s_t_u and we'll really be messed up then. - */ - /* Don't mess with this if we are still swapped out */ - if (!(thread->state & TH_SWAPPED_OUT)) - thread->sp_state = MK_SP_RUNNABLE; - -} + assert(!(thread->state & (TH_TERMINATE|TH_TERMINATE2))); + assert(thread->state & TH_WAIT); -void + + if (thread_unblock(thread, wresult)) { +#if SCHED_TRACE_THREAD_WAKEUPS + backtrace(&thread->thread_wakeup_bt[0], + (sizeof(thread->thread_wakeup_bt)/sizeof(uintptr_t))); +#endif + thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ); + } + + return (KERN_SUCCESS); +} + +/* + * Routine: thread_mark_wait_locked + * Purpose: + * Mark a thread as waiting. If, given the circumstances, + * it doesn't want to wait (i.e. already aborted), then + * indicate that in the return value. + * Conditions: + * at splsched() and thread is locked. + */ +__private_extern__ +wait_result_t thread_mark_wait_locked( - thread_t thread, - int interruptible) + thread_t thread, + wait_interrupt_t interruptible) { + boolean_t at_safe_point; - assert(thread == current_thread()); + assert(!(thread->state & (TH_WAIT|TH_IDLE|TH_UNINT|TH_TERMINATE2))); - thread->wait_result = -1; /* JMM - Needed for non-assert kernel */ - thread->state |= (interruptible && thread->interruptible) ? - TH_WAIT : (TH_WAIT | TH_UNINT); - thread->at_safe_point = (interruptible == THREAD_ABORTSAFE) && (thread->interruptible); - thread->sleep_stamp = sched_tick; -} + /* + * The thread may have certain types of interrupts/aborts masked + * off. Even if the wait location says these types of interrupts + * 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 = thread->options & TH_OPT_INTMASK; + at_safe_point = (interruptible == THREAD_ABORTSAFE); + 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); + thread->at_safe_point = at_safe_point; + return (thread->wait_result = THREAD_WAITING); + } + else + if (thread->sched_flags & TH_SFLAG_ABORTSAFELY) + thread->sched_flags &= ~TH_SFLAG_ABORTED_MASK; + + return (thread->wait_result = THREAD_INTERRUPTED); +} /* - * Routine: assert_wait_timeout + * Routine: thread_interrupt_level * Purpose: - * Assert that the thread intends to block, - * waiting for a timeout (no user known event). + * Set the maximum interruptible state for the + * current thread. The effective value of any + * interruptible flag passed into assert_wait + * will never exceed this. + * + * Useful for code that must not be interrupted, + * but which calls code that doesn't know that. + * Returns: + * The old interrupt level for the thread. */ -unsigned int assert_wait_timeout_event; - -void -assert_wait_timeout( - mach_msg_timeout_t msecs, - int interruptible) +__private_extern__ +wait_interrupt_t +thread_interrupt_level( + wait_interrupt_t new_level) { - spl_t s; + thread_t thread = current_thread(); + wait_interrupt_t result = thread->options & TH_OPT_INTMASK; + + thread->options = (thread->options & ~TH_OPT_INTMASK) | (new_level & TH_OPT_INTMASK); - assert_wait((event_t)&assert_wait_timeout_event, interruptible); - thread_set_timer(msecs, 1000*NSEC_PER_USEC); + return result; } /* @@ -548,7 +860,6 @@ assert_wait_possible(void) { thread_t thread; - extern unsigned int debug_mode; #if DEBUG if(debug_mode) return TRUE; /* Always succeed in debug mode */ @@ -556,7 +867,7 @@ assert_wait_possible(void) thread = current_thread(); - return (thread == NULL || wait_queue_assert_possible(thread)); + return (thread == NULL || waitq_wait_possible(thread)); } /* @@ -565,112 +876,346 @@ assert_wait_possible(void) * Assert that the current thread is about to go to * sleep until the specified event occurs. */ -void +wait_result_t assert_wait( event_t event, - int interruptible) + wait_interrupt_t interruptible) +{ + if (__improbable(event == NO_EVENT)) + panic("%s() called with NO_EVENT", __func__); + + 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); + + 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) +{ + 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__); + + struct waitq *waitq; + waitq = global_eventq(event); + + s = splsched(); + waitq_lock(waitq); + + clock_interval_to_deadline(interval, scale_factor, &deadline); + + 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; +} + +wait_result_t +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; + 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; + + clock_interval_to_absolutetime_interval(leeway, scale_factor, &slop); + + 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, slop, + thread); + + waitq_unlock(waitq); + splx(s); + return wresult; +} + +wait_result_t +assert_wait_deadline( + event_t event, + wait_interrupt_t interruptible, + uint64_t deadline) +{ + thread_t thread = current_thread(); + wait_result_t wresult; + spl_t s; + + if (__improbable(event == NO_EVENT)) + panic("%s() called with NO_EVENT", __func__); + + 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; +} + +wait_result_t +assert_wait_deadline_with_leeway( + event_t event, + wait_interrupt_t interruptible, + wait_timeout_urgency_t urgency, + uint64_t deadline, + uint64_t leeway) +{ + thread_t thread = current_thread(); + wait_result_t wresult; + spl_t s; + + if (__improbable(event == NO_EVENT)) + panic("%s() called with NO_EVENT", __func__); + + 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_isoncpu: + * + * 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. + */ +static inline boolean_t +thread_isoncpu(thread_t thread) { - register wait_queue_t wq; - register int index; + /* Not running or runnable */ + if (!(thread->state & TH_RUN)) + return (FALSE); - assert(event != NO_EVENT); - assert(assert_wait_possible()); + /* 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); + + /* + * 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); - index = wait_hash(event); - wq = &wait_queues[index]; - wait_queue_assert_wait(wq, - event, - interruptible); + /* + * 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_[un]stop(thread) - * Once a thread has blocked interruptibly (via assert_wait) prevent - * it from running until thread_unstop. + * thread_stop: * - * If someone else has already stopped the thread, wait for the - * stop to be cleared, and then stop it again. + * Force a preemption point for a thread and wait + * 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) * - * Return FALSE if interrupted. + * The thread must enter a wait state and stop via a + * separate means. * - * NOTE: thread_hold/thread_suspend should be called on the activation - * before calling thread_stop. TH_SUSP is only recognized when - * a thread blocks and only prevents clear_wait/thread_wakeup - * from restarting an interruptible wait. The wake_active flag is - * used to indicate that someone is waiting on the thread. + * Returns FALSE if interrupted. */ boolean_t thread_stop( - thread_t thread) + thread_t thread, + boolean_t until_not_runnable) { - spl_t s; + wait_result_t wresult; + spl_t s = splsched(); + boolean_t oncpu; - s = splsched(); wake_lock(thread); + thread_lock(thread); while (thread->state & TH_SUSP) { thread->wake_active = TRUE; - assert_wait((event_t)&thread->wake_active, THREAD_ABORTSAFE); + thread_unlock(thread); + + wresult = assert_wait(&thread->wake_active, THREAD_ABORTSAFE); wake_unlock(thread); splx(s); - thread_block((void (*)(void)) 0); - if (current_thread()->wait_result != THREAD_AWAKENED) + if (wresult == THREAD_WAITING) + wresult = thread_block(THREAD_CONTINUE_NULL); + + if (wresult != THREAD_AWAKENED) return (FALSE); s = splsched(); wake_lock(thread); + thread_lock(thread); } - thread_lock(thread); + thread->state |= TH_SUSP; - thread_unlock(thread); + while ((oncpu = thread_isoncpu(thread)) || + (until_not_runnable && (thread->state & TH_RUN))) { + processor_t processor; + + if (oncpu) { + assert(thread->state & TH_RUN); + processor = thread->chosen_processor; + cause_ast_check(processor); + } + + thread->wake_active = TRUE; + thread_unlock(thread); + + wresult = assert_wait(&thread->wake_active, THREAD_ABORTSAFE); + wake_unlock(thread); + splx(s); + + if (wresult == THREAD_WAITING) + wresult = thread_block(THREAD_CONTINUE_NULL); + + if (wresult != THREAD_AWAKENED) { + thread_unstop(thread); + return (FALSE); + } + + s = splsched(); + wake_lock(thread); + thread_lock(thread); + } + + thread_unlock(thread); wake_unlock(thread); splx(s); + + /* + * 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); } /* - * Clear TH_SUSP and if the thread has been stopped and is now runnable, - * put it back on the run queue. + * thread_unstop: + * + * Release a previous stop request and set + * the thread running if appropriate. + * + * Use only after a successful stop operation. */ void thread_unstop( - thread_t thread) + thread_t thread) { - sched_policy_t *policy; - sf_return_t sfr; - spl_t s; + spl_t s = splsched(); - s = splsched(); wake_lock(thread); thread_lock(thread); - if ((thread->state & (TH_RUN|TH_WAIT|TH_SUSP/*|TH_UNINT*/)) == TH_SUSP) { - thread->state = (thread->state & ~TH_SUSP) | TH_RUN; -#if THREAD_SWAPPER - if (thread->state & TH_SWAPPED_OUT) - thread_swapin(thread->top_act, FALSE); - else -#endif /* THREAD_SWAPPER */ - { - policy = &sched_policy[thread->policy]; - sfr = policy->sp_ops.sp_thread_unblock(policy, thread); - assert(sfr == SF_SUCCESS); - } - } - else + assert((thread->state & (TH_RUN|TH_WAIT|TH_SUSP)) != TH_SUSP); + if (thread->state & TH_SUSP) { thread->state &= ~TH_SUSP; if (thread->wake_active) { thread->wake_active = FALSE; thread_unlock(thread); + + thread_wakeup(&thread->wake_active); wake_unlock(thread); splx(s); - thread_wakeup((event_t)&thread->wake_active); return; } @@ -682,60 +1227,60 @@ thread_unstop( } /* - * Wait for the thread's RUN bit to clear + * thread_wait: + * + * Wait for a thread to stop running. (non-interruptible) + * */ -boolean_t +void thread_wait( - thread_t thread) + thread_t thread, + boolean_t until_not_runnable) { - spl_t s; + wait_result_t wresult; + boolean_t oncpu; + processor_t processor; + spl_t s = splsched(); - s = splsched(); wake_lock(thread); + thread_lock(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))) { - while (thread->state & (TH_RUN/*|TH_UNINT*/)) { - if (thread->last_processor != PROCESSOR_NULL) - cause_ast_check(thread->last_processor); + if (oncpu) { + assert(thread->state & TH_RUN); + processor = thread->chosen_processor; + cause_ast_check(processor); + } thread->wake_active = TRUE; - assert_wait((event_t)&thread->wake_active, THREAD_ABORTSAFE); + thread_unlock(thread); + + wresult = assert_wait(&thread->wake_active, THREAD_UNINT); wake_unlock(thread); splx(s); - thread_block((void (*)(void))0); - if (current_thread()->wait_result != THREAD_AWAKENED) - return (FALSE); + if (wresult == THREAD_WAITING) + thread_block(THREAD_CONTINUE_NULL); s = splsched(); wake_lock(thread); + thread_lock(thread); } + thread_unlock(thread); wake_unlock(thread); splx(s); - - return (TRUE); -} - - -/* - * thread_stop_wait(thread) - * Stop the thread then wait for it to block interruptibly - */ -boolean_t -thread_stop_wait( - thread_t thread) -{ - if (thread_stop(thread)) { - if (thread_wait(thread)) - return (TRUE); - - thread_unstop(thread); - } - - return (FALSE); } - /* * Routine: clear_wait_internal * @@ -747,23 +1292,47 @@ thread_stop_wait( * Conditions: * At splsched * the thread is locked. + * Returns: + * KERN_SUCCESS thread was rousted out a wait + * KERN_FAILURE thread was waiting but could not be rousted + * KERN_NOT_WAITING thread was not waiting */ -void +__private_extern__ kern_return_t clear_wait_internal( - thread_t thread, - int result) + thread_t thread, + wait_result_t wresult) { - /* - * If the thread isn't in a wait queue, just set it running. Otherwise, - * try to remove it from the queue and, if successful, then set it - * running. NEVER interrupt an uninterruptible thread. - */ - if (!((result == THREAD_INTERRUPTED) && (thread->state & TH_UNINT))) { - if (wait_queue_assert_possible(thread) || - (wait_queue_remove(thread) == KERN_SUCCESS)) { - thread_go_locked(thread, result); + uint32_t i = LockTimeOutUsec; + struct waitq *waitq = thread->waitq; + + do { + if (wresult == THREAD_INTERRUPTED && (thread->state & TH_UNINT)) + return (KERN_FAILURE); + + 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 (waitq != thread->waitq) + return KERN_NOT_WAITING; + continue; + } } - } + + /* 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, waitq, cpu_number()); + + return (KERN_FAILURE); } @@ -777,18 +1346,20 @@ clear_wait_internal( * thread thread to awaken * result Wakeup result the thread should see */ -void +kern_return_t clear_wait( - thread_t thread, - int result) + thread_t thread, + wait_result_t result) { + kern_return_t ret; spl_t s; s = splsched(); thread_lock(thread); - clear_wait_internal(thread, result); + ret = clear_wait_internal(thread, result); thread_unlock(thread); splx(s); + return ret; } @@ -799,1880 +1370,3508 @@ clear_wait( * and thread_wakeup_one. * */ -void +kern_return_t thread_wakeup_prim( - event_t event, - boolean_t one_thread, - int result) + event_t event, + boolean_t one_thread, + wait_result_t result) { - register wait_queue_t wq; - register int index; + if (__improbable(event == NO_EVENT)) + panic("%s() called with NO_EVENT", __func__); + + struct waitq *wq = global_eventq(event); - index = wait_hash(event); - wq = &wait_queues[index]; if (one_thread) - wait_queue_wakeup_one(wq, event, result); + return waitq_wakeup64_one(wq, CAST_EVENT64_T(event), result, WAITQ_ALL_PRIORITIES); else - wait_queue_wakeup_all(wq, event, result); + return waitq_wakeup64_all(wq, CAST_EVENT64_T(event), result, WAITQ_ALL_PRIORITIES); } /* - * thread_bind: - * - * Force a thread to execute on the specified processor. - * If the thread is currently executing, it may wait until its - * time slice is up before switching onto the specified processor. - * - * A processor of PROCESSOR_NULL causes the thread to be unbound. - * xxx - DO NOT export this to users. + * Wakeup a specified thread if and only if it's waiting for this event */ -void -thread_bind( - register thread_t thread, - processor_t processor) +kern_return_t +thread_wakeup_thread( + event_t event, + thread_t thread) { - spl_t s; + if (__improbable(event == NO_EVENT)) + panic("%s() called with NO_EVENT", __func__); - s = splsched(); - thread_lock(thread); - thread_bind_locked(thread, processor); - thread_unlock(thread); - splx(s); + struct waitq *wq = global_eventq(event); + + return waitq_wakeup64_thread(wq, CAST_EVENT64_T(event), thread, THREAD_AWAKENED); } /* - * Select a thread for this processor (the current processor) to run. - * May select the current thread, which must already be locked. + * Wakeup a thread waiting on an event and promote it to a priority. + * + * Requires woken thread to un-promote itself when done. */ -thread_t -thread_select( - register processor_t myprocessor) +kern_return_t +thread_wakeup_one_with_pri( + event_t event, + int priority) { - register thread_t thread; - processor_set_t pset; - register run_queue_t runq = &myprocessor->runq; - boolean_t other_runnable; - sched_policy_t *policy; - - /* - * Check for other non-idle runnable threads. - */ - myprocessor->first_quantum = TRUE; - pset = myprocessor->processor_set; - thread = current_thread(); + if (__improbable(event == NO_EVENT)) + panic("%s() called with NO_EVENT", __func__); -#if 0 /* CHECKME! */ - thread->unconsumed_quantum = myprocessor->quantum; -#endif + struct waitq *wq = global_eventq(event); - simple_lock(&runq->lock); - simple_lock(&pset->runq.lock); + return waitq_wakeup64_one(wq, CAST_EVENT64_T(event), THREAD_AWAKENED, priority); +} - other_runnable = runq->count > 0 || pset->runq.count > 0; +/* + * 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) +{ + if (__improbable(event == NO_EVENT)) + panic("%s() called with NO_EVENT", __func__); - if ( thread->state == TH_RUN && - (!other_runnable || - (runq->highq < thread->sched_pri && - pset->runq.highq < thread->sched_pri)) && - thread->processor_set == pset && - (thread->bound_processor == PROCESSOR_NULL || - thread->bound_processor == myprocessor) ) { + struct waitq *wq = global_eventq(event); - /* I am the highest priority runnable (non-idle) thread */ - simple_unlock(&pset->runq.lock); - simple_unlock(&runq->lock); + return waitq_wakeup64_identify(wq, CAST_EVENT64_T(event), THREAD_AWAKENED, priority); +} - /* Update the thread's meta-priority */ - policy = policy_id_to_sched_policy(thread->policy); - assert(policy != SCHED_POLICY_NULL); - (void)policy->sp_ops.sp_thread_update_mpri(policy, thread); - } - else - if (other_runnable) { - simple_unlock(&pset->runq.lock); - simple_unlock(&runq->lock); - thread = choose_thread(myprocessor); - } - else { - simple_unlock(&pset->runq.lock); - simple_unlock(&runq->lock); +/* + * 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. + * + * XXX - DO NOT export this to users - XXX + */ +processor_t +thread_bind( + processor_t processor) +{ + thread_t self = current_thread(); + processor_t prev; + spl_t s; - /* - * Nothing is runnable, so set this processor idle if it - * was running. If it was in an assignment or shutdown, - * leave it alone. Return its idle thread. - */ - simple_lock(&pset->idle_lock); - if (myprocessor->state == PROCESSOR_RUNNING) { - myprocessor->state = PROCESSOR_IDLE; - /* - * XXX Until it goes away, put master on end of queue, others - * XXX on front so master gets used last. - */ - if (myprocessor == master_processor) - queue_enter(&(pset->idle_queue), myprocessor, - processor_t, processor_queue); - else - queue_enter_first(&(pset->idle_queue), myprocessor, - processor_t, processor_queue); + s = splsched(); + thread_lock(self); - pset->idle_count++; - } - simple_unlock(&pset->idle_lock); + prev = thread_bind_internal(self, processor); - thread = myprocessor->idle_thread; - } + thread_unlock(self); + splx(s); - return (thread); + return (prev); } - /* - * Stop running the current thread and start running the new thread. - * If continuation is non-zero, and the current thread is blocked, - * then it will resume by executing continuation on a new stack. - * Returns TRUE if the hand-off succeeds. - * The reason parameter == AST_QUANTUM if the thread blocked - * because its quantum expired. - * Assumes splsched. + * thread_bind_internal: + * + * 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. + * + * 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 -__current_thread(void) +static processor_t +thread_bind_internal( + thread_t thread, + processor_t processor) { - return (current_thread()); + processor_t prev; + + /* */ + 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); + + 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); + + prev = thread->bound_processor; + thread->bound_processor = processor; + + return (prev); } -boolean_t -thread_invoke( - register thread_t old_thread, - register thread_t new_thread, - int reason, - void (*continuation)(void)) +/* + * 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). + */ + +/* + * 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; + +void +thread_vm_bind_group_add(void) { - sched_policy_t *policy; - sf_return_t sfr; - void (*lcont)(void); + thread_t self = current_thread(); - /* - * Mark thread interruptible. - */ - thread_lock(new_thread); - new_thread->state &= ~TH_UNINT; + thread_reference_internal(self); + self->options |= TH_OPT_SCHED_VM_GROUP; + + simple_lock(&sched_vm_group_list_lock); + 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); - if (cpu_data[cpu_number()].preemption_level != 1) - panic("thread_invoke: preemption_level %d\n", - cpu_data[cpu_number()].preemption_level); + thread_bind(master_processor); + /* Switch to bound processor if not already there */ + thread_block(THREAD_CONTINUE_NULL); +} + +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); - assert(thread_runnable(new_thread)); + s = splsched(); - assert(old_thread->continuation == (void (*)(void))0); + 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); - if ((old_thread->sched_mode & TH_MODE_REALTIME) && (!old_thread->stack_privilege)) { - old_thread->stack_privilege = old_thread->kernel_stack; + if (high_latency_observed && runnable_and_not_on_runq_observed) { + /* All the things we are looking for are true, stop looking */ + break; + } } - if (continuation != (void (*)()) 0) { - switch (new_thread->state & TH_STACK_STATE) { - case TH_STACK_HANDOFF: + splx(s); + 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 { /* - * If the old thread has stack privilege, we can't give - * his stack away. So go and get him one and treat this - * as a traditional context switch. + * 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 (old_thread->stack_privilege == current_stack()) - goto get_new_stack; + 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; + } + } + + 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); + + 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 */ + } + } + + if (removed) { + thread_run_queue_reinsert(thread, SCHED_PREEMPT | SCHED_TAILQ); + } + thread_unlock(thread); + } + splx(s); + } + + simple_unlock(&sched_vm_group_list_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 (DEVELOPMENT || DEBUG) +int sched_smt_balance = 1; +#endif + +#if __SMP__ +/* Invoked with pset locked, returns with pset unlocked */ +static void +sched_SMT_balance(processor_t cprocessor, processor_set_t cpset) { + processor_t ast_processor = NULL; + +#if (DEVELOPMENT || DEBUG) + if (__improbable(sched_smt_balance == 0)) + goto smt_balance_exit; +#endif + + assert(cprocessor == current_processor()); + if (cprocessor->is_SMT == FALSE) + goto smt_balance_exit; + + processor_t sib_processor = cprocessor->processor_secondary ? cprocessor->processor_secondary : cprocessor->processor_primary; + + /* 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; + + processor_t sprocessor; + + qe_foreach_element(sprocessor, &cpset->active_queue, processor_queue) { + if ((sprocessor->state == PROCESSOR_RUNNING) && + (sprocessor->processor_primary != sprocessor) && + (sprocessor->processor_primary->state == PROCESSOR_RUNNING) && + (sprocessor->current_pri < BASEPRI_RTQUEUES) && + ((cpset->pending_AST_cpu_mask & (1ULL << sprocessor->cpu_id)) == 0)) { + assert(sprocessor != cprocessor); + ast_processor = sprocessor; + break; + } + } + +smt_balance_exit: + pset_unlock(cpset); + 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); + cause_ast_check(ast_processor); + } +} +#endif /* __SMP__ */ + +/* + * 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; + + assert(processor == current_processor()); + assert((thread->state & (TH_RUN|TH_TERMINATE2)) == TH_RUN); + + do { /* - * Make the whole handoff/dispatch atomic to match the - * non-handoff case. + * Update the priority. */ - disable_preemption(); + if (SCHED(can_update_priority)(thread)) + SCHED(update_priority)(thread); + + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + + pset_lock(pset); + + 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 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) && !queue_empty(&pset->idle_queue) && !rt_runq.count) { + goto idle; + } + } + + rt_lock_lock(); /* - * Set up ast context of new thread and switch to its timer. + * 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. */ - new_thread->state &= ~(TH_STACK_HANDOFF|TH_UNINT); - new_thread->last_processor = current_processor(); - ast_context(new_thread->top_act, cpu_number()); - timer_switch(&new_thread->system_timer); - thread_unlock(new_thread); - old_thread->continuation = continuation; - stack_handoff(old_thread, new_thread); + if (((thread->state & (TH_TERMINATE|TH_IDLE|TH_WAIT|TH_RUN|TH_SUSP)) == TH_RUN) && + (thread->sched_pri >= BASEPRI_RTQUEUES || processor->processor_primary == processor) && + (thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == processor) && + (thread->affinity_set == AFFINITY_SET_NULL || thread->affinity_set->aset_pset == pset)) { + /* + * 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 > 0) { + thread_t next_rt = qe_queue_first(&rt_runq.queue, struct thread, runq_links); + + assert(next_rt->runq == THREAD_ON_RT_RUNQ); + + 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; + } + } + + /* This is still the best RT thread to run. */ + processor->deadline = thread->realtime.deadline; + + rt_lock_unlock(); + pset_unlock(pset); + + return (thread); + } + + if ((rt_runq.count == 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; + + rt_lock_unlock(); + pset_unlock(pset); + + return (thread); + } + } + + /* OK, so we're not going to run the current thread. Look at the RT queue. */ + if (rt_runq.count > 0) { + thread_t next_rt = qe_queue_first(&rt_runq.queue, struct thread, runq_links); + + assert(next_rt->runq == THREAD_ON_RT_RUNQ); + + if (__probable((next_rt->bound_processor == PROCESSOR_NULL || + (next_rt->bound_processor == processor)))) { +pick_new_rt_thread: + new_thread = qe_dequeue_head(&rt_runq.queue, struct thread, runq_links); + + new_thread->runq = PROCESSOR_NULL; + SCHED_STATS_RUNQ_CHANGE(&rt_runq.runq_stats, rt_runq.count); + rt_runq.count--; + + processor->deadline = new_thread->realtime.deadline; + + rt_lock_unlock(); + pset_unlock(pset); + + return (new_thread); + } + } + + processor->deadline = UINT64_MAX; + rt_lock_unlock(); + + /* No RT threads, so let's look at the regular threads. */ + if ((new_thread = SCHED(choose_thread)(processor, MINPRI, reason)) != THREAD_NULL) { + pset_unlock(pset); + return (new_thread); + } + +#if __SMP__ + if (SCHED(steal_thread_enabled)) { + /* + * No runnable threads, attempt to steal + * from other processors. Returns with pset lock dropped. + */ - wake_lock(old_thread); - thread_lock(old_thread); - act_machine_sv_free(old_thread->top_act); + if ((new_thread = SCHED(steal_thread)(pset)) != THREAD_NULL) { + return (new_thread); + } + + /* + * If other threads have appeared, shortcut + * around again. + */ + if (!SCHED(processor_queue_empty)(processor) || rt_runq.count > 0) + continue; - /* - * inline thread_dispatch but don't free stack + pset_lock(pset); + } +#endif + + idle: + /* + * Nothing is runnable, so set this processor idle if it + * was running. */ + if (processor->state == PROCESSOR_RUNNING) { + processor->state = PROCESSOR_IDLE; - switch (old_thread->state & (TH_RUN|TH_WAIT|TH_UNINT|TH_IDLE)) { - sched_policy_t *policy; - sf_return_t sfr; - - case TH_RUN | TH_UNINT: - case TH_RUN: - /* - * No reason to stop. Put back on a run queue. - */ - old_thread->state |= TH_STACK_HANDOFF; - - /* Get pointer to scheduling policy "object" */ - policy = &sched_policy[old_thread->policy]; - - /* Leave enqueueing thread up to scheduling policy */ - sfr = policy->sp_ops.sp_thread_dispatch(policy, old_thread); - assert(sfr == SF_SUCCESS); - break; - - case TH_RUN | TH_WAIT | TH_UNINT: - case TH_RUN | TH_WAIT: - old_thread->sleep_stamp = sched_tick; - /* fallthrough */ - - case TH_WAIT: /* this happens! */ - /* - * Waiting - */ - old_thread->state |= TH_STACK_HANDOFF; - old_thread->state &= ~TH_RUN; - if (old_thread->state & TH_TERMINATE) - thread_reaper_enqueue(old_thread); - - if (old_thread->wake_active) { - old_thread->wake_active = FALSE; - thread_unlock(old_thread); - wake_unlock(old_thread); - thread_wakeup((event_t)&old_thread->wake_active); - wake_lock(old_thread); - thread_lock(old_thread); - } - break; - - case TH_RUN | TH_IDLE: - /* - * Drop idle thread -- it is already in - * idle_thread_array. - */ - old_thread->state |= TH_STACK_HANDOFF; - break; - - default: - panic("State 0x%x \n",old_thread->state); - } - - /* Get pointer to scheduling policy "object" */ - policy = &sched_policy[old_thread->policy]; - - /* Indicate to sched policy that old thread has stopped execution */ - /*** ??? maybe use a macro -- rkc, 1/4/96 ***/ - sfr = policy->sp_ops.sp_thread_done(policy, old_thread); - assert(sfr == SF_SUCCESS); - thread_unlock(old_thread); - wake_unlock(old_thread); - thread_lock(new_thread); - - assert(thread_runnable(new_thread)); - - /* Get pointer to scheduling policy "object" */ - policy = &sched_policy[new_thread->policy]; - - /* Indicate to sched policy that new thread has started execution */ - /*** ??? maybe use a macro ***/ - sfr = policy->sp_ops.sp_thread_begin(policy, new_thread); - assert(sfr == SF_SUCCESS); - - lcont = new_thread->continuation; - new_thread->continuation = (void(*)(void))0; - - thread_unlock(new_thread); - enable_preemption(); - - counter_always(c_thread_invoke_hits++); - - if (new_thread->funnel_state & TH_FN_REFUNNEL) { - kern_return_t save_wait_result; - new_thread->funnel_state = 0; - save_wait_result = new_thread->wait_result; - KERNEL_DEBUG(0x6032428 | DBG_FUNC_NONE, new_thread->funnel_lock, 2, 0, 0, 0); - //mutex_lock(new_thread->funnel_lock); - funnel_lock(new_thread->funnel_lock); - KERNEL_DEBUG(0x6032430 | DBG_FUNC_NONE, new_thread->funnel_lock, 2, 0, 0, 0); - new_thread->funnel_state = TH_FN_OWNED; - new_thread->wait_result = save_wait_result; + if (processor->processor_primary == processor) { + re_queue_head(&pset->idle_queue, &processor->processor_queue); + } else { + re_queue_head(&pset->idle_secondary_queue, &processor->processor_queue); + } } - (void) spllo(); - assert(lcont); - call_continuation(lcont); - /*NOTREACHED*/ - return TRUE; +#if __SMP__ + /* Invoked with pset locked, returns with pset unlocked */ + sched_SMT_balance(processor, pset); +#else + pset_unlock(pset); +#endif - case TH_STACK_COMING_IN: +#if CONFIG_SCHED_IDLE_IN_PLACE /* - * waiting for a stack + * Choose idle thread if fast idle is not possible. */ - thread_swapin(new_thread); - thread_unlock(new_thread); - counter_always(c_thread_invoke_misses++); - return FALSE; + if (processor->processor_primary != processor) + return (processor->idle_thread); - case 0: - /* - * already has a stack - can't handoff - */ - if (new_thread == old_thread) { - - /* same thread but with continuation */ - counter(++c_thread_invoke_same); - thread_unlock(new_thread); - - if (old_thread->funnel_state & TH_FN_REFUNNEL) { - kern_return_t save_wait_result; - - old_thread->funnel_state = 0; - save_wait_result = old_thread->wait_result; - KERNEL_DEBUG(0x6032428 | DBG_FUNC_NONE, old_thread->funnel_lock, 3, 0, 0, 0); - funnel_lock(old_thread->funnel_lock); - KERNEL_DEBUG(0x6032430 | DBG_FUNC_NONE, old_thread->funnel_lock, 3, 0, 0, 0); - old_thread->funnel_state = TH_FN_OWNED; - old_thread->wait_result = save_wait_result; - } - (void) spllo(); - call_continuation(continuation); - /*NOTREACHED*/ - } - break; - } - } else { - /* - * check that the new thread has a stack - */ - if (new_thread->state & TH_STACK_STATE) { - get_new_stack: - /* has no stack. if not already waiting for one try to get one */ - if ((new_thread->state & TH_STACK_COMING_IN) || - /* not already waiting. nonblocking try to get one */ - !stack_alloc_try(new_thread, thread_continue)) - { - /* couldn't get one. schedule new thread to get a stack and - return failure so we can try another thread. */ - thread_swapin(new_thread); - thread_unlock(new_thread); - counter_always(c_thread_invoke_misses++); - return FALSE; - } - } else if (old_thread == new_thread) { - counter(++c_thread_invoke_same); - thread_unlock(new_thread); - return TRUE; - } - - /* new thread now has a stack. it has been setup to resume in - thread_continue so it can dispatch the old thread, deal with - funnelling and then go to it's true continuation point */ - } - - new_thread->state &= ~(TH_STACK_HANDOFF | TH_UNINT); + if ((thread->state & (TH_IDLE|TH_TERMINATE|TH_SUSP)) || !(thread->state & TH_WAIT) || thread->wake_active || thread->sched_pri >= BASEPRI_RTQUEUES) + return (processor->idle_thread); + + /* + * Perform idling activities directly without a + * context switch. Return dispatched thread, + * else check again for a runnable thread. + */ + new_thread = thread_select_idle(thread, processor); + +#else /* !CONFIG_SCHED_IDLE_IN_PLACE */ + + /* + * Do a full context switch to idle so that the current + * thread can start running on another processor without + * waiting for the fast-idled processor to wake up. + */ + new_thread = processor->idle_thread; + +#endif /* !CONFIG_SCHED_IDLE_IN_PLACE */ + + } while (new_thread == THREAD_NULL); + + return (new_thread); +} + +#if CONFIG_SCHED_IDLE_IN_PLACE +/* + * 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; + uint64_t arg1, arg2; + int urgency; + + sched_run_decr(thread); + + thread->state |= TH_IDLE; + processor->current_pri = IDLEPRI; + processor->current_thmode = TH_MODE_NONE; + processor->current_sfi_class = SFI_CLASS_KERNEL; + + /* Reload precise timing global policy to thread-local policy */ + thread->precise_user_kernel_time = use_precise_user_kernel_time(thread); + + thread_unlock(thread); /* - * Set up ast context of new thread and switch to its timer. + * Switch execution timing to processor idle thread. */ - new_thread->last_processor = current_processor(); - ast_context(new_thread->top_act, cpu_number()); - timer_switch(&new_thread->system_timer); - assert(thread_runnable(new_thread)); - + processor->last_dispatch = mach_absolute_time(); + +#ifdef CONFIG_MACH_APPROXIMATE_TIME + commpage_update_mach_approximate_time(processor->last_dispatch); +#endif + + thread->last_run_time = processor->last_dispatch; + thread_timer_event(processor->last_dispatch, &processor->idle_thread->system_timer); + PROCESSOR_DATA(processor, kernel_timer) = &processor->idle_thread->system_timer; + + /* + * Cancel the quantum timer while idling. + */ + timer_call_cancel(&processor->quantum_timer); + processor->first_timeslice = FALSE; + + (*thread->sched_call)(SCHED_CALL_BLOCK, thread); + + thread_tell_urgency(THREAD_URGENCY_NONE, 0, 0, 0, NULL); + + /* + * Enable interrupts and perform idling activities. No + * preemption due to TH_IDLE being set. + */ + spllo(); new_thread = processor_idle(thread, processor); + + /* + * Return at splsched. + */ + (*thread->sched_call)(SCHED_CALL_UNBLOCK, thread); + + thread_lock(thread); + /* - * N.B. On return from the call to switch_context, 'old_thread' - * points at the thread that yielded to us. Unfortunately, at - * this point, there are no simple_locks held, so if we are preempted - * before the call to thread_dispatch blocks preemption, it is - * possible for 'old_thread' to terminate, leaving us with a - * stale thread pointer. + * If awakened, switch to thread timer and start a new quantum. + * Otherwise skip; we will context switch to another thread or return here. */ - disable_preemption(); + 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; + + thread_quantum_init(thread); + processor->quantum_end = processor->last_dispatch + thread->quantum_remaining; + timer_call_enter1(&processor->quantum_timer, thread, processor->quantum_end, TIMER_CALL_SYS_CRITICAL | TIMER_CALL_LOCAL); + processor->first_timeslice = TRUE; + + thread->computation_epoch = processor->last_dispatch; + } + + thread->state &= ~TH_IDLE; + + urgency = thread_get_urgency(thread, &arg1, &arg2); + + thread_tell_urgency(urgency, arg1, arg2, 0, new_thread); + + sched_run_incr(thread); + + return (new_thread); +} +#endif /* CONFIG_SCHED_IDLE_IN_PLACE */ + +/* + * thread_invoke + * + * Called at splsched with neither thread locked. + * + * 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) + */ +static boolean_t +thread_invoke( + thread_t self, + thread_t thread, + ast_t reason) +{ + 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")); + } + + 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) + sched_timeshare_consider_maintenance(ctime); +#endif + + assert_thread_magic(self); + assert(self == current_thread()); + assert(self->runq == PROCESSOR_NULL); + assert((self->state & (TH_RUN|TH_TERMINATE2)) == TH_RUN); + + thread_lock(thread); + + 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); + + /* Allow realtime threads to hang onto a stack. */ + if ((self->sched_mode == TH_MODE_REALTIME) && !self->reserved_stack) + self->reserved_stack = self->kernel_stack; + + if (continuation != NULL) { + if (!thread->kernel_stack) { + /* + * If we are using a privileged stack, + * check to see whether we can exchange it with + * that of the other thread. + */ + if (self->kernel_stack == self->reserved_stack && !thread->reserved_stack) + goto need_stack; + + /* + * Context switch by performing a stack handoff. + */ + continuation = thread->continuation; + parameter = thread->parameter; + + processor = current_processor(); + processor->active_thread = thread; + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + if (thread->last_processor != processor && thread->last_processor != NULL) { + 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 = ctime; + self->last_run_time = ctime; + thread_timer_event(ctime, &thread->system_timer); + PROCESSOR_DATA(processor, kernel_timer) = &thread->system_timer; - thread_unlock(new_thread); + /* + * 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_switch(PROCESSOR_DATA(processor, current_state), + ctime, + PROCESSOR_DATA(processor, current_state)); + } + + 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); + + SCHED_STATS_CSW(processor, self->reason, self->sched_pri, thread->sched_pri); + + TLOG(1, "thread_invoke: calling stack_handoff\n"); + stack_handoff(self, thread); + + /* 'self' is now off core */ + assert(thread == current_thread()); + + DTRACE_SCHED(on__cpu); + +#if KPERF + kperf_on_cpu(thread, continuation, NULL); +#endif /* KPERF */ + + thread_dispatch(self, thread); + + thread->continuation = thread->parameter = NULL; + + counter(c_thread_invoke_hits++); + + (void) spllo(); - counter_always(c_thread_invoke_csw++); - current_task()->csw++; + assert(continuation); + call_continuation(continuation, parameter, thread->wait_result); + /*NOTREACHED*/ + } + 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); + + self->continuation = self->parameter = NULL; + + (void) spllo(); + + call_continuation(continuation, parameter, self->wait_result); + /*NOTREACHED*/ + } + } else { + /* + * Check that the other thread has a stack + */ + if (!thread->kernel_stack) { +need_stack: + if (!stack_alloc_try(thread)) { + counter(c_thread_invoke_misses++); + thread_unlock(thread); + thread_stack_enqueue(thread); + return (FALSE); + } + } else if (thread == self) { + ast_context(self); + counter(++c_thread_invoke_same); + thread_unlock(self); + + 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); + } + } + + /* + * Context switch by full context save. + */ + processor = current_processor(); + processor->active_thread = thread; + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + if (thread->last_processor != processor && thread->last_processor != NULL) { + 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++); + + self->reason = reason; + + processor->last_dispatch = ctime; + self->last_run_time = ctime; + thread_timer_event(ctime, &thread->system_timer); + PROCESSOR_DATA(processor, kernel_timer) = &thread->system_timer; + + /* + * 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_switch(PROCESSOR_DATA(processor, current_state), + ctime, + PROCESSOR_DATA(processor, current_state)); + } + + 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); + + /* + * 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. + */ + assert(continuation == self->continuation); + thread = machine_switch_context(self, continuation, thread); + assert(self == current_thread()); + TLOG(1,"thread_invoke: returning machine_switch_context: self %p continuation %p thread %p\n", self, continuation, thread); + + DTRACE_SCHED(on__cpu); + +#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; + + (void) spllo(); + + call_continuation(continuation, parameter, self->wait_result); + /*NOTREACHED*/ + } + + 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 = (volatile uint32_t) sched_run_buckets[TH_BUCKET_RUN]; + + /* + * 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). + * + */ + if ((sampled_sched_run_count == 1) && + (pset->pending_deferred_AST_cpu_mask)) { + qe_foreach_element_safe(active_processor, &pset->active_queue, processor_queue) { + /* + * 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->state == PROCESSOR_DISPATCHING) && + (pset->pending_deferred_AST_cpu_mask & (1ULL << active_processor->cpu_id)) && + (!(pset->pending_AST_cpu_mask & (1ULL << active_processor->cpu_id))) && + (active_processor != processor)) { + /* + * Squash all of the processor state back to some + * reasonable facsimile of PROCESSOR_IDLE. + * + * TODO: What queue policy do we actually want here? + * We want to promote selection of a good processor + * to run on. Do we want to enqueue at the head? + * The tail? At the (relative) old position in the + * queue? Or something else entirely? + */ + re_queue_head(&pset->idle_queue, &active_processor->processor_queue); + + assert(active_processor->next_thread == THREAD_NULL); + + active_processor->current_pri = IDLEPRI; + active_processor->current_thmode = TH_MODE_FIXED; + active_processor->current_sfi_class = SFI_CLASS_KERNEL; + active_processor->deadline = UINT64_MAX; + active_processor->state = PROCESSOR_IDLE; + pset->pending_deferred_AST_cpu_mask &= ~(1U << active_processor->cpu_id); + machine_signal_idle_cancel(active_processor); + } + + } + } + + pset_unlock(pset); +} +#else +/* We don't support deferred ASTs; everything is candycanes and sunshine. */ +#endif + +/* + * thread_dispatch: + * + * Handle threads at context switch. Re-dispatch other thread + * if still running, otherwise update run state and perform + * 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) +{ + processor_t processor = self->last_processor; + + assert(processor == current_processor()); + assert(self == current_thread()); + assert(thread != self); + + if (thread != THREAD_NULL) { + /* + * If blocked at a continuation, discard + * the stack. + */ + 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->t_ledger, + task_ledgers.cpu_time, consumed); + ledger_credit(thread->t_threadledger, + thread_ledgers.cpu_time, consumed); +#ifdef CONFIG_BANK + if (thread->t_bankledger) { + ledger_credit(thread->t_bankledger, + bank_ledgers.cpu_time, + (consumed - thread->t_deduct_bank_ledger_time)); + + } + thread->t_deduct_bank_ledger_time =0; +#endif + } + + wake_lock(thread); + thread_lock(thread); + + /* + * 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 (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) { + /* + * Cancel the deadline if the thread has + * consumed the entire quantum. + */ + if (thread->quantum_remaining == 0) { + thread->realtime.deadline = UINT64_MAX; + } + } 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->quantum_remaining < min_std_quantum) { + thread->reason |= AST_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->quantum_remaining = thread->quantum_remaining; + thread->reason |= AST_QUANTUM; + 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 runnable. + */ + thread->last_made_runnable_time = mach_approximate_time(); + + machine_thread_going_off_core(thread, FALSE); + + 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); + + 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); + } + + wake_unlock(thread); + } else { + /* + * Waiting. + */ + boolean_t should_terminate = FALSE; + uint32_t new_run_count; + + /* 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; + } + + thread->state &= ~TH_RUN; + thread->last_made_runnable_time = ~0ULL; + thread->chosen_processor = PROCESSOR_NULL; + + new_run_count = sched_run_decr(thread); + +#if CONFIG_SCHED_SFI + if ((thread->state & (TH_WAIT | TH_TERMINATE)) == TH_WAIT) { + if (thread->reason & AST_SFI) { + thread->wait_sfi_begin_time = processor->last_dispatch; + } + } +#endif + + machine_thread_going_off_core(thread, should_terminate); + + 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); + + (*thread->sched_call)(SCHED_CALL_BLOCK, thread); + + if (thread->wake_active) { + thread->wake_active = FALSE; + thread_unlock(thread); + + thread_wakeup(&thread->wake_active); + } else { + thread_unlock(thread); + } + + wake_unlock(thread); + + if (should_terminate) + thread_terminate_enqueue(thread); + } + } + } + + /* Update (new) current thread and reprogram quantum timer */ + thread_lock(self); + if (!(self->state & TH_IDLE)) { + uint64_t arg1, arg2; + int urgency; + uint64_t latency; + +#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); + latency = processor->last_dispatch - self->last_made_runnable_time; + + urgency = thread_get_urgency(self, &arg1, &arg2); + + thread_tell_urgency(urgency, arg1, arg2, latency, self); + + machine_thread_going_on_core(self, urgency, latency); + + /* + * Get a new quantum if none remaining. + */ + if (self->quantum_remaining == 0) { + thread_quantum_init(self); + } + + /* + * Set up quantum timer and timeslice. + */ + processor->quantum_end = processor->last_dispatch + self->quantum_remaining; + timer_call_enter1(&processor->quantum_timer, self, processor->quantum_end, TIMER_CALL_SYS_CRITICAL | TIMER_CALL_LOCAL); + + processor->first_timeslice = TRUE; + } else { + timer_call_cancel(&processor->quantum_timer); + processor->first_timeslice = FALSE; + + thread_tell_urgency(THREAD_URGENCY_NONE, 0, 0, 0, self); + machine_thread_going_on_core(self, THREAD_URGENCY_NONE, 0); + } + + self->computation_epoch = processor->last_dispatch; + self->reason = AST_NONE; + processor->starting_pri = self->sched_pri; + + thread_unlock(self); + +#if defined(CONFIG_SCHED_DEFERRED_AST) + /* + * TODO: Can we state that redispatching our old thread is also + * uninteresting? + */ + if ((((volatile uint32_t)sched_run_buckets[TH_BUCKET_RUN]) == 1) && + !(self->state & TH_IDLE)) { + pset_cancel_deferred_dispatch(processor->processor_set, processor); + } +#endif + +} + +/* + * thread_block_reason: + * + * Forces a reschedule, blocking the caller if a wait + * has been asserted. + * + * If a continuation is specified, then thread_invoke will + * attempt to discard the thread's kernel stack. When the + * thread resumes, it will execute the continuation function + * on a new kernel stack. + */ +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_t self = current_thread(); + processor_t processor; + thread_t new_thread; + spl_t s; + + counter(++c_thread_block_calls); + + s = splsched(); + + processor = current_processor(); + + /* If we're explicitly yielding, force a subsequent quantum */ + 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 (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, reason); + thread_unlock(self); + } while (!thread_invoke(self, new_thread, reason)); + + splx(s); + + return (self->wait_result); +} + +/* + * thread_block: + * + * Block the current thread if a wait has been asserted. + */ +wait_result_t +thread_block( + thread_continue_t continuation) +{ + return thread_block_reason(continuation, NULL, AST_NONE); +} + +wait_result_t +thread_block_parameter( + thread_continue_t continuation, + void *parameter) +{ + return thread_block_reason(continuation, parameter, AST_NONE); +} + +/* + * thread_run: + * + * Switch directly from the current thread to the + * new thread, handing off our quantum if appropriate. + * + * New thread must be runnable, and not on a run queue. + * + * Called at splsched. + */ +int +thread_run( + thread_t self, + thread_continue_t continuation, + void *parameter, + thread_t new_thread) +{ + ast_t handoff = AST_HANDOFF; + + self->continuation = continuation; + self->parameter = parameter; + + while (!thread_invoke(self, new_thread, handoff)) { + processor_t processor = current_processor(); + + thread_lock(self); + new_thread = thread_select(self, processor, AST_NONE); + thread_unlock(self); + handoff = AST_NONE; + } + + return (self->wait_result); +} + +/* + * thread_continue: + * + * Called at splsched when a thread first receives + * a new stack after a continuation. + */ +void +thread_continue( + thread_t thread) +{ + thread_t self = current_thread(); + thread_continue_t continuation; + void *parameter; + + DTRACE_SCHED(on__cpu); + + continuation = self->continuation; + parameter = self->parameter; + +#if KPERF + kperf_on_cpu(self, continuation, NULL); +#endif + + thread_dispatch(thread, self); + + self->continuation = self->parameter = NULL; + + if (thread != THREAD_NULL) + (void)spllo(); + + TLOG(1, "thread_continue: calling call_continuation \n"); + call_continuation(continuation, parameter, self->wait_result); + /*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: + * + * Initialize a run queue before first use. + */ +void +run_queue_init( + run_queue_t rq) +{ + rq->highq = NOPRI; + for (u_int i = 0; i < BITMAP_LEN(NRQS); i++) + rq->bitmap[i] = 0; + rq->urgency = rq->count = 0; + for (int i = 0; i < NRQS; i++) + queue_init(&rq->queues[i]); +} + +/* + * run_queue_dequeue: + * + * Perform a dequeue operation on a run queue, + * and return the resulting thread. + * + * The run queue must be locked (see thread_run_queue_remove() + * for more info), and not empty. + */ +thread_t +run_queue_dequeue( + run_queue_t rq, + integer_t options) +{ + thread_t thread; + queue_t queue = &rq->queues[rq->highq]; + + if (options & SCHED_HEADQ) { + thread = qe_dequeue_head(queue, struct thread, runq_links); + } else { + thread = qe_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 (SCHED(priority_is_urgent)(rq->highq)) { + rq->urgency--; assert(rq->urgency >= 0); + } + if (queue_empty(queue)) { + bitmap_clear(rq->bitmap, rq->highq); + rq->highq = bitmap_first(rq->bitmap, NRQS); + } + + return thread; +} + +/* + * run_queue_enqueue: + * + * Perform a enqueue operation on a run queue. + * + * The run queue must be locked (see thread_run_queue_remove() + * for more info). + */ +boolean_t +run_queue_enqueue( + run_queue_t rq, + thread_t thread, + integer_t options) +{ + queue_t queue = &rq->queues[thread->sched_pri]; + boolean_t result = FALSE; + + assert_thread_magic(thread); + + if (queue_empty(queue)) { + enqueue_tail(queue, &thread->runq_links); + + rq_bitmap_set(rq->bitmap, thread->sched_pri); + if (thread->sched_pri > rq->highq) { + rq->highq = thread->sched_pri; + result = TRUE; + } + } else { + if (options & SCHED_TAILQ) + enqueue_tail(queue, &thread->runq_links); + else + enqueue_head(queue, &thread->runq_links); + } + if (SCHED(priority_is_urgent)(thread->sched_pri)) + rq->urgency++; + SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count); + rq->count++; + + return (result); +} + +/* + * 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) +{ + assert(thread->runq != PROCESSOR_NULL); + assert_thread_magic(thread); + + remqueue(&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); + } + + if (queue_empty(&rq->queues[thread->sched_pri])) { + /* update run queue status */ + bitmap_clear(rq->bitmap, thread->sched_pri); + rq->highq = bitmap_first(rq->bitmap, NRQS); + } + + thread->runq = PROCESSOR_NULL; +} + +/* Assumes RT lock is not held, and acquires splsched/rt_lock itself */ +void +rt_runq_scan(sched_update_scan_context_t scan_context) +{ + spl_t s; + thread_t thread; + + s = splsched(); + rt_lock_lock(); + + qe_foreach_element_safe(thread, &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; + } + } + + rt_lock_unlock(); + splx(s); +} + + +/* + * realtime_queue_insert: + * + * Enqueue a thread for realtime execution. + */ +static boolean_t +realtime_queue_insert(thread_t thread) +{ + queue_t queue = &rt_runq.queue; + uint64_t deadline = thread->realtime.deadline; + boolean_t preempt = FALSE; + + rt_lock_lock(); + + 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; + } + } + } + + thread->runq = THREAD_ON_RT_RUNQ; + SCHED_STATS_RUNQ_CHANGE(&rt_runq.runq_stats, rt_runq.count); + rt_runq.count++; + + rt_lock_unlock(); + + return (preempt); +} + +/* + * realtime_setrun: + * + * Dispatch a thread for realtime execution. + * + * Thread must be locked. Associated pset must + * be locked, and is returned unlocked. + */ +static void +realtime_setrun( + processor_t processor, + thread_t thread) +{ + processor_set_t pset = processor->processor_set; + ast_t preempt; + + boolean_t do_signal_idle = FALSE, do_cause_ast = FALSE; + + thread->chosen_processor = processor; + + /* */ + assert(thread->bound_processor == PROCESSOR_NULL); + + /* + * Dispatch directly onto idle processor. + */ + if ( (thread->bound_processor == processor) + && processor->state == PROCESSOR_IDLE) { + re_queue_tail(&pset->active_queue, &processor->processor_queue); + + processor->next_thread = thread; + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + processor->deadline = thread->realtime.deadline; + processor->state = PROCESSOR_DISPATCHING; + + if (processor != current_processor()) { + if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) { + /* cleared on exit from main processor_idle() loop */ + pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id); + do_signal_idle = TRUE; + } + } + pset_unlock(pset); + + if (do_signal_idle) { + machine_signal_idle(processor); + } + return; + } + + if (processor->current_pri < BASEPRI_RTQUEUES) + preempt = (AST_PREEMPT | AST_URGENT); + else if (thread->realtime.deadline < processor->deadline) + preempt = (AST_PREEMPT | AST_URGENT); + else + preempt = AST_NONE; + + realtime_queue_insert(thread); + + if (preempt != AST_NONE) { + if (processor->state == PROCESSOR_IDLE) { + re_queue_tail(&pset->active_queue, &processor->processor_queue); + + processor->next_thread = THREAD_NULL; + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + processor->deadline = thread->realtime.deadline; + processor->state = PROCESSOR_DISPATCHING; + if (processor == current_processor()) { + ast_on(preempt); + } else { + if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) { + /* cleared on exit from main processor_idle() loop */ + pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id); + do_signal_idle = TRUE; + } + } + } else if (processor->state == PROCESSOR_DISPATCHING) { + if ((processor->next_thread == THREAD_NULL) && ((processor->current_pri < thread->sched_pri) || (processor->deadline > thread->realtime.deadline))) { + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + processor->deadline = thread->realtime.deadline; + } + } else { + if (processor == current_processor()) { + ast_on(preempt); + } else { + if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) { + /* cleared after IPI causes csw_check() to be called */ + pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id); + do_cause_ast = TRUE; + } + } + } + } else { + /* Selected processor was too busy, just keep thread enqueued and let other processors drain it naturally. */ + } + + pset_unlock(pset); + + if (do_signal_idle) { + machine_signal_idle(processor); + } else if (do_cause_ast) { + cause_ast_check(processor); + } +} + + +#if defined(CONFIG_SCHED_TIMESHARE_CORE) + +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; + ast_t preempt; + enum { eExitIdle, eInterruptRunning, eDoNothing } ipi_action = eDoNothing; + enum { eNoSignal, eDoSignal, eDoDeferredSignal } do_signal_idle = eNoSignal; + + boolean_t do_cause_ast = FALSE; + + thread->chosen_processor = processor; + + /* + * Dispatch directly onto idle processor. + */ + if ( (SCHED(direct_dispatch_to_idle_processors) || + thread->bound_processor == processor) + && processor->state == PROCESSOR_IDLE) { + + re_queue_tail(&pset->active_queue, &processor->processor_queue); + + processor->next_thread = thread; + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + processor->deadline = UINT64_MAX; + processor->state = PROCESSOR_DISPATCHING; + + if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) { + /* cleared on exit from main processor_idle() loop */ + pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id); + do_signal_idle = eDoSignal; + } + + pset_unlock(pset); + + if (do_signal_idle == eDoSignal) { + machine_signal_idle(processor); + } + + return; + } + + /* + * 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; + + SCHED(processor_enqueue)(processor, thread, options); + + if (preempt != AST_NONE) { + if (processor->state == PROCESSOR_IDLE) { + re_queue_tail(&pset->active_queue, &processor->processor_queue); + + processor->next_thread = THREAD_NULL; + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + processor->deadline = UINT64_MAX; + processor->state = PROCESSOR_DISPATCHING; + + ipi_action = eExitIdle; + } else if ( processor->state == PROCESSOR_DISPATCHING) { + if ((processor->next_thread == THREAD_NULL) && (processor->current_pri < thread->sched_pri)) { + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + processor->deadline = UINT64_MAX; + } + } 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 != current_processor() ) { + re_queue_tail(&pset->active_queue, &processor->processor_queue); + + processor->next_thread = THREAD_NULL; + processor->current_pri = thread->sched_pri; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class; + processor->deadline = UINT64_MAX; + processor->state = PROCESSOR_DISPATCHING; + + ipi_action = eExitIdle; + } + } + + switch (ipi_action) { + case eDoNothing: + break; + case eExitIdle: + if (processor == current_processor()) { + if (csw_check_locked(processor, pset, AST_NONE) != AST_NONE) + ast_on(preempt); + } else { +#if defined(CONFIG_SCHED_DEFERRED_AST) + if (!(pset->pending_deferred_AST_cpu_mask & (1ULL << processor->cpu_id)) && + !(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) { + /* cleared on exit from main processor_idle() loop */ + pset->pending_deferred_AST_cpu_mask |= (1ULL << processor->cpu_id); + do_signal_idle = eDoDeferredSignal; + } +#else + if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) { + /* cleared on exit from main processor_idle() loop */ + pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id); + do_signal_idle = eDoSignal; + } +#endif + } + break; + case eInterruptRunning: + if (processor == current_processor()) { + if (csw_check_locked(processor, pset, AST_NONE) != AST_NONE) + ast_on(preempt); + } else { + if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) { + /* cleared after IPI causes csw_check() to be called */ + pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id); + do_cause_ast = TRUE; + } + } + break; + } + + pset_unlock(pset); + + if (do_signal_idle == eDoSignal) { + machine_signal_idle(processor); + } +#if defined(CONFIG_SCHED_DEFERRED_AST) + else if (do_signal_idle == eDoDeferredSignal) { + /* + * TODO: The ability to cancel this signal could make + * sending it outside of the pset lock an issue. Do + * we need to address this? Or would the only fallout + * be that the core takes a signal? As long as we do + * not run the risk of having a core marked as signal + * outstanding, with no real signal outstanding, the + * only result should be that we fail to cancel some + * signals. + */ + machine_signal_idle_deferred(processor); + } +#endif + else if (do_cause_ast) { + cause_ast_check(processor); + } +} + +/* + * 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; + + 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 pset, + processor_t processor, + thread_t thread) +{ + processor_set_t nset, cset = pset; + + 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; + } + + /* + * 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 { + 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; + } + } + } + + /* + * 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_unpaired_primary_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_unpaired_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 { + + /* + * Choose an idle processor, in pset traversal order + */ + qe_foreach_element(processor, &cset->idle_queue, processor_queue) { + if (processor->is_recommended) + return processor; + } + + /* + * Otherwise, enumerate active and idle processors to find candidates + * with lower priority/etc. + */ + + qe_foreach_element(processor, &cset->active_queue, processor_queue) { + + if (!processor->is_recommended) { + continue; + } + + integer_t cpri = processor->current_pri; + 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 + */ + qe_foreach_element(processor, &cset->idle_secondary_queue, processor_queue) { + + if (!processor->is_recommended) { + continue; + } + + processor_t cprimary = processor->processor_primary; + + /* 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) { + integer_t primary_pri = cprimary->current_pri; + + if (primary_pri < lowest_unpaired_primary_priority) { + lowest_unpaired_primary_priority = primary_pri; + lp_unpaired_primary_processor = cprimary; + lp_unpaired_secondary_processor = processor; + } + } + } + + + if (thread->sched_pri >= BASEPRI_RTQUEUES) { + + /* + * For realtime threads, the most important aspect is + * scheduling latency, so we attempt to assign threads + * to good preemption candidates (assuming an idle primary + * processor was not available above). + */ + + if (thread->sched_pri > lowest_unpaired_primary_priority) { + /* Move to end of active queue so that the next thread doesn't also pick it */ + re_queue_tail(&cset->active_queue, &lp_unpaired_primary_processor->processor_queue); + return lp_unpaired_primary_processor; + } + if (thread->sched_pri > lowest_priority) { + /* Move to end of active queue so that the next thread doesn't also pick it */ + re_queue_tail(&cset->active_queue, &lp_processor->processor_queue); + return lp_processor; + } + if (thread->realtime.deadline < furthest_deadline) + return fd_processor; + + /* + * If all primary and secondary CPUs are busy with realtime + * threads with deadlines earlier than us, move on to next + * pset. + */ + } + else { + + if (thread->sched_pri > lowest_unpaired_primary_priority) { + /* Move to end of active queue so that the next thread doesn't also pick it */ + re_queue_tail(&cset->active_queue, &lp_unpaired_primary_processor->processor_queue); + return lp_unpaired_primary_processor; + } + if (thread->sched_pri > lowest_priority) { + /* Move to end of active queue so that the next thread doesn't also pick it */ + re_queue_tail(&cset->active_queue, &lp_processor->processor_queue); + return lp_processor; + } + + /* + * If all primary processor in this pset are running a higher + * priority thread, move on to next pset. Only when we have + * exhausted this search do we fall back to other heuristics. + */ + } - - thread_lock(old_thread); - old_thread->reason = reason; - assert(old_thread->runq == RUN_QUEUE_NULL); + /* + * Move onto the next processor set. + */ + nset = next_pset(cset); - if (continuation != (void (*)(void))0) - old_thread->continuation = continuation; + if (nset != pset) { + pset_unlock(cset); - /* Indicate to sched policy that old thread has stopped execution */ - policy = &sched_policy[old_thread->policy]; - /*** ??? maybe use a macro -- ***/ - sfr = policy->sp_ops.sp_thread_done(policy, old_thread); - assert(sfr == SF_SUCCESS); - thread_unlock(old_thread); + cset = nset; + pset_lock(cset); + } + } while (nset != pset); /* - * switch_context is machine-dependent. It does the - * machine-dependent components of a context-switch, like - * changing address spaces. It updates active_threads. + * Make sure that we pick a running processor, + * and that the correct processor set is locked. + * Since we may have unlock 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. */ - old_thread = switch_context(old_thread, continuation, new_thread); - - /* Now on new thread's stack. Set a local variable to refer to it. */ - new_thread = __current_thread(); - assert(old_thread != new_thread); + do { - assert(thread_runnable(new_thread)); + /* lowest_priority is evaluated in the main loops above */ + if (lp_unpaired_secondary_processor != PROCESSOR_NULL) { + processor = lp_unpaired_secondary_processor; + lp_unpaired_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 + */ + processor = master_processor; + } - thread_lock(new_thread); - assert(thread_runnable(new_thread)); - /* Indicate to sched policy that new thread has started execution */ - policy = &sched_policy[new_thread->policy]; - /*** ??? maybe use a macro -- rkc, 1/4/96 ***/ - sfr = policy->sp_ops.sp_thread_begin(policy, new_thread); - assert(sfr == SF_SUCCESS); - thread_unlock(new_thread); + /* + * Check that the correct processor set is + * returned locked. + */ + if (cset != processor->processor_set) { + pset_unlock(cset); + cset = processor->processor_set; + pset_lock(cset); + } - /* - * We're back. Now old_thread is the thread that resumed - * us, and we have to dispatch it. - */ - /* CHECKME! */ -// Code from OSF in Grenoble deleted the following fields. They were -// used in HPPA and 386 code, but not in the PPC for other than -// just setting and resetting. They didn't delete these lines from -// the MACH_RT builds, though, causing compile errors. I'm going -// to make a wild guess and assume we can just delete these. -#if 0 - if (old_thread->preempt == TH_NOT_PREEMPTABLE) { - /* - * Mark that we have been really preempted - */ - old_thread->preempt = TH_PREEMPTED; - } -#endif - thread_dispatch(old_thread); - enable_preemption(); + /* + * We must verify that the chosen processor is still available. + * master_processor is an exception, since we may need to preempt + * a running thread on it during processor shutdown (for sleep), + * and that thread needs to be enqueued on its runqueue to run + * when the processor is restarted. + */ + if (processor != master_processor && (processor->state == PROCESSOR_SHUTDOWN || processor->state == PROCESSOR_OFF_LINE)) + processor = PROCESSOR_NULL; - /* if we get here and 'continuation' is set that means the - * switch_context() path returned and did not call out - * to the continuation. we will do it manually here */ - if (continuation) { - call_continuation(continuation); - /* NOTREACHED */ - } + } while (processor == PROCESSOR_NULL); - return TRUE; + return (processor); } /* - * thread_continue: + * thread_setrun: + * + * Dispatch thread for execution, onto an idle + * processor or run queue, and signal a preemption + * as appropriate. * - * Called when the launching a new thread, at splsched(); + * Thread must be locked. */ void -thread_continue( - register thread_t old_thread) +thread_setrun( + thread_t thread, + integer_t options) { - register thread_t self; - register void (*continuation)(); - sched_policy_t *policy; - sf_return_t sfr; - - self = current_thread(); - - /* - * We must dispatch the old thread and then - * call the current thread's continuation. - * There might not be an old thread, if we are - * the first thread to run on this processor. - */ - if (old_thread != THREAD_NULL) { - thread_dispatch(old_thread); - - thread_lock(self); - - /* Get pointer to scheduling policy "object" */ - policy = &sched_policy[self->policy]; + processor_t processor; + processor_set_t pset; - /* Indicate to sched policy that new thread has started execution */ - /*** ??? maybe use a macro -- rkc, 1/4/96 ***/ - sfr = policy->sp_ops.sp_thread_begin(policy,self); - assert(sfr == SF_SUCCESS); - } else { - thread_lock(self); - } - - continuation = self->continuation; - self->continuation = (void (*)(void))0; - thread_unlock(self); + assert((thread->state & (TH_RUN|TH_WAIT|TH_UNINT|TH_TERMINATE|TH_TERMINATE2)) == TH_RUN); + assert(thread->runq == PROCESSOR_NULL); /* - * N.B. - the following is necessary, since thread_invoke() - * inhibits preemption on entry and reenables before it - * returns. Unfortunately, the first time a newly-created - * thread executes, it magically appears here, and never - * executes the enable_preemption() call in thread_invoke(). + * Update priority if needed. */ - enable_preemption(); - - if (self->funnel_state & TH_FN_REFUNNEL) { - kern_return_t save_wait_result; - self->funnel_state = 0; - save_wait_result = self->wait_result; - KERNEL_DEBUG(0x6032428 | DBG_FUNC_NONE, self->funnel_lock, 4, 0, 0, 0); - funnel_lock(self->funnel_lock); - KERNEL_DEBUG(0x6032430 | DBG_FUNC_NONE, self->funnel_lock, 4, 0, 0, 0); - self->wait_result = save_wait_result; - self->funnel_state = TH_FN_OWNED; - } - spllo(); - - assert(continuation); - (*continuation)(); - /*NOTREACHED*/ -} - -#if MACH_LDEBUG || MACH_KDB - -#define THREAD_LOG_SIZE 300 - -struct t64 { - unsigned long h; - unsigned long l; -}; + if (SCHED(can_update_priority)(thread)) + SCHED(update_priority)(thread); -struct { - struct t64 stamp; - thread_t thread; - long info1; - long info2; - long info3; - char * action; -} thread_log[THREAD_LOG_SIZE]; - -int thread_log_index; + thread->sfi_class = sfi_thread_classify(thread); -void check_thread_time(long n); + assert(thread->runq == PROCESSOR_NULL); +#if __SMP__ + if (thread->bound_processor == PROCESSOR_NULL) { + /* + * Unbound case. + */ + if (thread->affinity_set != AFFINITY_SET_NULL) { + /* + * Use affinity set policy hint. + */ + pset = thread->affinity_set->aset_pset; + pset_lock(pset); -int check_thread_time_crash; + processor = SCHED(choose_processor)(pset, PROCESSOR_NULL, thread); -#if 0 -void -check_thread_time(long us) -{ - struct t64 temp; + 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); - if (!check_thread_time_crash) - return; + 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; - temp = thread_log[0].stamp; - cyctm05_diff (&thread_log[1].stamp, &thread_log[0].stamp, &temp); + pset = task->pset_hint; + if (pset == PROCESSOR_SET_NULL) + pset = current_processor()->processor_set; - if (temp.l >= us && thread_log[1].info != 0x49) /* HACK!!! */ - panic ("check_thread_time"); -} -#endif + pset = choose_next_pset(pset); + pset_lock(pset); -void -log_thread_action(char * action, long info1, long info2, long info3) -{ - int i; - spl_t x; - static unsigned int tstamp; + processor = SCHED(choose_processor)(pset, PROCESSOR_NULL, thread); + task->pset_hint = processor->processor_set; - x = splhigh(); + 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 { + /* + * Bound case: + * + * Unconditionally dispatch on the processor. + */ + processor = thread->bound_processor; + pset = processor->processor_set; + pset_lock(pset); - for (i = THREAD_LOG_SIZE-1; i > 0; i--) { - thread_log[i] = thread_log[i-1]; + 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__ */ - thread_log[0].stamp.h = 0; - thread_log[0].stamp.l = tstamp++; - thread_log[0].thread = current_thread(); - thread_log[0].info1 = info1; - thread_log[0].info2 = info2; - thread_log[0].info3 = info3; - thread_log[0].action = action; -/* strcpy (&thread_log[0].action[0], action);*/ - - splx(x); + /* + * Dispatch the thread on the chosen processor. + * TODO: This should be based on sched_mode, not sched_pri + */ + if (thread->sched_pri >= BASEPRI_RTQUEUES) + realtime_setrun(processor, thread); + else + processor_setrun(processor, thread, options); } -#endif /* MACH_LDEBUG || MACH_KDB */ -#if MACH_KDB -#include -void db_show_thread_log(void); - -void -db_show_thread_log(void) +processor_set_t +task_choose_pset( + task_t task) { - int i; + processor_set_t pset = task->pset_hint; - db_printf ("%s %s %s %s %s %s\n", " Thread ", " Info1 ", " Info2 ", - " Info3 ", " Timestamp ", "Action"); + if (pset != PROCESSOR_SET_NULL) + pset = choose_next_pset(pset); - for (i = 0; i < THREAD_LOG_SIZE; i++) { - db_printf ("%08x %08x %08x %08x %08x/%08x %s\n", - thread_log[i].thread, - thread_log[i].info1, - thread_log[i].info2, - thread_log[i].info3, - thread_log[i].stamp.h, - thread_log[i].stamp.l, - thread_log[i].action); - } + return (pset); } -#endif /* MACH_KDB */ /* - * thread_block_reason: + * Check for a preemption point in + * the current context. * - * Block the current thread. If the thread is runnable - * then someone must have woken it up between its request - * to sleep and now. In this case, it goes back on a - * run queue. - * - * If a continuation is specified, then thread_block will - * attempt to discard the thread's kernel stack. When the - * thread resumes, it will execute the continuation function - * on a new kernel stack. + * Called at splsched with thread locked. */ -counter(mach_counter_t c_thread_block_calls = 0;) - -int -thread_block_reason( - void (*continuation)(void), - int reason) +ast_t +csw_check( + processor_t processor, + ast_t check_reason) { - register thread_t thread = current_thread(); - register processor_t myprocessor; - register thread_t new_thread; - spl_t s; + processor_set_t pset = processor->processor_set; + ast_t result; - counter(++c_thread_block_calls); + pset_lock(pset); - check_simple_locks(); + /* If we were sent a remote AST and interrupted a running processor, acknowledge it here with pset lock held */ + pset->pending_AST_cpu_mask &= ~(1ULL << processor->cpu_id); - machine_clock_assist(); + result = csw_check_locked(processor, pset, check_reason); - s = splsched(); + pset_unlock(pset); + + return result; +} + +/* + * Check for preemption at splsched with + * pset and thread locked + */ +ast_t +csw_check_locked( + processor_t processor, + processor_set_t pset __unused, + ast_t check_reason) +{ + ast_t result; + thread_t thread = processor->active_thread; - if ((thread->funnel_state & TH_FN_OWNED) && !(reason & AST_PREEMPT)) { - thread->funnel_state = TH_FN_REFUNNEL; - KERNEL_DEBUG(0x603242c | DBG_FUNC_NONE, thread->funnel_lock, 2, 0, 0, 0); - funnel_unlock(thread->funnel_lock); + if (processor->first_timeslice) { + if (rt_runq.count > 0) + return (check_reason | AST_PREEMPT | AST_URGENT); + } + else { + if (rt_runq.count > 0) { + if (BASEPRI_RTQUEUES > processor->current_pri) + return (check_reason | AST_PREEMPT | AST_URGENT); + else + return (check_reason | AST_PREEMPT); + } } - myprocessor = current_processor(); + result = SCHED(processor_csw_check)(processor); + if (result != AST_NONE) + return (check_reason | result | (thread_eager_preemption(thread) ? AST_URGENT : AST_NONE)); - thread_lock(thread); - if (thread->state & TH_ABORT) - clear_wait_internal(thread, THREAD_INTERRUPTED); +#if __SMP__ - /* Unconditionally remove either | both */ - ast_off(AST_QUANTUM|AST_BLOCK|AST_URGENT); + /* + * If the current thread is running on a processor that is no longer recommended, gently + * (non-urgently) get to a point and then block, and 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); - new_thread = thread_select(myprocessor); - assert(new_thread); - assert(thread_runnable(new_thread)); - thread_unlock(thread); - while (!thread_invoke(thread, new_thread, reason, continuation)) { - thread_lock(thread); - new_thread = thread_select(myprocessor); - assert(new_thread); - assert(thread_runnable(new_thread)); - thread_unlock(thread); - } + /* + * 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 + * Consider Capri in this scenario. + * + * if (!SCHED(processor_bound_count)(processor) && !queue_empty(&pset->idle_queue)) + * + * TODO: Alternatively - check if only primary is idle, or check if primary's pri is lower than mine. + */ - if (thread->funnel_state & TH_FN_REFUNNEL) { - kern_return_t save_wait_result; + if (processor->current_pri < BASEPRI_RTQUEUES && + processor->processor_primary != processor) + return (check_reason | AST_PREEMPT); +#endif - save_wait_result = thread->wait_result; - thread->funnel_state = 0; - KERNEL_DEBUG(0x6032428 | DBG_FUNC_NONE, thread->funnel_lock, 5, 0, 0, 0); - funnel_lock(thread->funnel_lock); - KERNEL_DEBUG(0x6032430 | DBG_FUNC_NONE, thread->funnel_lock, 5, 0, 0, 0); - thread->funnel_state = TH_FN_OWNED; - thread->wait_result = save_wait_result; - } + if (thread->state & TH_SUSP) + return (check_reason | AST_PREEMPT); - splx(s); +#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 - return thread->wait_result; + return (AST_NONE); } /* - * thread_block: + * set_sched_pri: * - * Now calls thread_block_reason() which forwards the - * the reason parameter to thread_invoke() so it can - * do the right thing if the thread's quantum expired. - */ -int -thread_block( - void (*continuation)(void)) -{ - return thread_block_reason(continuation, 0); -} - -/* - * thread_run: + * Set the scheduled priority of the specified thread. * - * Switch directly from the current thread to a specified - * thread. Both the current and new threads must be - * runnable. + * This may cause the thread to change queues. * - * Assumption: - * at splsched. - */ -int -thread_run( - thread_t old_thread, - void (*continuation)(void), - thread_t new_thread) -{ - while (!thread_invoke(old_thread, new_thread, 0, continuation)) { - register processor_t myprocessor = current_processor(); - thread_lock(old_thread); - new_thread = thread_select(myprocessor); - thread_unlock(old_thread); - } - return old_thread->wait_result; -} - -/* - * Dispatches a running thread that is not on a runq. - * Called at splsched. + * Thread must be locked. */ void -thread_dispatch( - register thread_t thread) +set_sched_pri( + thread_t thread, + int priority) { - sched_policy_t *policy; - sf_return_t sfr; + thread_t cthread = current_thread(); + boolean_t is_current_thread = (thread == cthread) ? TRUE : FALSE; + int curgency, nurgency; + uint64_t urgency_param1, urgency_param2; + boolean_t removed_from_runq = FALSE; + + /* If we're already at this priority, no need to mess with the runqueue */ + if (priority == thread->sched_pri) + return; - /* - * If we are discarding the thread's stack, we must do it - * before the thread has a chance to run. - */ - wake_lock(thread); - thread_lock(thread); + if (is_current_thread) { + assert(thread->runq == PROCESSOR_NULL); + curgency = thread_get_urgency(thread, &urgency_param1, &urgency_param2); + } else { + removed_from_runq = thread_run_queue_remove(thread); + } -#ifndef i386 - /* no continuations on i386 for now */ - if (thread->continuation != (void (*)())0) { - assert((thread->state & TH_STACK_STATE) == 0); - thread->state |= TH_STACK_HANDOFF; - stack_free(thread); - if (thread->top_act) { - act_machine_sv_free(thread->top_act); - } - } -#endif + thread->sched_pri = priority; - switch (thread->state & (TH_RUN|TH_WAIT|TH_UNINT|TH_IDLE)) { + KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY), + (uintptr_t)thread_tid(thread), + thread->base_pri, + thread->sched_pri, + 0, /* eventually, 'reason' */ + 0); - case TH_RUN | TH_UNINT: - case TH_RUN: - /* - * No reason to stop. Put back on a run queue. - */ - /* Leave enqueueing thread up to scheduling policy */ - policy = &sched_policy[thread->policy]; - /*** ??? maybe use a macro ***/ - sfr = policy->sp_ops.sp_thread_dispatch(policy, thread); - assert(sfr == SF_SUCCESS); - break; - - case TH_RUN | TH_WAIT | TH_UNINT: - case TH_RUN | TH_WAIT: - thread->sleep_stamp = sched_tick; - /* fallthrough */ - case TH_WAIT: /* this happens! */ - + if (is_current_thread) { + nurgency = thread_get_urgency(thread, &urgency_param1, &urgency_param2); /* - * Waiting + * 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. */ - thread->state &= ~TH_RUN; - if (thread->state & TH_TERMINATE) - thread_reaper_enqueue(thread); - - if (thread->wake_active) { - thread->wake_active = FALSE; - thread_unlock(thread); - wake_unlock(thread); - thread_wakeup((event_t)&thread->wake_active); - return; + if (nurgency != curgency) { + thread_tell_urgency(nurgency, urgency_param1, urgency_param2, 0, thread); + machine_thread_going_on_core(thread, nurgency, 0); } - break; - - case TH_RUN | TH_IDLE: - /* - * Drop idle thread -- it is already in - * idle_thread_array. - */ - break; + } - default: - panic("State 0x%x \n",thread->state); + /* TODO: Should this be TAILQ if it went down, HEADQ if it went up? */ + if (removed_from_runq) + thread_run_queue_reinsert(thread, SCHED_PREEMPT | SCHED_TAILQ); + else if (thread->state & TH_RUN) { + processor_t processor = thread->last_processor; + + if (is_current_thread) { + ast_t preempt; + + processor->current_pri = priority; + processor->current_thmode = thread->sched_mode; + processor->current_sfi_class = thread->sfi_class = sfi_thread_classify(thread); + if ((preempt = csw_check(processor, AST_NONE)) != AST_NONE) + ast_on(preempt); + } else if (processor != PROCESSOR_NULL && processor->active_thread == thread) + cause_ast_check(processor); } - thread_unlock(thread); - wake_unlock(thread); } /* - * Enqueue thread on run queue. Thread must be locked, - * and not already be on a run queue. + * 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. */ -int -run_queue_enqueue( - register run_queue_t rq, - register thread_t thread, - boolean_t tail) -{ - register int whichq; - int oldrqcount; - - whichq = thread->sched_pri; - assert(whichq >= MINPRI && whichq <= MAXPRI); +thread_t +thread_run_queue_remove_for_handoff(thread_t thread) { - simple_lock(&rq->lock); /* lock the run queue */ - assert(thread->runq == RUN_QUEUE_NULL); - if (tail) - enqueue_tail(&rq->queues[whichq], (queue_entry_t)thread); - else - enqueue_head(&rq->queues[whichq], (queue_entry_t)thread); + thread_t pulled_thread = THREAD_NULL; - setbit(MAXPRI - whichq, rq->bitmap); - if (whichq > rq->highq) - rq->highq = whichq; + thread_lock(thread); - oldrqcount = rq->count++; - thread->runq = rq; - thread->whichq = whichq; -#if DEBUG - thread_check(thread, rq); -#endif /* DEBUG */ - simple_unlock(&rq->lock); + /* + * 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. + */ + + 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; + } + + thread_unlock(thread); - return (oldrqcount); + return pulled_thread; } /* - * thread_setrun: + * thread_run_queue_remove: * - * Make thread runnable; dispatch directly onto an idle processor - * if possible. Else put on appropriate run queue (processor - * if bound, else processor set. Caller must have lock on thread. - * This is always called at splsched. - * The tail parameter, if TRUE || TAIL_Q, indicates that the - * thread should be placed at the tail of the runq. If - * FALSE || HEAD_Q the thread will be placed at the head of the - * appropriate runq. + * 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. */ -void -thread_setrun( - register thread_t new_thread, - boolean_t may_preempt, - boolean_t tail) +boolean_t +thread_run_queue_remove( + thread_t thread) { - register processor_t processor; - register run_queue_t runq; - register processor_set_t pset; - thread_t thread; - ast_t ast_flags = AST_BLOCK; + boolean_t removed = FALSE; + processor_t processor = thread->runq; - mp_disable_preemption(); + if ((thread->state & (TH_RUN|TH_WAIT)) == TH_WAIT) { + /* Thread isn't runnable */ + assert(thread->runq == PROCESSOR_NULL); + return FALSE; + } - assert(!(new_thread->state & TH_SWAPPED_OUT)); - assert(thread_runnable(new_thread)); - - /* - * Update priority if needed. - */ - if (new_thread->sched_stamp != sched_tick) - update_priority(new_thread); + if (processor == PROCESSOR_NULL) { + /* + * 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 (new_thread->policy & (POLICY_FIFO|POLICY_RR)) { - if ( new_thread->sched_pri >= (MAXPRI_KERNBAND - 2) && - kernel_preemption_mode == KERNEL_PREEMPT ) - ast_flags |= AST_URGENT; + return FALSE; } - - assert(new_thread->runq == RUN_QUEUE_NULL); - - /* - * Try to dispatch the thread directly onto an idle processor. - */ - if ((processor = new_thread->bound_processor) == PROCESSOR_NULL) { - /* - * Not bound, any processor in the processor set is ok. - */ - pset = new_thread->processor_set; - if (pset->idle_count > 0) { - simple_lock(&pset->idle_lock); - if (pset->idle_count > 0) { - processor = (processor_t) queue_first(&pset->idle_queue); - queue_remove(&(pset->idle_queue), processor, processor_t, - processor_queue); - pset->idle_count--; - processor->next_thread = new_thread; - processor->state = PROCESSOR_DISPATCHING; - simple_unlock(&pset->idle_lock); - if(processor->slot_num != cpu_number()) - machine_signal_idle(processor); - mp_enable_preemption(); - return; - } - simple_unlock(&pset->idle_lock); - } - - /* - * Preempt check - */ - runq = &pset->runq; - thread = current_thread(); - processor = current_processor(); - if ( may_preempt && - pset == processor->processor_set && - thread->sched_pri < new_thread->sched_pri ) { - /* - * XXX if we have a non-empty local runq or are - * XXX running a bound thread, ought to check for - * XXX another cpu running lower-pri thread to preempt. - */ - /* - * Turn off first_quantum to allow csw. - */ - processor->first_quantum = FALSE; + if (thread->sched_pri < BASEPRI_RTQUEUES) { + return SCHED(processor_queue_remove)(processor, thread); + } - ast_on(ast_flags); - } + rt_lock_lock(); + if (thread->runq != PROCESSOR_NULL) { /* - * Put us on the end of the runq, if we are not preempting - * or the guy we are preempting. + * Thread is on the RT run queue and we have a lock on + * that run queue. */ - run_queue_enqueue(runq, new_thread, tail); - } - else { - /* - * Bound, can only run on bound processor. Have to lock - * processor here because it may not be the current one. - */ - if (processor->state == PROCESSOR_IDLE) { - simple_lock(&processor->lock); - pset = processor->processor_set; - simple_lock(&pset->idle_lock); - if (processor->state == PROCESSOR_IDLE) { - queue_remove(&pset->idle_queue, processor, - processor_t, processor_queue); - pset->idle_count--; - processor->next_thread = new_thread; - processor->state = PROCESSOR_DISPATCHING; - simple_unlock(&pset->idle_lock); - simple_unlock(&processor->lock); - if(processor->slot_num != cpu_number()) - machine_signal_idle(processor); - mp_enable_preemption(); - return; - } - simple_unlock(&pset->idle_lock); - simple_unlock(&processor->lock); - } - - /* - * Cause ast on processor if processor is on line, and the - * currently executing thread is not bound to that processor - * (bound threads have implicit priority over non-bound threads). - * We also avoid sending the AST to the idle thread (if it got - * scheduled in the window between the 'if' above and here), - * since the idle_thread is bound. - */ - runq = &processor->runq; - thread = current_thread(); - if (processor == current_processor()) { - if ( thread->bound_processor == PROCESSOR_NULL || - thread->sched_pri < new_thread->sched_pri ) { - processor->first_quantum = FALSE; - ast_on(ast_flags); - } - run_queue_enqueue(runq, new_thread, tail); - } - else { - thread = cpu_data[processor->slot_num].active_thread; - if ( run_queue_enqueue(runq, new_thread, tail) == 0 && - processor->state != PROCESSOR_OFF_LINE && - thread && thread->bound_processor != processor ) - cause_ast_check(processor); - } + assert(thread->runq == THREAD_ON_RT_RUNQ); + + remqueue(&thread->runq_links); + SCHED_STATS_RUNQ_CHANGE(&rt_runq.runq_stats, rt_runq.count); + rt_runq.count--; + + thread->runq = PROCESSOR_NULL; + + removed = TRUE; } - mp_enable_preemption(); + rt_lock_unlock(); + + return (removed); } /* - * set_pri: + * Put the thread back where it goes after a thread_run_queue_remove * - * Set the priority of the specified thread to the specified - * priority. This may cause the thread to change queues. + * Thread must have been removed under the same thread lock hold * - * The thread *must* be locked by the caller. + * thread locked, at splsched */ void -set_pri( - thread_t thread, - int pri, - boolean_t resched) +thread_run_queue_reinsert(thread_t thread, integer_t options) { - register struct run_queue *rq; + assert(thread->runq == PROCESSOR_NULL); + + assert(thread->state & (TH_RUN)); + thread_setrun(thread, options); - rq = rem_runq(thread); - assert(thread->runq == RUN_QUEUE_NULL); - thread->sched_pri = pri; - if (rq != RUN_QUEUE_NULL) { - if (resched) - thread_setrun(thread, TRUE, TAIL_Q); - else - run_queue_enqueue(rq, thread, TAIL_Q); - } } -/* - * rem_runq: - * - * Remove a thread from its run queue. - * The run queue that the process was on is returned - * (or RUN_QUEUE_NULL if not on a run queue). Thread *must* be locked - * before calling this routine. Unusual locking protocol on runq - * field in thread structure makes this code interesting; see thread.h. - */ -run_queue_t -rem_runq( - thread_t thread) +void +sys_override_cpu_throttle(int flag) { - register struct run_queue *rq; + if (flag == CPU_THROTTLE_ENABLE) + cpu_throttle_enabled = 1; + if (flag == CPU_THROTTLE_DISABLE) + cpu_throttle_enabled = 0; +} - rq = thread->runq; - /* - * If rq is RUN_QUEUE_NULL, the thread will stay out of the - * run_queues because the caller locked the thread. Otherwise - * the thread is on a runq, but could leave. - */ - if (rq != RUN_QUEUE_NULL) { - simple_lock(&rq->lock); - if (rq == thread->runq) { - /* - * Thread is in a runq and we have a lock on - * that runq. - */ -#if DEBUG - thread_check(thread, rq); -#endif /* DEBUG */ - remqueue(&rq->queues[0], (queue_entry_t)thread); - rq->count--; - - 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); - } - thread->runq = RUN_QUEUE_NULL; - simple_unlock(&rq->lock); - } - else { - /* - * The thread left the runq before we could - * lock the runq. It is not on a runq now, and - * can't move again because this routine's - * caller locked the thread. - */ - assert(thread->runq == RUN_QUEUE_NULL); - simple_unlock(&rq->lock); - rq = RUN_QUEUE_NULL; - } - } +int +thread_get_urgency(thread_t thread, uint64_t *arg1, uint64_t *arg2) +{ + if (thread == NULL || (thread->state & TH_IDLE)) { + *arg1 = 0; + *arg2 = 0; + + return (THREAD_URGENCY_NONE); + } else if (thread->sched_mode == TH_MODE_REALTIME) { + *arg1 = thread->realtime.period; + *arg2 = thread->realtime.deadline; + + return (THREAD_URGENCY_REAL_TIME); + } else if (cpu_throttle_enabled && + ((thread->sched_pri <= MAXPRI_THROTTLE) && (thread->base_pri <= MAXPRI_THROTTLE))) { + /* + * Background urgency applied when thread priority is MAXPRI_THROTTLE or lower and thread is not promoted + */ + *arg1 = thread->sched_pri; + *arg2 = thread->base_pri; + + return (THREAD_URGENCY_BACKGROUND); + } else { + /* For otherwise unclassified threads, report throughput QoS + * parameters + */ + *arg1 = proc_get_effective_thread_policy(thread, TASK_POLICY_THROUGH_QOS); + *arg2 = proc_get_effective_task_policy(thread->task, TASK_POLICY_THROUGH_QOS); - return (rq); + return (THREAD_URGENCY_NORMAL); + } } /* - * choose_thread: - * - * Choose a thread to execute. The thread chosen is removed - * from its run queue. Note that this requires only that the runq - * lock be held. - * - * Strategy: - * Check processor runq first; if anything found, run it. - * Else check pset runq; if nothing found, return idle thread. + * This is the processor idle loop, which just looks for other threads + * to execute. Processor idle threads invoke this without supplying a + * current thread to idle without an asserted wait state. * - * Second line of strategy is implemented by choose_pset_thread. - * This is only called on processor startup and when thread_block - * thinks there's something in the processor runq. + * Returns a the next thread to execute if dispatched directly. */ + +#if 0 +#define IDLE_KERNEL_DEBUG_CONSTANT(...) KERNEL_DEBUG_CONSTANT(__VA_ARGS__) +#else +#define IDLE_KERNEL_DEBUG_CONSTANT(...) do { } while(0) +#endif + thread_t -choose_thread( - processor_t myprocessor) -{ - thread_t thread; - register queue_t q; - register run_queue_t runq; - processor_set_t pset; - - runq = &myprocessor->runq; - pset = myprocessor->processor_set; - - simple_lock(&runq->lock); - if (runq->count > 0 && runq->highq >= pset->runq.highq) { - q = runq->queues + runq->highq; -#if MACH_ASSERT - if (!queue_empty(q)) { -#endif /*MACH_ASSERT*/ - thread = (thread_t)q->next; - ((queue_entry_t)thread)->next->prev = q; - q->next = ((queue_entry_t)thread)->next; - thread->runq = RUN_QUEUE_NULL; - runq->count--; - if (queue_empty(q)) { - if (runq->highq != IDLEPRI) - clrbit(MAXPRI - runq->highq, runq->bitmap); - runq->highq = MAXPRI - ffsbit(runq->bitmap); - } - simple_unlock(&runq->lock); - return (thread); -#if MACH_ASSERT +processor_idle( + thread_t thread, + processor_t processor) +{ + processor_set_t pset = processor->processor_set; + thread_t new_thread; + int state; + (void)splsched(); + + 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); + + SCHED_STATS_CPU_IDLE_START(processor); + + timer_switch(&PROCESSOR_DATA(processor, system_state), + mach_absolute_time(), &PROCESSOR_DATA(processor, idle_state)); + PROCESSOR_DATA(processor, current_state) = &PROCESSOR_DATA(processor, idle_state); + + while (1) { + if (processor->state != PROCESSOR_IDLE) /* unsafe, but worst case we loop around once */ + break; + if (pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id)) + break; + if (processor->is_recommended) { + if (rt_runq.count) + break; + } else { + if (SCHED(processor_bound_count)(processor)) + break; } - panic("choose_thread"); -#endif /*MACH_ASSERT*/ - /*NOTREACHED*/ - } - simple_unlock(&runq->lock); - simple_lock(&pset->runq.lock); - return (choose_pset_thread(myprocessor, pset)); -} +#if CONFIG_SCHED_IDLE_IN_PLACE + if (thread != THREAD_NULL) { + /* Did idle-in-place thread wake up */ + if ((thread->state & (TH_WAIT|TH_SUSP)) != TH_WAIT || thread->wake_active) + break; + } +#endif + + IDLE_KERNEL_DEBUG_CONSTANT( + MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_NONE, (uintptr_t)thread_tid(thread), rt_runq.count, SCHED(processor_runq_count)(processor), -1, 0); + + machine_track_platform_idle(TRUE); + + machine_idle(); + + machine_track_platform_idle(FALSE); + + (void)splsched(); + IDLE_KERNEL_DEBUG_CONSTANT( + MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_NONE, (uintptr_t)thread_tid(thread), rt_runq.count, SCHED(processor_runq_count)(processor), -2, 0); -/* - * choose_pset_thread: choose a thread from processor_set runq or - * set processor idle and choose its idle thread. - * - * Caller must be at splsched and have a lock on the runq. This - * lock is released by this routine. myprocessor is always the current - * processor, and pset must be its processor set. - * This routine chooses and removes a thread from the runq if there - * is one (and returns it), else it sets the processor idle and - * returns its idle thread. - */ -thread_t -choose_pset_thread( - register processor_t myprocessor, - processor_set_t pset) -{ - register run_queue_t runq; - register thread_t thread; - register queue_t q; - - runq = &pset->runq; - if (runq->count > 0) { - q = runq->queues + runq->highq; -#if MACH_ASSERT - if (!queue_empty(q)) { -#endif /*MACH_ASSERT*/ - thread = (thread_t)q->next; - ((queue_entry_t)thread)->next->prev = q; - q->next = ((queue_entry_t)thread)->next; - thread->runq = RUN_QUEUE_NULL; - runq->count--; - if (queue_empty(q)) { - if (runq->highq != IDLEPRI) - clrbit(MAXPRI - runq->highq, runq->bitmap); - runq->highq = MAXPRI - ffsbit(runq->bitmap); - } - simple_unlock(&runq->lock); - return (thread); -#if MACH_ASSERT + 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; } - panic("choose_pset_thread"); -#endif /*MACH_ASSERT*/ - /*NOTREACHED*/ } - simple_unlock(&runq->lock); - /* - * Nothing is runnable, so set this processor idle if it - * was running. If it was in an assignment or shutdown, - * leave it alone. Return its idle thread. - */ - simple_lock(&pset->idle_lock); - if (myprocessor->state == PROCESSOR_RUNNING) { - myprocessor->state = PROCESSOR_IDLE; - /* - * XXX Until it goes away, put master on end of queue, others - * XXX on front so master gets used last. - */ - if (myprocessor == master_processor) - queue_enter(&(pset->idle_queue), myprocessor, - processor_t, processor_queue); - else - queue_enter_first(&(pset->idle_queue), myprocessor, - processor_t, processor_queue); + 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); - pset->idle_count++; - } - simple_unlock(&pset->idle_lock); + pset_lock(pset); - return (myprocessor->idle_thread); -} + /* If we were sent a remote AST and came out of idle, acknowledge it here with pset lock held */ + pset->pending_AST_cpu_mask &= ~(1ULL << processor->cpu_id); +#if defined(CONFIG_SCHED_DEFERRED_AST) + pset->pending_deferred_AST_cpu_mask &= ~(1ULL << processor->cpu_id); +#endif -/* - * no_dispatch_count counts number of times processors go non-idle - * without being dispatched. This should be very rare. - */ -int no_dispatch_count = 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; -/* - * This is the idle thread, which just looks for other threads - * to execute. - */ -void -idle_thread_continue(void) -{ - register processor_t myprocessor; - register volatile thread_t *threadp; - register volatile int *gcount; - register volatile int *lcount; - register thread_t new_thread; - register int state; - register processor_set_t pset; - int mycpu; + if ((new_thread != THREAD_NULL) && (SCHED(processor_queue_has_priority)(processor, new_thread->sched_pri, FALSE) || + (rt_runq.count > 0)) ) { + /* Something higher priority has popped up on the runqueue - redispatch this thread elsewhere */ + processor->current_pri = IDLEPRI; + processor->current_thmode = TH_MODE_FIXED; + processor->current_sfi_class = SFI_CLASS_KERNEL; + processor->deadline = UINT64_MAX; - mycpu = cpu_number(); - myprocessor = current_processor(); - threadp = (volatile thread_t *) &myprocessor->next_thread; - lcount = (volatile int *) &myprocessor->runq.count; + pset_unlock(pset); - for (;;) { -#ifdef MARK_CPU_IDLE - MARK_CPU_IDLE(mycpu); -#endif /* MARK_CPU_IDLE */ + thread_lock(new_thread); + KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_REDISPATCH), (uintptr_t)thread_tid(new_thread), new_thread->sched_pri, rt_runq.count, 0, 0); + thread_setrun(new_thread, SCHED_HEADQ); + thread_unlock(new_thread); - gcount = (volatile int *)&myprocessor->processor_set->runq.count; + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_END, + (uintptr_t)thread_tid(thread), state, 0, 0, 0); - (void)splsched(); - while ( (*threadp == (volatile thread_t)THREAD_NULL) && - (*gcount == 0) && (*lcount == 0) ) { + return (THREAD_NULL); + } - /* check for ASTs while we wait */ + pset_unlock(pset); - if (need_ast[mycpu] &~ (AST_SCHEDULING|AST_URGENT|AST_BSD|AST_BSD_INIT)) { - /* don't allow scheduling ASTs */ - need_ast[mycpu] &= ~(AST_SCHEDULING|AST_URGENT|AST_BSD|AST_BSD_INIT); - ast_taken(FALSE, AST_ALL, TRUE); /* back at spllo */ - } - else -#ifdef __ppc__ - machine_idle(); -#else - (void)spllo(); -#endif - machine_clock_assist(); + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_END, + (uintptr_t)thread_tid(thread), state, (uintptr_t)thread_tid(new_thread), 0, 0); - (void)splsched(); - } + return (new_thread); -#ifdef MARK_CPU_ACTIVE - (void)spllo(); - MARK_CPU_ACTIVE(mycpu); - (void)splsched(); -#endif /* MARK_CPU_ACTIVE */ + } else if (state == PROCESSOR_IDLE) { + re_queue_tail(&pset->active_queue, &processor->processor_queue); + processor->state = PROCESSOR_RUNNING; + processor->current_pri = IDLEPRI; + processor->current_thmode = TH_MODE_FIXED; + processor->current_sfi_class = SFI_CLASS_KERNEL; + processor->deadline = UINT64_MAX; + + } else if (state == PROCESSOR_SHUTDOWN) { /* - * This is not a switch statement to avoid the - * bounds checking code in the common case. + * Going off-line. Force a + * reschedule. */ - pset = myprocessor->processor_set; - simple_lock(&pset->idle_lock); -retry: - state = myprocessor->state; - if (state == PROCESSOR_DISPATCHING) { - /* - * Commmon case -- cpu dispatched. - */ - new_thread = *threadp; - *threadp = (volatile thread_t) THREAD_NULL; - myprocessor->state = PROCESSOR_RUNNING; - simple_unlock(&pset->idle_lock); - - thread_lock(new_thread); - simple_lock(&myprocessor->runq.lock); - simple_lock(&pset->runq.lock); - if ( myprocessor->runq.highq > new_thread->sched_pri || - pset->runq.highq > new_thread->sched_pri ) { - simple_unlock(&pset->runq.lock); - simple_unlock(&myprocessor->runq.lock); - - if (new_thread->bound_processor != PROCESSOR_NULL) - run_queue_enqueue(&myprocessor->runq, new_thread, HEAD_Q); - else - run_queue_enqueue(&pset->runq, new_thread, HEAD_Q); - thread_unlock(new_thread); - - counter(c_idle_thread_block++); - thread_block(idle_thread_continue); - } - else { - simple_unlock(&pset->runq.lock); - simple_unlock(&myprocessor->runq.lock); + if ((new_thread = processor->next_thread) != THREAD_NULL) { + processor->next_thread = THREAD_NULL; + processor->current_pri = IDLEPRI; + processor->current_thmode = TH_MODE_FIXED; + processor->current_sfi_class = SFI_CLASS_KERNEL; + processor->deadline = UINT64_MAX; - /* - * set up quantum for new thread. - */ - if (new_thread->policy & (POLICY_RR|POLICY_FIFO)) - myprocessor->quantum = new_thread->unconsumed_quantum; - else - myprocessor->quantum = pset->set_quantum; - thread_unlock(new_thread); + pset_unlock(pset); - myprocessor->first_quantum = TRUE; - counter(c_idle_thread_handoff++); - thread_run(myprocessor->idle_thread, - idle_thread_continue, new_thread); - } - } - else - if (state == PROCESSOR_IDLE) { - if (myprocessor->state != PROCESSOR_IDLE) { - /* - * Something happened, try again. - */ - goto retry; - } - /* - * Processor was not dispatched (Rare). - * Set it running again. - */ - no_dispatch_count++; - pset->idle_count--; - queue_remove(&pset->idle_queue, myprocessor, - processor_t, processor_queue); - myprocessor->state = PROCESSOR_RUNNING; - simple_unlock(&pset->idle_lock); + thread_lock(new_thread); + thread_setrun(new_thread, SCHED_HEADQ); + thread_unlock(new_thread); - counter(c_idle_thread_block++); - thread_block(idle_thread_continue); - } - else - if ( state == PROCESSOR_ASSIGN || - state == PROCESSOR_SHUTDOWN ) { - /* - * Changing processor sets, or going off-line. - * Release next_thread if there is one. Actual - * thread to run is on a runq. - */ - if ((new_thread = (thread_t)*threadp) != THREAD_NULL) { - *threadp = (volatile thread_t) THREAD_NULL; - simple_unlock(&pset->idle_lock); - thread_lock(new_thread); - thread_setrun(new_thread, FALSE, TAIL_Q); - thread_unlock(new_thread); - } else - simple_unlock(&pset->idle_lock); - - counter(c_idle_thread_block++); - thread_block(idle_thread_continue); + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_END, + (uintptr_t)thread_tid(thread), state, 0, 0, 0); + + return (THREAD_NULL); } - else { - simple_unlock(&pset->idle_lock); - printf("Bad processor state %d (Cpu %d)\n", - cpu_state(mycpu), mycpu); - panic("idle_thread"); + } - } + pset_unlock(pset); - (void)spllo(); - } + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, + MACHDBG_CODE(DBG_MACH_SCHED,MACH_IDLE) | DBG_FUNC_END, + (uintptr_t)thread_tid(thread), state, 0, 0, 0); + + return (THREAD_NULL); } +/* + * Each processor has a dedicated thread which + * executes the idle loop when there is no suitable + * previous context. + */ void idle_thread(void) { - thread_t self = current_thread(); + processor_t processor = current_processor(); + thread_t new_thread; + + 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); + /*NOTREACHED*/ + } + + thread_block((thread_continue_t)idle_thread); + /*NOTREACHED*/ +} + +kern_return_t +idle_thread_create( + processor_t processor) +{ + kern_return_t result; + thread_t thread; spl_t s; - stack_privilege(self); - thread_swappable(current_act(), FALSE); + result = kernel_thread_create((thread_continue_t)idle_thread, NULL, MAXPRI_KERNEL, &thread); + if (result != KERN_SUCCESS) + return (result); s = splsched(); - thread_lock(self); + thread_lock(thread); + thread->bound_processor = processor; + processor->idle_thread = thread; + 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); - self->priority = IDLEPRI; - self->sched_pri = self->priority; + return (KERN_SUCCESS); +} - thread_unlock(self); - splx(s); +/* + * sched_startup: + * + * Kicks off scheduler services. + * + * Called at splsched. + */ +void +sched_startup(void) +{ + kern_return_t result; + thread_t thread; - counter(c_idle_thread_block++); - thread_block((void(*)(void))0); - idle_thread_continue(); - /*NOTREACHED*/ + simple_lock_init(&sched_vm_group_list_lock, 0); + + + 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_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. + */ + thread_block(THREAD_CONTINUE_NULL); } -static AbsoluteTime sched_tick_interval, sched_tick_deadline; +#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; /* - * sched_tick_thread + * sched_init_thread: * - * Update the priorities of all threads periodically. + * Perform periodic bookkeeping functions about ten + * times per second. */ void -sched_tick_thread_continue(void) +sched_timeshare_maintenance_continue(void) { - AbsoluteTime abstime; -#if SIMPLE_CLOCK - int new_usec; -#endif /* SIMPLE_CLOCK */ + 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); - clock_get_uptime(&abstime); + /* 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; - sched_tick++; /* age usage one more time */ -#if SIMPLE_CLOCK /* - * Compensate for clock drift. sched_usec is an - * exponential average of the number of microseconds in - * a second. It decays in the same fashion as cpu_usage. + * Compute various averages. */ - new_usec = sched_usec_elapsed(); - sched_usec = (5*sched_usec + 3*new_usec)/8; -#endif /* SIMPLE_CLOCK */ + compute_averages(sched_tick_delta); /* - * Compute the scheduler load factors. + * Scan the run queues for threads which + * may need to be updated, and find the earliest runnable thread on the runqueue + * to report its latency. */ - compute_mach_factor(); + SCHED(thread_update_scan)(&scan_context); + + 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); /* - * Scan the run queues for runnable threads that need to - * have their priorities recalculated. + * Check to see if the special sched VM group needs attention. */ - do_thread_scan(); + sched_vm_group_maintenance(); + - clock_deadline_for_periodic_event(sched_tick_interval, abstime, - &sched_tick_deadline); + 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], 0, 0); - assert_wait((event_t)sched_tick_thread_continue, THREAD_INTERRUPTIBLE); - thread_set_timer_deadline(sched_tick_deadline); - thread_block(sched_tick_thread_continue); + 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; + +/* + * 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. + */ void -sched_tick_thread(void) -{ - thread_t self = current_thread(); - natural_t rate; - spl_t s; +sched_timeshare_consider_maintenance(uint64_t ctime) { + uint64_t ndeadline, deadline = sched_maintenance_deadline; - stack_privilege(self); - thread_swappable(self->top_act, FALSE); + if (__improbable(ctime >= deadline)) { + if (__improbable(current_thread() == sched_maintenance_thread)) + return; + OSMemoryBarrier(); - s = splsched(); - thread_lock(self); + ndeadline = ctime + sched_tick_interval; + + if (__probable(__sync_bool_compare_and_swap(&sched_maintenance_deadline, deadline, ndeadline))) { + thread_wakeup((event_t)sched_timeshare_maintenance_continue); + sched_maintenance_wakeups++; + } + } +} - self->priority = MAXPRI_STANDARD; - self->sched_pri = self->priority; +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ - thread_unlock(self); - splx(s); +void +sched_init_thread(void (*continuation)(void)) +{ + thread_block(THREAD_CONTINUE_NULL); + + thread_t thread = current_thread(); - rate = (1000 >> SCHED_TICK_SHIFT); - clock_interval_to_absolutetime_interval(rate, USEC_PER_SEC, - &sched_tick_interval); - clock_get_uptime(&sched_tick_deadline); + thread_set_thread_name(thread, "sched_maintenance_thread"); + + sched_maintenance_thread = thread; + + continuation(); - thread_block(sched_tick_thread_continue); /*NOTREACHED*/ } -#define MAX_STUCK_THREADS 128 +#if defined(CONFIG_SCHED_TIMESHARE_CORE) /* - * do_thread_scan: scan for stuck threads. A thread is stuck if - * it is runnable but its priority is so low that it has not - * run for several seconds. Its priority should be higher, but - * won't be until it runs and calls update_priority. The scanner - * finds these threads and does the updates. + * thread_update_scan / runq_scan: + * + * Scan the run queues to account for timesharing threads + * which need to be updated. * * Scanner runs in two passes. Pass one squirrels likely - * thread ids away in an array (takes out references for them). - * Pass two does the priority updates. This is necessary because - * the run queue lock is required for the candidate scan, but - * cannot be held during updates [set_pri will deadlock]. + * threads away in an array, pass two does the update. * - * Array length should be enough so that restart isn't necessary, - * but restart logic is included. Does not scan processor runqs. + * This is necessary because the run queue is locked for + * the candidate scan, but the thread is locked for the update. * + * Array should be sized to make forward progress, without + * disabling preemption for long periods. */ -thread_t stuck_threads[MAX_STUCK_THREADS]; -int stuck_count = 0; -/* - * do_runq_scan is the guts of pass 1. It scans a runq for - * stuck threads. A boolean is returned indicating whether - * a retry is needed. - */ +#define THREAD_UPDATE_SIZE 128 + +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 -do_runq_scan( - run_queue_t runq) +thread_update_add_thread(thread_t thread) { - register queue_t q; - register thread_t thread; - register int count; - spl_t s; - boolean_t result = FALSE; + if (thread_update_count == THREAD_UPDATE_SIZE) + return (FALSE); - s = splsched(); - simple_lock(&runq->lock); - if ((count = runq->count) > 0) { - q = runq->queues + runq->highq; - while (count > 0) { - queue_iterate(q, thread, thread_t, links) { - if ( !(thread->state & (TH_WAIT|TH_SUSP)) && - thread->policy == POLICY_TIMESHARE ) { - if (thread->sched_stamp != sched_tick) { - /* - * Stuck, save its id for later. - */ - if (stuck_count == MAX_STUCK_THREADS) { - /* - * !@#$% No more room. - */ - simple_unlock(&runq->lock); - splx(s); - - return (TRUE); - } - - /* - * Inline version of thread_reference - * XXX - lock ordering problem here: - * thread locks should be taken before runq - * locks: just try and get the thread's locks - * and ignore this thread if we fail, we might - * have better luck next time. - */ - if (simple_lock_try(&thread->lock)) { - thread->ref_count++; - thread_unlock(thread); - stuck_threads[stuck_count++] = thread; - } - else - result = TRUE; - } - } + thread_update_array[thread_update_count++] = thread; + thread_reference_internal(thread); + return (TRUE); +} - count--; - } +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; - q--; + 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); } - simple_unlock(&runq->lock); - splx(s); - return (result); + thread_update_count = 0; } -boolean_t thread_scan_enabled = TRUE; - -void -do_thread_scan(void) +/* + * Scan a runq for candidate threads. + * + * Returns TRUE if retry is needed. + */ +boolean_t +runq_scan( + run_queue_t runq, + sched_update_scan_context_t scan_context) { - register boolean_t restart_needed = FALSE; - register thread_t thread; - register processor_set_t pset = &default_pset; - register processor_t processor; - spl_t s; + int count = runq->count; + int queue_index; - if (!thread_scan_enabled) - return; + assert(count >= 0); - do { - restart_needed = do_runq_scan(&pset->runq); - if (!restart_needed) { - simple_lock(&pset->processors_lock); - processor = (processor_t)queue_first(&pset->processors); - while (!queue_end(&pset->processors, (queue_entry_t)processor)) { - if (restart_needed = do_runq_scan(&processor->runq)) - break; + if (count == 0) + return FALSE; - processor = (processor_t)queue_next(&processor->processors); + for (queue_index = bitmap_first(runq->bitmap, NRQS); + queue_index >= 0; + queue_index = bitmap_next(runq->bitmap, queue_index)) { + + thread_t thread; + queue_t queue = &runq->queues[queue_index]; + + qe_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; } - simple_unlock(&pset->processors_lock); - } - /* - * Ok, we now have a collection of candidates -- fix them. - */ - while (stuck_count > 0) { - thread = stuck_threads[--stuck_count]; - stuck_threads[stuck_count] = THREAD_NULL; - s = splsched(); - thread_lock(thread); - if (thread->policy == POLICY_TIMESHARE) { - if ( !(thread->state & (TH_WAIT|TH_SUSP)) && - thread->sched_stamp != sched_tick ) - update_priority(thread); + 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; + } } - thread_unlock(thread); - splx(s); - thread_deallocate(thread); - } - - } while (restart_needed); + count--; + } + } + + return FALSE; } - -/* - * Just in case someone doesn't use the macro - */ -#undef thread_wakeup -void -thread_wakeup( - event_t x); -void -thread_wakeup( - event_t x) +#endif /* CONFIG_SCHED_TIMESHARE_CORE */ + +boolean_t +thread_eager_preemption(thread_t thread) { - thread_wakeup_with_result(x, THREAD_AWAKENED); + return ((thread->sched_flags & TH_SFLAG_EAGERPREEMPT) != 0); } -boolean_t -thread_runnable( - thread_t thread) +void +thread_set_eager_preempt(thread_t thread) { - sched_policy_t *policy; + spl_t x; + processor_t p; + ast_t ast = AST_NONE; - /* Ask sched policy if thread is runnable */ - policy = policy_id_to_sched_policy(thread->policy); + x = splsched(); + p = current_processor(); - return ((policy != SCHED_POLICY_NULL)? - policy->sp_ops.sp_thread_runnable(policy, thread) : FALSE); -} + thread_lock(thread); + thread->sched_flags |= TH_SFLAG_EAGERPREEMPT; -#if DEBUG + if (thread == current_thread()) { -void -dump_processor_set( - processor_set_t ps) -{ - printf("processor_set: %08x\n",ps); - printf("idle_queue: %08x %08x, idle_count: 0x%x\n", - ps->idle_queue.next,ps->idle_queue.prev,ps->idle_count); - printf("processors: %08x %08x, processor_count: 0x%x\n", - ps->processors.next,ps->processors.prev,ps->processor_count); - printf("tasks: %08x %08x, task_count: 0x%x\n", - ps->tasks.next,ps->tasks.prev,ps->task_count); - printf("threads: %08x %08x, thread_count: 0x%x\n", - ps->threads.next,ps->threads.prev,ps->thread_count); - printf("ref_count: 0x%x, active: %x\n", - ps->ref_count,ps->active); - printf("pset_self: %08x, pset_name_self: %08x\n",ps->pset_self, ps->pset_name_self); - printf("max_priority: 0x%x, policies: 0x%x, set_quantum: 0x%x\n", - ps->max_priority, ps->policies, ps->set_quantum); -} - -#define processor_state(s) (((s)>PROCESSOR_SHUTDOWN)?"*unknown*":states[s]) + ast = csw_check(p, AST_NONE); + thread_unlock(thread); + if (ast != AST_NONE) { + (void) thread_block_reason(THREAD_CONTINUE_NULL, NULL, ast); + } + } else { + p = thread->last_processor; -void -dump_processor( - processor_t p) -{ - char *states[]={"OFF_LINE","RUNNING","IDLE","DISPATCHING", - "ASSIGN","SHUTDOWN"}; + if (p != PROCESSOR_NULL && p->state == PROCESSOR_RUNNING && + p->active_thread == thread) { + cause_ast_check(p); + } + + thread_unlock(thread); + } - printf("processor: %08x\n",p); - printf("processor_queue: %08x %08x\n", - p->processor_queue.next,p->processor_queue.prev); - printf("state: %8s, next_thread: %08x, idle_thread: %08x\n", - processor_state(p->state), p->next_thread, p->idle_thread); - printf("quantum: %u, first_quantum: %x, last_quantum: %u\n", - p->quantum, p->first_quantum, p->last_quantum); - printf("processor_set: %08x, processor_set_next: %08x\n", - p->processor_set, p->processor_set_next); - printf("processors: %08x %08x\n", p->processors.next,p->processors.prev); - printf("processor_self: %08x, slot_num: 0x%x\n", p->processor_self, p->slot_num); + splx(x); } void -dump_run_queue_struct( - run_queue_t rq) +thread_clear_eager_preempt(thread_t thread) { - char dump_buf[80]; - int i; - - for( i=0; i < NRQS; ) { - int j; - - printf("%6s",(i==0)?"runq:":""); - for( j=0; (j<8) && (i < NRQS); j++,i++ ) { - if( rq->queues[i].next == &rq->queues[i] ) - printf( " --------"); - else - printf(" %08x",rq->queues[i].next); - } - printf("\n"); - } - for( i=0; i < NRQBM; ) { - register unsigned int mask; - char *d=dump_buf; + spl_t x; - mask = ~0; - mask ^= (mask>>1); + x = splsched(); + thread_lock(thread); - do { - *d++ = ((rq->bitmap[i]&mask)?'r':'e'); - mask >>=1; - } while( mask ); - *d = '\0'; - printf("%8s%s\n",((i==0)?"bitmap:":""),dump_buf); - i++; - } - printf("highq: 0x%x, count: %u\n", rq->highq, rq->count); + thread->sched_flags &= ~TH_SFLAG_EAGERPREEMPT; + + thread_unlock(thread); + splx(x); } - + +/* + * Scheduling statistics + */ void -dump_run_queues( - run_queue_t runq) +sched_stats_handle_csw(processor_t processor, int reasons, int selfpri, int otherpri) { - register queue_t q1; - register int i; - register queue_entry_t e; + struct processor_sched_statistics *stats; + boolean_t to_realtime = FALSE; + + stats = &processor->processor_data.sched_stats; + stats->csw_count++; + + if (otherpri >= BASEPRI_REALTIME) { + stats->rt_sched_count++; + to_realtime = TRUE; + } + + if ((reasons & AST_PREEMPT) != 0) { + stats->preempt_count++; - q1 = runq->queues; - for (i = 0; i < NRQS; i++) { - if (q1->next != q1) { - int t_cnt; + if (selfpri >= BASEPRI_REALTIME) { + stats->preempted_rt_count++; + } - printf("[%u]",i); - for (t_cnt=0, e = q1->next; e != q1; e = e->next) { - printf("\t0x%08x",e); - if( (t_cnt = ++t_cnt%4) == 0 ) - printf("\n"); + if (to_realtime) { + stats->preempted_by_rt_count++; } - if( t_cnt ) - printf("\n"); - } - /* else - printf("[%u]\t\n",i); - */ - q1++; + } } void -checkrq( - run_queue_t rq, - char *msg) -{ - register queue_t q1; - register int i, j; - register queue_entry_t e; - register int highq; - - highq = NRQS; - j = 0; - q1 = rq->queues; - for (i = MAXPRI; i >= 0; i--) { - if (q1->next == q1) { - if (q1->prev != q1) { - panic("checkrq: empty at %s", msg); - } - } - else { - if (highq == -1) - highq = i; - - for (e = q1->next; e != q1; e = e->next) { - j++; - if (e->next->prev != e) - panic("checkrq-2 at %s", msg); - if (e->prev->next != e) - panic("checkrq-3 at %s", msg); - } - } - q1++; - } - if (j != rq->count) - panic("checkrq: count wrong at %s", msg); - if (rq->count != 0 && highq > rq->highq) - panic("checkrq: highq wrong at %s", msg); +sched_stats_handle_runq_change(struct runq_stats *stats, int old_count) +{ + uint64_t timestamp = mach_absolute_time(); + + stats->count_sum += (timestamp - stats->last_change_timestamp) * old_count; + stats->last_change_timestamp = timestamp; } +/* + * For calls from assembly code + */ +#undef thread_wakeup +void +thread_wakeup( + event_t x); + void -thread_check( - register thread_t thread, - register run_queue_t rq) +thread_wakeup( + event_t x) { - register int whichq = thread->sched_pri; - register queue_entry_t queue, entry; + thread_wakeup_with_result(x, THREAD_AWAKENED); +} - if (whichq < MINPRI || whichq > MAXPRI) - panic("thread_check: bad pri"); +boolean_t +preemption_enabled(void) +{ + return (get_preemption_level() == 0 && ml_get_interrupts_enabled()); +} - if (whichq != thread->whichq) - panic("thread_check: whichq"); +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); +} - queue = &rq->queues[whichq]; - entry = queue_first(queue); - while (!queue_end(queue, entry)) { - if (entry == (queue_entry_t)thread) - return; - entry = queue_next(entry); +kern_return_t +sched_work_interval_notify(thread_t thread, uint64_t work_interval_id, uint64_t start, uint64_t finish, uint64_t deadline, uint64_t next_start, uint32_t flags) +{ + int urgency; + uint64_t urgency_param1, urgency_param2; + spl_t s; + + if (work_interval_id == 0) { + return (KERN_INVALID_ARGUMENT); } - panic("thread_check: not found"); -} + assert(thread == current_thread()); -#endif /* DEBUG */ + thread_mtx_lock(thread); + if (thread->work_interval_id != work_interval_id) { + thread_mtx_unlock(thread); + return (KERN_INVALID_ARGUMENT); + } + thread_mtx_unlock(thread); + + s = splsched(); + thread_lock(thread); + urgency = thread_get_urgency(thread, &urgency_param1, &urgency_param2); + thread_unlock(thread); + splx(s); -#if MACH_KDB -#include -#define printf kdbprintf -extern int db_indent; -void db_sched(void); + machine_work_interval_notify(thread, work_interval_id, start, finish, deadline, next_start, urgency, flags); + return (KERN_SUCCESS); +} -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 no_dispatch %d\n", - c_idle_thread_handoff, - c_idle_thread_block, no_dispatch_count); - iprintf("Sched thread blocks: %d\n", c_sched_thread_block); -#endif /* MACH_COUNTERS */ - db_indent -= 2; -} -#endif /* MACH_KDB */ +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); +}